Branch data Line data Source code
1 : : /* SLP - Basic Block Vectorization
2 : : Copyright (C) 2007-2024 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 : : static bool vect_transform_slp_perm_load_1 (vec_info *, slp_tree,
57 : : load_permutation_t &,
58 : : const vec<tree> &,
59 : : gimple_stmt_iterator *,
60 : : poly_uint64, bool, bool,
61 : : unsigned *,
62 : : unsigned * = nullptr,
63 : : bool = false);
64 : : static int vectorizable_slp_permutation_1 (vec_info *, gimple_stmt_iterator *,
65 : : slp_tree, lane_permutation_t &,
66 : : vec<slp_tree> &, bool);
67 : : static bool vectorizable_slp_permutation (vec_info *, gimple_stmt_iterator *,
68 : : slp_tree, stmt_vector_for_cost *);
69 : : static void vect_print_slp_tree (dump_flags_t, dump_location_t, slp_tree);
70 : : static bool vect_slp_can_convert_to_external (const vec<stmt_vec_info> &);
71 : :
72 : : static object_allocator<_slp_tree> *slp_tree_pool;
73 : : static slp_tree slp_first_node;
74 : :
75 : : void
76 : 1068264 : vect_slp_init (void)
77 : : {
78 : 1068264 : slp_tree_pool = new object_allocator<_slp_tree> ("SLP nodes");
79 : 1068264 : }
80 : :
81 : : void
82 : 1068264 : vect_slp_fini (void)
83 : : {
84 : 1539213 : while (slp_first_node)
85 : 470949 : delete slp_first_node;
86 : 2136528 : delete slp_tree_pool;
87 : 1068264 : slp_tree_pool = NULL;
88 : 1068264 : }
89 : :
90 : : void *
91 : 6147437 : _slp_tree::operator new (size_t n)
92 : : {
93 : 6147437 : gcc_assert (n == sizeof (_slp_tree));
94 : 6147437 : return slp_tree_pool->allocate_raw ();
95 : : }
96 : :
97 : : void
98 : 6147437 : _slp_tree::operator delete (void *node, size_t n)
99 : : {
100 : 6147437 : gcc_assert (n == sizeof (_slp_tree));
101 : 6147437 : slp_tree_pool->remove_raw (node);
102 : 6147437 : }
103 : :
104 : :
105 : : /* Initialize a SLP node. */
106 : :
107 : 6147437 : _slp_tree::_slp_tree ()
108 : : {
109 : 6147437 : this->prev_node = NULL;
110 : 6147437 : if (slp_first_node)
111 : 5382321 : slp_first_node->prev_node = this;
112 : 6147437 : this->next_node = slp_first_node;
113 : 6147437 : slp_first_node = this;
114 : 6147437 : SLP_TREE_SCALAR_STMTS (this) = vNULL;
115 : 6147437 : SLP_TREE_SCALAR_OPS (this) = vNULL;
116 : 6147437 : SLP_TREE_VEC_DEFS (this) = vNULL;
117 : 6147437 : SLP_TREE_NUMBER_OF_VEC_STMTS (this) = 0;
118 : 6147437 : SLP_TREE_CHILDREN (this) = vNULL;
119 : 6147437 : SLP_TREE_LOAD_PERMUTATION (this) = vNULL;
120 : 6147437 : SLP_TREE_LANE_PERMUTATION (this) = vNULL;
121 : 6147437 : SLP_TREE_SIMD_CLONE_INFO (this) = vNULL;
122 : 6147437 : SLP_TREE_DEF_TYPE (this) = vect_uninitialized_def;
123 : 6147437 : SLP_TREE_CODE (this) = ERROR_MARK;
124 : 6147437 : this->ldst_lanes = false;
125 : 6147437 : SLP_TREE_VECTYPE (this) = NULL_TREE;
126 : 6147437 : SLP_TREE_REPRESENTATIVE (this) = NULL;
127 : 6147437 : SLP_TREE_MEMORY_ACCESS_TYPE (this) = VMAT_INVARIANT;
128 : 6147437 : SLP_TREE_REF_COUNT (this) = 1;
129 : 6147437 : this->failed = NULL;
130 : 6147437 : this->max_nunits = 1;
131 : 6147437 : this->lanes = 0;
132 : 6147437 : }
133 : :
134 : : /* Tear down a SLP node. */
135 : :
136 : 6147437 : _slp_tree::~_slp_tree ()
137 : : {
138 : 6147437 : if (this->prev_node)
139 : 3842161 : this->prev_node->next_node = this->next_node;
140 : : else
141 : 2305276 : slp_first_node = this->next_node;
142 : 6147437 : if (this->next_node)
143 : 4554452 : this->next_node->prev_node = this->prev_node;
144 : 6147437 : SLP_TREE_CHILDREN (this).release ();
145 : 6147437 : SLP_TREE_SCALAR_STMTS (this).release ();
146 : 6147437 : SLP_TREE_SCALAR_OPS (this).release ();
147 : 6147437 : SLP_TREE_VEC_DEFS (this).release ();
148 : 6147437 : SLP_TREE_LOAD_PERMUTATION (this).release ();
149 : 6147437 : SLP_TREE_LANE_PERMUTATION (this).release ();
150 : 6147437 : SLP_TREE_SIMD_CLONE_INFO (this).release ();
151 : 6147437 : if (this->failed)
152 : 1856591 : free (failed);
153 : 6147437 : }
154 : :
155 : : /* Push the single SSA definition in DEF to the vector of vector defs. */
156 : :
157 : : void
158 : 467670 : _slp_tree::push_vec_def (gimple *def)
159 : : {
160 : 467670 : if (gphi *phi = dyn_cast <gphi *> (def))
161 : 54500 : vec_defs.quick_push (gimple_phi_result (phi));
162 : : else
163 : : {
164 : 413170 : def_operand_p defop = single_ssa_def_operand (def, SSA_OP_ALL_DEFS);
165 : 413170 : vec_defs.quick_push (get_def_from_ptr (defop));
166 : : }
167 : 467670 : }
168 : :
169 : : /* Recursively free the memory allocated for the SLP tree rooted at NODE. */
170 : :
171 : : void
172 : 11320704 : vect_free_slp_tree (slp_tree node)
173 : : {
174 : 11320704 : int i;
175 : 11320704 : slp_tree child;
176 : :
177 : 11320704 : if (--SLP_TREE_REF_COUNT (node) != 0)
178 : 11320704 : return;
179 : :
180 : 8587839 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
181 : 2911351 : if (child)
182 : 2610280 : vect_free_slp_tree (child);
183 : :
184 : : /* If the node defines any SLP only patterns then those patterns are no
185 : : longer valid and should be removed. */
186 : 5676488 : stmt_vec_info rep_stmt_info = SLP_TREE_REPRESENTATIVE (node);
187 : 5676488 : if (rep_stmt_info && STMT_VINFO_SLP_VECT_ONLY_PATTERN (rep_stmt_info))
188 : : {
189 : 448 : stmt_vec_info stmt_info = vect_orig_stmt (rep_stmt_info);
190 : 448 : STMT_VINFO_IN_PATTERN_P (stmt_info) = false;
191 : 448 : STMT_SLP_TYPE (stmt_info) = STMT_SLP_TYPE (rep_stmt_info);
192 : : }
193 : :
194 : 5676488 : delete node;
195 : : }
196 : :
197 : : /* Return a location suitable for dumpings related to the SLP instance. */
198 : :
199 : : dump_user_location_t
200 : 3295584 : _slp_instance::location () const
201 : : {
202 : 3295584 : if (!root_stmts.is_empty ())
203 : 252656 : return root_stmts[0]->stmt;
204 : : else
205 : 3042928 : return SLP_TREE_SCALAR_STMTS (root)[0]->stmt;
206 : : }
207 : :
208 : :
209 : : /* Free the memory allocated for the SLP instance. */
210 : :
211 : : void
212 : 1201460 : vect_free_slp_instance (slp_instance instance)
213 : : {
214 : 1201460 : vect_free_slp_tree (SLP_INSTANCE_TREE (instance));
215 : 1201460 : SLP_INSTANCE_LOADS (instance).release ();
216 : 1201460 : SLP_INSTANCE_ROOT_STMTS (instance).release ();
217 : 1201460 : SLP_INSTANCE_REMAIN_DEFS (instance).release ();
218 : 1201460 : instance->subgraph_entries.release ();
219 : 1201460 : instance->cost_vec.release ();
220 : 1201460 : free (instance);
221 : 1201460 : }
222 : :
223 : :
224 : : /* Create an SLP node for SCALAR_STMTS. */
225 : :
226 : : slp_tree
227 : 70284 : vect_create_new_slp_node (unsigned nops, tree_code code)
228 : : {
229 : 70284 : slp_tree node = new _slp_tree;
230 : 70284 : SLP_TREE_SCALAR_STMTS (node) = vNULL;
231 : 70284 : SLP_TREE_CHILDREN (node).create (nops);
232 : 70284 : SLP_TREE_DEF_TYPE (node) = vect_internal_def;
233 : 70284 : SLP_TREE_CODE (node) = code;
234 : 70284 : return node;
235 : : }
236 : : /* Create an SLP node for SCALAR_STMTS. */
237 : :
238 : : static slp_tree
239 : 2691189 : vect_create_new_slp_node (slp_tree node,
240 : : vec<stmt_vec_info> scalar_stmts, unsigned nops)
241 : : {
242 : 2691189 : SLP_TREE_SCALAR_STMTS (node) = scalar_stmts;
243 : 2691189 : SLP_TREE_CHILDREN (node).create (nops);
244 : 2691189 : SLP_TREE_DEF_TYPE (node) = vect_internal_def;
245 : 2691189 : SLP_TREE_REPRESENTATIVE (node) = scalar_stmts[0];
246 : 2691189 : SLP_TREE_LANES (node) = scalar_stmts.length ();
247 : 2691189 : return node;
248 : : }
249 : :
250 : : /* Create an SLP node for SCALAR_STMTS. */
251 : :
252 : : static slp_tree
253 : 6192 : vect_create_new_slp_node (vec<stmt_vec_info> scalar_stmts, unsigned nops)
254 : : {
255 : 6192 : return vect_create_new_slp_node (new _slp_tree, scalar_stmts, nops);
256 : : }
257 : :
258 : : /* Create an SLP node for OPS. */
259 : :
260 : : static slp_tree
261 : 1523598 : vect_create_new_slp_node (slp_tree node, vec<tree> ops)
262 : : {
263 : 1523598 : SLP_TREE_SCALAR_OPS (node) = ops;
264 : 1523598 : SLP_TREE_DEF_TYPE (node) = vect_external_def;
265 : 0 : SLP_TREE_LANES (node) = ops.length ();
266 : 1523598 : return node;
267 : : }
268 : :
269 : : /* Create an SLP node for OPS. */
270 : :
271 : : static slp_tree
272 : 1523598 : vect_create_new_slp_node (vec<tree> ops)
273 : : {
274 : 1523598 : return vect_create_new_slp_node (new _slp_tree, ops);
275 : : }
276 : :
277 : :
278 : : /* This structure is used in creation of an SLP tree. Each instance
279 : : corresponds to the same operand in a group of scalar stmts in an SLP
280 : : node. */
281 : : typedef struct _slp_oprnd_info
282 : : {
283 : : /* Def-stmts for the operands. */
284 : : vec<stmt_vec_info> def_stmts;
285 : : /* Operands. */
286 : : vec<tree> ops;
287 : : /* Information about the first statement, its vector def-type, type, the
288 : : operand itself in case it's constant, and an indication if it's a pattern
289 : : stmt and gather/scatter info. */
290 : : tree first_op_type;
291 : : enum vect_def_type first_dt;
292 : : bool any_pattern;
293 : : bool first_gs_p;
294 : : gather_scatter_info first_gs_info;
295 : : } *slp_oprnd_info;
296 : :
297 : :
298 : : /* Allocate operands info for NOPS operands, and GROUP_SIZE def-stmts for each
299 : : operand. */
300 : : static vec<slp_oprnd_info>
301 : 2765120 : vect_create_oprnd_info (int nops, int group_size)
302 : : {
303 : 2765120 : int i;
304 : 2765120 : slp_oprnd_info oprnd_info;
305 : 2765120 : vec<slp_oprnd_info> oprnds_info;
306 : :
307 : 2765120 : oprnds_info.create (nops);
308 : 9794199 : for (i = 0; i < nops; i++)
309 : : {
310 : 4263959 : oprnd_info = XNEW (struct _slp_oprnd_info);
311 : 4263959 : oprnd_info->def_stmts.create (group_size);
312 : 4263959 : oprnd_info->ops.create (group_size);
313 : 4263959 : oprnd_info->first_dt = vect_uninitialized_def;
314 : 4263959 : oprnd_info->first_op_type = NULL_TREE;
315 : 4263959 : oprnd_info->any_pattern = false;
316 : 4263959 : oprnd_info->first_gs_p = false;
317 : 4263959 : oprnds_info.quick_push (oprnd_info);
318 : : }
319 : :
320 : 2765120 : return oprnds_info;
321 : : }
322 : :
323 : :
324 : : /* Free operands info. */
325 : :
326 : : static void
327 : 2763816 : vect_free_oprnd_info (vec<slp_oprnd_info> &oprnds_info)
328 : : {
329 : 2763816 : int i;
330 : 2763816 : slp_oprnd_info oprnd_info;
331 : :
332 : 7026471 : FOR_EACH_VEC_ELT (oprnds_info, i, oprnd_info)
333 : : {
334 : 4262655 : oprnd_info->def_stmts.release ();
335 : 4262655 : oprnd_info->ops.release ();
336 : 4262655 : XDELETE (oprnd_info);
337 : : }
338 : :
339 : 2763816 : oprnds_info.release ();
340 : 2763816 : }
341 : :
342 : : /* Return the execution frequency of NODE (so that a higher value indicates
343 : : a "more important" node when optimizing for speed). */
344 : :
345 : : static sreal
346 : 2713958 : vect_slp_node_weight (slp_tree node)
347 : : {
348 : 2713958 : stmt_vec_info stmt_info = vect_orig_stmt (SLP_TREE_REPRESENTATIVE (node));
349 : 2713958 : basic_block bb = gimple_bb (stmt_info->stmt);
350 : 2713958 : return bb->count.to_sreal_scale (ENTRY_BLOCK_PTR_FOR_FN (cfun)->count);
351 : : }
352 : :
353 : : /* Return true if STMTS contains a pattern statement. */
354 : :
355 : : static bool
356 : 23629 : vect_contains_pattern_stmt_p (vec<stmt_vec_info> stmts)
357 : : {
358 : 23629 : stmt_vec_info stmt_info;
359 : 23629 : unsigned int i;
360 : 77572 : FOR_EACH_VEC_ELT (stmts, i, stmt_info)
361 : 55941 : if (stmt_info && is_pattern_stmt_p (stmt_info))
362 : : return true;
363 : : return false;
364 : : }
365 : :
366 : : /* Return true when all lanes in the external or constant NODE have
367 : : the same value. */
368 : :
369 : : static bool
370 : 493385 : vect_slp_tree_uniform_p (slp_tree node)
371 : : {
372 : 493385 : gcc_assert (SLP_TREE_DEF_TYPE (node) == vect_constant_def
373 : : || SLP_TREE_DEF_TYPE (node) == vect_external_def);
374 : :
375 : : /* Pre-exsting vectors. */
376 : 852281 : if (SLP_TREE_SCALAR_OPS (node).is_empty ())
377 : : return false;
378 : :
379 : : unsigned i;
380 : : tree op, first = NULL_TREE;
381 : 1146217 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_OPS (node), i, op)
382 : 1011728 : if (!first)
383 : : first = op;
384 : 518343 : else if (!operand_equal_p (first, op, 0))
385 : : return false;
386 : :
387 : : return true;
388 : : }
389 : :
390 : : /* Find the place of the data-ref in STMT_INFO in the interleaving chain
391 : : that starts from FIRST_STMT_INFO. Return -1 if the data-ref is not a part
392 : : of the chain. */
393 : :
394 : : int
395 : 565087 : vect_get_place_in_interleaving_chain (stmt_vec_info stmt_info,
396 : : stmt_vec_info first_stmt_info)
397 : : {
398 : 565087 : stmt_vec_info next_stmt_info = first_stmt_info;
399 : 565087 : int result = 0;
400 : :
401 : 565087 : if (first_stmt_info != DR_GROUP_FIRST_ELEMENT (stmt_info))
402 : : return -1;
403 : :
404 : 1541746 : do
405 : : {
406 : 1541746 : if (next_stmt_info == stmt_info)
407 : : return result;
408 : 976659 : next_stmt_info = DR_GROUP_NEXT_ELEMENT (next_stmt_info);
409 : 976659 : if (next_stmt_info)
410 : 976659 : result += DR_GROUP_GAP (next_stmt_info);
411 : : }
412 : 976659 : while (next_stmt_info);
413 : :
414 : : return -1;
415 : : }
416 : :
417 : : /* Check whether it is possible to load COUNT elements of type ELT_TYPE
418 : : using the method implemented by duplicate_and_interleave. Return true
419 : : if so, returning the number of intermediate vectors in *NVECTORS_OUT
420 : : (if nonnull) and the type of each intermediate vector in *VECTOR_TYPE_OUT
421 : : (if nonnull). */
422 : :
423 : : bool
424 : 0 : can_duplicate_and_interleave_p (vec_info *vinfo, unsigned int count,
425 : : tree elt_type, unsigned int *nvectors_out,
426 : : tree *vector_type_out,
427 : : tree *permutes)
428 : : {
429 : 0 : tree base_vector_type = get_vectype_for_scalar_type (vinfo, elt_type, count);
430 : 0 : if (!base_vector_type || !VECTOR_MODE_P (TYPE_MODE (base_vector_type)))
431 : 0 : return false;
432 : :
433 : 0 : machine_mode base_vector_mode = TYPE_MODE (base_vector_type);
434 : 0 : poly_int64 elt_bytes = count * GET_MODE_UNIT_SIZE (base_vector_mode);
435 : 0 : unsigned int nvectors = 1;
436 : 0 : for (;;)
437 : : {
438 : 0 : scalar_int_mode int_mode;
439 : 0 : poly_int64 elt_bits = elt_bytes * BITS_PER_UNIT;
440 : 0 : if (int_mode_for_size (elt_bits, 1).exists (&int_mode))
441 : : {
442 : : /* Get the natural vector type for this SLP group size. */
443 : 0 : tree int_type = build_nonstandard_integer_type
444 : 0 : (GET_MODE_BITSIZE (int_mode), 1);
445 : 0 : tree vector_type
446 : 0 : = get_vectype_for_scalar_type (vinfo, int_type, count);
447 : 0 : poly_int64 half_nelts;
448 : 0 : if (vector_type
449 : 0 : && VECTOR_MODE_P (TYPE_MODE (vector_type))
450 : 0 : && known_eq (GET_MODE_SIZE (TYPE_MODE (vector_type)),
451 : : GET_MODE_SIZE (base_vector_mode))
452 : 0 : && multiple_p (GET_MODE_NUNITS (TYPE_MODE (vector_type)),
453 : : 2, &half_nelts))
454 : : {
455 : : /* Try fusing consecutive sequences of COUNT / NVECTORS elements
456 : : together into elements of type INT_TYPE and using the result
457 : : to build NVECTORS vectors. */
458 : 0 : poly_uint64 nelts = GET_MODE_NUNITS (TYPE_MODE (vector_type));
459 : 0 : vec_perm_builder sel1 (nelts, 2, 3);
460 : 0 : vec_perm_builder sel2 (nelts, 2, 3);
461 : :
462 : 0 : for (unsigned int i = 0; i < 3; ++i)
463 : : {
464 : 0 : sel1.quick_push (i);
465 : 0 : sel1.quick_push (i + nelts);
466 : 0 : sel2.quick_push (half_nelts + i);
467 : 0 : sel2.quick_push (half_nelts + i + nelts);
468 : : }
469 : 0 : vec_perm_indices indices1 (sel1, 2, nelts);
470 : 0 : vec_perm_indices indices2 (sel2, 2, nelts);
471 : 0 : machine_mode vmode = TYPE_MODE (vector_type);
472 : 0 : if (can_vec_perm_const_p (vmode, vmode, indices1)
473 : 0 : && can_vec_perm_const_p (vmode, vmode, indices2))
474 : : {
475 : 0 : if (nvectors_out)
476 : 0 : *nvectors_out = nvectors;
477 : 0 : if (vector_type_out)
478 : 0 : *vector_type_out = vector_type;
479 : 0 : if (permutes)
480 : : {
481 : 0 : permutes[0] = vect_gen_perm_mask_checked (vector_type,
482 : : indices1);
483 : 0 : permutes[1] = vect_gen_perm_mask_checked (vector_type,
484 : : indices2);
485 : : }
486 : 0 : return true;
487 : : }
488 : 0 : }
489 : : }
490 : 0 : if (!multiple_p (elt_bytes, 2, &elt_bytes))
491 : : return false;
492 : 0 : nvectors *= 2;
493 : : /* We need to be able to fuse COUNT / NVECTORS elements together. */
494 : 0 : if (!multiple_p (count, nvectors))
495 : : return false;
496 : : }
497 : : }
498 : :
499 : : /* Return true if DTA and DTB match. */
500 : :
501 : : static bool
502 : 16735114 : vect_def_types_match (enum vect_def_type dta, enum vect_def_type dtb)
503 : : {
504 : 16735114 : return (dta == dtb
505 : 360005 : || ((dta == vect_external_def || dta == vect_constant_def)
506 : 225335 : && (dtb == vect_external_def || dtb == vect_constant_def)));
507 : : }
508 : :
509 : : static const int cond_expr_maps[3][5] = {
510 : : { 4, -1, -2, 1, 2 },
511 : : { 4, -2, -1, 1, 2 },
512 : : { 4, -1, -2, 2, 1 }
513 : : };
514 : : static const int no_arg_map[] = { 0 };
515 : : static const int arg0_map[] = { 1, 0 };
516 : : static const int arg1_map[] = { 1, 1 };
517 : : static const int arg2_arg3_map[] = { 2, 2, 3 };
518 : : static const int arg1_arg3_map[] = { 2, 1, 3 };
519 : : static const int arg1_arg4_arg5_map[] = { 3, 1, 4, 5 };
520 : : static const int arg1_arg3_arg4_map[] = { 3, 1, 3, 4 };
521 : : static const int arg3_arg2_map[] = { 2, 3, 2 };
522 : : static const int op1_op0_map[] = { 2, 1, 0 };
523 : : static const int off_map[] = { 1, -3 };
524 : : static const int off_op0_map[] = { 2, -3, 0 };
525 : : static const int off_arg2_arg3_map[] = { 3, -3, 2, 3 };
526 : : static const int off_arg3_arg2_map[] = { 3, -3, 3, 2 };
527 : : static const int mask_call_maps[6][7] = {
528 : : { 1, 1, },
529 : : { 2, 1, 2, },
530 : : { 3, 1, 2, 3, },
531 : : { 4, 1, 2, 3, 4, },
532 : : { 5, 1, 2, 3, 4, 5, },
533 : : { 6, 1, 2, 3, 4, 5, 6 },
534 : : };
535 : :
536 : : /* For most SLP statements, there is a one-to-one mapping between
537 : : gimple arguments and child nodes. If that is not true for STMT,
538 : : return an array that contains:
539 : :
540 : : - the number of child nodes, followed by
541 : : - for each child node, the index of the argument associated with that node.
542 : : The special index -1 is the first operand of an embedded comparison and
543 : : the special index -2 is the second operand of an embedded comparison.
544 : : The special indes -3 is the offset of a gather as analyzed by
545 : : vect_check_gather_scatter.
546 : :
547 : : SWAP is as for vect_get_and_check_slp_defs. */
548 : :
549 : : static const int *
550 : 17886188 : vect_get_operand_map (const gimple *stmt, bool gather_scatter_p = false,
551 : : unsigned char swap = 0)
552 : : {
553 : 17886188 : if (auto assign = dyn_cast<const gassign *> (stmt))
554 : : {
555 : 16702968 : if (gimple_assign_rhs_code (assign) == COND_EXPR
556 : 16702968 : && COMPARISON_CLASS_P (gimple_assign_rhs1 (assign)))
557 : 0 : gcc_unreachable ();
558 : 16702968 : if (TREE_CODE_CLASS (gimple_assign_rhs_code (assign)) == tcc_comparison
559 : 16702968 : && swap)
560 : : return op1_op0_map;
561 : 16351674 : if (gather_scatter_p)
562 : 21704 : return (TREE_CODE (gimple_assign_lhs (assign)) != SSA_NAME
563 : 21704 : ? off_op0_map : off_map);
564 : : }
565 : 17513190 : gcc_assert (!swap);
566 : 17513190 : if (auto call = dyn_cast<const gcall *> (stmt))
567 : : {
568 : 140999 : if (gimple_call_internal_p (call))
569 : 72374 : switch (gimple_call_internal_fn (call))
570 : : {
571 : 12248 : case IFN_MASK_LOAD:
572 : 20016 : return gather_scatter_p ? off_arg2_arg3_map : arg2_arg3_map;
573 : :
574 : 0 : case IFN_GATHER_LOAD:
575 : 0 : return arg1_map;
576 : :
577 : 0 : case IFN_MASK_GATHER_LOAD:
578 : 0 : case IFN_MASK_LEN_GATHER_LOAD:
579 : 0 : return arg1_arg4_arg5_map;
580 : :
581 : 0 : case IFN_SCATTER_STORE:
582 : 0 : return arg1_arg3_map;
583 : :
584 : 0 : case IFN_MASK_SCATTER_STORE:
585 : 0 : case IFN_MASK_LEN_SCATTER_STORE:
586 : 0 : return arg1_arg3_arg4_map;
587 : :
588 : 7036 : case IFN_MASK_STORE:
589 : 12552 : return gather_scatter_p ? off_arg3_arg2_map : arg3_arg2_map;
590 : :
591 : 884 : case IFN_MASK_CALL:
592 : 884 : {
593 : 884 : unsigned nargs = gimple_call_num_args (call);
594 : 884 : if (nargs >= 2 && nargs <= 7)
595 : 884 : return mask_call_maps[nargs-2];
596 : : else
597 : : return nullptr;
598 : : }
599 : :
600 : 140 : case IFN_CLZ:
601 : 140 : case IFN_CTZ:
602 : 140 : return arg0_map;
603 : :
604 : 7292 : case IFN_GOMP_SIMD_LANE:
605 : 7292 : return no_arg_map;
606 : :
607 : : default:
608 : : break;
609 : : }
610 : : }
611 : : return nullptr;
612 : : }
613 : :
614 : : /* Return the SLP node child index for operand OP of STMT. */
615 : :
616 : : int
617 : 1303557 : vect_slp_child_index_for_operand (const gimple *stmt, int op,
618 : : bool gather_scatter_p)
619 : : {
620 : 1303557 : const int *opmap = vect_get_operand_map (stmt, gather_scatter_p);
621 : 1303557 : if (!opmap)
622 : : return op;
623 : 20745 : for (int i = 1; i < 1 + opmap[0]; ++i)
624 : 20745 : if (opmap[i] == op)
625 : 11200 : return i - 1;
626 : 0 : gcc_unreachable ();
627 : : }
628 : :
629 : : /* Get the defs for the rhs of STMT (collect them in OPRNDS_INFO), check that
630 : : they are of a valid type and that they match the defs of the first stmt of
631 : : the SLP group (stored in OPRNDS_INFO). This function tries to match stmts
632 : : by swapping operands of STMTS[STMT_NUM] when possible. Non-zero SWAP
633 : : indicates swap is required for cond_expr stmts. Specifically, SWAP
634 : : is 1 if STMT is cond and operands of comparison need to be swapped;
635 : : SWAP is 2 if STMT is cond and code of comparison needs to be inverted.
636 : :
637 : : If there was a fatal error return -1; if the error could be corrected by
638 : : swapping operands of father node of this one, return 1; if everything is
639 : : ok return 0. */
640 : : static int
641 : 12026210 : vect_get_and_check_slp_defs (vec_info *vinfo, unsigned char swap,
642 : : bool *skip_args,
643 : : vec<stmt_vec_info> stmts, unsigned stmt_num,
644 : : vec<slp_oprnd_info> *oprnds_info)
645 : : {
646 : 12026210 : stmt_vec_info stmt_info = stmts[stmt_num];
647 : 12026210 : tree oprnd;
648 : 12026210 : unsigned int i, number_of_oprnds;
649 : 12026210 : enum vect_def_type dt = vect_uninitialized_def;
650 : 12026210 : slp_oprnd_info oprnd_info;
651 : 12026210 : gather_scatter_info gs_info;
652 : 12026210 : unsigned int gs_op = -1u;
653 : 12026210 : unsigned int commutative_op = -1U;
654 : 12026210 : bool first = stmt_num == 0;
655 : :
656 : 12026210 : if (!stmt_info)
657 : : {
658 : 0 : for (auto oi : *oprnds_info)
659 : : {
660 : 0 : oi->def_stmts.quick_push (NULL);
661 : 0 : oi->ops.quick_push (NULL_TREE);
662 : : }
663 : : return 0;
664 : : }
665 : :
666 : 12026210 : if (!is_a<gcall *> (stmt_info->stmt)
667 : : && !is_a<gassign *> (stmt_info->stmt)
668 : : && !is_a<gphi *> (stmt_info->stmt))
669 : : return -1;
670 : :
671 : 12026210 : number_of_oprnds = gimple_num_args (stmt_info->stmt);
672 : 12026210 : const int *map
673 : 24052420 : = vect_get_operand_map (stmt_info->stmt,
674 : 12026210 : STMT_VINFO_GATHER_SCATTER_P (stmt_info), swap);
675 : 12026210 : if (map)
676 : 370375 : number_of_oprnds = *map++;
677 : 12026210 : if (gcall *stmt = dyn_cast <gcall *> (stmt_info->stmt))
678 : : {
679 : 41985 : if (gimple_call_internal_p (stmt))
680 : : {
681 : 23400 : internal_fn ifn = gimple_call_internal_fn (stmt);
682 : 23400 : commutative_op = first_commutative_argument (ifn);
683 : : }
684 : : }
685 : 11984225 : else if (gassign *stmt = dyn_cast <gassign *> (stmt_info->stmt))
686 : : {
687 : 14063159 : if (commutative_tree_code (gimple_assign_rhs_code (stmt)))
688 : 12026210 : commutative_op = 0;
689 : : }
690 : :
691 : 12026210 : bool swapped = (swap != 0);
692 : 12026210 : bool backedge = false;
693 : 12026210 : enum vect_def_type *dts = XALLOCAVEC (enum vect_def_type, number_of_oprnds);
694 : 33169982 : for (i = 0; i < number_of_oprnds; i++)
695 : : {
696 : 21146391 : oprnd_info = (*oprnds_info)[i];
697 : 21146391 : int opno = map ? map[i] : int (i);
698 : 21146391 : if (opno == -3)
699 : : {
700 : 11373 : gcc_assert (STMT_VINFO_GATHER_SCATTER_P (stmt_info));
701 : 11373 : if (!is_a <loop_vec_info> (vinfo)
702 : 11373 : || !vect_check_gather_scatter (stmt_info,
703 : : as_a <loop_vec_info> (vinfo),
704 : : first ? &oprnd_info->first_gs_info
705 : : : &gs_info))
706 : 2619 : return -1;
707 : :
708 : 11373 : if (first)
709 : : {
710 : 11183 : oprnd_info->first_gs_p = true;
711 : 11183 : oprnd = oprnd_info->first_gs_info.offset;
712 : : }
713 : : else
714 : : {
715 : 190 : gs_op = i;
716 : 190 : oprnd = gs_info.offset;
717 : : }
718 : : }
719 : 21135018 : else if (opno < 0)
720 : 0 : oprnd = TREE_OPERAND (gimple_arg (stmt_info->stmt, 0), -1 - opno);
721 : : else
722 : : {
723 : 21135018 : oprnd = gimple_arg (stmt_info->stmt, opno);
724 : 21135018 : if (gphi *stmt = dyn_cast <gphi *> (stmt_info->stmt))
725 : : {
726 : 1104370 : edge e = gimple_phi_arg_edge (stmt, opno);
727 : 2208740 : backedge = (is_a <bb_vec_info> (vinfo)
728 : 1577667 : ? e->flags & EDGE_DFS_BACK
729 : 473297 : : dominated_by_p (CDI_DOMINATORS, e->src,
730 : 473297 : gimple_bb (stmt_info->stmt)));
731 : : }
732 : : }
733 : 21146391 : if (TREE_CODE (oprnd) == VIEW_CONVERT_EXPR)
734 : 2411 : oprnd = TREE_OPERAND (oprnd, 0);
735 : :
736 : 21146391 : stmt_vec_info def_stmt_info;
737 : 21146391 : if (!vect_is_simple_use (oprnd, vinfo, &dts[i], &def_stmt_info))
738 : : {
739 : 1181 : if (dump_enabled_p ())
740 : 4 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
741 : : "Build SLP failed: can't analyze def for %T\n",
742 : : oprnd);
743 : :
744 : 1181 : return -1;
745 : : }
746 : :
747 : 21145210 : if (skip_args[i])
748 : : {
749 : 397781 : oprnd_info->def_stmts.quick_push (NULL);
750 : 397781 : oprnd_info->ops.quick_push (NULL_TREE);
751 : 397781 : oprnd_info->first_dt = vect_uninitialized_def;
752 : 397781 : continue;
753 : : }
754 : :
755 : 20747429 : oprnd_info->def_stmts.quick_push (def_stmt_info);
756 : 20747429 : oprnd_info->ops.quick_push (oprnd);
757 : :
758 : 20747429 : if (def_stmt_info
759 : 20747429 : && is_pattern_stmt_p (def_stmt_info))
760 : : {
761 : 332228 : if (STMT_VINFO_RELATED_STMT (vect_orig_stmt (def_stmt_info))
762 : : != def_stmt_info)
763 : 239082 : oprnd_info->any_pattern = true;
764 : : else
765 : : /* If we promote this to external use the original stmt def. */
766 : 93146 : oprnd_info->ops.last ()
767 : 186292 : = gimple_get_lhs (vect_orig_stmt (def_stmt_info)->stmt);
768 : : }
769 : :
770 : : /* If there's a extern def on a backedge make sure we can
771 : : code-generate at the region start.
772 : : ??? This is another case that could be fixed by adjusting
773 : : how we split the function but at the moment we'd have conflicting
774 : : goals there. */
775 : 20747429 : if (backedge
776 : 121855 : && dts[i] == vect_external_def
777 : 1462 : && is_a <bb_vec_info> (vinfo)
778 : 1462 : && TREE_CODE (oprnd) == SSA_NAME
779 : 1438 : && !SSA_NAME_IS_DEFAULT_DEF (oprnd)
780 : 20748867 : && !dominated_by_p (CDI_DOMINATORS, vinfo->bbs[0],
781 : 1438 : gimple_bb (SSA_NAME_DEF_STMT (oprnd))))
782 : : {
783 : 1438 : if (dump_enabled_p ())
784 : 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
785 : : "Build SLP failed: extern def %T only defined "
786 : : "on backedge\n", oprnd);
787 : 1438 : return -1;
788 : : }
789 : :
790 : 20745991 : if (first)
791 : : {
792 : 3872759 : tree type = TREE_TYPE (oprnd);
793 : 3872759 : dt = dts[i];
794 : :
795 : : /* For the swapping logic below force vect_reduction_def
796 : : for the reduction op in a SLP reduction group. */
797 : 3872759 : if (!STMT_VINFO_DATA_REF (stmt_info)
798 : 2777299 : && REDUC_GROUP_FIRST_ELEMENT (stmt_info)
799 : 1244 : && (int)i == STMT_VINFO_REDUC_IDX (stmt_info)
800 : 3873369 : && def_stmt_info)
801 : 610 : dts[i] = dt = vect_reduction_def;
802 : :
803 : : /* Check the types of the definition. */
804 : 3872759 : switch (dt)
805 : : {
806 : 3872759 : case vect_external_def:
807 : 3872759 : case vect_constant_def:
808 : 3872759 : case vect_internal_def:
809 : 3872759 : case vect_reduction_def:
810 : 3872759 : case vect_double_reduction_def:
811 : 3872759 : case vect_induction_def:
812 : 3872759 : case vect_nested_cycle:
813 : 3872759 : case vect_first_order_recurrence:
814 : 3872759 : break;
815 : :
816 : 0 : default:
817 : : /* FORNOW: Not supported. */
818 : 0 : if (dump_enabled_p ())
819 : 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
820 : : "Build SLP failed: illegal type of def %T\n",
821 : : oprnd);
822 : 0 : return -1;
823 : : }
824 : :
825 : 3872759 : oprnd_info->first_dt = dt;
826 : 3872759 : oprnd_info->first_op_type = type;
827 : : }
828 : : }
829 : 12023591 : if (first)
830 : : return 0;
831 : :
832 : : /* Now match the operand definition types to that of the first stmt. */
833 : 25820133 : for (i = 0; i < number_of_oprnds;)
834 : : {
835 : 16852553 : if (skip_args[i])
836 : : {
837 : 11018 : ++i;
838 : 11018 : continue;
839 : : }
840 : :
841 : 16841535 : oprnd_info = (*oprnds_info)[i];
842 : 16841535 : dt = dts[i];
843 : 16841535 : stmt_vec_info def_stmt_info = oprnd_info->def_stmts[stmt_num];
844 : 16841535 : oprnd = oprnd_info->ops[stmt_num];
845 : 16841535 : tree type = TREE_TYPE (oprnd);
846 : :
847 : 16841535 : if (!types_compatible_p (oprnd_info->first_op_type, type))
848 : : {
849 : 109216 : if (dump_enabled_p ())
850 : 170 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
851 : : "Build SLP failed: different operand types\n");
852 : 109216 : return 1;
853 : : }
854 : :
855 : 16732319 : if ((gs_op == i) != oprnd_info->first_gs_p)
856 : : {
857 : 0 : if (dump_enabled_p ())
858 : 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
859 : : "Build SLP failed: mixed gather and non-gather\n");
860 : 0 : return 1;
861 : : }
862 : 16732319 : else if (gs_op == i)
863 : : {
864 : 182 : if (!operand_equal_p (oprnd_info->first_gs_info.base,
865 : 182 : gs_info.base))
866 : : {
867 : 20 : if (dump_enabled_p ())
868 : 8 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
869 : : "Build SLP failed: different gather base\n");
870 : 20 : return 1;
871 : : }
872 : 162 : if (oprnd_info->first_gs_info.scale != gs_info.scale)
873 : : {
874 : 8 : if (dump_enabled_p ())
875 : 2 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
876 : : "Build SLP failed: different gather scale\n");
877 : 8 : return 1;
878 : : }
879 : : }
880 : :
881 : : /* Not first stmt of the group, check that the def-stmt/s match
882 : : the def-stmt/s of the first stmt. Allow different definition
883 : : types for reduction chains: the first stmt must be a
884 : : vect_reduction_def (a phi node), and the rest
885 : : end in the reduction chain. */
886 : 16732291 : if ((!vect_def_types_match (oprnd_info->first_dt, dt)
887 : 304553 : && !(oprnd_info->first_dt == vect_reduction_def
888 : 5123 : && !STMT_VINFO_DATA_REF (stmt_info)
889 : 5123 : && REDUC_GROUP_FIRST_ELEMENT (stmt_info)
890 : 1705 : && def_stmt_info
891 : 1699 : && !STMT_VINFO_DATA_REF (def_stmt_info)
892 : 1693 : && (REDUC_GROUP_FIRST_ELEMENT (def_stmt_info)
893 : : == REDUC_GROUP_FIRST_ELEMENT (stmt_info))))
894 : 16429208 : || (!STMT_VINFO_DATA_REF (stmt_info)
895 : 15130661 : && REDUC_GROUP_FIRST_ELEMENT (stmt_info)
896 : 4174 : && ((!def_stmt_info
897 : 4084 : || STMT_VINFO_DATA_REF (def_stmt_info)
898 : 7311 : || (REDUC_GROUP_FIRST_ELEMENT (def_stmt_info)
899 : : != REDUC_GROUP_FIRST_ELEMENT (stmt_info)))
900 : 4174 : != (oprnd_info->first_dt != vect_reduction_def))))
901 : : {
902 : : /* Try swapping operands if we got a mismatch. For BB
903 : : vectorization only in case it will clearly improve things. */
904 : 305107 : if (i == commutative_op && !swapped
905 : 303398 : && (!is_a <bb_vec_info> (vinfo)
906 : 2114 : || (!vect_def_types_match ((*oprnds_info)[i+1]->first_dt,
907 : 2114 : dts[i+1])
908 : 465 : && (vect_def_types_match (oprnd_info->first_dt, dts[i+1])
909 : : || vect_def_types_match
910 : 244 : ((*oprnds_info)[i+1]->first_dt, dts[i])))))
911 : : {
912 : 1709 : if (dump_enabled_p ())
913 : 338 : dump_printf_loc (MSG_NOTE, vect_location,
914 : : "trying swapped operands\n");
915 : 1709 : std::swap (dts[i], dts[i+1]);
916 : 1709 : std::swap ((*oprnds_info)[i]->def_stmts[stmt_num],
917 : 1709 : (*oprnds_info)[i+1]->def_stmts[stmt_num]);
918 : 1709 : std::swap ((*oprnds_info)[i]->ops[stmt_num],
919 : 1709 : (*oprnds_info)[i+1]->ops[stmt_num]);
920 : : /* After swapping some operands we lost track whether an
921 : : operand has any pattern defs so be conservative here. */
922 : 1709 : if ((*oprnds_info)[i]->any_pattern
923 : 1709 : || (*oprnds_info)[i+1]->any_pattern)
924 : 8 : (*oprnds_info)[i]->any_pattern
925 : 4 : = (*oprnds_info)[i+1]->any_pattern = true;
926 : 1709 : swapped = true;
927 : 1709 : continue;
928 : : }
929 : :
930 : 301689 : if (is_a <bb_vec_info> (vinfo)
931 : 288207 : && !oprnd_info->any_pattern
932 : 589696 : && number_of_oprnds > 1)
933 : : {
934 : : /* Now for commutative ops we should see whether we can
935 : : make the other operand matching. */
936 : 116079 : if (dump_enabled_p ())
937 : 149 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
938 : : "treating operand as external\n");
939 : 116079 : oprnd_info->first_dt = dt = vect_external_def;
940 : : }
941 : : else
942 : : {
943 : 185610 : if (dump_enabled_p ())
944 : 479 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
945 : : "Build SLP failed: different types\n");
946 : 185610 : return 1;
947 : : }
948 : : }
949 : :
950 : : /* Make sure to demote the overall operand to external. */
951 : 16544972 : if (dt == vect_external_def)
952 : 374792 : oprnd_info->first_dt = vect_external_def;
953 : : /* For a SLP reduction chain we want to duplicate the reduction to
954 : : each of the chain members. That gets us a sane SLP graph (still
955 : : the stmts are not 100% correct wrt the initial values). */
956 : 16170180 : else if ((dt == vect_internal_def
957 : 16170180 : || dt == vect_reduction_def)
958 : 15302318 : && oprnd_info->first_dt == vect_reduction_def
959 : 14456 : && !STMT_VINFO_DATA_REF (stmt_info)
960 : 14456 : && REDUC_GROUP_FIRST_ELEMENT (stmt_info)
961 : 1879 : && !STMT_VINFO_DATA_REF (def_stmt_info)
962 : 16172059 : && (REDUC_GROUP_FIRST_ELEMENT (def_stmt_info)
963 : : == REDUC_GROUP_FIRST_ELEMENT (stmt_info)))
964 : : {
965 : 1879 : oprnd_info->def_stmts[stmt_num] = oprnd_info->def_stmts[0];
966 : 1879 : oprnd_info->ops[stmt_num] = oprnd_info->ops[0];
967 : : }
968 : :
969 : 16544972 : ++i;
970 : : }
971 : :
972 : : /* Swap operands. */
973 : 8967580 : if (swapped)
974 : : {
975 : 121827 : if (dump_enabled_p ())
976 : 1011 : dump_printf_loc (MSG_NOTE, vect_location,
977 : : "swapped operands to match def types in %G",
978 : : stmt_info->stmt);
979 : : }
980 : :
981 : : return 0;
982 : : }
983 : :
984 : : /* Return true if call statements CALL1 and CALL2 are similar enough
985 : : to be combined into the same SLP group. */
986 : :
987 : : bool
988 : 25840 : compatible_calls_p (gcall *call1, gcall *call2)
989 : : {
990 : 25840 : unsigned int nargs = gimple_call_num_args (call1);
991 : 25840 : if (nargs != gimple_call_num_args (call2))
992 : : return false;
993 : :
994 : 22996 : if (gimple_call_combined_fn (call1) != gimple_call_combined_fn (call2))
995 : : return false;
996 : :
997 : 22996 : if (gimple_call_internal_p (call1))
998 : : {
999 : 6365 : if (!types_compatible_p (TREE_TYPE (gimple_call_lhs (call1)),
1000 : 6365 : TREE_TYPE (gimple_call_lhs (call2))))
1001 : : return false;
1002 : 12929 : for (unsigned int i = 0; i < nargs; ++i)
1003 : 6564 : if (!types_compatible_p (TREE_TYPE (gimple_call_arg (call1, i)),
1004 : 6564 : TREE_TYPE (gimple_call_arg (call2, i))))
1005 : : return false;
1006 : : }
1007 : : else
1008 : : {
1009 : 16631 : if (!operand_equal_p (gimple_call_fn (call1),
1010 : 16631 : gimple_call_fn (call2), 0))
1011 : : return false;
1012 : :
1013 : 32862 : if (gimple_call_fntype (call1) != gimple_call_fntype (call2))
1014 : : return false;
1015 : : }
1016 : :
1017 : : /* Check that any unvectorized arguments are equal. */
1018 : 17319 : if (const int *map = vect_get_operand_map (call1))
1019 : : {
1020 : 15 : unsigned int nkept = *map++;
1021 : 15 : unsigned int mapi = 0;
1022 : 57 : for (unsigned int i = 0; i < nargs; ++i)
1023 : 42 : if (mapi < nkept && map[mapi] == int (i))
1024 : 27 : mapi += 1;
1025 : 15 : else if (!operand_equal_p (gimple_call_arg (call1, i),
1026 : 15 : gimple_call_arg (call2, i)))
1027 : : return false;
1028 : : }
1029 : :
1030 : : return true;
1031 : : }
1032 : :
1033 : : /* A subroutine of vect_build_slp_tree for checking VECTYPE, which is the
1034 : : caller's attempt to find the vector type in STMT_INFO with the narrowest
1035 : : element type. Return true if VECTYPE is nonnull and if it is valid
1036 : : for STMT_INFO. When returning true, update MAX_NUNITS to reflect the
1037 : : number of units in VECTYPE. GROUP_SIZE and MAX_NUNITS are as for
1038 : : vect_build_slp_tree. */
1039 : :
1040 : : static bool
1041 : 4593249 : vect_record_max_nunits (vec_info *vinfo, stmt_vec_info stmt_info,
1042 : : unsigned int group_size,
1043 : : tree vectype, poly_uint64 *max_nunits)
1044 : : {
1045 : 4593249 : if (!vectype)
1046 : : {
1047 : 0 : if (dump_enabled_p ())
1048 : 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1049 : : "Build SLP failed: unsupported data-type in %G\n",
1050 : : stmt_info->stmt);
1051 : : /* Fatal mismatch. */
1052 : 0 : return false;
1053 : : }
1054 : :
1055 : : /* If populating the vector type requires unrolling then fail
1056 : : before adjusting *max_nunits for basic-block vectorization. */
1057 : 4593249 : if (is_a <bb_vec_info> (vinfo)
1058 : 4593249 : && !multiple_p (group_size, TYPE_VECTOR_SUBPARTS (vectype)))
1059 : : {
1060 : 130563 : if (dump_enabled_p ())
1061 : 34 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1062 : : "Build SLP failed: unrolling required "
1063 : : "in basic block SLP\n");
1064 : : /* Fatal mismatch. */
1065 : 130563 : return false;
1066 : : }
1067 : :
1068 : : /* In case of multiple types we need to detect the smallest type. */
1069 : 4462686 : vect_update_max_nunits (max_nunits, vectype);
1070 : 4462686 : return true;
1071 : : }
1072 : :
1073 : : /* Verify if the scalar stmts STMTS are isomorphic, require data
1074 : : permutation or are of unsupported types of operation. Return
1075 : : true if they are, otherwise return false and indicate in *MATCHES
1076 : : which stmts are not isomorphic to the first one. If MATCHES[0]
1077 : : is false then this indicates the comparison could not be
1078 : : carried out or the stmts will never be vectorized by SLP.
1079 : :
1080 : : Note COND_EXPR is possibly isomorphic to another one after swapping its
1081 : : operands. Set SWAP[i] to 1 if stmt I is COND_EXPR and isomorphic to
1082 : : the first stmt by swapping the two operands of comparison; set SWAP[i]
1083 : : to 2 if stmt I is isormorphic to the first stmt by inverting the code
1084 : : of comparison. Take A1 >= B1 ? X1 : Y1 as an exmple, it can be swapped
1085 : : to (B1 <= A1 ? X1 : Y1); or be inverted to (A1 < B1) ? Y1 : X1. */
1086 : :
1087 : : static bool
1088 : 4539102 : vect_build_slp_tree_1 (vec_info *vinfo, unsigned char *swap,
1089 : : vec<stmt_vec_info> stmts, unsigned int group_size,
1090 : : poly_uint64 *max_nunits, bool *matches,
1091 : : bool *two_operators, tree *node_vectype)
1092 : : {
1093 : 4539102 : unsigned int i;
1094 : 4539102 : stmt_vec_info first_stmt_info = stmts[0];
1095 : 4539102 : code_helper first_stmt_code = ERROR_MARK;
1096 : 4539102 : code_helper alt_stmt_code = ERROR_MARK;
1097 : 4539102 : code_helper first_cond_code = ERROR_MARK;
1098 : 4539102 : bool need_same_oprnds = false;
1099 : 4539102 : tree first_lhs = NULL_TREE;
1100 : 4539102 : tree first_op1 = NULL_TREE;
1101 : 4539102 : stmt_vec_info first_load = NULL, prev_first_load = NULL;
1102 : 4539102 : bool first_stmt_ldst_p = false;
1103 : 4539102 : bool first_stmt_phi_p = false;
1104 : 4539102 : int first_reduc_idx = -1;
1105 : 4539102 : bool maybe_soft_fail = false;
1106 : 4539102 : tree soft_fail_nunits_vectype = NULL_TREE;
1107 : :
1108 : 4539102 : tree vectype, nunits_vectype;
1109 : 4539102 : if (!vect_get_vector_types_for_stmt (vinfo, first_stmt_info, &vectype,
1110 : 4539102 : &nunits_vectype, group_size))
1111 : : {
1112 : : /* Fatal mismatch. */
1113 : 177053 : matches[0] = false;
1114 : 177053 : return false;
1115 : : }
1116 : : /* Record nunits required but continue analysis, producing matches[]
1117 : : as if nunits was not an issue. This allows splitting of groups
1118 : : to happen. */
1119 : 4362049 : if (nunits_vectype
1120 : 4362049 : && !vect_record_max_nunits (vinfo, first_stmt_info, group_size,
1121 : : nunits_vectype, max_nunits))
1122 : : {
1123 : 130563 : gcc_assert (is_a <bb_vec_info> (vinfo));
1124 : 130563 : maybe_soft_fail = true;
1125 : 130563 : soft_fail_nunits_vectype = nunits_vectype;
1126 : : }
1127 : :
1128 : 4362049 : gcc_assert (vectype);
1129 : 4362049 : *node_vectype = vectype;
1130 : :
1131 : : /* For every stmt in NODE find its def stmt/s. */
1132 : 4362049 : stmt_vec_info stmt_info;
1133 : 20279376 : FOR_EACH_VEC_ELT (stmts, i, stmt_info)
1134 : : {
1135 : 16131595 : bool ldst_p = false;
1136 : 16131595 : bool phi_p = false;
1137 : 16131595 : code_helper rhs_code = ERROR_MARK;
1138 : :
1139 : 16131595 : swap[i] = 0;
1140 : 16131595 : matches[i] = false;
1141 : 16131595 : if (!stmt_info)
1142 : : {
1143 : 70473 : matches[i] = true;
1144 : 15987800 : continue;
1145 : : }
1146 : :
1147 : 16061122 : gimple *stmt = stmt_info->stmt;
1148 : 16061122 : if (dump_enabled_p ())
1149 : 221316 : dump_printf_loc (MSG_NOTE, vect_location, "Build SLP for %G", stmt);
1150 : :
1151 : : /* Fail to vectorize statements marked as unvectorizable, throw
1152 : : or are volatile. */
1153 : 16061122 : if (!STMT_VINFO_VECTORIZABLE (stmt_info)
1154 : 15839407 : || stmt_can_throw_internal (cfun, stmt)
1155 : 31089029 : || gimple_has_volatile_ops (stmt))
1156 : : {
1157 : 228762 : if (dump_enabled_p ())
1158 : 231 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1159 : : "Build SLP failed: unvectorizable statement %G",
1160 : : stmt);
1161 : : /* ??? For BB vectorization we want to commutate operands in a way
1162 : : to shuffle all unvectorizable defs into one operand and have
1163 : : the other still vectorized. The following doesn't reliably
1164 : : work for this though but it's the easiest we can do here. */
1165 : 228762 : if (is_a <bb_vec_info> (vinfo) && i != 0)
1166 : 65604 : continue;
1167 : : /* Fatal mismatch. */
1168 : 163158 : matches[0] = false;
1169 : 163158 : return false;
1170 : : }
1171 : :
1172 : 15832360 : gcall *call_stmt = dyn_cast <gcall *> (stmt);
1173 : 15832360 : tree lhs = gimple_get_lhs (stmt);
1174 : 15832360 : if (lhs == NULL_TREE
1175 : 15832360 : && (!call_stmt
1176 : 1955 : || !gimple_call_internal_p (stmt)
1177 : 1955 : || !internal_store_fn_p (gimple_call_internal_fn (stmt))))
1178 : : {
1179 : 60 : if (dump_enabled_p ())
1180 : 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1181 : : "Build SLP failed: not GIMPLE_ASSIGN nor "
1182 : : "GIMPLE_CALL %G", stmt);
1183 : 60 : if (is_a <bb_vec_info> (vinfo) && i != 0)
1184 : 60 : continue;
1185 : : /* Fatal mismatch. */
1186 : 0 : matches[0] = false;
1187 : 0 : return false;
1188 : : }
1189 : :
1190 : 15832300 : if (call_stmt)
1191 : : {
1192 : 102322 : combined_fn cfn = gimple_call_combined_fn (call_stmt);
1193 : 102322 : if (cfn != CFN_LAST && cfn != CFN_MASK_CALL)
1194 : 51920 : rhs_code = cfn;
1195 : : else
1196 : : rhs_code = CALL_EXPR;
1197 : :
1198 : 102322 : if (cfn == CFN_MASK_LOAD
1199 : 102322 : || cfn == CFN_GATHER_LOAD
1200 : : || cfn == CFN_MASK_GATHER_LOAD
1201 : : || cfn == CFN_MASK_LEN_GATHER_LOAD
1202 : : || cfn == CFN_SCATTER_STORE
1203 : : || cfn == CFN_MASK_SCATTER_STORE
1204 : : || cfn == CFN_MASK_LEN_SCATTER_STORE)
1205 : : ldst_p = true;
1206 : : else if (cfn == CFN_MASK_STORE)
1207 : : {
1208 : : ldst_p = true;
1209 : : rhs_code = CFN_MASK_STORE;
1210 : : }
1211 : : else if (cfn == CFN_GOMP_SIMD_LANE)
1212 : : ;
1213 : 92882 : else if ((cfn != CFN_LAST
1214 : : && cfn != CFN_MASK_CALL
1215 : 42480 : && internal_fn_p (cfn)
1216 : 30209 : && !vectorizable_internal_fn_p (as_internal_fn (cfn)))
1217 : 92817 : || gimple_call_tail_p (call_stmt)
1218 : 92817 : || gimple_call_noreturn_p (call_stmt)
1219 : 185699 : || gimple_call_chain (call_stmt))
1220 : : {
1221 : 665 : if (dump_enabled_p ())
1222 : 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1223 : : "Build SLP failed: unsupported call type %G",
1224 : : (gimple *) call_stmt);
1225 : 665 : if (is_a <bb_vec_info> (vinfo) && i != 0)
1226 : 87 : continue;
1227 : : /* Fatal mismatch. */
1228 : 578 : matches[0] = false;
1229 : 578 : return false;
1230 : : }
1231 : : }
1232 : 15729978 : else if (gimple_code (stmt) == GIMPLE_PHI)
1233 : : {
1234 : : rhs_code = ERROR_MARK;
1235 : : phi_p = true;
1236 : : }
1237 : : else
1238 : : {
1239 : 14918478 : rhs_code = gimple_assign_rhs_code (stmt);
1240 : 14918478 : ldst_p = STMT_VINFO_DATA_REF (stmt_info) != nullptr;
1241 : : }
1242 : :
1243 : : /* Check the operation. */
1244 : 15831635 : if (i == 0)
1245 : : {
1246 : 4198313 : first_lhs = lhs;
1247 : 4198313 : first_stmt_code = rhs_code;
1248 : 4198313 : first_stmt_ldst_p = ldst_p;
1249 : 4198313 : first_stmt_phi_p = phi_p;
1250 : 4198313 : first_reduc_idx = STMT_VINFO_REDUC_IDX (stmt_info);
1251 : :
1252 : : /* Shift arguments should be equal in all the packed stmts for a
1253 : : vector shift with scalar shift operand. */
1254 : 4198313 : if (rhs_code == LSHIFT_EXPR || rhs_code == RSHIFT_EXPR
1255 : 4066954 : || rhs_code == LROTATE_EXPR
1256 : 8265260 : || rhs_code == RROTATE_EXPR)
1257 : : {
1258 : : /* First see if we have a vector/vector shift. */
1259 : 131505 : if (!directly_supported_p (rhs_code, vectype, optab_vector))
1260 : : {
1261 : : /* No vector/vector shift, try for a vector/scalar shift. */
1262 : 118699 : if (!directly_supported_p (rhs_code, vectype, optab_scalar))
1263 : : {
1264 : 29915 : if (dump_enabled_p ())
1265 : 457 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1266 : : "Build SLP failed: "
1267 : : "op not supported by target.\n");
1268 : 29915 : if (is_a <bb_vec_info> (vinfo) && i != 0)
1269 : : continue;
1270 : : /* Fatal mismatch. */
1271 : 29915 : matches[0] = false;
1272 : 29915 : return false;
1273 : : }
1274 : 88784 : need_same_oprnds = true;
1275 : 88784 : first_op1 = gimple_assign_rhs2 (stmt);
1276 : : }
1277 : : }
1278 : 4066808 : else if (rhs_code == WIDEN_LSHIFT_EXPR)
1279 : : {
1280 : 0 : need_same_oprnds = true;
1281 : 0 : first_op1 = gimple_assign_rhs2 (stmt);
1282 : : }
1283 : 4066808 : else if (!ldst_p
1284 : 4066808 : && rhs_code == BIT_FIELD_REF)
1285 : : {
1286 : 5710 : tree vec = TREE_OPERAND (gimple_assign_rhs1 (stmt), 0);
1287 : 5710 : if (!is_a <bb_vec_info> (vinfo)
1288 : 5628 : || TREE_CODE (vec) != SSA_NAME
1289 : : /* When the element types are not compatible we pun the
1290 : : source to the target vectype which requires equal size. */
1291 : 11326 : || ((!VECTOR_TYPE_P (TREE_TYPE (vec))
1292 : 4753 : || !types_compatible_p (TREE_TYPE (vectype),
1293 : 4753 : TREE_TYPE (TREE_TYPE (vec))))
1294 : 1056 : && !operand_equal_p (TYPE_SIZE (vectype),
1295 : 1056 : TYPE_SIZE (TREE_TYPE (vec)))))
1296 : : {
1297 : 802 : if (dump_enabled_p ())
1298 : 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1299 : : "Build SLP failed: "
1300 : : "BIT_FIELD_REF not supported\n");
1301 : : /* Fatal mismatch. */
1302 : 802 : matches[0] = false;
1303 : 802 : return false;
1304 : : }
1305 : : }
1306 : 4061098 : else if (rhs_code == CFN_DIV_POW2)
1307 : : {
1308 : 0 : need_same_oprnds = true;
1309 : 0 : first_op1 = gimple_call_arg (call_stmt, 1);
1310 : : }
1311 : 4061098 : else if (rhs_code == CFN_GOMP_SIMD_LANE)
1312 : : {
1313 : 3646 : need_same_oprnds = true;
1314 : 3646 : first_op1 = gimple_call_arg (call_stmt, 1);
1315 : : }
1316 : : }
1317 : : else
1318 : : {
1319 : 11634300 : if (first_reduc_idx != STMT_VINFO_REDUC_IDX (stmt_info)
1320 : : /* For SLP reduction groups the index isn't necessarily
1321 : : uniform but only that of the first stmt matters. */
1322 : 11633322 : && !(first_reduc_idx != -1
1323 : 1697 : && STMT_VINFO_REDUC_IDX (stmt_info) != -1
1324 : 1697 : && REDUC_GROUP_FIRST_ELEMENT (stmt_info)))
1325 : : {
1326 : 978 : if (dump_enabled_p ())
1327 : : {
1328 : 56 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1329 : : "Build SLP failed: different reduc_idx "
1330 : : "%d instead of %d in %G",
1331 : : STMT_VINFO_REDUC_IDX (stmt_info),
1332 : : first_reduc_idx, stmt);
1333 : : }
1334 : : /* Mismatch. */
1335 : 978 : continue;
1336 : : }
1337 : 11632344 : if (!ldst_p
1338 : 9254235 : && first_stmt_code != rhs_code
1339 : 13108387 : && alt_stmt_code == ERROR_MARK)
1340 : : alt_stmt_code = rhs_code;
1341 : 13093959 : if ((!ldst_p
1342 : 9254235 : && first_stmt_code != rhs_code
1343 : 1476043 : && (first_stmt_code != IMAGPART_EXPR
1344 : 112 : || rhs_code != REALPART_EXPR)
1345 : 1476031 : && (first_stmt_code != REALPART_EXPR
1346 : 238 : || rhs_code != IMAGPART_EXPR)
1347 : : /* Handle mismatches in plus/minus by computing both
1348 : : and merging the results. */
1349 : 1476028 : && !((first_stmt_code == PLUS_EXPR
1350 : 1373101 : || first_stmt_code == MINUS_EXPR)
1351 : 125514 : && (alt_stmt_code == PLUS_EXPR
1352 : 115609 : || alt_stmt_code == MINUS_EXPR)
1353 : 19316 : && rhs_code == alt_stmt_code)
1354 : 1456890 : && !(first_stmt_code.is_tree_code ()
1355 : 1351046 : && rhs_code.is_tree_code ()
1356 : 1240276 : && (TREE_CODE_CLASS (tree_code (first_stmt_code))
1357 : : == tcc_comparison)
1358 : 131170 : && (swap_tree_comparison (tree_code (first_stmt_code))
1359 : 131170 : == tree_code (rhs_code))))
1360 : : || (ldst_p
1361 : 4756218 : && (STMT_VINFO_GROUPED_ACCESS (stmt_info)
1362 : 2378109 : != STMT_VINFO_GROUPED_ACCESS (first_stmt_info)))
1363 : : || (ldst_p
1364 : 2331696 : && (STMT_VINFO_GATHER_SCATTER_P (stmt_info)
1365 : 2331696 : != STMT_VINFO_GATHER_SCATTER_P (first_stmt_info)))
1366 : 10170844 : || first_stmt_ldst_p != ldst_p
1367 : 21803188 : || first_stmt_phi_p != phi_p)
1368 : : {
1369 : 1461615 : if (dump_enabled_p ())
1370 : : {
1371 : 3236 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1372 : : "Build SLP failed: different operation "
1373 : : "in stmt %G", stmt);
1374 : 3236 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1375 : : "original stmt %G", first_stmt_info->stmt);
1376 : : }
1377 : : /* Mismatch. */
1378 : 1461615 : continue;
1379 : : }
1380 : :
1381 : 10172552 : if (!ldst_p
1382 : 7839135 : && first_stmt_code == BIT_FIELD_REF
1383 : 10175878 : && (TREE_OPERAND (gimple_assign_rhs1 (first_stmt_info->stmt), 0)
1384 : 5149 : != TREE_OPERAND (gimple_assign_rhs1 (stmt_info->stmt), 0)))
1385 : : {
1386 : 1823 : if (dump_enabled_p ())
1387 : 36 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1388 : : "Build SLP failed: different BIT_FIELD_REF "
1389 : : "arguments in %G", stmt);
1390 : : /* Mismatch. */
1391 : 1823 : continue;
1392 : : }
1393 : :
1394 : 10168906 : if (call_stmt
1395 : 25951 : && first_stmt_code != CFN_MASK_LOAD
1396 : 10194796 : && first_stmt_code != CFN_MASK_STORE)
1397 : : {
1398 : 25840 : if (!is_a <gcall *> (stmts[0]->stmt)
1399 : 25840 : || !compatible_calls_p (as_a <gcall *> (stmts[0]->stmt),
1400 : : call_stmt))
1401 : : {
1402 : 8521 : if (dump_enabled_p ())
1403 : 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1404 : : "Build SLP failed: different calls in %G",
1405 : : stmt);
1406 : : /* Mismatch. */
1407 : 8521 : continue;
1408 : : }
1409 : : }
1410 : :
1411 : 9965892 : if ((phi_p || gimple_could_trap_p (stmt_info->stmt))
1412 : 10802524 : && (gimple_bb (first_stmt_info->stmt)
1413 : 836632 : != gimple_bb (stmt_info->stmt)))
1414 : : {
1415 : 28404 : if (dump_enabled_p ())
1416 : 4 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1417 : : "Build SLP failed: different BB for PHI "
1418 : : "or possibly trapping operation in %G", stmt);
1419 : : /* Mismatch. */
1420 : 28404 : continue;
1421 : : }
1422 : :
1423 : 10131981 : if (need_same_oprnds)
1424 : : {
1425 : 51631 : tree other_op1 = gimple_arg (stmt, 1);
1426 : 51631 : if (!operand_equal_p (first_op1, other_op1, 0))
1427 : : {
1428 : 6735 : if (dump_enabled_p ())
1429 : 121 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1430 : : "Build SLP failed: different shift "
1431 : : "arguments in %G", stmt);
1432 : : /* Mismatch. */
1433 : 6735 : continue;
1434 : : }
1435 : : }
1436 : :
1437 : 10125837 : if (first_lhs
1438 : 10125246 : && lhs
1439 : 10125246 : && !types_compatible_p (TREE_TYPE (lhs), TREE_TYPE (first_lhs)))
1440 : : {
1441 : 591 : if (dump_enabled_p ())
1442 : 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1443 : : "Build SLP failed: different vector type "
1444 : : "in %G", stmt);
1445 : : /* Mismatch. */
1446 : 591 : continue;
1447 : : }
1448 : : }
1449 : :
1450 : : /* Grouped store or load. */
1451 : 14292251 : if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
1452 : : {
1453 : 3690141 : gcc_assert (ldst_p);
1454 : 3690141 : if (DR_IS_WRITE (STMT_VINFO_DATA_REF (stmt_info)))
1455 : : {
1456 : : /* Store. */
1457 : 3021510 : gcc_assert (rhs_code == CFN_MASK_STORE
1458 : : || REFERENCE_CLASS_P (lhs)
1459 : : || DECL_P (lhs));
1460 : : }
1461 : : else
1462 : : {
1463 : : /* Load. */
1464 : 668631 : first_load = DR_GROUP_FIRST_ELEMENT (stmt_info);
1465 : 668631 : if (prev_first_load)
1466 : : {
1467 : : /* Check that there are no loads from different interleaving
1468 : : chains in the same node. */
1469 : 316652 : if (prev_first_load != first_load)
1470 : : {
1471 : 24843 : if (dump_enabled_p ())
1472 : 1886 : dump_printf_loc (MSG_MISSED_OPTIMIZATION,
1473 : : vect_location,
1474 : : "Build SLP failed: different "
1475 : : "interleaving chains in one node %G",
1476 : : stmt);
1477 : : /* Mismatch. */
1478 : 24843 : continue;
1479 : : }
1480 : : }
1481 : : else
1482 : : prev_first_load = first_load;
1483 : : }
1484 : : }
1485 : : /* Non-grouped store or load. */
1486 : 10602110 : else if (ldst_p)
1487 : : {
1488 : 425704 : if (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info))
1489 : 251472 : && rhs_code != CFN_GATHER_LOAD
1490 : : && rhs_code != CFN_MASK_GATHER_LOAD
1491 : : && rhs_code != CFN_MASK_LEN_GATHER_LOAD
1492 : : && rhs_code != CFN_SCATTER_STORE
1493 : : && rhs_code != CFN_MASK_SCATTER_STORE
1494 : : && rhs_code != CFN_MASK_LEN_SCATTER_STORE
1495 : 251472 : && !STMT_VINFO_GATHER_SCATTER_P (stmt_info)
1496 : : /* Not grouped loads are handled as externals for BB
1497 : : vectorization. For loop vectorization we can handle
1498 : : splats the same we handle single element interleaving. */
1499 : 670768 : && (is_a <bb_vec_info> (vinfo)
1500 : 245064 : || stmt_info != first_stmt_info))
1501 : : {
1502 : : /* Not grouped load. */
1503 : 15655 : if (dump_enabled_p ())
1504 : 138 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1505 : : "Build SLP failed: not grouped load %G", stmt);
1506 : :
1507 : 15655 : if (i != 0)
1508 : 15655 : continue;
1509 : : /* Fatal mismatch. */
1510 : 0 : matches[0] = false;
1511 : 0 : return false;
1512 : : }
1513 : : }
1514 : : /* Not memory operation. */
1515 : : else
1516 : : {
1517 : 10176406 : if (!phi_p
1518 : 9505407 : && rhs_code.is_tree_code ()
1519 : 9461703 : && TREE_CODE_CLASS (tree_code (rhs_code)) != tcc_binary
1520 : 1253440 : && TREE_CODE_CLASS (tree_code (rhs_code)) != tcc_unary
1521 : 722793 : && TREE_CODE_CLASS (tree_code (rhs_code)) != tcc_expression
1522 : 663959 : && TREE_CODE_CLASS (tree_code (rhs_code)) != tcc_comparison
1523 : 66943 : && rhs_code != VIEW_CONVERT_EXPR
1524 : : && rhs_code != CALL_EXPR
1525 : 10176406 : && rhs_code != BIT_FIELD_REF)
1526 : : {
1527 : 19815 : if (dump_enabled_p ())
1528 : 9 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1529 : : "Build SLP failed: operation unsupported %G",
1530 : : stmt);
1531 : 19815 : if (is_a <bb_vec_info> (vinfo) && i != 0)
1532 : 0 : continue;
1533 : : /* Fatal mismatch. */
1534 : 19815 : matches[0] = false;
1535 : 19815 : return false;
1536 : : }
1537 : :
1538 : 10156591 : if (rhs_code == COND_EXPR)
1539 : : {
1540 : 56516 : tree cond_expr = gimple_assign_rhs1 (stmt);
1541 : 56516 : enum tree_code cond_code = TREE_CODE (cond_expr);
1542 : 56516 : enum tree_code swap_code = ERROR_MARK;
1543 : 56516 : enum tree_code invert_code = ERROR_MARK;
1544 : :
1545 : 56516 : if (i == 0)
1546 : 47641 : first_cond_code = TREE_CODE (cond_expr);
1547 : 8875 : else if (TREE_CODE_CLASS (cond_code) == tcc_comparison)
1548 : : {
1549 : 0 : bool honor_nans = HONOR_NANS (TREE_OPERAND (cond_expr, 0));
1550 : 0 : swap_code = swap_tree_comparison (cond_code);
1551 : 0 : invert_code = invert_tree_comparison (cond_code, honor_nans);
1552 : : }
1553 : :
1554 : 56516 : if (first_cond_code == cond_code)
1555 : : ;
1556 : : /* Isomorphic can be achieved by swapping. */
1557 : 0 : else if (first_cond_code == swap_code)
1558 : 0 : swap[i] = 1;
1559 : : /* Isomorphic can be achieved by inverting. */
1560 : 0 : else if (first_cond_code == invert_code)
1561 : 0 : swap[i] = 2;
1562 : : else
1563 : : {
1564 : 0 : if (dump_enabled_p ())
1565 : 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
1566 : : "Build SLP failed: different"
1567 : : " operation %G", stmt);
1568 : : /* Mismatch. */
1569 : 0 : continue;
1570 : : }
1571 : : }
1572 : :
1573 : 10156591 : if (rhs_code.is_tree_code ()
1574 : 9441888 : && TREE_CODE_CLASS ((tree_code)rhs_code) == tcc_comparison
1575 : 10753607 : && (swap_tree_comparison ((tree_code)first_stmt_code)
1576 : 597016 : == (tree_code)rhs_code))
1577 : 399183 : swap[i] = 1;
1578 : : }
1579 : :
1580 : 14231938 : matches[i] = true;
1581 : : }
1582 : :
1583 : 18427980 : for (i = 0; i < group_size; ++i)
1584 : 15033491 : if (!matches[i])
1585 : : return false;
1586 : :
1587 : : /* If we allowed a two-operation SLP node verify the target can cope
1588 : : with the permute we are going to use. */
1589 : 3394489 : if (alt_stmt_code != ERROR_MARK
1590 : 3394489 : && (!alt_stmt_code.is_tree_code ()
1591 : 53716 : || (TREE_CODE_CLASS (tree_code (alt_stmt_code)) != tcc_reference
1592 : 53716 : && TREE_CODE_CLASS (tree_code (alt_stmt_code)) != tcc_comparison)))
1593 : : {
1594 : 12123 : *two_operators = true;
1595 : : }
1596 : :
1597 : 3394489 : if (maybe_soft_fail)
1598 : : {
1599 : 130058 : unsigned HOST_WIDE_INT const_nunits;
1600 : 130058 : if (!TYPE_VECTOR_SUBPARTS
1601 : 130058 : (soft_fail_nunits_vectype).is_constant (&const_nunits)
1602 : 130058 : || const_nunits > group_size)
1603 : 0 : matches[0] = false;
1604 : : else
1605 : : {
1606 : : /* With constant vector elements simulate a mismatch at the
1607 : : point we need to split. */
1608 : 130058 : unsigned tail = group_size & (const_nunits - 1);
1609 : 130058 : memset (&matches[group_size - tail], 0, sizeof (bool) * tail);
1610 : : }
1611 : 130058 : return false;
1612 : : }
1613 : :
1614 : : return true;
1615 : : }
1616 : :
1617 : : /* Traits for the hash_set to record failed SLP builds for a stmt set.
1618 : : Note we never remove apart from at destruction time so we do not
1619 : : need a special value for deleted that differs from empty. */
1620 : : struct bst_traits
1621 : : {
1622 : : typedef vec <stmt_vec_info> value_type;
1623 : : typedef vec <stmt_vec_info> compare_type;
1624 : : static inline hashval_t hash (value_type);
1625 : : static inline bool equal (value_type existing, value_type candidate);
1626 : 378744795 : static inline bool is_empty (value_type x) { return !x.exists (); }
1627 : 86676788 : static inline bool is_deleted (value_type x) { return !x.exists (); }
1628 : : static const bool empty_zero_p = true;
1629 : 0 : static inline void mark_empty (value_type &x) { x.release (); }
1630 : : static inline void mark_deleted (value_type &x) { x.release (); }
1631 : 7234441 : static inline void remove (value_type &x) { x.release (); }
1632 : : };
1633 : : inline hashval_t
1634 : 75314664 : bst_traits::hash (value_type x)
1635 : : {
1636 : 75314664 : inchash::hash h;
1637 : 372967167 : for (unsigned i = 0; i < x.length (); ++i)
1638 : 297652503 : h.add_int (x[i] ? gimple_uid (x[i]->stmt) : -1);
1639 : 75314664 : return h.end ();
1640 : : }
1641 : : inline bool
1642 : 66433938 : bst_traits::equal (value_type existing, value_type candidate)
1643 : : {
1644 : 199301814 : if (existing.length () != candidate.length ())
1645 : : return false;
1646 : 69222483 : for (unsigned i = 0; i < existing.length (); ++i)
1647 : 65886230 : if (existing[i] != candidate[i])
1648 : : return false;
1649 : : return true;
1650 : : }
1651 : :
1652 : : typedef hash_map <vec <stmt_vec_info>, slp_tree,
1653 : : simple_hashmap_traits <bst_traits, slp_tree> >
1654 : : scalar_stmts_to_slp_tree_map_t;
1655 : :
1656 : : /* Release BST_MAP. */
1657 : :
1658 : : static void
1659 : 1391188 : release_scalar_stmts_to_slp_tree_map (scalar_stmts_to_slp_tree_map_t *bst_map)
1660 : : {
1661 : : /* The map keeps a reference on SLP nodes built, release that. */
1662 : 8625629 : for (scalar_stmts_to_slp_tree_map_t::iterator it = bst_map->begin ();
1663 : 15860070 : it != bst_map->end (); ++it)
1664 : 7234441 : if ((*it).second)
1665 : 7234441 : vect_free_slp_tree ((*it).second);
1666 : 1391188 : delete bst_map;
1667 : 1391188 : }
1668 : :
1669 : : /* ??? This was std::pair<std::pair<tree_code, vect_def_type>, tree>
1670 : : but then vec::insert does memmove and that's not compatible with
1671 : : std::pair. */
1672 : : struct chain_op_t
1673 : : {
1674 : 3328529 : chain_op_t (tree_code code_, vect_def_type dt_, tree op_)
1675 : 3328529 : : code (code_), dt (dt_), op (op_) {}
1676 : : tree_code code;
1677 : : vect_def_type dt;
1678 : : tree op;
1679 : : };
1680 : :
1681 : : /* Comparator for sorting associatable chains. */
1682 : :
1683 : : static int
1684 : 7964241 : dt_sort_cmp (const void *op1_, const void *op2_, void *)
1685 : : {
1686 : 7964241 : auto *op1 = (const chain_op_t *) op1_;
1687 : 7964241 : auto *op2 = (const chain_op_t *) op2_;
1688 : 7964241 : if (op1->dt != op2->dt)
1689 : 959392 : return (int)op1->dt - (int)op2->dt;
1690 : 7004849 : return (int)op1->code - (int)op2->code;
1691 : : }
1692 : :
1693 : : /* Linearize the associatable expression chain at START with the
1694 : : associatable operation CODE (where PLUS_EXPR also allows MINUS_EXPR),
1695 : : filling CHAIN with the result and using WORKLIST as intermediate storage.
1696 : : CODE_STMT and ALT_CODE_STMT are filled with the first stmt using CODE
1697 : : or MINUS_EXPR. *CHAIN_STMTS if not NULL is filled with all computation
1698 : : stmts, starting with START. */
1699 : :
1700 : : static void
1701 : 1485744 : vect_slp_linearize_chain (vec_info *vinfo,
1702 : : vec<std::pair<tree_code, gimple *> > &worklist,
1703 : : vec<chain_op_t> &chain,
1704 : : enum tree_code code, gimple *start,
1705 : : gimple *&code_stmt, gimple *&alt_code_stmt,
1706 : : vec<gimple *> *chain_stmts)
1707 : : {
1708 : : /* For each lane linearize the addition/subtraction (or other
1709 : : uniform associatable operation) expression tree. */
1710 : 1485744 : worklist.safe_push (std::make_pair (code, start));
1711 : 3328529 : while (!worklist.is_empty ())
1712 : : {
1713 : 1842785 : auto entry = worklist.pop ();
1714 : 1842785 : gassign *stmt = as_a <gassign *> (entry.second);
1715 : 1842785 : enum tree_code in_code = entry.first;
1716 : 3685570 : enum tree_code this_code = gimple_assign_rhs_code (stmt);
1717 : : /* Pick some stmts suitable for SLP_TREE_REPRESENTATIVE. */
1718 : 1842785 : if (!code_stmt
1719 : 1842785 : && gimple_assign_rhs_code (stmt) == code)
1720 : 1255303 : code_stmt = stmt;
1721 : 587482 : else if (!alt_code_stmt
1722 : 587482 : && gimple_assign_rhs_code (stmt) == MINUS_EXPR)
1723 : 301609 : alt_code_stmt = stmt;
1724 : 1842785 : if (chain_stmts)
1725 : 1836618 : chain_stmts->safe_push (stmt);
1726 : 5528355 : for (unsigned opnum = 1; opnum <= 2; ++opnum)
1727 : : {
1728 : 3685570 : tree op = gimple_op (stmt, opnum);
1729 : 3685570 : vect_def_type dt;
1730 : 3685570 : stmt_vec_info def_stmt_info;
1731 : 3685570 : bool res = vect_is_simple_use (op, vinfo, &dt, &def_stmt_info);
1732 : 3685570 : gcc_assert (res);
1733 : 3685570 : if (dt == vect_internal_def
1734 : 3685570 : && is_pattern_stmt_p (def_stmt_info))
1735 : 998 : op = gimple_get_lhs (def_stmt_info->stmt);
1736 : 3685570 : gimple *use_stmt;
1737 : 3685570 : use_operand_p use_p;
1738 : 3685570 : if (dt == vect_internal_def
1739 : 3449473 : && single_imm_use (op, &use_p, &use_stmt)
1740 : 2169800 : && is_gimple_assign (def_stmt_info->stmt)
1741 : 5683805 : && (gimple_assign_rhs_code (def_stmt_info->stmt) == code
1742 : 1665105 : || (code == PLUS_EXPR
1743 : 793517 : && (gimple_assign_rhs_code (def_stmt_info->stmt)
1744 : : == MINUS_EXPR))))
1745 : : {
1746 : 357041 : tree_code op_def_code = this_code;
1747 : 357041 : if (op_def_code == MINUS_EXPR && opnum == 1)
1748 : 53134 : op_def_code = PLUS_EXPR;
1749 : 357041 : if (in_code == MINUS_EXPR)
1750 : 193 : op_def_code = op_def_code == PLUS_EXPR ? MINUS_EXPR : PLUS_EXPR;
1751 : 357041 : worklist.safe_push (std::make_pair (op_def_code,
1752 : 357041 : def_stmt_info->stmt));
1753 : : }
1754 : : else
1755 : : {
1756 : 3328529 : tree_code op_def_code = this_code;
1757 : 3328529 : if (op_def_code == MINUS_EXPR && opnum == 1)
1758 : 252417 : op_def_code = PLUS_EXPR;
1759 : 3328529 : if (in_code == MINUS_EXPR)
1760 : 4227 : op_def_code = op_def_code == PLUS_EXPR ? MINUS_EXPR : PLUS_EXPR;
1761 : 3328529 : chain.safe_push (chain_op_t (op_def_code, dt, op));
1762 : : }
1763 : : }
1764 : : }
1765 : 1485744 : }
1766 : :
1767 : : static slp_tree
1768 : : vect_build_slp_tree_2 (vec_info *vinfo, slp_tree node,
1769 : : vec<stmt_vec_info> stmts, unsigned int group_size,
1770 : : poly_uint64 *max_nunits,
1771 : : bool *matches, unsigned *limit, unsigned *tree_size,
1772 : : scalar_stmts_to_slp_tree_map_t *bst_map);
1773 : :
1774 : : static slp_tree
1775 : 4885685 : vect_build_slp_tree (vec_info *vinfo,
1776 : : vec<stmt_vec_info> stmts, unsigned int group_size,
1777 : : poly_uint64 *max_nunits,
1778 : : bool *matches, unsigned *limit, unsigned *tree_size,
1779 : : scalar_stmts_to_slp_tree_map_t *bst_map)
1780 : : {
1781 : 4885685 : if (slp_tree *leader = bst_map->get (stmts))
1782 : : {
1783 : 345135 : if (dump_enabled_p ())
1784 : 18721 : dump_printf_loc (MSG_NOTE, vect_location, "re-using %sSLP tree %p\n",
1785 : 18721 : !(*leader)->failed ? "" : "failed ",
1786 : : (void *) *leader);
1787 : 345135 : if (!(*leader)->failed)
1788 : : {
1789 : 299123 : SLP_TREE_REF_COUNT (*leader)++;
1790 : 299123 : vect_update_max_nunits (max_nunits, (*leader)->max_nunits);
1791 : 299123 : stmts.release ();
1792 : 299123 : return *leader;
1793 : : }
1794 : 46012 : memcpy (matches, (*leader)->failed, sizeof (bool) * group_size);
1795 : 46012 : return NULL;
1796 : : }
1797 : :
1798 : : /* Single-lane SLP doesn't have the chance of run-away, do not account
1799 : : it to the limit. */
1800 : 4540550 : if (stmts.length () > 1)
1801 : : {
1802 : 2984284 : if (*limit == 0)
1803 : : {
1804 : 1347 : if (dump_enabled_p ())
1805 : 48 : dump_printf_loc (MSG_NOTE, vect_location,
1806 : : "SLP discovery limit exceeded\n");
1807 : 1347 : memset (matches, 0, sizeof (bool) * group_size);
1808 : 1347 : return NULL;
1809 : : }
1810 : 2982937 : --*limit;
1811 : : }
1812 : :
1813 : : /* Seed the bst_map with a stub node to be filled by vect_build_slp_tree_2
1814 : : so we can pick up backedge destinations during discovery. */
1815 : 4539203 : slp_tree res = new _slp_tree;
1816 : 4539203 : SLP_TREE_DEF_TYPE (res) = vect_internal_def;
1817 : 4539203 : SLP_TREE_SCALAR_STMTS (res) = stmts;
1818 : 4539203 : bst_map->put (stmts.copy (), res);
1819 : :
1820 : 4539203 : if (dump_enabled_p ())
1821 : 150439 : dump_printf_loc (MSG_NOTE, vect_location,
1822 : : "starting SLP discovery for node %p\n", (void *) res);
1823 : :
1824 : 4539203 : poly_uint64 this_max_nunits = 1;
1825 : 4539203 : slp_tree res_ = vect_build_slp_tree_2 (vinfo, res, stmts, group_size,
1826 : : &this_max_nunits,
1827 : : matches, limit, tree_size, bst_map);
1828 : 4539203 : if (!res_)
1829 : : {
1830 : 1856591 : if (dump_enabled_p ())
1831 : 8453 : dump_printf_loc (MSG_NOTE, vect_location,
1832 : : "SLP discovery for node %p failed\n", (void *) res);
1833 : : /* Mark the node invalid so we can detect those when still in use
1834 : : as backedge destinations. */
1835 : 1856591 : SLP_TREE_SCALAR_STMTS (res) = vNULL;
1836 : 1856591 : SLP_TREE_DEF_TYPE (res) = vect_uninitialized_def;
1837 : 1856591 : res->failed = XNEWVEC (bool, group_size);
1838 : 1856591 : if (flag_checking)
1839 : : {
1840 : : unsigned i;
1841 : 3541086 : for (i = 0; i < group_size; ++i)
1842 : 3541086 : if (!matches[i])
1843 : : break;
1844 : 1856591 : gcc_assert (i < group_size);
1845 : : }
1846 : 1856591 : memcpy (res->failed, matches, sizeof (bool) * group_size);
1847 : : }
1848 : : else
1849 : : {
1850 : 2682612 : if (dump_enabled_p ())
1851 : 141986 : dump_printf_loc (MSG_NOTE, vect_location,
1852 : : "SLP discovery for node %p succeeded\n",
1853 : : (void *) res);
1854 : 2682612 : gcc_assert (res_ == res);
1855 : 2682612 : res->max_nunits = this_max_nunits;
1856 : 2682612 : vect_update_max_nunits (max_nunits, this_max_nunits);
1857 : : /* Keep a reference for the bst_map use. */
1858 : 2682612 : SLP_TREE_REF_COUNT (res)++;
1859 : : }
1860 : : return res_;
1861 : : }
1862 : :
1863 : : /* Helper for building an associated SLP node chain. */
1864 : :
1865 : : static void
1866 : 165 : vect_slp_build_two_operator_nodes (slp_tree perm, tree vectype,
1867 : : slp_tree op0, slp_tree op1,
1868 : : stmt_vec_info oper1, stmt_vec_info oper2,
1869 : : vec<std::pair<unsigned, unsigned> > lperm)
1870 : : {
1871 : 165 : unsigned group_size = SLP_TREE_LANES (op1);
1872 : :
1873 : 165 : slp_tree child1 = new _slp_tree;
1874 : 165 : SLP_TREE_DEF_TYPE (child1) = vect_internal_def;
1875 : 165 : SLP_TREE_VECTYPE (child1) = vectype;
1876 : 165 : SLP_TREE_LANES (child1) = group_size;
1877 : 165 : SLP_TREE_CHILDREN (child1).create (2);
1878 : 165 : SLP_TREE_CHILDREN (child1).quick_push (op0);
1879 : 165 : SLP_TREE_CHILDREN (child1).quick_push (op1);
1880 : 165 : SLP_TREE_REPRESENTATIVE (child1) = oper1;
1881 : :
1882 : 165 : slp_tree child2 = new _slp_tree;
1883 : 165 : SLP_TREE_DEF_TYPE (child2) = vect_internal_def;
1884 : 165 : SLP_TREE_VECTYPE (child2) = vectype;
1885 : 165 : SLP_TREE_LANES (child2) = group_size;
1886 : 165 : SLP_TREE_CHILDREN (child2).create (2);
1887 : 165 : SLP_TREE_CHILDREN (child2).quick_push (op0);
1888 : 165 : SLP_TREE_REF_COUNT (op0)++;
1889 : 165 : SLP_TREE_CHILDREN (child2).quick_push (op1);
1890 : 165 : SLP_TREE_REF_COUNT (op1)++;
1891 : 165 : SLP_TREE_REPRESENTATIVE (child2) = oper2;
1892 : :
1893 : 165 : SLP_TREE_DEF_TYPE (perm) = vect_internal_def;
1894 : 165 : SLP_TREE_CODE (perm) = VEC_PERM_EXPR;
1895 : 165 : SLP_TREE_VECTYPE (perm) = vectype;
1896 : 165 : SLP_TREE_LANES (perm) = group_size;
1897 : : /* ??? We should set this NULL but that's not expected. */
1898 : 165 : SLP_TREE_REPRESENTATIVE (perm) = oper1;
1899 : 165 : SLP_TREE_LANE_PERMUTATION (perm) = lperm;
1900 : 165 : SLP_TREE_CHILDREN (perm).quick_push (child1);
1901 : 165 : SLP_TREE_CHILDREN (perm).quick_push (child2);
1902 : 165 : }
1903 : :
1904 : : /* Recursively build an SLP tree starting from NODE.
1905 : : Fail (and return a value not equal to zero) if def-stmts are not
1906 : : isomorphic, require data permutation or are of unsupported types of
1907 : : operation. Otherwise, return 0.
1908 : : The value returned is the depth in the SLP tree where a mismatch
1909 : : was found. */
1910 : :
1911 : : static slp_tree
1912 : 4539203 : vect_build_slp_tree_2 (vec_info *vinfo, slp_tree node,
1913 : : vec<stmt_vec_info> stmts, unsigned int group_size,
1914 : : poly_uint64 *max_nunits,
1915 : : bool *matches, unsigned *limit, unsigned *tree_size,
1916 : : scalar_stmts_to_slp_tree_map_t *bst_map)
1917 : : {
1918 : 4539203 : unsigned nops, i, this_tree_size = 0;
1919 : 4539203 : poly_uint64 this_max_nunits = *max_nunits;
1920 : :
1921 : 4539203 : matches[0] = false;
1922 : :
1923 : 4539203 : stmt_vec_info stmt_info = stmts[0];
1924 : 4539203 : if (!is_a<gcall *> (stmt_info->stmt)
1925 : : && !is_a<gassign *> (stmt_info->stmt)
1926 : : && !is_a<gphi *> (stmt_info->stmt))
1927 : : return NULL;
1928 : :
1929 : 4539102 : nops = gimple_num_args (stmt_info->stmt);
1930 : 4539102 : if (const int *map = vect_get_operand_map (stmt_info->stmt,
1931 : 4539102 : STMT_VINFO_GATHER_SCATTER_P
1932 : : (stmt_info)))
1933 : 19008 : nops = map[0];
1934 : :
1935 : : /* If the SLP node is a PHI (induction or reduction), terminate
1936 : : the recursion. */
1937 : 4539102 : bool *skip_args = XALLOCAVEC (bool, nops);
1938 : 4539102 : memset (skip_args, 0, sizeof (bool) * nops);
1939 : 4539102 : if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
1940 : 1641613 : if (gphi *stmt = dyn_cast <gphi *> (stmt_info->stmt))
1941 : : {
1942 : 231200 : tree scalar_type = TREE_TYPE (PHI_RESULT (stmt));
1943 : 231200 : tree vectype = get_vectype_for_scalar_type (vinfo, scalar_type,
1944 : : group_size);
1945 : 231200 : if (!vect_record_max_nunits (vinfo, stmt_info, group_size, vectype,
1946 : : max_nunits))
1947 : : return NULL;
1948 : :
1949 : 231200 : vect_def_type def_type = STMT_VINFO_DEF_TYPE (stmt_info);
1950 : 231200 : if (def_type == vect_induction_def)
1951 : : {
1952 : : /* Induction PHIs are not cycles but walk the initial
1953 : : value. Only for inner loops through, for outer loops
1954 : : we need to pick up the value from the actual PHIs
1955 : : to more easily support peeling and epilogue vectorization. */
1956 : 160178 : class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
1957 : 160178 : if (!nested_in_vect_loop_p (loop, stmt_info))
1958 : 159391 : skip_args[loop_preheader_edge (loop)->dest_idx] = true;
1959 : : else
1960 : : loop = loop->inner;
1961 : 160178 : skip_args[loop_latch_edge (loop)->dest_idx] = true;
1962 : : }
1963 : 71022 : else if (def_type == vect_reduction_def
1964 : : || def_type == vect_double_reduction_def
1965 : : || def_type == vect_nested_cycle
1966 : 71022 : || def_type == vect_first_order_recurrence)
1967 : : {
1968 : : /* Else def types have to match. */
1969 : : stmt_vec_info other_info;
1970 : : bool all_same = true;
1971 : 141875 : FOR_EACH_VEC_ELT (stmts, i, other_info)
1972 : : {
1973 : 72084 : if (STMT_VINFO_DEF_TYPE (other_info) != def_type)
1974 : 1505927 : return NULL;
1975 : 72084 : if (other_info != stmt_info)
1976 : 756 : all_same = false;
1977 : : }
1978 : 69791 : class loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
1979 : : /* Reduction initial values are not explicitely represented. */
1980 : 69791 : if (def_type != vect_first_order_recurrence
1981 : 69791 : && gimple_bb (stmt_info->stmt) == loop->header)
1982 : 66828 : skip_args[loop_preheader_edge (loop)->dest_idx] = true;
1983 : : /* Reduction chain backedge defs are filled manually.
1984 : : ??? Need a better way to identify a SLP reduction chain PHI.
1985 : : Or a better overall way to SLP match those. */
1986 : 69791 : if (stmts.length () > 1
1987 : 69791 : && all_same && def_type == vect_reduction_def)
1988 : 366 : skip_args[loop_latch_edge (loop)->dest_idx] = true;
1989 : : }
1990 : 1231 : else if (def_type != vect_internal_def)
1991 : : return NULL;
1992 : : }
1993 : :
1994 : :
1995 : 4539102 : bool two_operators = false;
1996 : 4539102 : unsigned char *swap = XALLOCAVEC (unsigned char, group_size);
1997 : 4539102 : tree vectype = NULL_TREE;
1998 : 4539102 : if (!vect_build_slp_tree_1 (vinfo, swap, stmts, group_size,
1999 : : &this_max_nunits, matches, &two_operators,
2000 : : &vectype))
2001 : : return NULL;
2002 : :
2003 : : /* If the SLP node is a load, terminate the recursion unless masked. */
2004 : 3264431 : if (STMT_VINFO_DATA_REF (stmt_info)
2005 : 1581719 : && DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info)))
2006 : : {
2007 : 506459 : if (STMT_VINFO_GATHER_SCATTER_P (stmt_info))
2008 : : gcc_assert (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info)));
2009 : : else
2010 : : {
2011 : 500241 : *max_nunits = this_max_nunits;
2012 : 500241 : (*tree_size)++;
2013 : 500241 : node = vect_create_new_slp_node (node, stmts, 0);
2014 : 500241 : SLP_TREE_VECTYPE (node) = vectype;
2015 : : /* And compute the load permutation. Whether it is actually
2016 : : a permutation depends on the unrolling factor which is
2017 : : decided later. */
2018 : 500241 : vec<unsigned> load_permutation;
2019 : 500241 : int j;
2020 : 500241 : stmt_vec_info load_info;
2021 : 500241 : load_permutation.create (group_size);
2022 : 500241 : stmt_vec_info first_stmt_info
2023 : 500241 : = STMT_VINFO_GROUPED_ACCESS (stmt_info)
2024 : 500241 : ? DR_GROUP_FIRST_ELEMENT (stmt_info) : stmt_info;
2025 : 500241 : bool any_permute = false;
2026 : 1357754 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), j, load_info)
2027 : : {
2028 : 857513 : int load_place;
2029 : 857513 : if (! load_info)
2030 : : {
2031 : 70473 : if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
2032 : : load_place = j;
2033 : : else
2034 : : load_place = 0;
2035 : : }
2036 : 787040 : else if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
2037 : 563202 : load_place = vect_get_place_in_interleaving_chain
2038 : 563202 : (load_info, first_stmt_info);
2039 : : else
2040 : : load_place = 0;
2041 : 633675 : gcc_assert (load_place != -1);
2042 : 857513 : any_permute |= load_place != j;
2043 : 857513 : load_permutation.quick_push (load_place);
2044 : : }
2045 : :
2046 : 500241 : if (gcall *stmt = dyn_cast <gcall *> (stmt_info->stmt))
2047 : : {
2048 : 2413 : gcc_assert (gimple_call_internal_p (stmt, IFN_MASK_LOAD));
2049 : 2413 : bool has_gaps = false;
2050 : 2413 : if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
2051 : 217 : for (stmt_vec_info si = DR_GROUP_NEXT_ELEMENT (first_stmt_info);
2052 : 326 : si; si = DR_GROUP_NEXT_ELEMENT (si))
2053 : 109 : if (DR_GROUP_GAP (si) != 1)
2054 : 24 : has_gaps = true;
2055 : : /* We cannot handle permuted masked loads directly, see
2056 : : PR114375. We cannot handle strided masked loads or masked
2057 : : loads with gaps unless the mask is uniform. */
2058 : 2413 : if ((STMT_VINFO_GROUPED_ACCESS (stmt_info)
2059 : 217 : && (DR_GROUP_GAP (first_stmt_info) != 0
2060 : 91 : || (has_gaps
2061 : 24 : && STMT_VINFO_SLP_VECT_ONLY (first_stmt_info))))
2062 : 4676 : || STMT_VINFO_STRIDED_P (stmt_info))
2063 : : {
2064 : 176 : load_permutation.release ();
2065 : 176 : matches[0] = false;
2066 : 498042 : return NULL;
2067 : : }
2068 : :
2069 : : /* For permuted masked loads do an unpermuted masked load of
2070 : : the whole group followed by a SLP permute node. */
2071 : 2237 : if (any_permute
2072 : 2237 : || (STMT_VINFO_GROUPED_ACCESS (stmt_info)
2073 : 43 : && DR_GROUP_SIZE (first_stmt_info) != group_size))
2074 : : {
2075 : : /* Discover the whole unpermuted load. */
2076 : 38 : vec<stmt_vec_info> stmts2;
2077 : 38 : unsigned dr_group_size = STMT_VINFO_GROUPED_ACCESS (stmt_info)
2078 : 70 : ? DR_GROUP_SIZE (first_stmt_info) : 1;
2079 : 38 : stmts2.create (dr_group_size);
2080 : 38 : stmts2.quick_grow_cleared (dr_group_size);
2081 : 38 : unsigned i = 0;
2082 : 38 : for (stmt_vec_info si = first_stmt_info;
2083 : 120 : si; si = DR_GROUP_NEXT_ELEMENT (si))
2084 : : {
2085 : 82 : if (si != first_stmt_info)
2086 : 44 : for (unsigned k = 1; k < DR_GROUP_GAP (si); ++k)
2087 : 0 : stmts2[i++] = NULL;
2088 : 82 : stmts2[i++] = si;
2089 : : }
2090 : 38 : bool *matches2 = XALLOCAVEC (bool, dr_group_size);
2091 : 38 : slp_tree unperm_load
2092 : 38 : = vect_build_slp_tree (vinfo, stmts2, dr_group_size,
2093 : : &this_max_nunits, matches2, limit,
2094 : 38 : &this_tree_size, bst_map);
2095 : : /* When we are able to do the full masked load emit that
2096 : : followed by 'node' being the desired final permutation. */
2097 : 38 : if (unperm_load)
2098 : : {
2099 : 28 : gcc_assert
2100 : : (!SLP_TREE_LOAD_PERMUTATION (unperm_load).exists ());
2101 : 28 : lane_permutation_t lperm;
2102 : 28 : lperm.create (group_size);
2103 : 68 : for (unsigned j = 0; j < load_permutation.length (); ++j)
2104 : 40 : lperm.quick_push
2105 : 40 : (std::make_pair (0, load_permutation[j]));
2106 : 28 : SLP_TREE_CODE (node) = VEC_PERM_EXPR;
2107 : 28 : SLP_TREE_CHILDREN (node).safe_push (unperm_load);
2108 : 28 : SLP_TREE_LANE_PERMUTATION (node) = lperm;
2109 : 28 : load_permutation.release ();
2110 : 28 : return node;
2111 : : }
2112 : 10 : stmts2.release ();
2113 : 10 : load_permutation.release ();
2114 : 10 : matches[0] = false;
2115 : 10 : return NULL;
2116 : : }
2117 : 2199 : load_permutation.release ();
2118 : : }
2119 : : else
2120 : : {
2121 : 497828 : if (!any_permute
2122 : 425892 : && STMT_VINFO_GROUPED_ACCESS (stmt_info)
2123 : 707369 : && group_size == DR_GROUP_SIZE (first_stmt_info))
2124 : 140990 : load_permutation.release ();
2125 : 497828 : SLP_TREE_LOAD_PERMUTATION (node) = load_permutation;
2126 : 497828 : return node;
2127 : : }
2128 : : }
2129 : : }
2130 : 2757972 : else if (gimple_assign_single_p (stmt_info->stmt)
2131 : 2155332 : && !gimple_vuse (stmt_info->stmt)
2132 : 2765581 : && gimple_assign_rhs_code (stmt_info->stmt) == BIT_FIELD_REF)
2133 : : {
2134 : : /* vect_build_slp_tree_2 determined all BIT_FIELD_REFs reference
2135 : : the same SSA name vector of a compatible type to vectype. */
2136 : 2323 : vec<std::pair<unsigned, unsigned> > lperm = vNULL;
2137 : 2323 : tree vec = TREE_OPERAND (gimple_assign_rhs1 (stmt_info->stmt), 0);
2138 : 2323 : stmt_vec_info estmt_info;
2139 : 7641 : FOR_EACH_VEC_ELT (stmts, i, estmt_info)
2140 : : {
2141 : 5321 : gassign *estmt = as_a <gassign *> (estmt_info->stmt);
2142 : 5321 : tree bfref = gimple_assign_rhs1 (estmt);
2143 : 5321 : HOST_WIDE_INT lane;
2144 : 5321 : if (!known_eq (bit_field_size (bfref),
2145 : : tree_to_poly_uint64 (TYPE_SIZE (TREE_TYPE (vectype))))
2146 : 10639 : || !constant_multiple_p (bit_field_offset (bfref),
2147 : 5318 : bit_field_size (bfref), &lane))
2148 : : {
2149 : 3 : lperm.release ();
2150 : 3 : matches[0] = false;
2151 : 3 : return NULL;
2152 : : }
2153 : 5318 : lperm.safe_push (std::make_pair (0, (unsigned)lane));
2154 : : }
2155 : 2320 : slp_tree vnode = vect_create_new_slp_node (vNULL);
2156 : 2320 : if (operand_equal_p (TYPE_SIZE (vectype), TYPE_SIZE (TREE_TYPE (vec))))
2157 : : /* ??? We record vectype here but we hide eventually necessary
2158 : : punning and instead rely on code generation to materialize
2159 : : VIEW_CONVERT_EXPRs as necessary. We instead should make
2160 : : this explicit somehow. */
2161 : 643 : SLP_TREE_VECTYPE (vnode) = vectype;
2162 : : else
2163 : : {
2164 : : /* For different size but compatible elements we can still
2165 : : use VEC_PERM_EXPR without punning. */
2166 : 1677 : gcc_assert (VECTOR_TYPE_P (TREE_TYPE (vec))
2167 : : && types_compatible_p (TREE_TYPE (vectype),
2168 : : TREE_TYPE (TREE_TYPE (vec))));
2169 : 1677 : SLP_TREE_VECTYPE (vnode) = TREE_TYPE (vec);
2170 : : }
2171 : 2320 : auto nunits = TYPE_VECTOR_SUBPARTS (SLP_TREE_VECTYPE (vnode));
2172 : 2320 : unsigned HOST_WIDE_INT const_nunits;
2173 : 2320 : if (nunits.is_constant (&const_nunits))
2174 : 2320 : SLP_TREE_LANES (vnode) = const_nunits;
2175 : 2320 : SLP_TREE_VEC_DEFS (vnode).safe_push (vec);
2176 : : /* We are always building a permutation node even if it is an identity
2177 : : permute to shield the rest of the vectorizer from the odd node
2178 : : representing an actual vector without any scalar ops.
2179 : : ??? We could hide it completely with making the permute node
2180 : : external? */
2181 : 2320 : node = vect_create_new_slp_node (node, stmts, 1);
2182 : 2320 : SLP_TREE_CODE (node) = VEC_PERM_EXPR;
2183 : 2320 : SLP_TREE_LANE_PERMUTATION (node) = lperm;
2184 : 2320 : SLP_TREE_VECTYPE (node) = vectype;
2185 : 2320 : SLP_TREE_CHILDREN (node).quick_push (vnode);
2186 : 2320 : return node;
2187 : : }
2188 : : /* When discovery reaches an associatable operation see whether we can
2189 : : improve that to match up lanes in a way superior to the operand
2190 : : swapping code which at most looks at two defs.
2191 : : ??? For BB vectorization we cannot do the brute-force search
2192 : : for matching as we can succeed by means of builds from scalars
2193 : : and have no good way to "cost" one build against another. */
2194 : 2755649 : else if (is_a <loop_vec_info> (vinfo)
2195 : : /* Do not bother for single-lane SLP. */
2196 : 1277214 : && group_size > 1
2197 : : /* ??? We don't handle !vect_internal_def defs below. */
2198 : 44907 : && STMT_VINFO_DEF_TYPE (stmt_info) == vect_internal_def
2199 : : /* ??? Do not associate a reduction, this will wreck REDUC_IDX
2200 : : mapping as long as that exists on the stmt_info level. */
2201 : 40042 : && STMT_VINFO_REDUC_IDX (stmt_info) == -1
2202 : 39642 : && is_gimple_assign (stmt_info->stmt)
2203 : 39438 : && (associative_tree_code (gimple_assign_rhs_code (stmt_info->stmt))
2204 : 31773 : || gimple_assign_rhs_code (stmt_info->stmt) == MINUS_EXPR)
2205 : 2765007 : && ((FLOAT_TYPE_P (vectype) && flag_associative_math)
2206 : 7998 : || (INTEGRAL_TYPE_P (TREE_TYPE (vectype))
2207 : 5864 : && TYPE_OVERFLOW_WRAPS (TREE_TYPE (vectype)))))
2208 : : {
2209 : : /* See if we have a chain of (mixed) adds or subtracts or other
2210 : : associatable ops. */
2211 : 4517 : enum tree_code code = gimple_assign_rhs_code (stmt_info->stmt);
2212 : 4517 : if (code == MINUS_EXPR)
2213 : 703 : code = PLUS_EXPR;
2214 : 4517 : stmt_vec_info other_op_stmt_info = NULL;
2215 : 4517 : stmt_vec_info op_stmt_info = NULL;
2216 : 4517 : unsigned chain_len = 0;
2217 : 4517 : auto_vec<chain_op_t> chain;
2218 : 4517 : auto_vec<std::pair<tree_code, gimple *> > worklist;
2219 : 4517 : auto_vec<vec<chain_op_t> > chains (group_size);
2220 : 4517 : auto_vec<slp_tree, 4> children;
2221 : 4517 : bool hard_fail = true;
2222 : 5360 : for (unsigned lane = 0; lane < group_size; ++lane)
2223 : : {
2224 : 5054 : if (!stmts[lane])
2225 : : {
2226 : : /* ??? Below we require lane zero is present. */
2227 : 0 : if (lane == 0)
2228 : : {
2229 : : hard_fail = false;
2230 : 4211 : break;
2231 : : }
2232 : 0 : chains.quick_push (vNULL);
2233 : 0 : continue;
2234 : : }
2235 : : /* For each lane linearize the addition/subtraction (or other
2236 : : uniform associatable operation) expression tree. */
2237 : 5054 : gimple *op_stmt = NULL, *other_op_stmt = NULL;
2238 : 5054 : vect_slp_linearize_chain (vinfo, worklist, chain, code,
2239 : 5054 : stmts[lane]->stmt, op_stmt, other_op_stmt,
2240 : : NULL);
2241 : 5054 : if (!op_stmt_info && op_stmt)
2242 : 3967 : op_stmt_info = vinfo->lookup_stmt (op_stmt);
2243 : 5054 : if (!other_op_stmt_info && other_op_stmt)
2244 : 727 : other_op_stmt_info = vinfo->lookup_stmt (other_op_stmt);
2245 : 5054 : if (chain.length () == 2)
2246 : : {
2247 : : /* In a chain of just two elements resort to the regular
2248 : : operand swapping scheme. Likewise if we run into a
2249 : : length mismatch process regularly as well as we did not
2250 : : process the other lanes we cannot report a good hint what
2251 : : lanes to try swapping in the parent. */
2252 : : hard_fail = false;
2253 : : break;
2254 : : }
2255 : 843 : else if (chain_len == 0)
2256 : 343 : chain_len = chain.length ();
2257 : 1000 : else if (chain.length () != chain_len)
2258 : : {
2259 : : /* ??? Here we could slip in magic to compensate with
2260 : : neutral operands. */
2261 : 0 : matches[lane] = false;
2262 : 0 : if (lane != group_size - 1)
2263 : 0 : matches[0] = false;
2264 : : break;
2265 : : }
2266 : 843 : chains.quick_push (chain.copy ());
2267 : 843 : chain.truncate (0);
2268 : : }
2269 : 9034 : if (chains.length () == group_size)
2270 : : {
2271 : : /* We cannot yet use SLP_TREE_CODE to communicate the operation. */
2272 : 306 : if (!op_stmt_info)
2273 : : {
2274 : 15 : hard_fail = false;
2275 : 15 : goto out;
2276 : : }
2277 : : /* Now we have a set of chains with the same length. */
2278 : : /* 1. pre-sort according to def_type and operation. */
2279 : 1057 : for (unsigned lane = 0; lane < group_size; ++lane)
2280 : 1532 : chains[lane].stablesort (dt_sort_cmp, vinfo);
2281 : 291 : if (dump_enabled_p ())
2282 : : {
2283 : 123 : dump_printf_loc (MSG_NOTE, vect_location,
2284 : : "pre-sorted chains of %s\n",
2285 : : get_tree_code_name (code));
2286 : 522 : for (unsigned lane = 0; lane < group_size; ++lane)
2287 : : {
2288 : 399 : if (!stmts[lane])
2289 : 0 : dump_printf (MSG_NOTE, "--");
2290 : : else
2291 : 1822 : for (unsigned opnum = 0; opnum < chain_len; ++opnum)
2292 : 2846 : dump_printf (MSG_NOTE, "%s %T ",
2293 : 1423 : get_tree_code_name (chains[lane][opnum].code),
2294 : 1423 : chains[lane][opnum].op);
2295 : 399 : dump_printf (MSG_NOTE, "\n");
2296 : : }
2297 : : }
2298 : : /* 2. try to build children nodes, associating as necessary. */
2299 : : /* 2a. prepare and perform early checks to avoid eating into
2300 : : discovery limit unnecessarily. */
2301 : 291 : vect_def_type *dts = XALLOCAVEC (vect_def_type, chain_len);
2302 : 1226 : for (unsigned n = 0; n < chain_len; ++n)
2303 : : {
2304 : 935 : vect_def_type dt = chains[0][n].dt;
2305 : 935 : unsigned lane;
2306 : 3497 : for (lane = 0; lane < group_size; ++lane)
2307 : 5124 : if (stmts[lane] && chains[lane][n].dt != dt)
2308 : : {
2309 : 0 : if (dt == vect_constant_def
2310 : 0 : && chains[lane][n].dt == vect_external_def)
2311 : : dt = vect_external_def;
2312 : 0 : else if (dt == vect_external_def
2313 : 0 : && chains[lane][n].dt == vect_constant_def)
2314 : : ;
2315 : : else
2316 : : break;
2317 : : }
2318 : 935 : if (lane != group_size)
2319 : : {
2320 : 0 : if (dump_enabled_p ())
2321 : 0 : dump_printf_loc (MSG_NOTE, vect_location,
2322 : : "giving up on chain due to mismatched "
2323 : : "def types\n");
2324 : 0 : matches[lane] = false;
2325 : 0 : if (lane != group_size - 1)
2326 : 0 : matches[0] = false;
2327 : 0 : goto out;
2328 : : }
2329 : 935 : dts[n] = dt;
2330 : 935 : if (dt == vect_constant_def
2331 : 935 : || dt == vect_external_def)
2332 : : {
2333 : : /* Check whether we can build the invariant. If we can't
2334 : : we never will be able to. */
2335 : 78 : tree type = TREE_TYPE (chains[0][n].op);
2336 : 935 : if (!GET_MODE_SIZE (vinfo->vector_mode).is_constant ()
2337 : : && (TREE_CODE (type) == BOOLEAN_TYPE
2338 : : || !can_duplicate_and_interleave_p (vinfo, group_size,
2339 : : type)))
2340 : : {
2341 : : matches[0] = false;
2342 : : goto out;
2343 : : }
2344 : : }
2345 : 857 : else if (dt != vect_internal_def)
2346 : : {
2347 : : /* Not sure, we might need sth special.
2348 : : gcc.dg/vect/pr96854.c,
2349 : : gfortran.dg/vect/fast-math-pr37021.f90
2350 : : and gfortran.dg/vect/pr61171.f trigger. */
2351 : : /* Soft-fail for now. */
2352 : 0 : hard_fail = false;
2353 : 0 : goto out;
2354 : : }
2355 : : }
2356 : : /* 2b. do the actual build. */
2357 : 1144 : for (unsigned n = 0; n < chain_len; ++n)
2358 : : {
2359 : 880 : vect_def_type dt = dts[n];
2360 : 880 : unsigned lane;
2361 : 880 : if (dt == vect_constant_def
2362 : 880 : || dt == vect_external_def)
2363 : : {
2364 : 78 : vec<tree> ops;
2365 : 78 : ops.create (group_size);
2366 : 387 : for (lane = 0; lane < group_size; ++lane)
2367 : 231 : if (stmts[lane])
2368 : 231 : ops.quick_push (chains[lane][n].op);
2369 : : else
2370 : 0 : ops.quick_push (NULL_TREE);
2371 : 78 : slp_tree child = vect_create_new_slp_node (ops);
2372 : 78 : SLP_TREE_DEF_TYPE (child) = dt;
2373 : 78 : children.safe_push (child);
2374 : : }
2375 : : else
2376 : : {
2377 : 802 : vec<stmt_vec_info> op_stmts;
2378 : 802 : op_stmts.create (group_size);
2379 : 802 : slp_tree child = NULL;
2380 : : /* Brute-force our way. We have to consider a lane
2381 : : failing after fixing an earlier fail up in the
2382 : : SLP discovery recursion. So track the current
2383 : : permute per lane. */
2384 : 802 : unsigned *perms = XALLOCAVEC (unsigned, group_size);
2385 : 802 : memset (perms, 0, sizeof (unsigned) * group_size);
2386 : 931 : do
2387 : : {
2388 : 931 : op_stmts.truncate (0);
2389 : 4437 : for (lane = 0; lane < group_size; ++lane)
2390 : 2575 : if (stmts[lane])
2391 : 2575 : op_stmts.quick_push
2392 : 2575 : (vinfo->lookup_def (chains[lane][n].op));
2393 : : else
2394 : 0 : op_stmts.quick_push (NULL);
2395 : 931 : child = vect_build_slp_tree (vinfo, op_stmts,
2396 : : group_size, &this_max_nunits,
2397 : : matches, limit,
2398 : : &this_tree_size, bst_map);
2399 : : /* ??? We're likely getting too many fatal mismatches
2400 : : here so maybe we want to ignore them (but then we
2401 : : have no idea which lanes fatally mismatched). */
2402 : 931 : if (child || !matches[0])
2403 : : break;
2404 : : /* Swap another lane we have not yet matched up into
2405 : : lanes that did not match. If we run out of
2406 : : permute possibilities for a lane terminate the
2407 : : search. */
2408 : 405 : bool term = false;
2409 : 405 : for (lane = 1; lane < group_size; ++lane)
2410 : 276 : if (!matches[lane])
2411 : : {
2412 : 220 : if (n + perms[lane] + 1 == chain_len)
2413 : : {
2414 : : term = true;
2415 : : break;
2416 : : }
2417 : 193 : if (dump_enabled_p ())
2418 : 119 : dump_printf_loc (MSG_NOTE, vect_location,
2419 : : "swapping operand %d and %d "
2420 : : "of lane %d\n",
2421 : : n, n + perms[lane] + 1, lane);
2422 : 386 : std::swap (chains[lane][n],
2423 : 193 : chains[lane][n + perms[lane] + 1]);
2424 : 193 : perms[lane]++;
2425 : : }
2426 : 156 : if (term)
2427 : : break;
2428 : : }
2429 : : while (1);
2430 : 802 : if (!child)
2431 : : {
2432 : 27 : if (dump_enabled_p ())
2433 : 20 : dump_printf_loc (MSG_NOTE, vect_location,
2434 : : "failed to match up op %d\n", n);
2435 : 27 : op_stmts.release ();
2436 : 27 : if (lane != group_size - 1)
2437 : 12 : matches[0] = false;
2438 : : else
2439 : 15 : matches[lane] = false;
2440 : 27 : goto out;
2441 : : }
2442 : 775 : if (dump_enabled_p ())
2443 : : {
2444 : 312 : dump_printf_loc (MSG_NOTE, vect_location,
2445 : : "matched up op %d to\n", n);
2446 : 312 : vect_print_slp_tree (MSG_NOTE, vect_location, child);
2447 : : }
2448 : 775 : children.safe_push (child);
2449 : : }
2450 : : }
2451 : : /* 3. build SLP nodes to combine the chain. */
2452 : 936 : for (unsigned lane = 0; lane < group_size; ++lane)
2453 : 1360 : if (stmts[lane] && chains[lane][0].code != code)
2454 : : {
2455 : : /* See if there's any alternate all-PLUS entry. */
2456 : : unsigned n;
2457 : 8 : for (n = 1; n < chain_len; ++n)
2458 : : {
2459 : 36 : for (lane = 0; lane < group_size; ++lane)
2460 : 56 : if (stmts[lane] && chains[lane][n].code != code)
2461 : : break;
2462 : 8 : if (lane == group_size)
2463 : : break;
2464 : : }
2465 : 8 : if (n != chain_len)
2466 : : {
2467 : : /* Swap that in at first position. */
2468 : 8 : std::swap (children[0], children[n]);
2469 : 36 : for (lane = 0; lane < group_size; ++lane)
2470 : 28 : if (stmts[lane])
2471 : 28 : std::swap (chains[lane][0], chains[lane][n]);
2472 : : }
2473 : : else
2474 : : {
2475 : : /* ??? When this triggers and we end up with two
2476 : : vect_constant/external_def up-front things break (ICE)
2477 : : spectacularly finding an insertion place for the
2478 : : all-constant op. We should have a fully
2479 : : vect_internal_def operand though(?) so we can swap
2480 : : that into first place and then prepend the all-zero
2481 : : constant. */
2482 : 0 : if (dump_enabled_p ())
2483 : 0 : dump_printf_loc (MSG_NOTE, vect_location,
2484 : : "inserting constant zero to compensate "
2485 : : "for (partially) negated first "
2486 : : "operand\n");
2487 : 0 : chain_len++;
2488 : 0 : for (lane = 0; lane < group_size; ++lane)
2489 : 0 : if (stmts[lane])
2490 : 0 : chains[lane].safe_insert
2491 : 0 : (0, chain_op_t (code, vect_constant_def, NULL_TREE));
2492 : 0 : vec<tree> zero_ops;
2493 : 0 : zero_ops.create (group_size);
2494 : 0 : zero_ops.quick_push (build_zero_cst (TREE_TYPE (vectype)));
2495 : 0 : for (lane = 1; lane < group_size; ++lane)
2496 : 0 : if (stmts[lane])
2497 : 0 : zero_ops.quick_push (zero_ops[0]);
2498 : : else
2499 : 0 : zero_ops.quick_push (NULL_TREE);
2500 : 0 : slp_tree zero = vect_create_new_slp_node (zero_ops);
2501 : 0 : SLP_TREE_DEF_TYPE (zero) = vect_constant_def;
2502 : 0 : children.safe_insert (0, zero);
2503 : : }
2504 : : break;
2505 : : }
2506 : 852 : for (unsigned i = 1; i < children.length (); ++i)
2507 : : {
2508 : 588 : slp_tree op0 = children[i - 1];
2509 : 588 : slp_tree op1 = children[i];
2510 : 588 : bool this_two_op = false;
2511 : 2011 : for (unsigned lane = 0; lane < group_size; ++lane)
2512 : 3176 : if (stmts[lane] && chains[lane][i].code != chains[0][i].code)
2513 : : {
2514 : : this_two_op = true;
2515 : : break;
2516 : : }
2517 : 588 : slp_tree child;
2518 : 588 : if (i == children.length () - 1)
2519 : 264 : child = vect_create_new_slp_node (node, stmts, 2);
2520 : : else
2521 : 324 : child = vect_create_new_slp_node (2, ERROR_MARK);
2522 : 588 : if (this_two_op)
2523 : : {
2524 : 165 : vec<std::pair<unsigned, unsigned> > lperm;
2525 : 165 : lperm.create (group_size);
2526 : 603 : for (unsigned lane = 0; lane < group_size; ++lane)
2527 : 438 : lperm.quick_push (std::make_pair
2528 : 438 : (chains[lane][i].code != chains[0][i].code, lane));
2529 : 330 : vect_slp_build_two_operator_nodes (child, vectype, op0, op1,
2530 : 165 : (chains[0][i].code == code
2531 : : ? op_stmt_info
2532 : : : other_op_stmt_info),
2533 : 165 : (chains[0][i].code == code
2534 : : ? other_op_stmt_info
2535 : : : op_stmt_info),
2536 : : lperm);
2537 : : }
2538 : : else
2539 : : {
2540 : 423 : SLP_TREE_DEF_TYPE (child) = vect_internal_def;
2541 : 423 : SLP_TREE_VECTYPE (child) = vectype;
2542 : 423 : SLP_TREE_LANES (child) = group_size;
2543 : 423 : SLP_TREE_CHILDREN (child).quick_push (op0);
2544 : 423 : SLP_TREE_CHILDREN (child).quick_push (op1);
2545 : 423 : SLP_TREE_REPRESENTATIVE (child)
2546 : 846 : = (chains[0][i].code == code
2547 : 423 : ? op_stmt_info : other_op_stmt_info);
2548 : : }
2549 : 588 : children[i] = child;
2550 : : }
2551 : 264 : *tree_size += this_tree_size + 1;
2552 : 264 : *max_nunits = this_max_nunits;
2553 : 1255 : while (!chains.is_empty ())
2554 : 700 : chains.pop ().release ();
2555 : : return node;
2556 : : }
2557 : 4211 : out:
2558 : 4253 : if (dump_enabled_p ())
2559 : 3106 : dump_printf_loc (MSG_NOTE, vect_location,
2560 : : "failed to line up SLP graph by re-associating "
2561 : : "operations in lanes%s\n",
2562 : : !hard_fail ? " trying regular discovery" : "");
2563 : 4254 : while (!children.is_empty ())
2564 : 1 : vect_free_slp_tree (children.pop ());
2565 : 4396 : while (!chains.is_empty ())
2566 : 143 : chains.pop ().release ();
2567 : : /* Hard-fail, otherwise we might run into quadratic processing of the
2568 : : chains starting one stmt into the chain again. */
2569 : 4253 : if (hard_fail)
2570 : : return NULL;
2571 : : /* Fall thru to normal processing. */
2572 : 4517 : }
2573 : :
2574 : : /* Get at the operands, verifying they are compatible. */
2575 : 2763775 : vec<slp_oprnd_info> oprnds_info = vect_create_oprnd_info (nops, group_size);
2576 : 2763775 : slp_oprnd_info oprnd_info;
2577 : 14787366 : FOR_EACH_VEC_ELT (stmts, i, stmt_info)
2578 : : {
2579 : 12026210 : int res = vect_get_and_check_slp_defs (vinfo, swap[i], skip_args,
2580 : : stmts, i, &oprnds_info);
2581 : 12026210 : if (res != 0)
2582 : 592327 : matches[(res == -1) ? 0 : i] = false;
2583 : 12026210 : if (!matches[0])
2584 : : break;
2585 : : }
2586 : 14452402 : for (i = 0; i < group_size; ++i)
2587 : 11919782 : if (!matches[i])
2588 : : {
2589 : 231155 : vect_free_oprnd_info (oprnds_info);
2590 : 231155 : return NULL;
2591 : : }
2592 : 7597860 : swap = NULL;
2593 : :
2594 : 7597860 : bool has_two_operators_perm = false;
2595 : 15195720 : auto_vec<unsigned> two_op_perm_indices[2];
2596 : 2532620 : vec<stmt_vec_info> two_op_scalar_stmts[2] = {vNULL, vNULL};
2597 : :
2598 : 2543782 : if (two_operators && oprnds_info.length () == 2 && group_size > 2)
2599 : : {
2600 : 1345 : unsigned idx = 0;
2601 : 1345 : hash_map<gimple *, unsigned> seen;
2602 : 1345 : vec<slp_oprnd_info> new_oprnds_info
2603 : 1345 : = vect_create_oprnd_info (1, group_size);
2604 : 1345 : bool success = true;
2605 : :
2606 : 1345 : enum tree_code code = ERROR_MARK;
2607 : 1345 : if (oprnds_info[0]->def_stmts[0]
2608 : 1345 : && is_a<gassign *> (oprnds_info[0]->def_stmts[0]->stmt))
2609 : 1244 : code = gimple_assign_rhs_code (oprnds_info[0]->def_stmts[0]->stmt);
2610 : :
2611 : 4033 : for (unsigned j = 0; j < group_size; ++j)
2612 : : {
2613 : 9859 : FOR_EACH_VEC_ELT (oprnds_info, i, oprnd_info)
2614 : : {
2615 : 7171 : stmt_vec_info stmt_info = oprnd_info->def_stmts[j];
2616 : 6935 : if (!stmt_info || !stmt_info->stmt
2617 : 6935 : || !is_a<gassign *> (stmt_info->stmt)
2618 : 6932 : || gimple_assign_rhs_code (stmt_info->stmt) != code
2619 : 13315 : || skip_args[i])
2620 : : {
2621 : 1027 : success = false;
2622 : 1027 : break;
2623 : : }
2624 : :
2625 : 6144 : bool exists;
2626 : 6144 : unsigned &stmt_idx
2627 : 6144 : = seen.get_or_insert (stmt_info->stmt, &exists);
2628 : :
2629 : 6144 : if (!exists)
2630 : : {
2631 : 5192 : new_oprnds_info[0]->def_stmts.safe_push (stmt_info);
2632 : 5192 : new_oprnds_info[0]->ops.safe_push (oprnd_info->ops[j]);
2633 : 5192 : stmt_idx = idx;
2634 : 5192 : idx++;
2635 : : }
2636 : :
2637 : 6144 : two_op_perm_indices[i].safe_push (stmt_idx);
2638 : : }
2639 : :
2640 : 3715 : if (!success)
2641 : : break;
2642 : : }
2643 : :
2644 : 1345 : if (success && idx == group_size)
2645 : : {
2646 : 41 : if (dump_enabled_p ())
2647 : : {
2648 : 0 : dump_printf_loc (MSG_NOTE, vect_location,
2649 : : "Replace two_operators operands:\n");
2650 : :
2651 : 0 : FOR_EACH_VEC_ELT (oprnds_info, i, oprnd_info)
2652 : : {
2653 : 0 : dump_printf_loc (MSG_NOTE, vect_location,
2654 : : "Operand %u:\n", i);
2655 : 0 : for (unsigned j = 0; j < group_size; j++)
2656 : 0 : dump_printf_loc (MSG_NOTE, vect_location, "\tstmt %u %G",
2657 : 0 : j, oprnd_info->def_stmts[j]->stmt);
2658 : : }
2659 : :
2660 : 0 : dump_printf_loc (MSG_NOTE, vect_location,
2661 : : "With a single operand:\n");
2662 : 0 : for (unsigned j = 0; j < group_size; j++)
2663 : 0 : dump_printf_loc (MSG_NOTE, vect_location, "\tstmt %u %G",
2664 : 0 : j, new_oprnds_info[0]->def_stmts[j]->stmt);
2665 : : }
2666 : :
2667 : 41 : two_op_scalar_stmts[0].safe_splice (oprnds_info[0]->def_stmts);
2668 : 41 : two_op_scalar_stmts[1].safe_splice (oprnds_info[1]->def_stmts);
2669 : :
2670 : 41 : new_oprnds_info[0]->first_op_type = oprnds_info[0]->first_op_type;
2671 : 41 : new_oprnds_info[0]->first_dt = oprnds_info[0]->first_dt;
2672 : 41 : new_oprnds_info[0]->any_pattern = oprnds_info[0]->any_pattern;
2673 : 41 : new_oprnds_info[0]->first_gs_p = oprnds_info[0]->first_gs_p;
2674 : 41 : new_oprnds_info[0]->first_gs_info = oprnds_info[0]->first_gs_info;
2675 : :
2676 : 41 : vect_free_oprnd_info (oprnds_info);
2677 : 41 : oprnds_info = new_oprnds_info;
2678 : 41 : nops = 1;
2679 : 41 : has_two_operators_perm = true;
2680 : : }
2681 : 1345 : }
2682 : :
2683 : 5065240 : auto_vec<slp_tree, 4> children;
2684 : :
2685 : 2532620 : stmt_info = stmts[0];
2686 : :
2687 : : /* Create SLP_TREE nodes for the definition node/s. */
2688 : 6156165 : FOR_EACH_VEC_ELT (oprnds_info, i, oprnd_info)
2689 : : {
2690 : 3872837 : slp_tree child = nullptr;
2691 : 3872837 : unsigned int j;
2692 : :
2693 : : /* We're skipping certain operands from processing, for example
2694 : : outer loop reduction initial defs. */
2695 : 3872837 : if (skip_args[i])
2696 : : {
2697 : 386763 : children.safe_push (NULL);
2698 : 3623545 : continue;
2699 : : }
2700 : :
2701 : 3486074 : if (oprnd_info->first_dt == vect_uninitialized_def)
2702 : : {
2703 : : /* COND_EXPR have one too many eventually if the condition
2704 : : is a SSA name. */
2705 : 0 : gcc_assert (i == 3 && nops == 4);
2706 : 0 : continue;
2707 : : }
2708 : :
2709 : 3486074 : if (is_a <bb_vec_info> (vinfo)
2710 : 1650754 : && oprnd_info->first_dt == vect_internal_def
2711 : 4386297 : && !oprnd_info->any_pattern)
2712 : : {
2713 : : /* For BB vectorization, if all defs are the same do not
2714 : : bother to continue the build along the single-lane
2715 : : graph but use a splat of the scalar value. */
2716 : 855415 : stmt_vec_info first_def = oprnd_info->def_stmts[0];
2717 : 909354 : for (j = 1; j < group_size; ++j)
2718 : 863986 : if (oprnd_info->def_stmts[j] != first_def)
2719 : : break;
2720 : 855415 : if (j == group_size
2721 : : /* But avoid doing this for loads where we may be
2722 : : able to CSE things, unless the stmt is not
2723 : : vectorizable. */
2724 : 855415 : && (!STMT_VINFO_VECTORIZABLE (first_def)
2725 : 56593 : || !gimple_vuse (first_def->stmt)))
2726 : : {
2727 : 36677 : if (dump_enabled_p ())
2728 : 94 : dump_printf_loc (MSG_NOTE, vect_location,
2729 : : "Using a splat of the uniform operand %G",
2730 : : first_def->stmt);
2731 : 36677 : oprnd_info->first_dt = vect_external_def;
2732 : : }
2733 : : }
2734 : :
2735 : 3486074 : if (oprnd_info->first_dt == vect_external_def
2736 : 3486074 : || oprnd_info->first_dt == vect_constant_def)
2737 : : {
2738 : 1259284 : if (!GET_MODE_SIZE (vinfo->vector_mode).is_constant ())
2739 : : {
2740 : : tree op0;
2741 : : tree uniform_val = op0 = oprnd_info->ops[0];
2742 : : for (j = 1; j < oprnd_info->ops.length (); ++j)
2743 : : if (oprnd_info->ops[j]
2744 : : && !operand_equal_p (uniform_val, oprnd_info->ops[j]))
2745 : : {
2746 : : uniform_val = NULL_TREE;
2747 : : break;
2748 : : }
2749 : : if (!uniform_val
2750 : : && !can_duplicate_and_interleave_p (vinfo,
2751 : : oprnd_info->ops.length (),
2752 : : TREE_TYPE (op0)))
2753 : : {
2754 : : matches[j] = false;
2755 : : if (dump_enabled_p ())
2756 : : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
2757 : : "Build SLP failed: invalid type of def "
2758 : : "for variable-length SLP %T\n", op0);
2759 : : goto fail;
2760 : : }
2761 : : }
2762 : 1259284 : slp_tree invnode = vect_create_new_slp_node (oprnd_info->ops);
2763 : 1259284 : SLP_TREE_DEF_TYPE (invnode) = oprnd_info->first_dt;
2764 : 1259284 : oprnd_info->ops = vNULL;
2765 : 1259284 : children.safe_push (invnode);
2766 : 1259284 : continue;
2767 : 1259284 : }
2768 : :
2769 : : /* When we have a masked load with uniform mask discover this
2770 : : as a single-lane mask with a splat permute. This way we can
2771 : : recognize this as a masked load-lane by stripping the splat. */
2772 : 2226790 : if (is_a <gcall *> (STMT_VINFO_STMT (stmt_info))
2773 : 33351 : && gimple_call_internal_p (STMT_VINFO_STMT (stmt_info),
2774 : : IFN_MASK_LOAD)
2775 : 4910 : && STMT_VINFO_GROUPED_ACCESS (stmt_info)
2776 : 2226822 : && ! STMT_VINFO_SLP_VECT_ONLY (DR_GROUP_FIRST_ELEMENT (stmt_info)))
2777 : : {
2778 : 0 : vec<stmt_vec_info> def_stmts2;
2779 : 0 : def_stmts2.create (1);
2780 : 0 : def_stmts2.quick_push (oprnd_info->def_stmts[0]);
2781 : 0 : child = vect_build_slp_tree (vinfo, def_stmts2, 1,
2782 : : &this_max_nunits,
2783 : : matches, limit,
2784 : : &this_tree_size, bst_map);
2785 : 0 : if (child)
2786 : : {
2787 : 0 : slp_tree pnode = vect_create_new_slp_node (1, VEC_PERM_EXPR);
2788 : 0 : SLP_TREE_VECTYPE (pnode) = SLP_TREE_VECTYPE (child);
2789 : 0 : SLP_TREE_LANES (pnode) = group_size;
2790 : 0 : SLP_TREE_SCALAR_STMTS (pnode).create (group_size);
2791 : 0 : SLP_TREE_LANE_PERMUTATION (pnode).create (group_size);
2792 : 0 : for (unsigned k = 0; k < group_size; ++k)
2793 : : {
2794 : 0 : SLP_TREE_SCALAR_STMTS (pnode)
2795 : 0 : .quick_push (oprnd_info->def_stmts[0]);
2796 : 0 : SLP_TREE_LANE_PERMUTATION (pnode)
2797 : 0 : .quick_push (std::make_pair (0u, 0u));
2798 : : }
2799 : 0 : SLP_TREE_CHILDREN (pnode).quick_push (child);
2800 : 0 : pnode->max_nunits = child->max_nunits;
2801 : 0 : children.safe_push (pnode);
2802 : 0 : oprnd_info->def_stmts = vNULL;
2803 : 0 : continue;
2804 : 0 : }
2805 : : else
2806 : 0 : def_stmts2.release ();
2807 : : }
2808 : :
2809 : 2226790 : if ((child = vect_build_slp_tree (vinfo, oprnd_info->def_stmts,
2810 : : group_size, &this_max_nunits,
2811 : : matches, limit,
2812 : : &this_tree_size, bst_map)) != NULL)
2813 : : {
2814 : 1709588 : oprnd_info->def_stmts = vNULL;
2815 : 1709588 : children.safe_push (child);
2816 : 1709588 : continue;
2817 : : }
2818 : :
2819 : : /* If the SLP build for operand zero failed and operand zero
2820 : : and one can be commutated try that for the scalar stmts
2821 : : that failed the match. */
2822 : 517202 : if (i == 0
2823 : : /* A first scalar stmt mismatch signals a fatal mismatch. */
2824 : 381956 : && matches[0]
2825 : : /* ??? For COND_EXPRs we can swap the comparison operands
2826 : : as well as the arms under some constraints. */
2827 : 229053 : && nops == 2
2828 : 129062 : && oprnds_info[1]->first_dt == vect_internal_def
2829 : 75044 : && is_gimple_assign (stmt_info->stmt)
2830 : : /* Swapping operands for reductions breaks assumptions later on. */
2831 : 573175 : && STMT_VINFO_REDUC_IDX (stmt_info) == -1)
2832 : : {
2833 : : /* See whether we can swap the matching or the non-matching
2834 : : stmt operands. */
2835 : : bool swap_not_matching = true;
2836 : 69334 : do
2837 : : {
2838 : 7057256 : for (j = 0; j < group_size; ++j)
2839 : : {
2840 : 7013367 : if (matches[j] != !swap_not_matching)
2841 : 75552 : continue;
2842 : 6937815 : stmt_vec_info stmt_info = stmts[j];
2843 : : /* Verify if we can swap operands of this stmt. */
2844 : 6937815 : gassign *stmt = dyn_cast <gassign *> (stmt_info->stmt);
2845 : 6937815 : if (!stmt
2846 : 6937815 : || !commutative_tree_code (gimple_assign_rhs_code (stmt)))
2847 : : {
2848 : 25445 : if (!swap_not_matching)
2849 : 11940 : goto fail;
2850 : : swap_not_matching = false;
2851 : : break;
2852 : : }
2853 : : }
2854 : : }
2855 : 57394 : while (j != group_size);
2856 : :
2857 : : /* Swap mismatched definition stmts. */
2858 : 43889 : if (dump_enabled_p ())
2859 : 348 : dump_printf_loc (MSG_NOTE, vect_location,
2860 : : "Re-trying with swapped operands of stmts ");
2861 : 7017390 : for (j = 0; j < group_size; ++j)
2862 : 6973501 : if (matches[j] == !swap_not_matching)
2863 : : {
2864 : 13824476 : std::swap (oprnds_info[0]->def_stmts[j],
2865 : 6912238 : oprnds_info[1]->def_stmts[j]);
2866 : 13824476 : std::swap (oprnds_info[0]->ops[j],
2867 : 6912238 : oprnds_info[1]->ops[j]);
2868 : 6912238 : if (dump_enabled_p ())
2869 : 971 : dump_printf (MSG_NOTE, "%d ", j);
2870 : : }
2871 : 43889 : if (dump_enabled_p ())
2872 : 348 : dump_printf (MSG_NOTE, "\n");
2873 : : /* After swapping some operands we lost track whether an
2874 : : operand has any pattern defs so be conservative here. */
2875 : 84884 : if (oprnds_info[0]->any_pattern || oprnds_info[1]->any_pattern)
2876 : 3065 : oprnds_info[0]->any_pattern = oprnds_info[1]->any_pattern = true;
2877 : : /* And try again with scratch 'matches' ... */
2878 : 43889 : bool *tem = XALLOCAVEC (bool, group_size);
2879 : 43889 : if ((child = vect_build_slp_tree (vinfo, oprnd_info->def_stmts,
2880 : : group_size, &this_max_nunits,
2881 : : tem, limit,
2882 : : &this_tree_size, bst_map)) != NULL)
2883 : : {
2884 : 7503 : oprnd_info->def_stmts = vNULL;
2885 : 7503 : children.safe_push (child);
2886 : 7503 : continue;
2887 : : }
2888 : : }
2889 : 509699 : fail:
2890 : :
2891 : : /* If the SLP build failed and we analyze a basic-block
2892 : : simply treat nodes we fail to build as externally defined
2893 : : (and thus build vectors from the scalar defs).
2894 : : The cost model will reject outright expensive cases.
2895 : : ??? This doesn't treat cases where permutation ultimatively
2896 : : fails (or we don't try permutation below). Ideally we'd
2897 : : even compute a permutation that will end up with the maximum
2898 : : SLP tree size... */
2899 : 509699 : if (is_a <bb_vec_info> (vinfo)
2900 : : /* ??? Rejecting patterns this way doesn't work. We'd have to
2901 : : do extra work to cancel the pattern so the uses see the
2902 : : scalar version. */
2903 : 474416 : && !is_pattern_stmt_p (stmt_info)
2904 : 957282 : && !oprnd_info->any_pattern)
2905 : : {
2906 : : /* But if there's a leading vector sized set of matching stmts
2907 : : fail here so we can split the group. This matches the condition
2908 : : vect_analyze_slp_instance uses. */
2909 : : /* ??? We might want to split here and combine the results to support
2910 : : multiple vector sizes better. */
2911 : 731399 : for (j = 0; j < group_size; ++j)
2912 : 731399 : if (!matches[j])
2913 : : break;
2914 : 566839 : if (!known_ge (j, TYPE_VECTOR_SUBPARTS (vectype))
2915 : 447249 : && vect_slp_can_convert_to_external (oprnd_info->def_stmts))
2916 : : {
2917 : 260407 : if (dump_enabled_p ())
2918 : 445 : dump_printf_loc (MSG_NOTE, vect_location,
2919 : : "Building vector operands from scalars\n");
2920 : 260407 : this_tree_size++;
2921 : 260407 : child = vect_create_new_slp_node (oprnd_info->ops);
2922 : 260407 : children.safe_push (child);
2923 : 260407 : oprnd_info->ops = vNULL;
2924 : 260407 : continue;
2925 : : }
2926 : : }
2927 : :
2928 : 249292 : gcc_assert (child == NULL);
2929 : 289702 : FOR_EACH_VEC_ELT (children, j, child)
2930 : 40410 : if (child)
2931 : 40410 : vect_free_slp_tree (child);
2932 : 249292 : vect_free_oprnd_info (oprnds_info);
2933 : 249292 : return NULL;
2934 : : }
2935 : :
2936 : 2283328 : vect_free_oprnd_info (oprnds_info);
2937 : :
2938 : : /* If we have all children of a child built up from uniform scalars
2939 : : or does more than one possibly expensive vector construction then
2940 : : just throw that away, causing it built up from scalars.
2941 : : The exception is the SLP node for the vector store. */
2942 : 2283328 : if (is_a <bb_vec_info> (vinfo)
2943 : 1037853 : && !STMT_VINFO_GROUPED_ACCESS (stmt_info)
2944 : : /* ??? Rejecting patterns this way doesn't work. We'd have to
2945 : : do extra work to cancel the pattern so the uses see the
2946 : : scalar version. */
2947 : 2669318 : && !is_pattern_stmt_p (stmt_info))
2948 : : {
2949 : : slp_tree child;
2950 : : unsigned j;
2951 : : bool all_uniform_p = true;
2952 : : unsigned n_vector_builds = 0;
2953 : 1066345 : FOR_EACH_VEC_ELT (children, j, child)
2954 : : {
2955 : 705338 : if (!child)
2956 : : ;
2957 : 705338 : else if (SLP_TREE_DEF_TYPE (child) == vect_internal_def)
2958 : : all_uniform_p = false;
2959 : 490912 : else if (!vect_slp_tree_uniform_p (child))
2960 : : {
2961 : 357387 : all_uniform_p = false;
2962 : 357387 : if (SLP_TREE_DEF_TYPE (child) == vect_external_def)
2963 : 324551 : n_vector_builds++;
2964 : : }
2965 : : }
2966 : 361007 : if (all_uniform_p
2967 : 361007 : || n_vector_builds > 1
2968 : 626922 : || (n_vector_builds == children.length ()
2969 : 22190 : && is_a <gphi *> (stmt_info->stmt)))
2970 : : {
2971 : : /* Roll back. */
2972 : 101156 : matches[0] = false;
2973 : 312430 : FOR_EACH_VEC_ELT (children, j, child)
2974 : 211274 : if (child)
2975 : 211274 : vect_free_slp_tree (child);
2976 : :
2977 : 101156 : if (dump_enabled_p ())
2978 : 114 : dump_printf_loc (MSG_NOTE, vect_location,
2979 : : "Building parent vector operands from "
2980 : : "scalars instead\n");
2981 : 101156 : return NULL;
2982 : : }
2983 : : }
2984 : :
2985 : 2182172 : *tree_size += this_tree_size + 1;
2986 : 2182172 : *max_nunits = this_max_nunits;
2987 : :
2988 : 2182172 : if (two_operators)
2989 : : {
2990 : : /* ??? We'd likely want to either cache in bst_map sth like
2991 : : { a+b, NULL, a+b, NULL } and { NULL, a-b, NULL, a-b } or
2992 : : the true { a+b, a+b, a+b, a+b } ... but there we don't have
2993 : : explicit stmts to put in so the keying on 'stmts' doesn't
2994 : : work (but we have the same issue with nodes that use 'ops'). */
2995 : :
2996 : 3915 : if (has_two_operators_perm)
2997 : : {
2998 : 21 : slp_tree child = children[0];
2999 : 21 : children.truncate (0);
3000 : 63 : for (i = 0; i < 2; i++)
3001 : : {
3002 : 42 : slp_tree pnode
3003 : 42 : = vect_create_new_slp_node (two_op_scalar_stmts[i], 2);
3004 : 42 : SLP_TREE_CODE (pnode) = VEC_PERM_EXPR;
3005 : 42 : SLP_TREE_VECTYPE (pnode) = vectype;
3006 : 42 : SLP_TREE_CHILDREN (pnode).quick_push (child);
3007 : 42 : SLP_TREE_CHILDREN (pnode).quick_push (child);
3008 : 42 : lane_permutation_t& perm = SLP_TREE_LANE_PERMUTATION (pnode);
3009 : 42 : children.safe_push (pnode);
3010 : :
3011 : 466 : for (unsigned j = 0; j < stmts.length (); j++)
3012 : 424 : perm.safe_push (std::make_pair (0, two_op_perm_indices[i][j]));
3013 : : }
3014 : :
3015 : 21 : SLP_TREE_REF_COUNT (child) += 4;
3016 : : }
3017 : :
3018 : 3915 : slp_tree one = new _slp_tree;
3019 : 3915 : slp_tree two = new _slp_tree;
3020 : 3915 : SLP_TREE_DEF_TYPE (one) = vect_internal_def;
3021 : 3915 : SLP_TREE_DEF_TYPE (two) = vect_internal_def;
3022 : 3915 : SLP_TREE_VECTYPE (one) = vectype;
3023 : 3915 : SLP_TREE_VECTYPE (two) = vectype;
3024 : 3915 : SLP_TREE_CHILDREN (one).safe_splice (children);
3025 : 3915 : SLP_TREE_CHILDREN (two).safe_splice (children);
3026 : 3915 : slp_tree child;
3027 : 15660 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (two), i, child)
3028 : 7830 : SLP_TREE_REF_COUNT (child)++;
3029 : :
3030 : : /* Here we record the original defs since this
3031 : : node represents the final lane configuration. */
3032 : 3915 : node = vect_create_new_slp_node (node, stmts, 2);
3033 : 3915 : SLP_TREE_VECTYPE (node) = vectype;
3034 : 3915 : SLP_TREE_CODE (node) = VEC_PERM_EXPR;
3035 : 3915 : SLP_TREE_CHILDREN (node).quick_push (one);
3036 : 3915 : SLP_TREE_CHILDREN (node).quick_push (two);
3037 : 3915 : gassign *stmt = as_a <gassign *> (stmts[0]->stmt);
3038 : 3915 : enum tree_code code0 = gimple_assign_rhs_code (stmt);
3039 : 3915 : enum tree_code ocode = ERROR_MARK;
3040 : 3915 : stmt_vec_info ostmt_info;
3041 : 3915 : unsigned j = 0;
3042 : 14717 : FOR_EACH_VEC_ELT (stmts, i, ostmt_info)
3043 : : {
3044 : 10802 : gassign *ostmt = as_a <gassign *> (ostmt_info->stmt);
3045 : 10802 : if (gimple_assign_rhs_code (ostmt) != code0)
3046 : : {
3047 : 5402 : SLP_TREE_LANE_PERMUTATION (node).safe_push (std::make_pair (1, i));
3048 : 5402 : ocode = gimple_assign_rhs_code (ostmt);
3049 : 5402 : j = i;
3050 : : }
3051 : : else
3052 : 5400 : SLP_TREE_LANE_PERMUTATION (node).safe_push (std::make_pair (0, i));
3053 : : }
3054 : :
3055 : 3915 : SLP_TREE_CODE (one) = code0;
3056 : 3915 : SLP_TREE_CODE (two) = ocode;
3057 : 3915 : SLP_TREE_LANES (one) = stmts.length ();
3058 : 3915 : SLP_TREE_LANES (two) = stmts.length ();
3059 : 3915 : SLP_TREE_REPRESENTATIVE (one) = stmts[0];
3060 : 3915 : SLP_TREE_REPRESENTATIVE (two) = stmts[j];
3061 : :
3062 : 3915 : return node;
3063 : : }
3064 : :
3065 : 2178257 : node = vect_create_new_slp_node (node, stmts, nops);
3066 : 2178257 : SLP_TREE_VECTYPE (node) = vectype;
3067 : 2178257 : SLP_TREE_CHILDREN (node).splice (children);
3068 : 2178257 : return node;
3069 : 7597860 : }
3070 : :
3071 : : /* Dump a single SLP tree NODE. */
3072 : :
3073 : : static void
3074 : 455891 : vect_print_slp_tree (dump_flags_t dump_kind, dump_location_t loc,
3075 : : slp_tree node)
3076 : : {
3077 : 455891 : unsigned i, j;
3078 : 455891 : slp_tree child;
3079 : 455891 : stmt_vec_info stmt_info;
3080 : 455891 : tree op;
3081 : :
3082 : 455891 : dump_metadata_t metadata (dump_kind, loc.get_impl_location ());
3083 : 455891 : dump_user_location_t user_loc = loc.get_user_location ();
3084 : 455891 : dump_printf_loc (metadata, user_loc,
3085 : : "node%s %p (max_nunits=" HOST_WIDE_INT_PRINT_UNSIGNED
3086 : : ", refcnt=%u)",
3087 : 455891 : SLP_TREE_DEF_TYPE (node) == vect_external_def
3088 : : ? " (external)"
3089 : : : (SLP_TREE_DEF_TYPE (node) == vect_constant_def
3090 : 441078 : ? " (constant)"
3091 : : : ""), (void *) node,
3092 : 455891 : estimated_poly_value (node->max_nunits),
3093 : : SLP_TREE_REF_COUNT (node));
3094 : 455891 : if (SLP_TREE_VECTYPE (node))
3095 : 383588 : dump_printf (metadata, " %T", SLP_TREE_VECTYPE (node));
3096 : 455891 : dump_printf (metadata, "\n");
3097 : 455891 : if (SLP_TREE_DEF_TYPE (node) == vect_internal_def)
3098 : : {
3099 : 369697 : if (SLP_TREE_CODE (node) == VEC_PERM_EXPR)
3100 : 14452 : dump_printf_loc (metadata, user_loc, "op: VEC_PERM_EXPR\n");
3101 : : else
3102 : 355245 : dump_printf_loc (metadata, user_loc, "op template: %G",
3103 : 355245 : SLP_TREE_REPRESENTATIVE (node)->stmt);
3104 : : }
3105 : 455891 : if (SLP_TREE_SCALAR_STMTS (node).exists ())
3106 : 874226 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
3107 : 512692 : if (stmt_info)
3108 : 505470 : dump_printf_loc (metadata, user_loc, "\t%sstmt %u %G",
3109 : 505470 : STMT_VINFO_LIVE_P (stmt_info) ? "[l] " : "",
3110 : : i, stmt_info->stmt);
3111 : : else
3112 : 7222 : dump_printf_loc (metadata, user_loc, "\tstmt %u ---\n", i);
3113 : : else
3114 : : {
3115 : 94357 : dump_printf_loc (metadata, user_loc, "\t{ ");
3116 : 299446 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_OPS (node), i, op)
3117 : 110732 : dump_printf (metadata, "%T%s ", op,
3118 : 110732 : i < SLP_TREE_SCALAR_OPS (node).length () - 1 ? "," : "");
3119 : 94357 : dump_printf (metadata, "}\n");
3120 : : }
3121 : 455891 : if (SLP_TREE_LOAD_PERMUTATION (node).exists ())
3122 : : {
3123 : 62753 : dump_printf_loc (metadata, user_loc, "\tload permutation {");
3124 : 204340 : FOR_EACH_VEC_ELT (SLP_TREE_LOAD_PERMUTATION (node), i, j)
3125 : 78834 : dump_printf (dump_kind, " %u", j);
3126 : 62753 : dump_printf (dump_kind, " }\n");
3127 : : }
3128 : 455891 : if (SLP_TREE_LANE_PERMUTATION (node).exists ())
3129 : : {
3130 : 14460 : dump_printf_loc (metadata, user_loc, "\tlane permutation {");
3131 : 69337 : for (i = 0; i < SLP_TREE_LANE_PERMUTATION (node).length (); ++i)
3132 : 40417 : dump_printf (dump_kind, " %u[%u]",
3133 : 40417 : SLP_TREE_LANE_PERMUTATION (node)[i].first,
3134 : 40417 : SLP_TREE_LANE_PERMUTATION (node)[i].second);
3135 : 14460 : dump_printf (dump_kind, " }%s\n",
3136 : 14460 : node->ldst_lanes ? " (load-lanes)" : "");
3137 : : }
3138 : 455891 : if (SLP_TREE_CHILDREN (node).is_empty ())
3139 : 170377 : return;
3140 : 285514 : dump_printf_loc (metadata, user_loc, "\tchildren");
3141 : 1040256 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
3142 : 469228 : dump_printf (dump_kind, " %p", (void *)child);
3143 : 285514 : dump_printf (dump_kind, "%s\n",
3144 : 285514 : node->ldst_lanes && !SLP_TREE_LANE_PERMUTATION (node).exists ()
3145 : : ? " (store-lanes)" : "");
3146 : : }
3147 : :
3148 : : DEBUG_FUNCTION void
3149 : 0 : debug (slp_tree node)
3150 : : {
3151 : 0 : debug_dump_context ctx;
3152 : 0 : vect_print_slp_tree (MSG_NOTE,
3153 : 0 : dump_location_t::from_location_t (UNKNOWN_LOCATION),
3154 : : node);
3155 : 0 : }
3156 : :
3157 : : /* Recursive helper for the dot producer below. */
3158 : :
3159 : : static void
3160 : 0 : dot_slp_tree (FILE *f, slp_tree node, hash_set<slp_tree> &visited)
3161 : : {
3162 : 0 : if (visited.add (node))
3163 : : return;
3164 : :
3165 : 0 : fprintf (f, "\"%p\" [label=\"", (void *)node);
3166 : 0 : vect_print_slp_tree (MSG_NOTE,
3167 : 0 : dump_location_t::from_location_t (UNKNOWN_LOCATION),
3168 : : node);
3169 : 0 : fprintf (f, "\"];\n");
3170 : :
3171 : :
3172 : 0 : for (slp_tree child : SLP_TREE_CHILDREN (node))
3173 : 0 : fprintf (f, "\"%p\" -> \"%p\";", (void *)node, (void *)child);
3174 : :
3175 : 0 : for (slp_tree child : SLP_TREE_CHILDREN (node))
3176 : 0 : if (child)
3177 : 0 : dot_slp_tree (f, child, visited);
3178 : : }
3179 : :
3180 : : DEBUG_FUNCTION void
3181 : 0 : dot_slp_tree (const char *fname, slp_tree node)
3182 : : {
3183 : 0 : FILE *f = fopen (fname, "w");
3184 : 0 : fprintf (f, "digraph {\n");
3185 : 0 : fflush (f);
3186 : 0 : {
3187 : 0 : debug_dump_context ctx (f);
3188 : 0 : hash_set<slp_tree> visited;
3189 : 0 : dot_slp_tree (f, node, visited);
3190 : 0 : }
3191 : 0 : fflush (f);
3192 : 0 : fprintf (f, "}\n");
3193 : 0 : fclose (f);
3194 : 0 : }
3195 : :
3196 : : DEBUG_FUNCTION void
3197 : 0 : dot_slp_tree (const char *fname, const vec<slp_instance> &slp_instances)
3198 : : {
3199 : 0 : FILE *f = fopen (fname, "w");
3200 : 0 : fprintf (f, "digraph {\n");
3201 : 0 : fflush (f);
3202 : 0 : {
3203 : 0 : debug_dump_context ctx (f);
3204 : 0 : hash_set<slp_tree> visited;
3205 : 0 : for (auto inst : slp_instances)
3206 : 0 : dot_slp_tree (f, SLP_INSTANCE_TREE (inst), visited);
3207 : 0 : }
3208 : 0 : fflush (f);
3209 : 0 : fprintf (f, "}\n");
3210 : 0 : fclose (f);
3211 : 0 : }
3212 : :
3213 : : /* Dump a slp tree NODE using flags specified in DUMP_KIND. */
3214 : :
3215 : : static void
3216 : 496441 : vect_print_slp_graph (dump_flags_t dump_kind, dump_location_t loc,
3217 : : slp_tree node, hash_set<slp_tree> &visited)
3218 : : {
3219 : 496441 : unsigned i;
3220 : 496441 : slp_tree child;
3221 : :
3222 : 496441 : if (visited.add (node))
3223 : 496441 : return;
3224 : :
3225 : 455579 : vect_print_slp_tree (dump_kind, loc, node);
3226 : :
3227 : 1380027 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
3228 : 468869 : if (child)
3229 : 417829 : vect_print_slp_graph (dump_kind, loc, child, visited);
3230 : : }
3231 : :
3232 : : static void
3233 : 47477 : vect_print_slp_graph (dump_flags_t dump_kind, dump_location_t loc,
3234 : : slp_tree entry)
3235 : : {
3236 : 47477 : hash_set<slp_tree> visited;
3237 : 47477 : vect_print_slp_graph (dump_kind, loc, entry, visited);
3238 : 47477 : }
3239 : :
3240 : : DEBUG_FUNCTION void
3241 : 0 : debug (slp_instance instance)
3242 : : {
3243 : 0 : debug_dump_context ctx;
3244 : 0 : vect_print_slp_graph (MSG_NOTE,
3245 : 0 : dump_location_t::from_location_t (UNKNOWN_LOCATION),
3246 : : SLP_INSTANCE_TREE (instance));
3247 : 0 : }
3248 : :
3249 : : /* Mark the tree rooted at NODE with PURE_SLP. */
3250 : :
3251 : : static void
3252 : 4628258 : vect_mark_slp_stmts (vec_info *vinfo, slp_tree node,
3253 : : hash_set<slp_tree> &visited)
3254 : : {
3255 : 4628258 : int i;
3256 : 4628258 : stmt_vec_info stmt_info;
3257 : 4628258 : slp_tree child;
3258 : :
3259 : 4628258 : if (SLP_TREE_DEF_TYPE (node) != vect_internal_def)
3260 : : return;
3261 : :
3262 : 3244855 : if (visited.add (node))
3263 : : return;
3264 : :
3265 : 8021294 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
3266 : 5009190 : if (stmt_info)
3267 : : {
3268 : 4870914 : STMT_SLP_TYPE (stmt_info) = pure_slp;
3269 : : /* ??? For .MASK_LOAD and .MASK_STORE detected as load/store-lanes
3270 : : when there is the mask_conversion pattern applied we have lost the
3271 : : alternate lanes of the uniform mask which nevertheless
3272 : : have separate pattern defs. To not confuse hybrid
3273 : : analysis we mark those as covered as well here. */
3274 : 4870914 : if (node->ldst_lanes)
3275 : 5009190 : if (gcall *call = dyn_cast <gcall *> (stmt_info->stmt))
3276 : 0 : if (gimple_call_internal_p (call, IFN_MASK_LOAD)
3277 : 0 : || gimple_call_internal_p (call, IFN_MASK_STORE))
3278 : : {
3279 : 0 : tree mask = gimple_call_arg (call,
3280 : : internal_fn_mask_index
3281 : 0 : (gimple_call_internal_fn (call)));
3282 : 0 : if (TREE_CODE (mask) == SSA_NAME)
3283 : 0 : if (stmt_vec_info mask_info = vinfo->lookup_def (mask))
3284 : : {
3285 : 0 : mask_info = vect_stmt_to_vectorize (mask_info);
3286 : 0 : STMT_SLP_TYPE (mask_info) = pure_slp;
3287 : : }
3288 : : }
3289 : : }
3290 : :
3291 : 6937142 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
3292 : 3925038 : if (child)
3293 : 3432888 : vect_mark_slp_stmts (vinfo, child, visited);
3294 : : }
3295 : :
3296 : : static void
3297 : 1195370 : vect_mark_slp_stmts (vec_info *vinfo, slp_tree node)
3298 : : {
3299 : 1195370 : hash_set<slp_tree> visited;
3300 : 1195370 : vect_mark_slp_stmts (vinfo, node, visited);
3301 : 1195370 : }
3302 : :
3303 : : /* Mark the statements of the tree rooted at NODE as relevant (vect_used). */
3304 : :
3305 : : static void
3306 : 2211451 : vect_mark_slp_stmts_relevant (slp_tree node, hash_set<slp_tree> &visited)
3307 : : {
3308 : 2211451 : int i;
3309 : 2211451 : stmt_vec_info stmt_info;
3310 : 2211451 : slp_tree child;
3311 : :
3312 : 2211451 : if (SLP_TREE_DEF_TYPE (node) != vect_internal_def)
3313 : : return;
3314 : :
3315 : 1313778 : if (visited.add (node))
3316 : : return;
3317 : :
3318 : 4111149 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
3319 : 2885613 : if (stmt_info)
3320 : : {
3321 : 2885613 : gcc_assert (!STMT_VINFO_RELEVANT (stmt_info)
3322 : : || STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_scope);
3323 : 2885613 : STMT_VINFO_RELEVANT (stmt_info) = vect_used_in_scope;
3324 : : }
3325 : :
3326 : 2680714 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
3327 : 1455178 : if (child)
3328 : 1455178 : vect_mark_slp_stmts_relevant (child, visited);
3329 : : }
3330 : :
3331 : : static void
3332 : 756273 : vect_mark_slp_stmts_relevant (slp_tree node)
3333 : : {
3334 : 756273 : hash_set<slp_tree> visited;
3335 : 756273 : vect_mark_slp_stmts_relevant (node, visited);
3336 : 756273 : }
3337 : :
3338 : :
3339 : : /* Gather loads in the SLP graph NODE and populate the INST loads array. */
3340 : :
3341 : : static void
3342 : 7776667 : vect_gather_slp_loads (vec<slp_tree> &loads, slp_tree node,
3343 : : hash_set<slp_tree> &visited)
3344 : : {
3345 : 7776667 : if (!node || visited.add (node))
3346 : 1358781 : return;
3347 : :
3348 : 6417886 : if (SLP_TREE_DEF_TYPE (node) != vect_internal_def)
3349 : : return;
3350 : :
3351 : 4586217 : if (SLP_TREE_CODE (node) != VEC_PERM_EXPR)
3352 : : {
3353 : 4421895 : stmt_vec_info stmt_info = SLP_TREE_REPRESENTATIVE (node);
3354 : 4421895 : if (STMT_VINFO_DATA_REF (stmt_info)
3355 : 1858957 : && DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info)))
3356 : 834079 : loads.safe_push (node);
3357 : : }
3358 : :
3359 : : unsigned i;
3360 : : slp_tree child;
3361 : 10722327 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
3362 : 6136110 : vect_gather_slp_loads (loads, child, visited);
3363 : : }
3364 : :
3365 : :
3366 : : /* Find the last store in SLP INSTANCE. */
3367 : :
3368 : : stmt_vec_info
3369 : 2685148 : vect_find_last_scalar_stmt_in_slp (slp_tree node)
3370 : : {
3371 : 2685148 : stmt_vec_info last = NULL;
3372 : 2685148 : stmt_vec_info stmt_vinfo;
3373 : :
3374 : 9671610 : for (int i = 0; SLP_TREE_SCALAR_STMTS (node).iterate (i, &stmt_vinfo); i++)
3375 : 6986462 : if (stmt_vinfo)
3376 : : {
3377 : 6986462 : stmt_vinfo = vect_orig_stmt (stmt_vinfo);
3378 : 6986462 : last = last ? get_later_stmt (stmt_vinfo, last) : stmt_vinfo;
3379 : : }
3380 : :
3381 : 2685148 : return last;
3382 : : }
3383 : :
3384 : : /* Find the first stmt in NODE. */
3385 : :
3386 : : stmt_vec_info
3387 : 540000 : vect_find_first_scalar_stmt_in_slp (slp_tree node)
3388 : : {
3389 : 540000 : stmt_vec_info first = NULL;
3390 : 540000 : stmt_vec_info stmt_vinfo;
3391 : :
3392 : 1802075 : for (int i = 0; SLP_TREE_SCALAR_STMTS (node).iterate (i, &stmt_vinfo); i++)
3393 : 1262075 : if (stmt_vinfo)
3394 : : {
3395 : 1259547 : stmt_vinfo = vect_orig_stmt (stmt_vinfo);
3396 : 1259547 : if (!first
3397 : 1259547 : || get_later_stmt (stmt_vinfo, first) == first)
3398 : : first = stmt_vinfo;
3399 : : }
3400 : :
3401 : 540000 : return first;
3402 : : }
3403 : :
3404 : : /* Splits a group of stores, currently beginning at FIRST_VINFO, into
3405 : : two groups: one (still beginning at FIRST_VINFO) of size GROUP1_SIZE
3406 : : (also containing the first GROUP1_SIZE stmts, since stores are
3407 : : consecutive), the second containing the remainder.
3408 : : Return the first stmt in the second group. */
3409 : :
3410 : : static stmt_vec_info
3411 : 145705 : vect_split_slp_store_group (stmt_vec_info first_vinfo, unsigned group1_size)
3412 : : {
3413 : 145705 : gcc_assert (DR_GROUP_FIRST_ELEMENT (first_vinfo) == first_vinfo);
3414 : 145705 : gcc_assert (group1_size > 0);
3415 : 145705 : int group2_size = DR_GROUP_SIZE (first_vinfo) - group1_size;
3416 : 145705 : gcc_assert (group2_size > 0);
3417 : 145705 : DR_GROUP_SIZE (first_vinfo) = group1_size;
3418 : :
3419 : 145705 : stmt_vec_info stmt_info = first_vinfo;
3420 : 488310 : for (unsigned i = group1_size; i > 1; i--)
3421 : : {
3422 : 342605 : stmt_info = DR_GROUP_NEXT_ELEMENT (stmt_info);
3423 : 342605 : gcc_assert (DR_GROUP_GAP (stmt_info) == 1);
3424 : : }
3425 : : /* STMT is now the last element of the first group. */
3426 : 145705 : stmt_vec_info group2 = DR_GROUP_NEXT_ELEMENT (stmt_info);
3427 : 145705 : DR_GROUP_NEXT_ELEMENT (stmt_info) = 0;
3428 : :
3429 : 145705 : DR_GROUP_SIZE (group2) = group2_size;
3430 : 408976 : for (stmt_info = group2; stmt_info;
3431 : 263271 : stmt_info = DR_GROUP_NEXT_ELEMENT (stmt_info))
3432 : : {
3433 : 263271 : DR_GROUP_FIRST_ELEMENT (stmt_info) = group2;
3434 : 263271 : gcc_assert (DR_GROUP_GAP (stmt_info) == 1);
3435 : : }
3436 : :
3437 : : /* For the second group, the DR_GROUP_GAP is that before the original group,
3438 : : plus skipping over the first vector. */
3439 : 145705 : DR_GROUP_GAP (group2) = DR_GROUP_GAP (first_vinfo) + group1_size;
3440 : :
3441 : : /* DR_GROUP_GAP of the first group now has to skip over the second group too. */
3442 : 145705 : DR_GROUP_GAP (first_vinfo) += group2_size;
3443 : :
3444 : 145705 : if (dump_enabled_p ())
3445 : 59 : dump_printf_loc (MSG_NOTE, vect_location, "Split group into %d and %d\n",
3446 : : group1_size, group2_size);
3447 : :
3448 : 145705 : return group2;
3449 : : }
3450 : :
3451 : : /* Calculate the unrolling factor for an SLP instance with GROUP_SIZE
3452 : : statements and a vector of NUNITS elements. */
3453 : :
3454 : : static poly_uint64
3455 : 2838589 : calculate_unrolling_factor (poly_uint64 nunits, unsigned int group_size)
3456 : : {
3457 : 2838589 : return exact_div (common_multiple (nunits, group_size), group_size);
3458 : : }
3459 : :
3460 : : /* Helper that checks to see if a node is a load node. */
3461 : :
3462 : : static inline bool
3463 : 88 : vect_is_slp_load_node (slp_tree root)
3464 : : {
3465 : 88 : return (SLP_TREE_CODE (root) != VEC_PERM_EXPR
3466 : 76 : && SLP_TREE_DEF_TYPE (root) == vect_internal_def
3467 : 64 : && STMT_VINFO_GROUPED_ACCESS (SLP_TREE_REPRESENTATIVE (root))
3468 : 128 : && DR_IS_READ (STMT_VINFO_DATA_REF (SLP_TREE_REPRESENTATIVE (root))));
3469 : : }
3470 : :
3471 : :
3472 : : /* Helper function of optimize_load_redistribution that performs the operation
3473 : : recursively. */
3474 : :
3475 : : static slp_tree
3476 : 8600 : optimize_load_redistribution_1 (scalar_stmts_to_slp_tree_map_t *bst_map,
3477 : : vec_info *vinfo, unsigned int group_size,
3478 : : hash_map<slp_tree, slp_tree> *load_map,
3479 : : slp_tree root)
3480 : : {
3481 : 8600 : if (slp_tree *leader = load_map->get (root))
3482 : 644 : return *leader;
3483 : :
3484 : 7956 : slp_tree node;
3485 : 7956 : unsigned i;
3486 : :
3487 : : /* For now, we don't know anything about externals so do not do anything. */
3488 : 7956 : if (!root || SLP_TREE_DEF_TYPE (root) != vect_internal_def)
3489 : : return NULL;
3490 : 5516 : else if (SLP_TREE_CODE (root) == VEC_PERM_EXPR)
3491 : : {
3492 : : /* First convert this node into a load node and add it to the leaves
3493 : : list and flatten the permute from a lane to a load one. If it's
3494 : : unneeded it will be elided later. */
3495 : 68 : vec<stmt_vec_info> stmts;
3496 : 68 : stmts.create (SLP_TREE_LANES (root));
3497 : 68 : lane_permutation_t lane_perm = SLP_TREE_LANE_PERMUTATION (root);
3498 : 108 : for (unsigned j = 0; j < lane_perm.length (); j++)
3499 : : {
3500 : 88 : std::pair<unsigned, unsigned> perm = lane_perm[j];
3501 : 88 : node = SLP_TREE_CHILDREN (root)[perm.first];
3502 : :
3503 : 88 : if (!vect_is_slp_load_node (node)
3504 : 88 : || SLP_TREE_CHILDREN (node).exists ())
3505 : : {
3506 : 48 : stmts.release ();
3507 : 48 : goto next;
3508 : : }
3509 : :
3510 : 40 : stmts.quick_push (SLP_TREE_SCALAR_STMTS (node)[perm.second]);
3511 : : }
3512 : :
3513 : 20 : if (dump_enabled_p ())
3514 : 0 : dump_printf_loc (MSG_NOTE, vect_location,
3515 : : "converting stmts on permute node %p\n",
3516 : : (void *) root);
3517 : :
3518 : 20 : bool *matches = XALLOCAVEC (bool, group_size);
3519 : 20 : poly_uint64 max_nunits = 1;
3520 : 20 : unsigned tree_size = 0, limit = 1;
3521 : 20 : node = vect_build_slp_tree (vinfo, stmts, group_size, &max_nunits,
3522 : : matches, &limit, &tree_size, bst_map);
3523 : 20 : if (!node)
3524 : 0 : stmts.release ();
3525 : :
3526 : 20 : load_map->put (root, node);
3527 : 20 : return node;
3528 : : }
3529 : :
3530 : 5448 : next:
3531 : 5496 : load_map->put (root, NULL);
3532 : :
3533 : 12728 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (root), i , node)
3534 : : {
3535 : 7232 : slp_tree value
3536 : 7232 : = optimize_load_redistribution_1 (bst_map, vinfo, group_size, load_map,
3537 : : node);
3538 : 7232 : if (value)
3539 : : {
3540 : 20 : SLP_TREE_REF_COUNT (value)++;
3541 : 20 : SLP_TREE_CHILDREN (root)[i] = value;
3542 : : /* ??? We know the original leafs of the replaced nodes will
3543 : : be referenced by bst_map, only the permutes created by
3544 : : pattern matching are not. */
3545 : 20 : if (SLP_TREE_REF_COUNT (node) == 1)
3546 : 20 : load_map->remove (node);
3547 : 20 : vect_free_slp_tree (node);
3548 : : }
3549 : : }
3550 : :
3551 : : return NULL;
3552 : : }
3553 : :
3554 : : /* Temporary workaround for loads not being CSEd during SLP build. This
3555 : : function will traverse the SLP tree rooted in ROOT for INSTANCE and find
3556 : : VEC_PERM nodes that blend vectors from multiple nodes that all read from the
3557 : : same DR such that the final operation is equal to a permuted load. Such
3558 : : NODES are then directly converted into LOADS themselves. The nodes are
3559 : : CSEd using BST_MAP. */
3560 : :
3561 : : static void
3562 : 1251 : optimize_load_redistribution (scalar_stmts_to_slp_tree_map_t *bst_map,
3563 : : vec_info *vinfo, unsigned int group_size,
3564 : : hash_map<slp_tree, slp_tree> *load_map,
3565 : : slp_tree root)
3566 : : {
3567 : 1251 : slp_tree node;
3568 : 1251 : unsigned i;
3569 : :
3570 : 2619 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (root), i , node)
3571 : : {
3572 : 1368 : slp_tree value
3573 : 1368 : = optimize_load_redistribution_1 (bst_map, vinfo, group_size, load_map,
3574 : : node);
3575 : 1368 : if (value)
3576 : : {
3577 : 0 : SLP_TREE_REF_COUNT (value)++;
3578 : 0 : SLP_TREE_CHILDREN (root)[i] = value;
3579 : : /* ??? We know the original leafs of the replaced nodes will
3580 : : be referenced by bst_map, only the permutes created by
3581 : : pattern matching are not. */
3582 : 0 : if (SLP_TREE_REF_COUNT (node) == 1)
3583 : 0 : load_map->remove (node);
3584 : 0 : vect_free_slp_tree (node);
3585 : : }
3586 : : }
3587 : 1251 : }
3588 : :
3589 : : /* Helper function of vect_match_slp_patterns.
3590 : :
3591 : : Attempts to match patterns against the slp tree rooted in REF_NODE using
3592 : : VINFO. Patterns are matched in post-order traversal.
3593 : :
3594 : : If matching is successful the value in REF_NODE is updated and returned, if
3595 : : not then it is returned unchanged. */
3596 : :
3597 : : static bool
3598 : 4566958 : vect_match_slp_patterns_2 (slp_tree *ref_node, vec_info *vinfo,
3599 : : slp_tree_to_load_perm_map_t *perm_cache,
3600 : : slp_compat_nodes_map_t *compat_cache,
3601 : : hash_set<slp_tree> *visited)
3602 : : {
3603 : 4566958 : unsigned i;
3604 : 4566958 : slp_tree node = *ref_node;
3605 : 4566958 : bool found_p = false;
3606 : 4566958 : if (!node || visited->add (node))
3607 : 658933 : return false;
3608 : :
3609 : : slp_tree child;
3610 : 7273523 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
3611 : 3365498 : found_p |= vect_match_slp_patterns_2 (&SLP_TREE_CHILDREN (node)[i],
3612 : : vinfo, perm_cache, compat_cache,
3613 : : visited);
3614 : :
3615 : 11724075 : for (unsigned x = 0; x < num__slp_patterns; x++)
3616 : : {
3617 : 7816050 : vect_pattern *pattern
3618 : 7816050 : = slp_patterns[x] (perm_cache, compat_cache, ref_node);
3619 : 7816050 : if (pattern)
3620 : : {
3621 : 448 : pattern->build (vinfo);
3622 : 448 : delete pattern;
3623 : 448 : found_p = true;
3624 : : }
3625 : : }
3626 : :
3627 : : return found_p;
3628 : : }
3629 : :
3630 : : /* Applies pattern matching to the given SLP tree rooted in REF_NODE using
3631 : : vec_info VINFO.
3632 : :
3633 : : The modified tree is returned. Patterns are tried in order and multiple
3634 : : patterns may match. */
3635 : :
3636 : : static bool
3637 : 1201460 : vect_match_slp_patterns (slp_instance instance, vec_info *vinfo,
3638 : : hash_set<slp_tree> *visited,
3639 : : slp_tree_to_load_perm_map_t *perm_cache,
3640 : : slp_compat_nodes_map_t *compat_cache)
3641 : : {
3642 : 1201460 : DUMP_VECT_SCOPE ("vect_match_slp_patterns");
3643 : 1201460 : slp_tree *ref_node = &SLP_INSTANCE_TREE (instance);
3644 : :
3645 : 1201460 : if (dump_enabled_p ())
3646 : 32217 : dump_printf_loc (MSG_NOTE, vect_location,
3647 : : "Analyzing SLP tree %p for patterns\n",
3648 : 32217 : (void *) SLP_INSTANCE_TREE (instance));
3649 : :
3650 : 1201460 : return vect_match_slp_patterns_2 (ref_node, vinfo, perm_cache, compat_cache,
3651 : 1201460 : visited);
3652 : : }
3653 : :
3654 : : /* STMT_INFO is a store group of size GROUP_SIZE that we are considering
3655 : : vectorizing with VECTYPE that might be NULL. MASKED_P indicates whether
3656 : : the stores are masked.
3657 : : Return true if we could use IFN_STORE_LANES instead and if that appears
3658 : : to be the better approach. */
3659 : :
3660 : : static bool
3661 : 4208 : vect_slp_prefer_store_lanes_p (vec_info *vinfo, stmt_vec_info stmt_info,
3662 : : tree vectype, bool masked_p,
3663 : : unsigned int group_size,
3664 : : unsigned int new_group_size)
3665 : : {
3666 : 4208 : if (!vectype)
3667 : : {
3668 : 4208 : tree scalar_type = TREE_TYPE (DR_REF (STMT_VINFO_DATA_REF (stmt_info)));
3669 : 4208 : vectype = get_vectype_for_scalar_type (vinfo, scalar_type);
3670 : : }
3671 : 4208 : if (!vectype)
3672 : : return false;
3673 : : /* Allow the split if one of the two new groups would operate on full
3674 : : vectors *within* rather than across one scalar loop iteration.
3675 : : This is purely a heuristic, but it should work well for group
3676 : : sizes of 3 and 4, where the possible splits are:
3677 : :
3678 : : 3->2+1: OK if the vector has exactly two elements
3679 : : 4->2+2: Likewise
3680 : : 4->3+1: Less clear-cut. */
3681 : 4208 : if (multiple_p (group_size - new_group_size, TYPE_VECTOR_SUBPARTS (vectype))
3682 : 2643 : || multiple_p (new_group_size, TYPE_VECTOR_SUBPARTS (vectype)))
3683 : 1565 : return false;
3684 : 2643 : return vect_store_lanes_supported (vectype, group_size, masked_p) != IFN_LAST;
3685 : : }
3686 : :
3687 : : /* Analyze an SLP instance starting from a group of grouped stores. Call
3688 : : vect_build_slp_tree to build a tree of packed stmts if possible.
3689 : : Return FALSE if it's impossible to SLP any stmt in the loop. */
3690 : :
3691 : : static bool
3692 : : vect_analyze_slp_instance (vec_info *vinfo,
3693 : : scalar_stmts_to_slp_tree_map_t *bst_map,
3694 : : stmt_vec_info stmt_info, slp_instance_kind kind,
3695 : : unsigned max_tree_size, unsigned *limit,
3696 : : bool force_single_lane);
3697 : :
3698 : : /* Build an interleaving scheme for the store sources RHS_NODES from
3699 : : SCALAR_STMTS. */
3700 : :
3701 : : static slp_tree
3702 : 6102 : vect_build_slp_store_interleaving (vec<slp_tree> &rhs_nodes,
3703 : : vec<stmt_vec_info> &scalar_stmts,
3704 : : poly_uint64 max_nunits)
3705 : : {
3706 : 6102 : unsigned int group_size = scalar_stmts.length ();
3707 : 12204 : slp_tree node = vect_create_new_slp_node (scalar_stmts,
3708 : 6102 : SLP_TREE_CHILDREN
3709 : : (rhs_nodes[0]).length ());
3710 : 6102 : SLP_TREE_VECTYPE (node) = SLP_TREE_VECTYPE (rhs_nodes[0]);
3711 : 6102 : node->max_nunits = max_nunits;
3712 : 6102 : for (unsigned l = 0;
3713 : 12239 : l < SLP_TREE_CHILDREN (rhs_nodes[0]).length (); ++l)
3714 : : {
3715 : : /* And a permute merging all RHS SLP trees. */
3716 : 6137 : slp_tree perm = vect_create_new_slp_node (rhs_nodes.length (),
3717 : 6137 : VEC_PERM_EXPR);
3718 : 6137 : SLP_TREE_CHILDREN (node).quick_push (perm);
3719 : 6137 : SLP_TREE_LANE_PERMUTATION (perm).create (group_size);
3720 : 6137 : SLP_TREE_VECTYPE (perm) = SLP_TREE_VECTYPE (node);
3721 : 6137 : perm->max_nunits = max_nunits;
3722 : 6137 : SLP_TREE_LANES (perm) = group_size;
3723 : : /* ??? We should set this NULL but that's not expected. */
3724 : 6137 : SLP_TREE_REPRESENTATIVE (perm)
3725 : 6137 : = SLP_TREE_REPRESENTATIVE (SLP_TREE_CHILDREN (rhs_nodes[0])[l]);
3726 : 24143 : for (unsigned j = 0; j < rhs_nodes.length (); ++j)
3727 : : {
3728 : 18006 : SLP_TREE_CHILDREN (perm)
3729 : 18006 : .quick_push (SLP_TREE_CHILDREN (rhs_nodes[j])[l]);
3730 : 18006 : SLP_TREE_CHILDREN (rhs_nodes[j])[l]->refcnt++;
3731 : 18006 : for (unsigned k = 0;
3732 : 38213 : k < SLP_TREE_SCALAR_STMTS (rhs_nodes[j]).length (); ++k)
3733 : : {
3734 : : /* ??? We should populate SLP_TREE_SCALAR_STMTS
3735 : : or SLP_TREE_SCALAR_OPS but then we might have
3736 : : a mix of both in our children. */
3737 : 20207 : SLP_TREE_LANE_PERMUTATION (perm)
3738 : 20207 : .quick_push (std::make_pair (j, k));
3739 : : }
3740 : : }
3741 : :
3742 : : /* Now we have a single permute node but we cannot code-generate
3743 : : the case with more than two inputs.
3744 : : Perform pairwise reduction, reducing the two inputs
3745 : : with the least number of lanes to one and then repeat until
3746 : : we end up with two inputs. That scheme makes sure we end
3747 : : up with permutes satisfying the restriction of requiring at
3748 : : most two vector inputs to produce a single vector output
3749 : : when the number of lanes is even. */
3750 : 11869 : while (SLP_TREE_CHILDREN (perm).length () > 2)
3751 : : {
3752 : : /* When we have three equal sized groups left the pairwise
3753 : : reduction does not result in a scheme that avoids using
3754 : : three vectors. Instead merge the first two groups
3755 : : to the final size with do-not-care elements (chosen
3756 : : from the first group) and then merge with the third.
3757 : : { A0, B0, x, A1, B1, x, ... }
3758 : : -> { A0, B0, C0, A1, B1, C1, ... }
3759 : : This handles group size of three (and at least
3760 : : power-of-two multiples of that). */
3761 : 5732 : if (SLP_TREE_CHILDREN (perm).length () == 3
3762 : 2620 : && (SLP_TREE_LANES (SLP_TREE_CHILDREN (perm)[0])
3763 : 2620 : == SLP_TREE_LANES (SLP_TREE_CHILDREN (perm)[1]))
3764 : 5732 : && (SLP_TREE_LANES (SLP_TREE_CHILDREN (perm)[0])
3765 : 1778 : == SLP_TREE_LANES (SLP_TREE_CHILDREN (perm)[2])))
3766 : : {
3767 : 1591 : int ai = 0;
3768 : 1591 : int bi = 1;
3769 : 1591 : slp_tree a = SLP_TREE_CHILDREN (perm)[ai];
3770 : 1591 : slp_tree b = SLP_TREE_CHILDREN (perm)[bi];
3771 : 1591 : unsigned n = SLP_TREE_LANES (perm);
3772 : :
3773 : 1591 : slp_tree permab = vect_create_new_slp_node (2, VEC_PERM_EXPR);
3774 : 1591 : SLP_TREE_LANES (permab) = n;
3775 : 1591 : SLP_TREE_LANE_PERMUTATION (permab).create (n);
3776 : 1591 : SLP_TREE_VECTYPE (permab) = SLP_TREE_VECTYPE (perm);
3777 : 1591 : permab->max_nunits = max_nunits;
3778 : : /* ??? Should be NULL but that's not expected. */
3779 : 1591 : SLP_TREE_REPRESENTATIVE (permab) = SLP_TREE_REPRESENTATIVE (perm);
3780 : 1591 : SLP_TREE_CHILDREN (permab).quick_push (a);
3781 : 3193 : for (unsigned k = 0; k < SLP_TREE_LANES (a); ++k)
3782 : 1602 : SLP_TREE_LANE_PERMUTATION (permab)
3783 : 1602 : .quick_push (std::make_pair (0, k));
3784 : 1591 : SLP_TREE_CHILDREN (permab).quick_push (b);
3785 : 3193 : for (unsigned k = 0; k < SLP_TREE_LANES (b); ++k)
3786 : 1602 : SLP_TREE_LANE_PERMUTATION (permab)
3787 : 1602 : .quick_push (std::make_pair (1, k));
3788 : : /* Push the do-not-care lanes. */
3789 : 3193 : for (unsigned k = 0; k < SLP_TREE_LANES (a); ++k)
3790 : 1602 : SLP_TREE_LANE_PERMUTATION (permab)
3791 : 1602 : .quick_push (std::make_pair (0, k));
3792 : :
3793 : : /* Put the merged node into 'perm', in place of a. */
3794 : 1591 : SLP_TREE_CHILDREN (perm)[ai] = permab;
3795 : : /* Adjust the references to b in the permutation
3796 : : of perm and to the later children which we'll
3797 : : remove. */
3798 : 6397 : for (unsigned k = 0; k < SLP_TREE_LANES (perm); ++k)
3799 : : {
3800 : 4806 : std::pair<unsigned, unsigned> &p
3801 : 4806 : = SLP_TREE_LANE_PERMUTATION (perm)[k];
3802 : 4806 : if (p.first == (unsigned) bi)
3803 : : {
3804 : 1602 : p.first = ai;
3805 : 1602 : p.second += SLP_TREE_LANES (a);
3806 : : }
3807 : 3204 : else if (p.first > (unsigned) bi)
3808 : 1602 : p.first--;
3809 : : }
3810 : 1591 : SLP_TREE_CHILDREN (perm).ordered_remove (bi);
3811 : 1591 : break;
3812 : : }
3813 : :
3814 : : /* Pick the two nodes with the least number of lanes,
3815 : : prefer the earliest candidate and maintain ai < bi. */
3816 : : int ai = -1;
3817 : : int bi = -1;
3818 : 35360 : for (unsigned ci = 0; ci < SLP_TREE_CHILDREN (perm).length (); ++ci)
3819 : : {
3820 : 31219 : if (ai == -1)
3821 : 4141 : ai = ci;
3822 : 27078 : else if (bi == -1)
3823 : 4141 : bi = ci;
3824 : 22937 : else if ((SLP_TREE_LANES (SLP_TREE_CHILDREN (perm)[ci])
3825 : 22937 : < SLP_TREE_LANES (SLP_TREE_CHILDREN (perm)[ai]))
3826 : 22937 : || (SLP_TREE_LANES (SLP_TREE_CHILDREN (perm)[ci])
3827 : 18682 : < SLP_TREE_LANES (SLP_TREE_CHILDREN (perm)[bi])))
3828 : : {
3829 : 9376 : if (SLP_TREE_LANES (SLP_TREE_CHILDREN (perm)[ai])
3830 : 4688 : <= SLP_TREE_LANES (SLP_TREE_CHILDREN (perm)[bi]))
3831 : 2174 : bi = ci;
3832 : : else
3833 : : {
3834 : 2514 : ai = bi;
3835 : 2514 : bi = ci;
3836 : : }
3837 : : }
3838 : : }
3839 : :
3840 : : /* Produce a merge of nodes ai and bi. */
3841 : 4141 : slp_tree a = SLP_TREE_CHILDREN (perm)[ai];
3842 : 4141 : slp_tree b = SLP_TREE_CHILDREN (perm)[bi];
3843 : 4141 : unsigned n = SLP_TREE_LANES (a) + SLP_TREE_LANES (b);
3844 : 4141 : slp_tree permab = vect_create_new_slp_node (2, VEC_PERM_EXPR);
3845 : 4141 : SLP_TREE_LANES (permab) = n;
3846 : 4141 : SLP_TREE_LANE_PERMUTATION (permab).create (n);
3847 : 4141 : SLP_TREE_VECTYPE (permab) = SLP_TREE_VECTYPE (perm);
3848 : 4141 : permab->max_nunits = max_nunits;
3849 : : /* ??? Should be NULL but that's not expected. */
3850 : 4141 : SLP_TREE_REPRESENTATIVE (permab) = SLP_TREE_REPRESENTATIVE (perm);
3851 : 4141 : SLP_TREE_CHILDREN (permab).quick_push (a);
3852 : 10667 : for (unsigned k = 0; k < SLP_TREE_LANES (a); ++k)
3853 : 6526 : SLP_TREE_LANE_PERMUTATION (permab)
3854 : 6526 : .quick_push (std::make_pair (0, k));
3855 : 4141 : SLP_TREE_CHILDREN (permab).quick_push (b);
3856 : 10175 : for (unsigned k = 0; k < SLP_TREE_LANES (b); ++k)
3857 : 6034 : SLP_TREE_LANE_PERMUTATION (permab)
3858 : 6034 : .quick_push (std::make_pair (1, k));
3859 : :
3860 : : /* Put the merged node into 'perm', in place of a. */
3861 : 4141 : SLP_TREE_CHILDREN (perm)[ai] = permab;
3862 : : /* Adjust the references to b in the permutation
3863 : : of perm and to the later children which we'll
3864 : : remove. */
3865 : 55734 : for (unsigned k = 0; k < SLP_TREE_LANES (perm); ++k)
3866 : : {
3867 : 51593 : std::pair<unsigned, unsigned> &p
3868 : 51593 : = SLP_TREE_LANE_PERMUTATION (perm)[k];
3869 : 51593 : if (p.first == (unsigned) bi)
3870 : : {
3871 : 6034 : p.first = ai;
3872 : 6034 : p.second += SLP_TREE_LANES (a);
3873 : : }
3874 : 45559 : else if (p.first > (unsigned) bi)
3875 : 18738 : p.first--;
3876 : : }
3877 : 4141 : SLP_TREE_CHILDREN (perm).ordered_remove (bi);
3878 : : }
3879 : : }
3880 : :
3881 : 6102 : return node;
3882 : : }
3883 : :
3884 : : /* Analyze an SLP instance starting from SCALAR_STMTS which are a group
3885 : : of KIND. Return true if successful. */
3886 : :
3887 : : static bool
3888 : 2568337 : vect_build_slp_instance (vec_info *vinfo,
3889 : : slp_instance_kind kind,
3890 : : vec<stmt_vec_info> &scalar_stmts,
3891 : : vec<stmt_vec_info> &root_stmt_infos,
3892 : : vec<tree> &remain,
3893 : : unsigned max_tree_size, unsigned *limit,
3894 : : scalar_stmts_to_slp_tree_map_t *bst_map,
3895 : : /* ??? We need stmt_info for group splitting. */
3896 : : stmt_vec_info stmt_info_,
3897 : : bool force_single_lane)
3898 : : {
3899 : : /* If there's no budget left bail out early. */
3900 : 2568337 : if (*limit == 0)
3901 : : return false;
3902 : :
3903 : 2543007 : if (kind == slp_inst_kind_ctor)
3904 : : {
3905 : 7666 : if (dump_enabled_p ())
3906 : 66 : dump_printf_loc (MSG_NOTE, vect_location,
3907 : : "Analyzing vectorizable constructor: %G\n",
3908 : 33 : root_stmt_infos[0]->stmt);
3909 : : }
3910 : 2535341 : else if (kind == slp_inst_kind_gcond)
3911 : : {
3912 : 63413 : if (dump_enabled_p ())
3913 : 3246 : dump_printf_loc (MSG_NOTE, vect_location,
3914 : : "Analyzing vectorizable control flow: %G",
3915 : 1623 : root_stmt_infos[0]->stmt);
3916 : : }
3917 : :
3918 : 2543007 : if (dump_enabled_p ())
3919 : : {
3920 : 35720 : dump_printf_loc (MSG_NOTE, vect_location,
3921 : : "Starting SLP discovery for\n");
3922 : 91162 : for (unsigned i = 0; i < scalar_stmts.length (); ++i)
3923 : 110884 : dump_printf_loc (MSG_NOTE, vect_location,
3924 : 55442 : " %G", scalar_stmts[i]->stmt);
3925 : : }
3926 : :
3927 : : /* Build the tree for the SLP instance. */
3928 : 2543007 : unsigned int group_size = scalar_stmts.length ();
3929 : 2543007 : bool *matches = XALLOCAVEC (bool, group_size);
3930 : 2543007 : poly_uint64 max_nunits = 1;
3931 : 2543007 : unsigned tree_size = 0;
3932 : 2543007 : unsigned i;
3933 : :
3934 : 2543007 : slp_tree node = NULL;
3935 : 2543007 : if (group_size > 1 && force_single_lane)
3936 : : {
3937 : 1683 : matches[0] = true;
3938 : 1683 : matches[1] = false;
3939 : : }
3940 : : else
3941 : 2541324 : node = vect_build_slp_tree (vinfo, scalar_stmts, group_size,
3942 : : &max_nunits, matches, limit,
3943 : : &tree_size, bst_map);
3944 : 2543007 : if (node != NULL)
3945 : : {
3946 : : /* Calculate the unrolling factor based on the smallest type. */
3947 : 1195358 : poly_uint64 unrolling_factor
3948 : 1195358 : = calculate_unrolling_factor (max_nunits, group_size);
3949 : :
3950 : 1195358 : if (maybe_ne (unrolling_factor, 1U)
3951 : 1195358 : && is_a <bb_vec_info> (vinfo))
3952 : : {
3953 : 0 : unsigned HOST_WIDE_INT const_max_nunits;
3954 : 0 : if (!max_nunits.is_constant (&const_max_nunits)
3955 : 0 : || const_max_nunits > group_size)
3956 : : {
3957 : 0 : if (dump_enabled_p ())
3958 : 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
3959 : : "Build SLP failed: store group "
3960 : : "size not a multiple of the vector size "
3961 : : "in basic block SLP\n");
3962 : 0 : vect_free_slp_tree (node);
3963 : 0 : return false;
3964 : : }
3965 : : /* Fatal mismatch. */
3966 : 0 : if (dump_enabled_p ())
3967 : 0 : dump_printf_loc (MSG_NOTE, vect_location,
3968 : : "SLP discovery succeeded but node needs "
3969 : : "splitting\n");
3970 : 0 : memset (matches, true, group_size);
3971 : 0 : matches[group_size / const_max_nunits * const_max_nunits] = false;
3972 : 0 : vect_free_slp_tree (node);
3973 : : }
3974 : : else
3975 : : {
3976 : : /* Create a new SLP instance. */
3977 : 1195358 : slp_instance new_instance = XNEW (class _slp_instance);
3978 : 1195358 : SLP_INSTANCE_TREE (new_instance) = node;
3979 : 1195358 : SLP_INSTANCE_LOADS (new_instance) = vNULL;
3980 : 1195358 : SLP_INSTANCE_ROOT_STMTS (new_instance) = root_stmt_infos;
3981 : 1195358 : SLP_INSTANCE_REMAIN_DEFS (new_instance) = remain;
3982 : 1195358 : SLP_INSTANCE_KIND (new_instance) = kind;
3983 : 1195358 : new_instance->reduc_phis = NULL;
3984 : 1195358 : new_instance->cost_vec = vNULL;
3985 : 1195358 : new_instance->subgraph_entries = vNULL;
3986 : :
3987 : 1195358 : if (dump_enabled_p ())
3988 : 31387 : dump_printf_loc (MSG_NOTE, vect_location,
3989 : : "SLP size %u vs. limit %u.\n",
3990 : : tree_size, max_tree_size);
3991 : :
3992 : : /* Fixup SLP reduction chains. */
3993 : 1195358 : if (kind == slp_inst_kind_reduc_chain)
3994 : : {
3995 : : /* If this is a reduction chain with a conversion in front
3996 : : amend the SLP tree with a node for that. */
3997 : 343 : gimple *scalar_def
3998 : 343 : = vect_orig_stmt (scalar_stmts[group_size - 1])->stmt;
3999 : 343 : if (STMT_VINFO_DEF_TYPE (scalar_stmts[0]) != vect_reduction_def)
4000 : : {
4001 : : /* Get at the conversion stmt - we know it's the single use
4002 : : of the last stmt of the reduction chain. */
4003 : 48 : use_operand_p use_p;
4004 : 48 : bool r = single_imm_use (gimple_assign_lhs (scalar_def),
4005 : : &use_p, &scalar_def);
4006 : 48 : gcc_assert (r);
4007 : 48 : stmt_vec_info next_info = vinfo->lookup_stmt (scalar_def);
4008 : 48 : next_info = vect_stmt_to_vectorize (next_info);
4009 : 48 : scalar_stmts = vNULL;
4010 : 48 : scalar_stmts.create (group_size);
4011 : 150 : for (unsigned i = 0; i < group_size; ++i)
4012 : 102 : scalar_stmts.quick_push (next_info);
4013 : 48 : slp_tree conv = vect_create_new_slp_node (scalar_stmts, 1);
4014 : 48 : SLP_TREE_VECTYPE (conv) = STMT_VINFO_VECTYPE (next_info);
4015 : 48 : SLP_TREE_CHILDREN (conv).quick_push (node);
4016 : 48 : SLP_INSTANCE_TREE (new_instance) = conv;
4017 : : /* We also have to fake this conversion stmt as SLP reduction
4018 : : group so we don't have to mess with too much code
4019 : : elsewhere. */
4020 : 48 : REDUC_GROUP_FIRST_ELEMENT (next_info) = next_info;
4021 : 48 : REDUC_GROUP_NEXT_ELEMENT (next_info) = NULL;
4022 : : }
4023 : : /* Fill the backedge child of the PHI SLP node. The
4024 : : general matching code cannot find it because the
4025 : : scalar code does not reflect how we vectorize the
4026 : : reduction. */
4027 : 343 : use_operand_p use_p;
4028 : 343 : imm_use_iterator imm_iter;
4029 : 343 : class loop *loop = LOOP_VINFO_LOOP (as_a <loop_vec_info> (vinfo));
4030 : 1092 : FOR_EACH_IMM_USE_FAST (use_p, imm_iter,
4031 : : gimple_get_lhs (scalar_def))
4032 : : /* There are exactly two non-debug uses, the reduction
4033 : : PHI and the loop-closed PHI node. */
4034 : 749 : if (!is_gimple_debug (USE_STMT (use_p))
4035 : 749 : && gimple_bb (USE_STMT (use_p)) == loop->header)
4036 : : {
4037 : 343 : auto_vec<stmt_vec_info, 64> phis (group_size);
4038 : 343 : stmt_vec_info phi_info
4039 : 343 : = vinfo->lookup_stmt (USE_STMT (use_p));
4040 : 2128 : for (unsigned i = 0; i < group_size; ++i)
4041 : 1785 : phis.quick_push (phi_info);
4042 : 343 : slp_tree *phi_node = bst_map->get (phis);
4043 : 343 : unsigned dest_idx = loop_latch_edge (loop)->dest_idx;
4044 : 686 : SLP_TREE_CHILDREN (*phi_node)[dest_idx]
4045 : 343 : = SLP_INSTANCE_TREE (new_instance);
4046 : 343 : SLP_INSTANCE_TREE (new_instance)->refcnt++;
4047 : 343 : }
4048 : : }
4049 : :
4050 : 1195358 : vinfo->slp_instances.safe_push (new_instance);
4051 : :
4052 : : /* ??? We've replaced the old SLP_INSTANCE_GROUP_SIZE with
4053 : : the number of scalar stmts in the root in a few places.
4054 : : Verify that assumption holds. */
4055 : 2390716 : gcc_assert (SLP_TREE_SCALAR_STMTS (SLP_INSTANCE_TREE (new_instance))
4056 : : .length () == group_size);
4057 : :
4058 : 1195358 : if (dump_enabled_p ())
4059 : : {
4060 : 31387 : dump_printf_loc (MSG_NOTE, vect_location,
4061 : : "Final SLP tree for instance %p:\n",
4062 : : (void *) new_instance);
4063 : 31387 : vect_print_slp_graph (MSG_NOTE, vect_location,
4064 : : SLP_INSTANCE_TREE (new_instance));
4065 : : }
4066 : :
4067 : 1195358 : return true;
4068 : : }
4069 : : }
4070 : : /* Failed to SLP. */
4071 : :
4072 : 1347649 : stmt_vec_info stmt_info = stmt_info_;
4073 : : /* Try to break the group up into pieces. */
4074 : 1347649 : if (*limit > 0 && kind == slp_inst_kind_store)
4075 : : {
4076 : : /* ??? We could delay all the actual splitting of store-groups
4077 : : until after SLP discovery of the original group completed.
4078 : : Then we can recurse to vect_build_slp_instance directly. */
4079 : 1027455 : for (i = 0; i < group_size; i++)
4080 : 1027455 : if (!matches[i])
4081 : : break;
4082 : :
4083 : : /* For basic block SLP, try to break the group up into multiples of
4084 : : a vector size. */
4085 : 352863 : if (is_a <bb_vec_info> (vinfo)
4086 : 352863 : && (i > 1 && i < group_size))
4087 : : {
4088 : : /* Free the allocated memory. */
4089 : 143817 : scalar_stmts.release ();
4090 : :
4091 : 143817 : tree scalar_type
4092 : 143817 : = TREE_TYPE (DR_REF (STMT_VINFO_DATA_REF (stmt_info)));
4093 : 287634 : tree vectype = get_vectype_for_scalar_type (vinfo, scalar_type,
4094 : 143817 : 1 << floor_log2 (i));
4095 : 143817 : unsigned HOST_WIDE_INT const_nunits;
4096 : 143817 : if (vectype
4097 : 143817 : && TYPE_VECTOR_SUBPARTS (vectype).is_constant (&const_nunits))
4098 : : {
4099 : : /* Split into two groups at the first vector boundary. */
4100 : 143817 : gcc_assert ((const_nunits & (const_nunits - 1)) == 0);
4101 : 143817 : unsigned group1_size = i & ~(const_nunits - 1);
4102 : :
4103 : 143817 : if (dump_enabled_p ())
4104 : 57 : dump_printf_loc (MSG_NOTE, vect_location,
4105 : : "Splitting SLP group at stmt %u\n", i);
4106 : 143817 : stmt_vec_info rest = vect_split_slp_store_group (stmt_info,
4107 : : group1_size);
4108 : 143817 : bool res = vect_analyze_slp_instance (vinfo, bst_map, stmt_info,
4109 : : kind, max_tree_size,
4110 : : limit, false);
4111 : : /* Split the rest at the failure point and possibly
4112 : : re-analyze the remaining matching part if it has
4113 : : at least two lanes. */
4114 : 143817 : if (group1_size < i
4115 : 4656 : && (i + 1 < group_size
4116 : 2790 : || i - group1_size > 1))
4117 : : {
4118 : 1888 : stmt_vec_info rest2 = rest;
4119 : 1888 : rest = vect_split_slp_store_group (rest, i - group1_size);
4120 : 1888 : if (i - group1_size > 1)
4121 : 47 : res |= vect_analyze_slp_instance (vinfo, bst_map, rest2,
4122 : : kind, max_tree_size,
4123 : : limit, false);
4124 : : }
4125 : : /* Re-analyze the non-matching tail if it has at least
4126 : : two lanes. */
4127 : 143817 : if (i + 1 < group_size)
4128 : 20542 : res |= vect_analyze_slp_instance (vinfo, bst_map,
4129 : : rest, kind, max_tree_size,
4130 : : limit, false);
4131 : 143817 : return res;
4132 : : }
4133 : : }
4134 : :
4135 : : /* For loop vectorization split the RHS into arbitrary pieces of
4136 : : size >= 1. */
4137 : 209046 : else if (is_a <loop_vec_info> (vinfo)
4138 : 209046 : && (group_size != 1 && i < group_size))
4139 : : {
4140 : 6336 : gcall *call = dyn_cast <gcall *> (stmt_info->stmt);
4141 : 35 : bool masked_p = call
4142 : 35 : && gimple_call_internal_p (call)
4143 : 35 : && internal_fn_mask_index (gimple_call_internal_fn (call)) != -1;
4144 : : /* There are targets that cannot do even/odd interleaving schemes
4145 : : so they absolutely need to use load/store-lanes. For now
4146 : : force single-lane SLP for them - they would be happy with
4147 : : uniform power-of-two lanes (but depending on element size),
4148 : : but even if we can use 'i' as indicator we would need to
4149 : : backtrack when later lanes fail to discover with the same
4150 : : granularity. We cannot turn any of strided or scatter store
4151 : : into store-lanes. */
4152 : : /* ??? If this is not in sync with what get_load_store_type
4153 : : later decides the SLP representation is not good for other
4154 : : store vectorization methods. */
4155 : 6336 : bool want_store_lanes
4156 : 6336 : = (! STMT_VINFO_GATHER_SCATTER_P (stmt_info)
4157 : 6336 : && ! STMT_VINFO_STRIDED_P (stmt_info)
4158 : 4256 : && ! STMT_VINFO_SLP_VECT_ONLY (stmt_info)
4159 : 4245 : && compare_step_with_zero (vinfo, stmt_info) > 0
4160 : 10544 : && vect_slp_prefer_store_lanes_p (vinfo, stmt_info, NULL_TREE,
4161 : 12672 : masked_p, group_size, 1));
4162 : 6336 : if (want_store_lanes || force_single_lane)
4163 : : i = 1;
4164 : :
4165 : : /* A fatal discovery fail doesn't always mean single-lane SLP
4166 : : isn't a possibility, so try. */
4167 : 4653 : if (i == 0)
4168 : : i = 1;
4169 : :
4170 : 6336 : if (dump_enabled_p ())
4171 : 868 : dump_printf_loc (MSG_NOTE, vect_location,
4172 : : "Splitting SLP group at stmt %u\n", i);
4173 : :
4174 : : /* Analyze the stored values and pinch them together with
4175 : : a permute node so we can preserve the whole store group. */
4176 : 6336 : auto_vec<slp_tree> rhs_nodes;
4177 : 6336 : poly_uint64 max_nunits = 1;
4178 : :
4179 : 6336 : unsigned int rhs_common_nlanes = 0;
4180 : 6336 : unsigned int start = 0, end = i;
4181 : 28250 : while (start < group_size)
4182 : : {
4183 : 22148 : gcc_assert (end - start >= 1);
4184 : 22148 : vec<stmt_vec_info> substmts;
4185 : 22148 : substmts.create (end - start);
4186 : 67755 : for (unsigned j = start; j < end; ++j)
4187 : 45607 : substmts.quick_push (scalar_stmts[j]);
4188 : 22148 : max_nunits = 1;
4189 : 22148 : node = vect_build_slp_tree (vinfo, substmts, end - start,
4190 : : &max_nunits,
4191 : : matches, limit, &tree_size, bst_map);
4192 : 22148 : if (node)
4193 : : {
4194 : 17918 : rhs_nodes.safe_push (node);
4195 : 17918 : vect_update_max_nunits (&max_nunits, node->max_nunits);
4196 : 17918 : if (start == 0)
4197 : 6102 : rhs_common_nlanes = SLP_TREE_LANES (node);
4198 : 11816 : else if (rhs_common_nlanes != SLP_TREE_LANES (node))
4199 : 1387 : rhs_common_nlanes = 0;
4200 : 17918 : start = end;
4201 : 17918 : if (want_store_lanes || force_single_lane)
4202 : 5043 : end = start + 1;
4203 : : else
4204 : : end = group_size;
4205 : : }
4206 : : else
4207 : : {
4208 : 4230 : substmts.release ();
4209 : 4230 : if (end - start == 1)
4210 : : {
4211 : : /* Single-lane discovery failed. Free ressources. */
4212 : 234 : for (auto node : rhs_nodes)
4213 : 0 : vect_free_slp_tree (node);
4214 : 234 : scalar_stmts.release ();
4215 : 234 : if (dump_enabled_p ())
4216 : 38 : dump_printf_loc (MSG_NOTE, vect_location,
4217 : : "SLP discovery failed\n");
4218 : 234 : return false;
4219 : : }
4220 : :
4221 : : /* ??? It really happens that we soft-fail SLP
4222 : : build at a mismatch but the matching part hard-fails
4223 : : later. As we know we arrived here with a group
4224 : : larger than one try a group of size one! */
4225 : 3996 : if (!matches[0])
4226 : 66 : end = start + 1;
4227 : : else
4228 : 8968 : for (unsigned j = start; j < end; j++)
4229 : 8968 : if (!matches[j - start])
4230 : : {
4231 : : end = j;
4232 : : break;
4233 : : }
4234 : : }
4235 : : }
4236 : :
4237 : : /* Now re-assess whether we want store lanes in case the
4238 : : discovery ended up producing all single-lane RHSs. */
4239 : 6102 : if (! want_store_lanes
4240 : 6102 : && rhs_common_nlanes == 1
4241 : 5119 : && ! STMT_VINFO_GATHER_SCATTER_P (stmt_info)
4242 : 5119 : && ! STMT_VINFO_STRIDED_P (stmt_info)
4243 : 3432 : && ! STMT_VINFO_SLP_VECT_ONLY (stmt_info)
4244 : 3421 : && compare_step_with_zero (vinfo, stmt_info) > 0
4245 : 9500 : && (vect_store_lanes_supported (SLP_TREE_VECTYPE (rhs_nodes[0]),
4246 : : group_size, masked_p)
4247 : : != IFN_LAST))
4248 : : want_store_lanes = true;
4249 : :
4250 : : /* Now we assume we can build the root SLP node from all stores. */
4251 : 6102 : if (want_store_lanes)
4252 : : {
4253 : : /* For store-lanes feed the store node with all RHS nodes
4254 : : in order. */
4255 : 0 : node = vect_create_new_slp_node (scalar_stmts,
4256 : 0 : SLP_TREE_CHILDREN
4257 : : (rhs_nodes[0]).length ());
4258 : 0 : SLP_TREE_VECTYPE (node) = SLP_TREE_VECTYPE (rhs_nodes[0]);
4259 : 0 : node->max_nunits = max_nunits;
4260 : 0 : node->ldst_lanes = true;
4261 : 0 : SLP_TREE_CHILDREN (node)
4262 : 0 : .reserve_exact (SLP_TREE_CHILDREN (rhs_nodes[0]).length ()
4263 : 0 : + rhs_nodes.length () - 1);
4264 : : /* First store value and possibly mask. */
4265 : 0 : SLP_TREE_CHILDREN (node)
4266 : 0 : .splice (SLP_TREE_CHILDREN (rhs_nodes[0]));
4267 : : /* Rest of the store values. All mask nodes are the same,
4268 : : this should be guaranteed by dataref group discovery. */
4269 : 0 : for (unsigned j = 1; j < rhs_nodes.length (); ++j)
4270 : 0 : SLP_TREE_CHILDREN (node)
4271 : 0 : .quick_push (SLP_TREE_CHILDREN (rhs_nodes[j])[0]);
4272 : 0 : for (slp_tree child : SLP_TREE_CHILDREN (node))
4273 : 0 : child->refcnt++;
4274 : : }
4275 : : else
4276 : 6102 : node = vect_build_slp_store_interleaving (rhs_nodes, scalar_stmts,
4277 : : max_nunits);
4278 : :
4279 : 24020 : while (!rhs_nodes.is_empty ())
4280 : 17918 : vect_free_slp_tree (rhs_nodes.pop ());
4281 : :
4282 : : /* Create a new SLP instance. */
4283 : 6102 : slp_instance new_instance = XNEW (class _slp_instance);
4284 : 6102 : SLP_INSTANCE_TREE (new_instance) = node;
4285 : 6102 : SLP_INSTANCE_LOADS (new_instance) = vNULL;
4286 : 6102 : SLP_INSTANCE_ROOT_STMTS (new_instance) = root_stmt_infos;
4287 : 6102 : SLP_INSTANCE_REMAIN_DEFS (new_instance) = remain;
4288 : 6102 : SLP_INSTANCE_KIND (new_instance) = kind;
4289 : 6102 : new_instance->reduc_phis = NULL;
4290 : 6102 : new_instance->cost_vec = vNULL;
4291 : 6102 : new_instance->subgraph_entries = vNULL;
4292 : :
4293 : 6102 : if (dump_enabled_p ())
4294 : 830 : dump_printf_loc (MSG_NOTE, vect_location,
4295 : : "SLP size %u vs. limit %u.\n",
4296 : : tree_size, max_tree_size);
4297 : :
4298 : 6102 : vinfo->slp_instances.safe_push (new_instance);
4299 : :
4300 : : /* ??? We've replaced the old SLP_INSTANCE_GROUP_SIZE with
4301 : : the number of scalar stmts in the root in a few places.
4302 : : Verify that assumption holds. */
4303 : 12204 : gcc_assert (SLP_TREE_SCALAR_STMTS (SLP_INSTANCE_TREE (new_instance))
4304 : : .length () == group_size);
4305 : :
4306 : 6102 : if (dump_enabled_p ())
4307 : : {
4308 : 830 : dump_printf_loc (MSG_NOTE, vect_location,
4309 : : "Final SLP tree for instance %p:\n",
4310 : : (void *) new_instance);
4311 : 830 : vect_print_slp_graph (MSG_NOTE, vect_location,
4312 : : SLP_INSTANCE_TREE (new_instance));
4313 : : }
4314 : 6102 : return true;
4315 : 6336 : }
4316 : : else
4317 : : /* Free the allocated memory. */
4318 : 202710 : scalar_stmts.release ();
4319 : :
4320 : : /* Even though the first vector did not all match, we might be able to SLP
4321 : : (some) of the remainder. FORNOW ignore this possibility. */
4322 : : }
4323 : : else
4324 : : /* Free the allocated memory. */
4325 : 994786 : scalar_stmts.release ();
4326 : :
4327 : : /* Failed to SLP. */
4328 : 1197496 : if (dump_enabled_p ())
4329 : 3408 : dump_printf_loc (MSG_NOTE, vect_location, "SLP discovery failed\n");
4330 : : return false;
4331 : : }
4332 : :
4333 : :
4334 : : /* Analyze an SLP instance starting from a group of grouped stores. Call
4335 : : vect_build_slp_tree to build a tree of packed stmts if possible.
4336 : : Return FALSE if it's impossible to SLP any stmt in the loop. */
4337 : :
4338 : : static bool
4339 : 1011528 : vect_analyze_slp_instance (vec_info *vinfo,
4340 : : scalar_stmts_to_slp_tree_map_t *bst_map,
4341 : : stmt_vec_info stmt_info,
4342 : : slp_instance_kind kind,
4343 : : unsigned max_tree_size, unsigned *limit,
4344 : : bool force_single_lane)
4345 : : {
4346 : 1011528 : vec<stmt_vec_info> scalar_stmts;
4347 : :
4348 : 1011528 : if (is_a <bb_vec_info> (vinfo))
4349 : 994076 : vect_location = stmt_info->stmt;
4350 : :
4351 : 1011528 : stmt_vec_info next_info = stmt_info;
4352 : 1011528 : if (kind == slp_inst_kind_store)
4353 : : {
4354 : : /* Collect the stores and store them in scalar_stmts. */
4355 : 1010918 : scalar_stmts.create (DR_GROUP_SIZE (stmt_info));
4356 : 5004904 : while (next_info)
4357 : : {
4358 : 2983068 : scalar_stmts.quick_push (vect_stmt_to_vectorize (next_info));
4359 : 2983068 : next_info = DR_GROUP_NEXT_ELEMENT (next_info);
4360 : : }
4361 : : }
4362 : 610 : else if (kind == slp_inst_kind_reduc_chain)
4363 : : {
4364 : : /* Collect the reduction stmts and store them in scalar_stmts. */
4365 : 610 : scalar_stmts.create (REDUC_GROUP_SIZE (stmt_info));
4366 : 3784 : while (next_info)
4367 : : {
4368 : 2564 : scalar_stmts.quick_push (vect_stmt_to_vectorize (next_info));
4369 : 2564 : next_info = REDUC_GROUP_NEXT_ELEMENT (next_info);
4370 : : }
4371 : : /* Mark the first element of the reduction chain as reduction to properly
4372 : : transform the node. In the reduction analysis phase only the last
4373 : : element of the chain is marked as reduction. */
4374 : 610 : STMT_VINFO_DEF_TYPE (stmt_info)
4375 : 610 : = STMT_VINFO_DEF_TYPE (scalar_stmts.last ());
4376 : 610 : STMT_VINFO_REDUC_DEF (vect_orig_stmt (stmt_info))
4377 : 668 : = STMT_VINFO_REDUC_DEF (vect_orig_stmt (scalar_stmts.last ()));
4378 : : }
4379 : : else
4380 : 0 : gcc_unreachable ();
4381 : :
4382 : 1011528 : vec<stmt_vec_info> roots = vNULL;
4383 : 1011528 : vec<tree> remain = vNULL;
4384 : : /* Build the tree for the SLP instance. */
4385 : 1012138 : bool res = vect_build_slp_instance (vinfo, kind, scalar_stmts,
4386 : : roots, remain,
4387 : : max_tree_size, limit, bst_map,
4388 : : kind == slp_inst_kind_store
4389 : : ? stmt_info : NULL, force_single_lane);
4390 : :
4391 : : /* ??? If this is slp_inst_kind_store and the above succeeded here's
4392 : : where we should do store group splitting. */
4393 : :
4394 : 1011528 : return res;
4395 : : }
4396 : :
4397 : : /* qsort comparator ordering SLP load nodes. */
4398 : :
4399 : : static int
4400 : 1093229 : vllp_cmp (const void *a_, const void *b_)
4401 : : {
4402 : 1093229 : const slp_tree a = *(const slp_tree *)a_;
4403 : 1093229 : const slp_tree b = *(const slp_tree *)b_;
4404 : 1093229 : stmt_vec_info a0 = SLP_TREE_SCALAR_STMTS (a)[0];
4405 : 1093229 : stmt_vec_info b0 = SLP_TREE_SCALAR_STMTS (b)[0];
4406 : 1093229 : if (STMT_VINFO_GROUPED_ACCESS (a0)
4407 : 540171 : && STMT_VINFO_GROUPED_ACCESS (b0)
4408 : 1607809 : && DR_GROUP_FIRST_ELEMENT (a0) == DR_GROUP_FIRST_ELEMENT (b0))
4409 : : {
4410 : : /* Same group, order after lanes used. */
4411 : 201871 : if (SLP_TREE_LANES (a) < SLP_TREE_LANES (b))
4412 : : return 1;
4413 : 199050 : else if (SLP_TREE_LANES (a) > SLP_TREE_LANES (b))
4414 : : return -1;
4415 : : else
4416 : : {
4417 : : /* Try to order loads using the same lanes together, breaking
4418 : : the tie with the lane number that first differs. */
4419 : 195797 : if (!SLP_TREE_LOAD_PERMUTATION (a).exists ()
4420 : 195797 : && !SLP_TREE_LOAD_PERMUTATION (b).exists ())
4421 : : return 0;
4422 : 195797 : else if (SLP_TREE_LOAD_PERMUTATION (a).exists ()
4423 : 195797 : && !SLP_TREE_LOAD_PERMUTATION (b).exists ())
4424 : : return 1;
4425 : 195162 : else if (!SLP_TREE_LOAD_PERMUTATION (a).exists ()
4426 : 195162 : && SLP_TREE_LOAD_PERMUTATION (b).exists ())
4427 : : return -1;
4428 : : else
4429 : : {
4430 : 194567 : for (unsigned i = 0; i < SLP_TREE_LANES (a); ++i)
4431 : 194567 : if (SLP_TREE_LOAD_PERMUTATION (a)[i]
4432 : 194567 : != SLP_TREE_LOAD_PERMUTATION (b)[i])
4433 : : {
4434 : : /* In-order lane first, that's what the above case for
4435 : : no permutation does. */
4436 : 194278 : if (SLP_TREE_LOAD_PERMUTATION (a)[i] == i)
4437 : : return -1;
4438 : 118506 : else if (SLP_TREE_LOAD_PERMUTATION (b)[i] == i)
4439 : : return 1;
4440 : 67763 : else if (SLP_TREE_LOAD_PERMUTATION (a)[i]
4441 : 67763 : < SLP_TREE_LOAD_PERMUTATION (b)[i])
4442 : : return -1;
4443 : : else
4444 : : return 1;
4445 : : }
4446 : : return 0;
4447 : : }
4448 : : }
4449 : : }
4450 : : else /* Different groups or non-groups. */
4451 : : {
4452 : : /* Order groups as their first element to keep them together. */
4453 : 891358 : if (STMT_VINFO_GROUPED_ACCESS (a0))
4454 : 891358 : a0 = DR_GROUP_FIRST_ELEMENT (a0);
4455 : 891358 : if (STMT_VINFO_GROUPED_ACCESS (b0))
4456 : 891358 : b0 = DR_GROUP_FIRST_ELEMENT (b0);
4457 : 891358 : if (a0 == b0)
4458 : : return 0;
4459 : : /* Tie using UID. */
4460 : 891334 : else if (gimple_uid (STMT_VINFO_STMT (a0))
4461 : 891334 : < gimple_uid (STMT_VINFO_STMT (b0)))
4462 : : return -1;
4463 : : else
4464 : : {
4465 : 380061 : gcc_assert (gimple_uid (STMT_VINFO_STMT (a0))
4466 : : != gimple_uid (STMT_VINFO_STMT (b0)));
4467 : : return 1;
4468 : : }
4469 : : }
4470 : : }
4471 : :
4472 : : /* Process the set of LOADS that are all from the same dataref group. */
4473 : :
4474 : : static void
4475 : 59558 : vect_lower_load_permutations (loop_vec_info loop_vinfo,
4476 : : scalar_stmts_to_slp_tree_map_t *bst_map,
4477 : : const array_slice<slp_tree> &loads,
4478 : : bool force_single_lane)
4479 : : {
4480 : : /* We at this point want to lower without a fixed VF or vector
4481 : : size in mind which means we cannot actually compute whether we
4482 : : need three or more vectors for a load permutation yet. So always
4483 : : lower. */
4484 : 59558 : stmt_vec_info first
4485 : 59558 : = DR_GROUP_FIRST_ELEMENT (SLP_TREE_SCALAR_STMTS (loads[0])[0]);
4486 : 59558 : unsigned group_lanes = DR_GROUP_SIZE (first);
4487 : :
4488 : : /* Verify if all load permutations can be implemented with a suitably
4489 : : large element load-lanes operation. */
4490 : 59558 : unsigned ld_lanes_lanes = SLP_TREE_LANES (loads[0]);
4491 : 59558 : if (STMT_VINFO_STRIDED_P (first)
4492 : 57649 : || compare_step_with_zero (loop_vinfo, first) <= 0
4493 : 57403 : || exact_log2 (ld_lanes_lanes) == -1
4494 : : /* ??? For now only support the single-lane case as there is
4495 : : missing support on the store-lane side and code generation
4496 : : isn't up to the task yet. */
4497 : 55532 : || ld_lanes_lanes != 1
4498 : 109828 : || vect_load_lanes_supported (SLP_TREE_VECTYPE (loads[0]),
4499 : : group_lanes / ld_lanes_lanes,
4500 : : false) == IFN_LAST)
4501 : : ld_lanes_lanes = 0;
4502 : : else
4503 : : /* Verify the loads access the same number of lanes aligned to
4504 : : ld_lanes_lanes. */
4505 : 0 : for (slp_tree load : loads)
4506 : : {
4507 : 0 : if (SLP_TREE_LANES (load) != ld_lanes_lanes)
4508 : : {
4509 : : ld_lanes_lanes = 0;
4510 : : break;
4511 : : }
4512 : 0 : unsigned first = SLP_TREE_LOAD_PERMUTATION (load)[0];
4513 : 0 : if (first % ld_lanes_lanes != 0)
4514 : : {
4515 : : ld_lanes_lanes = 0;
4516 : : break;
4517 : : }
4518 : 0 : for (unsigned i = 1; i < SLP_TREE_LANES (load); ++i)
4519 : : if (SLP_TREE_LOAD_PERMUTATION (load)[i] != first + i)
4520 : : {
4521 : : ld_lanes_lanes = 0;
4522 : : break;
4523 : : }
4524 : : }
4525 : :
4526 : : /* Only a power-of-two number of lanes matches interleaving with N levels.
4527 : : ??? An even number of lanes could be reduced to 1<<ceil_log2(N)-1 lanes
4528 : : at each step. */
4529 : 90034 : if (ld_lanes_lanes == 0 && exact_log2 (group_lanes) == -1 && group_lanes != 3)
4530 : : return;
4531 : :
4532 : 136709 : for (slp_tree load : loads)
4533 : : {
4534 : : /* Leave masked or gather loads alone for now. */
4535 : 84223 : if (!SLP_TREE_CHILDREN (load).is_empty ())
4536 : 33678 : continue;
4537 : :
4538 : : /* We want to pattern-match special cases here and keep those
4539 : : alone. Candidates are splats and load-lane. */
4540 : :
4541 : : /* We need to lower only loads of less than half of the groups
4542 : : lanes, including duplicate lanes. Note this leaves nodes
4543 : : with a non-1:1 load permutation around instead of canonicalizing
4544 : : those into a load and a permute node. Removing this early
4545 : : check would do such canonicalization. */
4546 : 84201 : if (SLP_TREE_LANES (load) >= (group_lanes + 1) / 2
4547 : 30255 : && ld_lanes_lanes == 0)
4548 : 30255 : continue;
4549 : :
4550 : : /* Build the permute to get the original load permutation order. */
4551 : 53946 : bool contiguous = true;
4552 : 53946 : lane_permutation_t final_perm;
4553 : 53946 : final_perm.create (SLP_TREE_LANES (load));
4554 : 108413 : for (unsigned i = 0; i < SLP_TREE_LANES (load); ++i)
4555 : : {
4556 : 54467 : final_perm.quick_push
4557 : 54467 : (std::make_pair (0, SLP_TREE_LOAD_PERMUTATION (load)[i]));
4558 : 54467 : if (i != 0
4559 : 54467 : && (SLP_TREE_LOAD_PERMUTATION (load)[i]
4560 : 521 : != SLP_TREE_LOAD_PERMUTATION (load)[i-1] + 1))
4561 : : contiguous = false;
4562 : : }
4563 : :
4564 : : /* When the load permutation accesses a contiguous unpermuted,
4565 : : power-of-two aligned and sized chunk leave the load alone.
4566 : : We can likely (re-)load it more efficiently rather than
4567 : : extracting it from the larger load.
4568 : : ??? Long-term some of the lowering should move to where
4569 : : the vector types involved are fixed. */
4570 : 57347 : if (!force_single_lane
4571 : 53946 : && ld_lanes_lanes == 0
4572 : 30745 : && contiguous
4573 : 30540 : && (SLP_TREE_LANES (load) > 1 || loads.size () == 1)
4574 : 7186 : && pow2p_hwi (SLP_TREE_LANES (load))
4575 : 7186 : && pow2p_hwi (group_lanes)
4576 : 3401 : && SLP_TREE_LOAD_PERMUTATION (load)[0] % SLP_TREE_LANES (load) == 0
4577 : 57347 : && group_lanes % SLP_TREE_LANES (load) == 0)
4578 : : {
4579 : 3401 : final_perm.release ();
4580 : 3401 : continue;
4581 : : }
4582 : :
4583 : : /* First build (and possibly re-use) a load node for the
4584 : : unpermuted group. Gaps in the middle and on the end are
4585 : : represented with NULL stmts. */
4586 : 50545 : vec<stmt_vec_info> stmts;
4587 : 50545 : stmts.create (group_lanes);
4588 : 170871 : for (stmt_vec_info s = first; s; s = DR_GROUP_NEXT_ELEMENT (s))
4589 : : {
4590 : 120326 : if (s != first)
4591 : 73362 : for (unsigned i = 1; i < DR_GROUP_GAP (s); ++i)
4592 : 3581 : stmts.quick_push (NULL);
4593 : 120326 : stmts.quick_push (s);
4594 : : }
4595 : 136699 : for (unsigned i = 0; i < DR_GROUP_GAP (first); ++i)
4596 : 86154 : stmts.quick_push (NULL);
4597 : 50545 : poly_uint64 max_nunits = 1;
4598 : 50545 : bool *matches = XALLOCAVEC (bool, group_lanes);
4599 : 50545 : unsigned limit = 1;
4600 : 50545 : unsigned tree_size = 0;
4601 : 50545 : slp_tree l0 = vect_build_slp_tree (loop_vinfo, stmts,
4602 : : group_lanes,
4603 : : &max_nunits, matches, &limit,
4604 : 50545 : &tree_size, bst_map);
4605 : 50545 : gcc_assert (!SLP_TREE_LOAD_PERMUTATION (l0).exists ());
4606 : :
4607 : 50545 : if (ld_lanes_lanes != 0)
4608 : : {
4609 : : /* ??? If this is not in sync with what get_load_store_type
4610 : : later decides the SLP representation is not good for other
4611 : : store vectorization methods. */
4612 : 0 : l0->ldst_lanes = true;
4613 : 0 : load->ldst_lanes = true;
4614 : : }
4615 : :
4616 : 166115 : while (1)
4617 : : {
4618 : 108330 : unsigned group_lanes = SLP_TREE_LANES (l0);
4619 : 108330 : if (ld_lanes_lanes != 0
4620 : 108330 : || SLP_TREE_LANES (load) >= (group_lanes + 1) / 2)
4621 : : break;
4622 : :
4623 : : /* Try to lower by reducing the group to half its size using an
4624 : : interleaving scheme. For this try to compute whether all
4625 : : elements needed for this load are in even or odd elements of
4626 : : an even/odd decomposition with N consecutive elements.
4627 : : Thus { e, e, o, o, e, e, o, o } woud be an even/odd decomposition
4628 : : with N == 2. */
4629 : : /* ??? Only an even number of lanes can be handed this way, but the
4630 : : fallback below could work for any number. We have to make sure
4631 : : to round up in that case. */
4632 : 57785 : gcc_assert ((group_lanes & 1) == 0 || group_lanes == 3);
4633 : 14174 : unsigned even = 0, odd = 0;
4634 : 14174 : if ((group_lanes & 1) == 0)
4635 : : {
4636 : 14174 : even = (1 << ceil_log2 (group_lanes)) - 1;
4637 : 14174 : odd = even;
4638 : 57138 : for (auto l : final_perm)
4639 : : {
4640 : 14616 : even &= ~l.second;
4641 : 14616 : odd &= l.second;
4642 : : }
4643 : : }
4644 : :
4645 : : /* Now build an even or odd extraction from the unpermuted load. */
4646 : 57785 : lane_permutation_t perm;
4647 : 57785 : perm.create ((group_lanes + 1) / 2);
4648 : 57785 : unsigned even_level = even ? 1 << ctz_hwi (even) : 0;
4649 : 57785 : unsigned odd_level = odd ? 1 << ctz_hwi (odd) : 0;
4650 : 57785 : if (even_level
4651 : 13498 : && group_lanes % (2 * even_level) == 0
4652 : : /* ??? When code generating permutes we do not try to pun
4653 : : to larger component modes so level != 1 isn't a natural
4654 : : even/odd extract. Prefer one if possible. */
4655 : 13498 : && (even_level == 1 || !odd_level || odd_level != 1))
4656 : : {
4657 : : /* { 0, 1, ... 4, 5 ..., } */
4658 : 67008 : for (unsigned i = 0; i < group_lanes / 2 / even_level; ++i)
4659 : 111127 : for (unsigned j = 0; j < even_level; ++j)
4660 : 55646 : perm.quick_push (std::make_pair (0, 2 * i * even_level + j));
4661 : : }
4662 : 44287 : else if (odd_level)
4663 : : {
4664 : : /* { ..., 2, 3, ... 6, 7 } */
4665 : 2647 : gcc_assert (group_lanes % (2 * odd_level) == 0);
4666 : 11915 : for (unsigned i = 0; i < group_lanes / 2 / odd_level; ++i)
4667 : 18536 : for (unsigned j = 0; j < odd_level; ++j)
4668 : 9268 : perm.quick_push
4669 : 9268 : (std::make_pair (0, (2 * i + 1) * odd_level + j));
4670 : : }
4671 : : else
4672 : : {
4673 : : /* As fallback extract all used lanes and fill to half the
4674 : : group size by repeating the last element.
4675 : : ??? This is quite a bad strathegy for re-use - we could
4676 : : brute force our way to find more optimal filling lanes to
4677 : : maximize re-use when looking at all loads from the group. */
4678 : 43611 : auto_bitmap l;
4679 : 174444 : for (auto p : final_perm)
4680 : 43611 : bitmap_set_bit (l, p.second);
4681 : 43611 : unsigned i = 0;
4682 : 43611 : bitmap_iterator bi;
4683 : 87222 : EXECUTE_IF_SET_IN_BITMAP (l, 0, i, bi)
4684 : 43611 : perm.quick_push (std::make_pair (0, i));
4685 : 174444 : while (perm.length () < (group_lanes + 1) / 2)
4686 : 43611 : perm.quick_push (perm.last ());
4687 : 43611 : }
4688 : :
4689 : : /* Update final_perm with the intermediate permute. */
4690 : 116012 : for (unsigned i = 0; i < final_perm.length (); ++i)
4691 : : {
4692 : 58227 : unsigned l = final_perm[i].second;
4693 : 58227 : unsigned j;
4694 : 65308 : for (j = 0; j < perm.length (); ++j)
4695 : 65308 : if (perm[j].second == l)
4696 : : {
4697 : 58227 : final_perm[i].second = j;
4698 : 58227 : break;
4699 : : }
4700 : 58227 : gcc_assert (j < perm.length ());
4701 : : }
4702 : :
4703 : : /* And create scalar stmts. */
4704 : 57785 : vec<stmt_vec_info> perm_stmts;
4705 : 57785 : perm_stmts.create (perm.length ());
4706 : 209921 : for (unsigned i = 0; i < perm.length (); ++i)
4707 : 152136 : perm_stmts.quick_push (SLP_TREE_SCALAR_STMTS (l0)[perm[i].second]);
4708 : :
4709 : 57785 : slp_tree p = vect_create_new_slp_node (1, VEC_PERM_EXPR);
4710 : 57785 : SLP_TREE_CHILDREN (p).quick_push (l0);
4711 : 57785 : SLP_TREE_LANE_PERMUTATION (p) = perm;
4712 : 57785 : SLP_TREE_VECTYPE (p) = SLP_TREE_VECTYPE (load);
4713 : 57785 : SLP_TREE_LANES (p) = perm.length ();
4714 : 57785 : SLP_TREE_REPRESENTATIVE (p) = SLP_TREE_REPRESENTATIVE (load);
4715 : : /* ??? As we have scalar stmts for this intermediate permute we
4716 : : could CSE it via bst_map but we do not want to pick up
4717 : : another SLP node with a load permutation. We instead should
4718 : : have a "local" CSE map here. */
4719 : 57785 : SLP_TREE_SCALAR_STMTS (p) = perm_stmts;
4720 : :
4721 : : /* We now have a node for (group_lanes + 1) / 2 lanes. */
4722 : 57785 : l0 = p;
4723 : 57785 : }
4724 : :
4725 : : /* And finally from the ordered reduction node create the
4726 : : permute to shuffle the lanes into the original load-permutation
4727 : : order. We replace the original load node with this. */
4728 : 50545 : SLP_TREE_CODE (load) = VEC_PERM_EXPR;
4729 : 50545 : SLP_TREE_LOAD_PERMUTATION (load).release ();
4730 : 50545 : SLP_TREE_LANE_PERMUTATION (load) = final_perm;
4731 : 50545 : SLP_TREE_CHILDREN (load).create (1);
4732 : 50545 : SLP_TREE_CHILDREN (load).quick_push (l0);
4733 : : }
4734 : : }
4735 : :
4736 : : /* Transform SLP loads in the SLP graph created by SLP discovery to
4737 : : group loads from the same group and lower load permutations that
4738 : : are unlikely to be supported into a series of permutes.
4739 : : In the degenerate case of having only single-lane SLP instances
4740 : : this should result in a series of permute nodes emulating an
4741 : : interleaving scheme. */
4742 : :
4743 : : static void
4744 : 241907 : vect_lower_load_permutations (loop_vec_info loop_vinfo,
4745 : : scalar_stmts_to_slp_tree_map_t *bst_map,
4746 : : bool force_single_lane)
4747 : : {
4748 : : /* Gather and sort loads across all instances. */
4749 : 241907 : hash_set<slp_tree> visited;
4750 : 241907 : auto_vec<slp_tree> loads;
4751 : 1154808 : for (auto inst : loop_vinfo->slp_instances)
4752 : 439097 : vect_gather_slp_loads (loads, SLP_INSTANCE_TREE (inst), visited);
4753 : 241907 : if (loads.is_empty ())
4754 : 64986 : return;
4755 : 176921 : loads.qsort (vllp_cmp);
4756 : :
4757 : : /* Now process each dataref group separately. */
4758 : 176921 : unsigned firsti = 0;
4759 : 320222 : for (unsigned i = 1; i < loads.length (); ++i)
4760 : : {
4761 : 143301 : slp_tree first = loads[firsti];
4762 : 143301 : slp_tree next = loads[i];
4763 : 143301 : stmt_vec_info a0 = SLP_TREE_SCALAR_STMTS (first)[0];
4764 : 143301 : stmt_vec_info b0 = SLP_TREE_SCALAR_STMTS (next)[0];
4765 : 143301 : if (STMT_VINFO_GROUPED_ACCESS (a0)
4766 : 60794 : && STMT_VINFO_GROUPED_ACCESS (b0)
4767 : 200204 : && DR_GROUP_FIRST_ELEMENT (a0) == DR_GROUP_FIRST_ELEMENT (b0))
4768 : 36050 : continue;
4769 : : /* Now we have one or multiple SLP loads of the same group from
4770 : : firsti to i - 1. */
4771 : 107251 : if (STMT_VINFO_GROUPED_ACCESS (a0))
4772 : 24744 : vect_lower_load_permutations (loop_vinfo, bst_map,
4773 : 24744 : make_array_slice (&loads[firsti],
4774 : : i - firsti),
4775 : : force_single_lane);
4776 : : firsti = i;
4777 : : }
4778 : 353842 : if (firsti < loads.length ()
4779 : 353842 : && STMT_VINFO_GROUPED_ACCESS (SLP_TREE_SCALAR_STMTS (loads[firsti])[0]))
4780 : 34814 : vect_lower_load_permutations (loop_vinfo, bst_map,
4781 : 34814 : make_array_slice (&loads[firsti],
4782 : 34814 : loads.length () - firsti),
4783 : : force_single_lane);
4784 : 241907 : }
4785 : :
4786 : : /* Check if there are stmts in the loop can be vectorized using SLP. Build SLP
4787 : : trees of packed scalar stmts if SLP is possible. */
4788 : :
4789 : : opt_result
4790 : 837809 : vect_analyze_slp (vec_info *vinfo, unsigned max_tree_size,
4791 : : bool force_single_lane)
4792 : : {
4793 : 837809 : loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo);
4794 : 837809 : unsigned int i;
4795 : 837809 : stmt_vec_info first_element;
4796 : 837809 : slp_instance instance;
4797 : :
4798 : 837809 : DUMP_VECT_SCOPE ("vect_analyze_slp");
4799 : :
4800 : 837809 : unsigned limit = max_tree_size;
4801 : :
4802 : 837809 : scalar_stmts_to_slp_tree_map_t *bst_map
4803 : 837809 : = new scalar_stmts_to_slp_tree_map_t ();
4804 : :
4805 : : /* Find SLP sequences starting from groups of grouped stores. */
4806 : 2522130 : FOR_EACH_VEC_ELT (vinfo->grouped_stores, i, first_element)
4807 : 846512 : vect_analyze_slp_instance (vinfo, bst_map, first_element,
4808 : : slp_inst_kind_store, max_tree_size, &limit,
4809 : : force_single_lane);
4810 : :
4811 : : /* For loops also start SLP discovery from non-grouped stores. */
4812 : 837809 : if (loop_vinfo)
4813 : : {
4814 : : data_reference_p dr;
4815 : 827191 : FOR_EACH_VEC_ELT (vinfo->shared->datarefs, i, dr)
4816 : 585284 : if (DR_IS_WRITE (dr))
4817 : : {
4818 : 240088 : stmt_vec_info stmt_info = vinfo->lookup_dr (dr)->stmt;
4819 : : /* Grouped stores are already handled above. */
4820 : 240088 : if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
4821 : 65856 : continue;
4822 : 174232 : vec<stmt_vec_info> stmts;
4823 : 174232 : vec<stmt_vec_info> roots = vNULL;
4824 : 174232 : vec<tree> remain = vNULL;
4825 : 174232 : stmts.create (1);
4826 : 174232 : stmts.quick_push (stmt_info);
4827 : 174232 : vect_build_slp_instance (vinfo, slp_inst_kind_store,
4828 : : stmts, roots, remain, max_tree_size,
4829 : : &limit, bst_map, NULL, force_single_lane);
4830 : : }
4831 : : }
4832 : :
4833 : 837809 : if (bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (vinfo))
4834 : : {
4835 : 1720862 : for (unsigned i = 0; i < bb_vinfo->roots.length (); ++i)
4836 : : {
4837 : 1124960 : vect_location = bb_vinfo->roots[i].roots[0]->stmt;
4838 : : /* Apply patterns. */
4839 : 3525284 : for (unsigned j = 0; j < bb_vinfo->roots[i].stmts.length (); ++j)
4840 : 4800648 : bb_vinfo->roots[i].stmts[j]
4841 : 2456423 : = vect_stmt_to_vectorize (bb_vinfo->roots[i].stmts[j]);
4842 : 1124960 : if (vect_build_slp_instance (bb_vinfo, bb_vinfo->roots[i].kind,
4843 : 1124960 : bb_vinfo->roots[i].stmts,
4844 : 1124960 : bb_vinfo->roots[i].roots,
4845 : 1124960 : bb_vinfo->roots[i].remain,
4846 : : max_tree_size, &limit, bst_map, NULL,
4847 : : false))
4848 : : {
4849 : 110503 : bb_vinfo->roots[i].stmts = vNULL;
4850 : 110503 : bb_vinfo->roots[i].roots = vNULL;
4851 : 110503 : bb_vinfo->roots[i].remain = vNULL;
4852 : : }
4853 : : }
4854 : : }
4855 : :
4856 : 837809 : if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
4857 : : {
4858 : : /* Find SLP sequences starting from reduction chains. */
4859 : 242521 : FOR_EACH_VEC_ELT (loop_vinfo->reduction_chains, i, first_element)
4860 : 614 : if (! STMT_VINFO_RELEVANT_P (first_element)
4861 : 4 : && ! STMT_VINFO_LIVE_P (first_element))
4862 : : ;
4863 : 610 : else if (force_single_lane
4864 : 610 : || ! vect_analyze_slp_instance (vinfo, bst_map, first_element,
4865 : : slp_inst_kind_reduc_chain,
4866 : : max_tree_size, &limit,
4867 : : force_single_lane))
4868 : : {
4869 : : /* Dissolve reduction chain group. */
4870 : 267 : stmt_vec_info vinfo = first_element;
4871 : 267 : stmt_vec_info last = NULL;
4872 : 1046 : while (vinfo)
4873 : : {
4874 : 779 : stmt_vec_info next = REDUC_GROUP_NEXT_ELEMENT (vinfo);
4875 : 779 : REDUC_GROUP_FIRST_ELEMENT (vinfo) = NULL;
4876 : 779 : REDUC_GROUP_NEXT_ELEMENT (vinfo) = NULL;
4877 : 779 : last = vinfo;
4878 : 779 : vinfo = next;
4879 : : }
4880 : 267 : STMT_VINFO_DEF_TYPE (first_element) = vect_internal_def;
4881 : : /* It can be still vectorized as part of an SLP reduction. */
4882 : 267 : loop_vinfo->reductions.safe_push (last);
4883 : : }
4884 : :
4885 : : /* Find SLP sequences starting from groups of reductions. */
4886 : 241907 : if (loop_vinfo->reductions.length () > 0)
4887 : : {
4888 : : /* Collect reduction statements we can combine into
4889 : : a SLP reduction. */
4890 : 46253 : vec<stmt_vec_info> scalar_stmts;
4891 : 46253 : scalar_stmts.create (loop_vinfo->reductions.length ());
4892 : 206570 : for (auto next_info : loop_vinfo->reductions)
4893 : : {
4894 : 67811 : next_info = vect_stmt_to_vectorize (next_info);
4895 : 67811 : if ((STMT_VINFO_RELEVANT_P (next_info)
4896 : 11 : || STMT_VINFO_LIVE_P (next_info))
4897 : : /* ??? Make sure we didn't skip a conversion around a
4898 : : reduction path. In that case we'd have to reverse
4899 : : engineer that conversion stmt following the chain using
4900 : : reduc_idx and from the PHI using reduc_def. */
4901 : 67800 : && (STMT_VINFO_DEF_TYPE (next_info) == vect_reduction_def
4902 : 67800 : || (STMT_VINFO_DEF_TYPE (next_info)
4903 : : == vect_double_reduction_def)))
4904 : : {
4905 : : /* Do not discover SLP reductions combining lane-reducing
4906 : : ops, that will fail later. */
4907 : 67722 : if (!force_single_lane
4908 : 67722 : && !lane_reducing_stmt_p (STMT_VINFO_STMT (next_info)))
4909 : 50082 : scalar_stmts.quick_push (next_info);
4910 : : else
4911 : : {
4912 : : /* Do SLP discovery for single-lane reductions. */
4913 : 17640 : vec<stmt_vec_info> stmts;
4914 : 17640 : vec<stmt_vec_info> roots = vNULL;
4915 : 17640 : vec<tree> remain = vNULL;
4916 : 17640 : stmts.create (1);
4917 : 17640 : stmts.quick_push (next_info);
4918 : 17640 : vect_build_slp_instance (vinfo,
4919 : : slp_inst_kind_reduc_group,
4920 : : stmts, roots, remain,
4921 : : max_tree_size, &limit,
4922 : : bst_map, NULL,
4923 : : force_single_lane);
4924 : : }
4925 : : }
4926 : : }
4927 : : /* Save for re-processing on failure. */
4928 : 46253 : vec<stmt_vec_info> saved_stmts = scalar_stmts.copy ();
4929 : 46253 : vec<stmt_vec_info> roots = vNULL;
4930 : 46253 : vec<tree> remain = vNULL;
4931 : 46253 : if (scalar_stmts.length () <= 1
4932 : 46253 : || !vect_build_slp_instance (loop_vinfo,
4933 : : slp_inst_kind_reduc_group,
4934 : : scalar_stmts, roots, remain,
4935 : : max_tree_size, &limit, bst_map,
4936 : : NULL, force_single_lane))
4937 : : {
4938 : 45933 : if (scalar_stmts.length () <= 1)
4939 : 45933 : scalar_stmts.release ();
4940 : : /* Do SLP discovery for single-lane reductions. */
4941 : 158980 : for (auto stmt_info : saved_stmts)
4942 : : {
4943 : 49127 : vec<stmt_vec_info> stmts;
4944 : 49127 : vec<stmt_vec_info> roots = vNULL;
4945 : 49127 : vec<tree> remain = vNULL;
4946 : 49127 : stmts.create (1);
4947 : 49127 : stmts.quick_push (vect_stmt_to_vectorize (stmt_info));
4948 : 49127 : vect_build_slp_instance (vinfo,
4949 : : slp_inst_kind_reduc_group,
4950 : : stmts, roots, remain,
4951 : : max_tree_size, &limit,
4952 : : bst_map, NULL, force_single_lane);
4953 : : }
4954 : 45933 : saved_stmts.release ();
4955 : : }
4956 : : }
4957 : :
4958 : : /* Make sure to vectorize only-live stmts, usually inductions. */
4959 : 1031041 : for (edge e : get_loop_exit_edges (LOOP_VINFO_LOOP (loop_vinfo)))
4960 : 584666 : for (auto gsi = gsi_start_phis (e->dest); !gsi_end_p (gsi);
4961 : 279346 : gsi_next (&gsi))
4962 : : {
4963 : 279346 : gphi *lc_phi = *gsi;
4964 : 279346 : tree def = gimple_phi_arg_def_from_edge (lc_phi, e);
4965 : 279346 : stmt_vec_info stmt_info;
4966 : 279346 : if (TREE_CODE (def) == SSA_NAME
4967 : 270632 : && !virtual_operand_p (def)
4968 : 106522 : && (stmt_info = loop_vinfo->lookup_def (def))
4969 : 104428 : && ((stmt_info = vect_stmt_to_vectorize (stmt_info)), true)
4970 : 104428 : && STMT_VINFO_RELEVANT (stmt_info) == vect_used_only_live
4971 : 87584 : && STMT_VINFO_LIVE_P (stmt_info)
4972 : 366930 : && (STMT_VINFO_DEF_TYPE (stmt_info) == vect_induction_def
4973 : 81366 : || (STMT_VINFO_DEF_TYPE (stmt_info) == vect_internal_def
4974 : 12118 : && STMT_VINFO_REDUC_IDX (stmt_info) == -1)))
4975 : : {
4976 : 18234 : vec<stmt_vec_info> stmts;
4977 : 18234 : vec<stmt_vec_info> roots = vNULL;
4978 : 18234 : vec<tree> remain = vNULL;
4979 : 18234 : stmts.create (1);
4980 : 18234 : stmts.quick_push (vect_stmt_to_vectorize (stmt_info));
4981 : 18234 : vect_build_slp_instance (vinfo,
4982 : : slp_inst_kind_reduc_group,
4983 : : stmts, roots, remain,
4984 : : max_tree_size, &limit,
4985 : : bst_map, NULL, force_single_lane);
4986 : : }
4987 : 241907 : }
4988 : :
4989 : : /* Find SLP sequences starting from gconds. */
4990 : 410570 : for (auto cond : LOOP_VINFO_LOOP_CONDS (loop_vinfo))
4991 : : {
4992 : 63413 : auto cond_info = loop_vinfo->lookup_stmt (cond);
4993 : :
4994 : 63413 : cond_info = vect_stmt_to_vectorize (cond_info);
4995 : 63413 : vec<stmt_vec_info> roots = vNULL;
4996 : 63413 : roots.safe_push (cond_info);
4997 : 63413 : gimple *stmt = STMT_VINFO_STMT (cond_info);
4998 : 63413 : tree args0 = gimple_cond_lhs (stmt);
4999 : 63413 : tree args1 = gimple_cond_rhs (stmt);
5000 : :
5001 : : /* These should be enforced by cond lowering. */
5002 : 63413 : gcc_assert (gimple_cond_code (stmt) == NE_EXPR);
5003 : 63413 : gcc_assert (zerop (args1));
5004 : :
5005 : : /* An argument without a loop def will be codegened from vectorizing the
5006 : : root gcond itself. As such we don't need to try to build an SLP tree
5007 : : from them. It's highly likely that the resulting SLP tree here if both
5008 : : arguments have a def will be incompatible, but we rely on it being split
5009 : : later on. */
5010 : 63413 : auto varg = loop_vinfo->lookup_def (args0);
5011 : 63413 : vec<stmt_vec_info> stmts;
5012 : 63413 : vec<tree> remain = vNULL;
5013 : 63413 : stmts.create (1);
5014 : 63413 : stmts.quick_push (vect_stmt_to_vectorize (varg));
5015 : :
5016 : 63413 : vect_build_slp_instance (vinfo, slp_inst_kind_gcond,
5017 : : stmts, roots, remain,
5018 : : max_tree_size, &limit,
5019 : : bst_map, NULL, force_single_lane);
5020 : : }
5021 : :
5022 : : /* Find and create slp instances for inductions that have been forced
5023 : : live due to early break. */
5024 : 241907 : edge latch_e = loop_latch_edge (LOOP_VINFO_LOOP (loop_vinfo));
5025 : 452692 : for (auto stmt_info : LOOP_VINFO_EARLY_BREAKS_LIVE_IVS (loop_vinfo))
5026 : : {
5027 : 105535 : vec<stmt_vec_info> stmts;
5028 : 105535 : vec<stmt_vec_info> roots = vNULL;
5029 : 105535 : vec<tree> remain = vNULL;
5030 : 105535 : gphi *lc_phi = as_a<gphi *> (STMT_VINFO_STMT (stmt_info));
5031 : 105535 : tree def = gimple_phi_arg_def_from_edge (lc_phi, latch_e);
5032 : 105535 : stmt_vec_info lc_info = loop_vinfo->lookup_def (def);
5033 : 105535 : stmts.create (1);
5034 : 105565 : stmts.quick_push (vect_stmt_to_vectorize (lc_info));
5035 : 105535 : vect_build_slp_instance (vinfo, slp_inst_kind_reduc_group,
5036 : : stmts, roots, remain,
5037 : : max_tree_size, &limit,
5038 : : bst_map, NULL, force_single_lane);
5039 : : }
5040 : : }
5041 : :
5042 : 837809 : hash_set<slp_tree> visited_patterns;
5043 : 837809 : slp_tree_to_load_perm_map_t perm_cache;
5044 : 837809 : slp_compat_nodes_map_t compat_cache;
5045 : :
5046 : : /* See if any patterns can be found in the SLP tree. */
5047 : 837809 : bool pattern_found = false;
5048 : 2877078 : FOR_EACH_VEC_ELT (LOOP_VINFO_SLP_INSTANCES (vinfo), i, instance)
5049 : 1201460 : pattern_found |= vect_match_slp_patterns (instance, vinfo,
5050 : : &visited_patterns, &perm_cache,
5051 : : &compat_cache);
5052 : :
5053 : : /* If any were found optimize permutations of loads. */
5054 : 837809 : if (pattern_found)
5055 : : {
5056 : 224 : hash_map<slp_tree, slp_tree> load_map;
5057 : 1699 : FOR_EACH_VEC_ELT (LOOP_VINFO_SLP_INSTANCES (vinfo), i, instance)
5058 : : {
5059 : 1251 : slp_tree root = SLP_INSTANCE_TREE (instance);
5060 : 1251 : optimize_load_redistribution (bst_map, vinfo, SLP_TREE_LANES (root),
5061 : : &load_map, root);
5062 : : }
5063 : 224 : }
5064 : :
5065 : : /* Check whether we should force some SLP instances to use load/store-lanes
5066 : : and do so by forcing SLP re-discovery with single lanes. We used
5067 : : to cancel SLP when this applied to all instances in a loop but now
5068 : : we decide this per SLP instance. It's important to do this only
5069 : : after SLP pattern recognition. */
5070 : 837809 : if (is_a <loop_vec_info> (vinfo))
5071 : 681004 : FOR_EACH_VEC_ELT (LOOP_VINFO_SLP_INSTANCES (vinfo), i, instance)
5072 : 439097 : if (SLP_INSTANCE_KIND (instance) == slp_inst_kind_store
5073 : 186509 : && !SLP_INSTANCE_TREE (instance)->ldst_lanes)
5074 : : {
5075 : 186509 : slp_tree slp_root = SLP_INSTANCE_TREE (instance);
5076 : 186509 : int group_size = SLP_TREE_LANES (slp_root);
5077 : 186509 : tree vectype = SLP_TREE_VECTYPE (slp_root);
5078 : :
5079 : 186509 : stmt_vec_info rep_info = SLP_TREE_REPRESENTATIVE (slp_root);
5080 : 186509 : gimple *rep = STMT_VINFO_STMT (rep_info);
5081 : 186509 : bool masked = (is_gimple_call (rep)
5082 : 1678 : && gimple_call_internal_p (rep)
5083 : 188187 : && internal_fn_mask_index
5084 : 1678 : (gimple_call_internal_fn (rep)) != -1);
5085 : 186509 : if (!STMT_VINFO_GROUPED_ACCESS (rep_info)
5086 : 16540 : || slp_root->ldst_lanes
5087 : 203049 : || (vect_store_lanes_supported (vectype, group_size, masked)
5088 : : == IFN_LAST))
5089 : 186509 : continue;
5090 : :
5091 : 0 : auto_vec<slp_tree> loads;
5092 : 0 : hash_set<slp_tree> visited;
5093 : 0 : vect_gather_slp_loads (loads, slp_root, visited);
5094 : :
5095 : : /* Check whether any load in the SLP instance is possibly
5096 : : permuted. */
5097 : 0 : bool loads_permuted = false;
5098 : 0 : slp_tree load_node;
5099 : 0 : unsigned j;
5100 : 0 : FOR_EACH_VEC_ELT (loads, j, load_node)
5101 : : {
5102 : 0 : if (!SLP_TREE_LOAD_PERMUTATION (load_node).exists ())
5103 : 0 : continue;
5104 : : unsigned k;
5105 : : stmt_vec_info load_info;
5106 : 0 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (load_node), k, load_info)
5107 : 0 : if (SLP_TREE_LOAD_PERMUTATION (load_node)[k] != k)
5108 : : {
5109 : : loads_permuted = true;
5110 : : break;
5111 : : }
5112 : : }
5113 : :
5114 : : /* If the loads and stores can use load/store-lanes force re-discovery
5115 : : with single lanes. */
5116 : 0 : if (loads_permuted)
5117 : : {
5118 : 0 : bool can_use_lanes = true;
5119 : 0 : FOR_EACH_VEC_ELT (loads, j, load_node)
5120 : 0 : if (STMT_VINFO_GROUPED_ACCESS
5121 : : (SLP_TREE_REPRESENTATIVE (load_node)))
5122 : : {
5123 : 0 : stmt_vec_info stmt_vinfo = DR_GROUP_FIRST_ELEMENT
5124 : : (SLP_TREE_REPRESENTATIVE (load_node));
5125 : 0 : rep = STMT_VINFO_STMT (stmt_vinfo);
5126 : 0 : masked = (is_gimple_call (rep)
5127 : 0 : && gimple_call_internal_p (rep)
5128 : 0 : && internal_fn_mask_index
5129 : 0 : (gimple_call_internal_fn (rep)));
5130 : : /* Use SLP for strided accesses (or if we can't
5131 : : load-lanes). */
5132 : 0 : if (STMT_VINFO_STRIDED_P (stmt_vinfo)
5133 : 0 : || compare_step_with_zero (vinfo, stmt_vinfo) <= 0
5134 : 0 : || vect_load_lanes_supported
5135 : 0 : (STMT_VINFO_VECTYPE (stmt_vinfo),
5136 : 0 : DR_GROUP_SIZE (stmt_vinfo), masked) == IFN_LAST
5137 : : /* ??? During SLP re-discovery with a single lane
5138 : : a masked grouped load will appear permuted and
5139 : : discovery will fail. We have to rework this
5140 : : on the discovery side - for now avoid ICEing. */
5141 : 0 : || masked)
5142 : : {
5143 : : can_use_lanes = false;
5144 : : break;
5145 : : }
5146 : : }
5147 : :
5148 : 0 : if (can_use_lanes)
5149 : : {
5150 : 0 : if (dump_enabled_p ())
5151 : 0 : dump_printf_loc (MSG_NOTE, vect_location,
5152 : : "SLP instance %p can use load/store-lanes,"
5153 : : " re-discovering with single-lanes\n",
5154 : : (void *) instance);
5155 : :
5156 : 0 : stmt_vec_info stmt_info = SLP_TREE_REPRESENTATIVE (slp_root);
5157 : :
5158 : 0 : vect_free_slp_instance (instance);
5159 : 0 : limit = max_tree_size;
5160 : 0 : bool res = vect_analyze_slp_instance (vinfo, bst_map,
5161 : : stmt_info,
5162 : : slp_inst_kind_store,
5163 : : max_tree_size, &limit,
5164 : : true);
5165 : 0 : gcc_assert (res);
5166 : 0 : auto new_inst = LOOP_VINFO_SLP_INSTANCES (vinfo).pop ();
5167 : 0 : LOOP_VINFO_SLP_INSTANCES (vinfo)[i] = new_inst;
5168 : : }
5169 : : }
5170 : 0 : }
5171 : :
5172 : : /* When we end up with load permutations that we cannot possibly handle,
5173 : : like those requiring three vector inputs, lower them using interleaving
5174 : : like schemes. */
5175 : 837809 : if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
5176 : : {
5177 : 241907 : vect_lower_load_permutations (loop_vinfo, bst_map, force_single_lane);
5178 : 241907 : if (dump_enabled_p ())
5179 : : {
5180 : 19888 : dump_printf_loc (MSG_NOTE, vect_location,
5181 : : "SLP graph after lowering permutations:\n");
5182 : 19888 : hash_set<slp_tree> visited;
5183 : 90328 : FOR_EACH_VEC_ELT (LOOP_VINFO_SLP_INSTANCES (vinfo), i, instance)
5184 : 30974 : vect_print_slp_graph (MSG_NOTE, vect_location,
5185 : : SLP_INSTANCE_TREE (instance), visited);
5186 : 19888 : }
5187 : : }
5188 : :
5189 : 837809 : release_scalar_stmts_to_slp_tree_map (bst_map);
5190 : :
5191 : 837809 : if (pattern_found && dump_enabled_p ())
5192 : : {
5193 : 22 : dump_printf_loc (MSG_NOTE, vect_location,
5194 : : "Pattern matched SLP tree\n");
5195 : 22 : hash_set<slp_tree> visited;
5196 : 142 : FOR_EACH_VEC_ELT (LOOP_VINFO_SLP_INSTANCES (vinfo), i, instance)
5197 : 76 : vect_print_slp_graph (MSG_NOTE, vect_location,
5198 : : SLP_INSTANCE_TREE (instance), visited);
5199 : 22 : }
5200 : :
5201 : 837809 : return opt_result::success ();
5202 : 837809 : }
5203 : :
5204 : : /* Estimates the cost of inserting layout changes into the SLP graph.
5205 : : It can also say that the insertion is impossible. */
5206 : :
5207 : : struct slpg_layout_cost
5208 : : {
5209 : 8248368 : slpg_layout_cost () = default;
5210 : : slpg_layout_cost (sreal, bool);
5211 : :
5212 : 323486 : static slpg_layout_cost impossible () { return { sreal::max (), 0 }; }
5213 : 3543348 : bool is_possible () const { return depth != sreal::max (); }
5214 : :
5215 : : bool operator== (const slpg_layout_cost &) const;
5216 : : bool operator!= (const slpg_layout_cost &) const;
5217 : :
5218 : : bool is_better_than (const slpg_layout_cost &, bool) const;
5219 : :
5220 : : void add_parallel_cost (const slpg_layout_cost &);
5221 : : void add_serial_cost (const slpg_layout_cost &);
5222 : : void split (unsigned int);
5223 : :
5224 : : /* The longest sequence of layout changes needed during any traversal
5225 : : of the partition dag, weighted by execution frequency.
5226 : :
5227 : : This is the most important metric when optimizing for speed, since
5228 : : it helps to ensure that we keep the number of operations on
5229 : : critical paths to a minimum. */
5230 : : sreal depth = 0;
5231 : :
5232 : : /* An estimate of the total number of operations needed. It is weighted by
5233 : : execution frequency when optimizing for speed but not when optimizing for
5234 : : size. In order to avoid double-counting, a node with a fanout of N will
5235 : : distribute 1/N of its total cost to each successor.
5236 : :
5237 : : This is the most important metric when optimizing for size, since
5238 : : it helps to keep the total number of operations to a minimum, */
5239 : : sreal total = 0;
5240 : : };
5241 : :
5242 : : /* Construct costs for a node with weight WEIGHT. A higher weight
5243 : : indicates more frequent execution. IS_FOR_SIZE is true if we are
5244 : : optimizing for size rather than speed. */
5245 : :
5246 : 805348 : slpg_layout_cost::slpg_layout_cost (sreal weight, bool is_for_size)
5247 : 806097 : : depth (weight), total (is_for_size && weight > 0 ? 1 : weight)
5248 : : {
5249 : 805348 : }
5250 : :
5251 : : bool
5252 : 0 : slpg_layout_cost::operator== (const slpg_layout_cost &other) const
5253 : : {
5254 : 0 : return depth == other.depth && total == other.total;
5255 : : }
5256 : :
5257 : : bool
5258 : 0 : slpg_layout_cost::operator!= (const slpg_layout_cost &other) const
5259 : : {
5260 : 0 : return !operator== (other);
5261 : : }
5262 : :
5263 : : /* Return true if these costs are better than OTHER. IS_FOR_SIZE is
5264 : : true if we are optimizing for size rather than speed. */
5265 : :
5266 : : bool
5267 : 216317 : slpg_layout_cost::is_better_than (const slpg_layout_cost &other,
5268 : : bool is_for_size) const
5269 : : {
5270 : 216317 : if (is_for_size)
5271 : : {
5272 : 327 : if (total != other.total)
5273 : 130 : return total < other.total;
5274 : 197 : return depth < other.depth;
5275 : : }
5276 : : else
5277 : : {
5278 : 215990 : if (depth != other.depth)
5279 : 98978 : return depth < other.depth;
5280 : 117012 : return total < other.total;
5281 : : }
5282 : : }
5283 : :
5284 : : /* Increase the costs to account for something with cost INPUT_COST
5285 : : happening in parallel with the current costs. */
5286 : :
5287 : : void
5288 : 232588 : slpg_layout_cost::add_parallel_cost (const slpg_layout_cost &input_cost)
5289 : : {
5290 : 232588 : depth = std::max (depth, input_cost.depth);
5291 : 232588 : total += input_cost.total;
5292 : 232588 : }
5293 : :
5294 : : /* Increase the costs to account for something with cost INPUT_COST
5295 : : happening in series with the current costs. */
5296 : :
5297 : : void
5298 : 1002299 : slpg_layout_cost::add_serial_cost (const slpg_layout_cost &other)
5299 : : {
5300 : 1002299 : depth += other.depth;
5301 : 1002299 : total += other.total;
5302 : 1002299 : }
5303 : :
5304 : : /* Split the total cost among TIMES successors or predecessors. */
5305 : :
5306 : : void
5307 : 773581 : slpg_layout_cost::split (unsigned int times)
5308 : : {
5309 : 773581 : if (times > 1)
5310 : 250609 : total /= times;
5311 : 773581 : }
5312 : :
5313 : : /* Information about one node in the SLP graph, for use during
5314 : : vect_optimize_slp_pass. */
5315 : :
5316 : : struct slpg_vertex
5317 : : {
5318 : 7982617 : slpg_vertex (slp_tree node_) : node (node_) {}
5319 : :
5320 : : /* The node itself. */
5321 : : slp_tree node;
5322 : :
5323 : : /* Which partition the node belongs to, or -1 if none. Nodes outside of
5324 : : partitions are flexible; they can have whichever layout consumers
5325 : : want them to have. */
5326 : : int partition = -1;
5327 : :
5328 : : /* The number of nodes that directly use the result of this one
5329 : : (i.e. the number of nodes that count this one as a child). */
5330 : : unsigned int out_degree = 0;
5331 : :
5332 : : /* The execution frequency of the node. */
5333 : : sreal weight = 0;
5334 : :
5335 : : /* The total execution frequency of all nodes that directly use the
5336 : : result of this one. */
5337 : : sreal out_weight = 0;
5338 : : };
5339 : :
5340 : : /* Information about one partition of the SLP graph, for use during
5341 : : vect_optimize_slp_pass. */
5342 : :
5343 : 2631804 : struct slpg_partition_info
5344 : : {
5345 : : /* The nodes in the partition occupy indices [NODE_BEGIN, NODE_END)
5346 : : of m_partitioned_nodes. */
5347 : : unsigned int node_begin = 0;
5348 : : unsigned int node_end = 0;
5349 : :
5350 : : /* Which layout we've chosen to use for this partition, or -1 if
5351 : : we haven't picked one yet. */
5352 : : int layout = -1;
5353 : :
5354 : : /* The number of predecessors and successors in the partition dag.
5355 : : The predecessors always have lower partition numbers and the
5356 : : successors always have higher partition numbers.
5357 : :
5358 : : Note that the directions of these edges are not necessarily the
5359 : : same as in the data flow graph. For example, if an SCC has separate
5360 : : partitions for an inner loop and an outer loop, the inner loop's
5361 : : partition will have at least two incoming edges from the outer loop's
5362 : : partition: one for a live-in value and one for a live-out value.
5363 : : In data flow terms, one of these edges would also be from the outer loop
5364 : : to the inner loop, but the other would be in the opposite direction. */
5365 : : unsigned int in_degree = 0;
5366 : : unsigned int out_degree = 0;
5367 : : };
5368 : :
5369 : : /* Information about the costs of using a particular layout for a
5370 : : particular partition. It can also say that the combination is
5371 : : impossible. */
5372 : :
5373 : : struct slpg_partition_layout_costs
5374 : : {
5375 : 1041059 : bool is_possible () const { return internal_cost.is_possible (); }
5376 : 37276 : void mark_impossible () { internal_cost = slpg_layout_cost::impossible (); }
5377 : :
5378 : : /* The costs inherited from predecessor partitions. */
5379 : : slpg_layout_cost in_cost;
5380 : :
5381 : : /* The inherent cost of the layout within the node itself. For example,
5382 : : this is nonzero for a load if choosing a particular layout would require
5383 : : the load to permute the loaded elements. It is nonzero for a
5384 : : VEC_PERM_EXPR if the permutation cannot be eliminated or converted
5385 : : to full-vector moves. */
5386 : : slpg_layout_cost internal_cost;
5387 : :
5388 : : /* The costs inherited from successor partitions. */
5389 : : slpg_layout_cost out_cost;
5390 : : };
5391 : :
5392 : : /* This class tries to optimize the layout of vectors in order to avoid
5393 : : unnecessary shuffling. At the moment, the set of possible layouts are
5394 : : restricted to bijective permutations.
5395 : :
5396 : : The goal of the pass depends on whether we're optimizing for size or
5397 : : for speed. When optimizing for size, the goal is to reduce the overall
5398 : : number of layout changes (including layout changes implied by things
5399 : : like load permutations). When optimizing for speed, the goal is to
5400 : : reduce the maximum latency attributable to layout changes on any
5401 : : non-cyclical path through the data flow graph.
5402 : :
5403 : : For example, when optimizing a loop nest for speed, we will prefer
5404 : : to make layout changes outside of a loop rather than inside of a loop,
5405 : : and will prefer to make layout changes in parallel rather than serially,
5406 : : even if that increases the overall number of layout changes.
5407 : :
5408 : : The high-level procedure is:
5409 : :
5410 : : (1) Build a graph in which edges go from uses (parents) to definitions
5411 : : (children).
5412 : :
5413 : : (2) Divide the graph into a dag of strongly-connected components (SCCs).
5414 : :
5415 : : (3) When optimizing for speed, partition the nodes in each SCC based
5416 : : on their containing cfg loop. When optimizing for size, treat
5417 : : each SCC as a single partition.
5418 : :
5419 : : This gives us a dag of partitions. The goal is now to assign a
5420 : : layout to each partition.
5421 : :
5422 : : (4) Construct a set of vector layouts that are worth considering.
5423 : : Record which nodes must keep their current layout.
5424 : :
5425 : : (5) Perform a forward walk over the partition dag (from loads to stores)
5426 : : accumulating the "forward" cost of using each layout. When visiting
5427 : : each partition, assign a tentative choice of layout to the partition
5428 : : and use that choice when calculating the cost of using a different
5429 : : layout in successor partitions.
5430 : :
5431 : : (6) Perform a backward walk over the partition dag (from stores to loads),
5432 : : accumulating the "backward" cost of using each layout. When visiting
5433 : : each partition, make a final choice of layout for that partition based
5434 : : on the accumulated forward costs (from (5)) and backward costs
5435 : : (from (6)).
5436 : :
5437 : : (7) Apply the chosen layouts to the SLP graph.
5438 : :
5439 : : For example, consider the SLP statements:
5440 : :
5441 : : S1: a_1 = load
5442 : : loop:
5443 : : S2: a_2 = PHI<a_1, a_3>
5444 : : S3: b_1 = load
5445 : : S4: a_3 = a_2 + b_1
5446 : : exit:
5447 : : S5: a_4 = PHI<a_3>
5448 : : S6: store a_4
5449 : :
5450 : : S2 and S4 form an SCC and are part of the same loop. Every other
5451 : : statement is in a singleton SCC. In this example there is a one-to-one
5452 : : mapping between SCCs and partitions and the partition dag looks like this;
5453 : :
5454 : : S1 S3
5455 : : \ /
5456 : : S2+S4
5457 : : |
5458 : : S5
5459 : : |
5460 : : S6
5461 : :
5462 : : S2, S3 and S4 will have a higher execution frequency than the other
5463 : : statements, so when optimizing for speed, the goal is to avoid any
5464 : : layout changes:
5465 : :
5466 : : - within S3
5467 : : - within S2+S4
5468 : : - on the S3->S2+S4 edge
5469 : :
5470 : : For example, if S3 was originally a reversing load, the goal of the
5471 : : pass is to make it an unreversed load and change the layout on the
5472 : : S1->S2+S4 and S2+S4->S5 edges to compensate. (Changing the layout
5473 : : on S1->S2+S4 and S5->S6 would also be acceptable.)
5474 : :
5475 : : The difference between SCCs and partitions becomes important if we
5476 : : add an outer loop:
5477 : :
5478 : : S1: a_1 = ...
5479 : : loop1:
5480 : : S2: a_2 = PHI<a_1, a_6>
5481 : : S3: b_1 = load
5482 : : S4: a_3 = a_2 + b_1
5483 : : loop2:
5484 : : S5: a_4 = PHI<a_3, a_5>
5485 : : S6: c_1 = load
5486 : : S7: a_5 = a_4 + c_1
5487 : : exit2:
5488 : : S8: a_6 = PHI<a_5>
5489 : : S9: store a_6
5490 : : exit1:
5491 : :
5492 : : Here, S2, S4, S5, S7 and S8 form a single SCC. However, when optimizing
5493 : : for speed, we usually do not want restrictions in the outer loop to "infect"
5494 : : the decision for the inner loop. For example, if an outer-loop node
5495 : : in the SCC contains a statement with a fixed layout, that should not
5496 : : prevent the inner loop from using a different layout. Conversely,
5497 : : the inner loop should not dictate a layout to the outer loop: if the
5498 : : outer loop does a lot of computation, then it may not be efficient to
5499 : : do all of that computation in the inner loop's preferred layout.
5500 : :
5501 : : So when optimizing for speed, we partition the SCC into S2+S4+S8 (outer)
5502 : : and S5+S7 (inner). We also try to arrange partitions so that:
5503 : :
5504 : : - the partition for an outer loop comes before the partition for
5505 : : an inner loop
5506 : :
5507 : : - if a sibling loop A dominates a sibling loop B, A's partition
5508 : : comes before B's
5509 : :
5510 : : This gives the following partition dag for the example above:
5511 : :
5512 : : S1 S3
5513 : : \ /
5514 : : S2+S4+S8 S6
5515 : : | \\ /
5516 : : | S5+S7
5517 : : |
5518 : : S9
5519 : :
5520 : : There are two edges from S2+S4+S8 to S5+S7: one for the edge S4->S5 and
5521 : : one for a reversal of the edge S7->S8.
5522 : :
5523 : : The backward walk picks a layout for S5+S7 before S2+S4+S8. The choice
5524 : : for S2+S4+S8 therefore has to balance the cost of using the outer loop's
5525 : : preferred layout against the cost of changing the layout on entry to the
5526 : : inner loop (S4->S5) and on exit from the inner loop (S7->S8 reversed).
5527 : :
5528 : : Although this works well when optimizing for speed, it has the downside
5529 : : when optimizing for size that the choice of layout for S5+S7 is completely
5530 : : independent of S9, which lessens the chance of reducing the overall number
5531 : : of permutations. We therefore do not partition SCCs when optimizing
5532 : : for size.
5533 : :
5534 : : To give a concrete example of the difference between optimizing
5535 : : for size and speed, consider:
5536 : :
5537 : : a[0] = (b[1] << c[3]) - d[1];
5538 : : a[1] = (b[0] << c[2]) - d[0];
5539 : : a[2] = (b[3] << c[1]) - d[3];
5540 : : a[3] = (b[2] << c[0]) - d[2];
5541 : :
5542 : : There are three different layouts here: one for a, one for b and d,
5543 : : and one for c. When optimizing for speed it is better to permute each
5544 : : of b, c and d into the order required by a, since those permutations
5545 : : happen in parallel. But when optimizing for size, it is better to:
5546 : :
5547 : : - permute c into the same order as b
5548 : : - do the arithmetic
5549 : : - permute the result into the order required by a
5550 : :
5551 : : This gives 2 permutations rather than 3. */
5552 : :
5553 : : class vect_optimize_slp_pass
5554 : : {
5555 : : public:
5556 : 553379 : vect_optimize_slp_pass (vec_info *vinfo) : m_vinfo (vinfo) {}
5557 : : void run ();
5558 : :
5559 : : private:
5560 : : /* Graph building. */
5561 : : struct loop *containing_loop (slp_tree);
5562 : : bool is_cfg_latch_edge (graph_edge *);
5563 : : void build_vertices (hash_set<slp_tree> &, slp_tree);
5564 : : void build_vertices ();
5565 : : void build_graph ();
5566 : :
5567 : : /* Partitioning. */
5568 : : void create_partitions ();
5569 : : template<typename T> void for_each_partition_edge (unsigned int, T);
5570 : :
5571 : : /* Layout selection. */
5572 : : bool is_compatible_layout (slp_tree, unsigned int);
5573 : : int change_layout_cost (slp_tree, unsigned int, unsigned int);
5574 : : slpg_partition_layout_costs &partition_layout_costs (unsigned int,
5575 : : unsigned int);
5576 : : void change_vec_perm_layout (slp_tree, lane_permutation_t &,
5577 : : int, unsigned int);
5578 : : int internal_node_cost (slp_tree, int, unsigned int);
5579 : : void start_choosing_layouts ();
5580 : :
5581 : : /* Cost propagation. */
5582 : : slpg_layout_cost edge_layout_cost (graph_edge *, unsigned int,
5583 : : unsigned int, unsigned int);
5584 : : slpg_layout_cost total_in_cost (unsigned int);
5585 : : slpg_layout_cost forward_cost (graph_edge *, unsigned int, unsigned int);
5586 : : slpg_layout_cost backward_cost (graph_edge *, unsigned int, unsigned int);
5587 : : void forward_pass ();
5588 : : void backward_pass ();
5589 : :
5590 : : /* Rematerialization. */
5591 : : slp_tree get_result_with_layout (slp_tree, unsigned int);
5592 : : void materialize ();
5593 : :
5594 : : /* Clean-up. */
5595 : : void remove_redundant_permutations ();
5596 : :
5597 : : /* Masked load lanes discovery. */
5598 : : void decide_masked_load_lanes ();
5599 : :
5600 : : void dump ();
5601 : :
5602 : : vec_info *m_vinfo;
5603 : :
5604 : : /* True if we should optimize the graph for size, false if we should
5605 : : optimize it for speed. (It wouldn't be easy to make this decision
5606 : : more locally.) */
5607 : : bool m_optimize_size;
5608 : :
5609 : : /* A graph of all SLP nodes, with edges leading from uses to definitions.
5610 : : In other words, a node's predecessors are its slp_tree parents and
5611 : : a node's successors are its slp_tree children. */
5612 : : graph *m_slpg = nullptr;
5613 : :
5614 : : /* The vertices of M_SLPG, indexed by slp_tree::vertex. */
5615 : : auto_vec<slpg_vertex> m_vertices;
5616 : :
5617 : : /* The list of all leaves of M_SLPG. such as external definitions, constants,
5618 : : and loads. */
5619 : : auto_vec<int> m_leafs;
5620 : :
5621 : : /* This array has one entry for every vector layout that we're considering.
5622 : : Element 0 is null and indicates "no change". Other entries describe
5623 : : permutations that are inherent in the current graph and that we would
5624 : : like to reverse if possible.
5625 : :
5626 : : For example, a permutation { 1, 2, 3, 0 } means that something has
5627 : : effectively been permuted in that way, such as a load group
5628 : : { a[1], a[2], a[3], a[0] } (viewed as a permutation of a[0:3]).
5629 : : We'd then like to apply the reverse permutation { 3, 0, 1, 2 }
5630 : : in order to put things "back" in order. */
5631 : : auto_vec<vec<unsigned> > m_perms;
5632 : :
5633 : : /* A partitioning of the nodes for which a layout must be chosen.
5634 : : Each partition represents an <SCC, cfg loop> pair; that is,
5635 : : nodes in different SCCs belong to different partitions, and nodes
5636 : : within an SCC can be further partitioned according to a containing
5637 : : cfg loop. Partition <SCC1, L1> comes before <SCC2, L2> if:
5638 : :
5639 : : - SCC1 != SCC2 and SCC1 is a predecessor of SCC2 in a forward walk
5640 : : from leaves (such as loads) to roots (such as stores).
5641 : :
5642 : : - SCC1 == SCC2 and L1's header strictly dominates L2's header. */
5643 : : auto_vec<slpg_partition_info> m_partitions;
5644 : :
5645 : : /* The list of all nodes for which a layout must be chosen. Nodes for
5646 : : partition P come before the nodes for partition P+1. Nodes within a
5647 : : partition are in reverse postorder. */
5648 : : auto_vec<unsigned int> m_partitioned_nodes;
5649 : :
5650 : : /* Index P * num-layouts + L contains the cost of using layout L
5651 : : for partition P. */
5652 : : auto_vec<slpg_partition_layout_costs> m_partition_layout_costs;
5653 : :
5654 : : /* Index N * num-layouts + L, if nonnull, is a node that provides the
5655 : : original output of node N adjusted to have layout L. */
5656 : : auto_vec<slp_tree> m_node_layouts;
5657 : : };
5658 : :
5659 : : /* Fill the vertices and leafs vector with all nodes in the SLP graph.
5660 : : Also record whether we should optimize anything for speed rather
5661 : : than size. */
5662 : :
5663 : : void
5664 : 8576149 : vect_optimize_slp_pass::build_vertices (hash_set<slp_tree> &visited,
5665 : : slp_tree node)
5666 : : {
5667 : 8576149 : unsigned i;
5668 : 8576149 : slp_tree child;
5669 : :
5670 : 8576149 : if (visited.add (node))
5671 : 8576149 : return;
5672 : :
5673 : 7982617 : if (stmt_vec_info rep = SLP_TREE_REPRESENTATIVE (node))
5674 : : {
5675 : 5839312 : basic_block bb = gimple_bb (vect_orig_stmt (rep)->stmt);
5676 : 5428196 : if (optimize_bb_for_speed_p (bb))
5677 : 5313656 : m_optimize_size = false;
5678 : : }
5679 : :
5680 : 7982617 : node->vertex = m_vertices.length ();
5681 : 7982617 : m_vertices.safe_push (slpg_vertex (node));
5682 : :
5683 : 7982617 : bool leaf = true;
5684 : 7982617 : bool force_leaf = false;
5685 : 14926712 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
5686 : 6944095 : if (child)
5687 : : {
5688 : 6173229 : leaf = false;
5689 : 6173229 : build_vertices (visited, child);
5690 : : }
5691 : : else
5692 : : force_leaf = true;
5693 : : /* Since SLP discovery works along use-def edges all cycles have an
5694 : : entry - but there's the exception of cycles where we do not handle
5695 : : the entry explicitely (but with a NULL SLP node), like some reductions
5696 : : and inductions. Force those SLP PHIs to act as leafs to make them
5697 : : backwards reachable. */
5698 : 7982617 : if (leaf || force_leaf)
5699 : 3882142 : m_leafs.safe_push (node->vertex);
5700 : : }
5701 : :
5702 : : /* Fill the vertices and leafs vector with all nodes in the SLP graph. */
5703 : :
5704 : : void
5705 : 1106758 : vect_optimize_slp_pass::build_vertices ()
5706 : : {
5707 : 1106758 : hash_set<slp_tree> visited;
5708 : 1106758 : unsigned i;
5709 : 1106758 : slp_instance instance;
5710 : 1106758 : m_vertices.truncate (0);
5711 : 1106758 : m_leafs.truncate (0);
5712 : 5723194 : FOR_EACH_VEC_ELT (m_vinfo->slp_instances, i, instance)
5713 : 2402920 : build_vertices (visited, SLP_INSTANCE_TREE (instance));
5714 : 1106758 : }
5715 : :
5716 : : /* Apply (reverse) bijectite PERM to VEC. */
5717 : :
5718 : : template <class T>
5719 : : static void
5720 : 139364 : vect_slp_permute (vec<unsigned> perm,
5721 : : vec<T> &vec, bool reverse)
5722 : : {
5723 : 139364 : auto_vec<T, 64> saved;
5724 : 139364 : saved.create (vec.length ());
5725 : 477902 : for (unsigned i = 0; i < vec.length (); ++i)
5726 : 338538 : saved.quick_push (vec[i]);
5727 : :
5728 : 139364 : if (reverse)
5729 : : {
5730 : 948246 : for (unsigned i = 0; i < vec.length (); ++i)
5731 : 337470 : vec[perm[i]] = saved[i];
5732 : 476360 : for (unsigned i = 0; i < vec.length (); ++i)
5733 : 577246 : gcc_assert (vec[perm[i]] == saved[i]);
5734 : : }
5735 : : else
5736 : : {
5737 : 3084 : for (unsigned i = 0; i < vec.length (); ++i)
5738 : 1068 : vec[i] = saved[perm[i]];
5739 : 140432 : for (unsigned i = 0; i < vec.length (); ++i)
5740 : 1602 : gcc_assert (vec[i] == saved[perm[i]]);
5741 : : }
5742 : 139364 : }
5743 : :
5744 : : /* Return the cfg loop that contains NODE. */
5745 : :
5746 : : struct loop *
5747 : 3009082 : vect_optimize_slp_pass::containing_loop (slp_tree node)
5748 : : {
5749 : 3009082 : stmt_vec_info rep = SLP_TREE_REPRESENTATIVE (node);
5750 : 3009082 : if (!rep)
5751 : 5335 : return ENTRY_BLOCK_PTR_FOR_FN (cfun)->loop_father;
5752 : 3218665 : return gimple_bb (vect_orig_stmt (rep)->stmt)->loop_father;
5753 : : }
5754 : :
5755 : : /* Return true if UD (an edge from a use to a definition) is associated
5756 : : with a loop latch edge in the cfg. */
5757 : :
5758 : : bool
5759 : 6173229 : vect_optimize_slp_pass::is_cfg_latch_edge (graph_edge *ud)
5760 : : {
5761 : 6173229 : slp_tree use = m_vertices[ud->src].node;
5762 : 6173229 : slp_tree def = m_vertices[ud->dest].node;
5763 : 6173229 : if ((SLP_TREE_DEF_TYPE (use) != vect_internal_def
5764 : 6173229 : || SLP_TREE_CODE (use) == VEC_PERM_EXPR)
5765 : 5890742 : || SLP_TREE_DEF_TYPE (def) != vect_internal_def)
5766 : : return false;
5767 : :
5768 : 3349670 : stmt_vec_info use_rep = vect_orig_stmt (SLP_TREE_REPRESENTATIVE (use));
5769 : 3349670 : return (is_a<gphi *> (use_rep->stmt)
5770 : 289572 : && bb_loop_header_p (gimple_bb (use_rep->stmt))
5771 : 3508970 : && containing_loop (def) == containing_loop (use));
5772 : : }
5773 : :
5774 : : /* Build the graph. Mark edges that correspond to cfg loop latch edges with
5775 : : a nonnull data field. */
5776 : :
5777 : : void
5778 : 1106758 : vect_optimize_slp_pass::build_graph ()
5779 : : {
5780 : 1106758 : m_optimize_size = true;
5781 : 1106758 : build_vertices ();
5782 : :
5783 : 2213516 : m_slpg = new_graph (m_vertices.length ());
5784 : 11302891 : for (slpg_vertex &v : m_vertices)
5785 : 24032886 : for (slp_tree child : SLP_TREE_CHILDREN (v.node))
5786 : 6944095 : if (child)
5787 : : {
5788 : 6173229 : graph_edge *ud = add_edge (m_slpg, v.node->vertex, child->vertex);
5789 : 6173229 : if (is_cfg_latch_edge (ud))
5790 : 151626 : ud->data = this;
5791 : : }
5792 : 1106758 : }
5793 : :
5794 : : /* Return true if E corresponds to a loop latch edge in the cfg. */
5795 : :
5796 : : static bool
5797 : 3162285 : skip_cfg_latch_edges (graph_edge *e)
5798 : : {
5799 : 3162285 : return e->data;
5800 : : }
5801 : :
5802 : : /* Create the node partitions. */
5803 : :
5804 : : void
5805 : 553379 : vect_optimize_slp_pass::create_partitions ()
5806 : : {
5807 : : /* Calculate a postorder of the graph, ignoring edges that correspond
5808 : : to natural latch edges in the cfg. Reading the vector from the end
5809 : : to the beginning gives the reverse postorder. */
5810 : 553379 : auto_vec<int> initial_rpo;
5811 : 1106758 : graphds_dfs (m_slpg, &m_leafs[0], m_leafs.length (), &initial_rpo,
5812 : : false, NULL, skip_cfg_latch_edges);
5813 : 1660137 : gcc_assert (initial_rpo.length () == m_vertices.length ());
5814 : :
5815 : : /* Calculate the strongly connected components of the graph. */
5816 : 553379 : auto_vec<int> scc_grouping;
5817 : 553379 : unsigned int num_sccs = graphds_scc (m_slpg, NULL, NULL, &scc_grouping);
5818 : :
5819 : : /* Create a new index order in which all nodes from the same SCC are
5820 : : consecutive. Use scc_pos to record the index of the first node in
5821 : : each SCC. */
5822 : 553379 : auto_vec<unsigned int> scc_pos (num_sccs);
5823 : 553379 : int last_component = -1;
5824 : 553379 : unsigned int node_count = 0;
5825 : 5651305 : for (unsigned int node_i : scc_grouping)
5826 : : {
5827 : 3991168 : if (last_component != m_slpg->vertices[node_i].component)
5828 : : {
5829 : 3903040 : last_component = m_slpg->vertices[node_i].component;
5830 : 7806080 : gcc_assert (last_component == int (scc_pos.length ()));
5831 : 3903040 : scc_pos.quick_push (node_count);
5832 : : }
5833 : 3991168 : node_count += 1;
5834 : : }
5835 : 1106758 : gcc_assert (node_count == initial_rpo.length ()
5836 : : && last_component + 1 == int (num_sccs));
5837 : :
5838 : : /* Use m_partitioned_nodes to group nodes into SCC order, with the nodes
5839 : : inside each SCC following the RPO we calculated above. The fact that
5840 : : we ignored natural latch edges when calculating the RPO should ensure
5841 : : that, for natural loop nests:
5842 : :
5843 : : - the first node that we encounter in a cfg loop is the loop header phi
5844 : : - the loop header phis are in dominance order
5845 : :
5846 : : Arranging for this is an optimization (see below) rather than a
5847 : : correctness issue. Unnatural loops with a tangled mess of backedges
5848 : : will still work correctly, but might give poorer results.
5849 : :
5850 : : Also update scc_pos so that it gives 1 + the index of the last node
5851 : : in the SCC. */
5852 : 553379 : m_partitioned_nodes.safe_grow (node_count);
5853 : 5097926 : for (unsigned int old_i = initial_rpo.length (); old_i-- > 0;)
5854 : : {
5855 : 3991168 : unsigned int node_i = initial_rpo[old_i];
5856 : 3991168 : unsigned int new_i = scc_pos[m_slpg->vertices[node_i].component]++;
5857 : 3991168 : m_partitioned_nodes[new_i] = node_i;
5858 : : }
5859 : :
5860 : : /* When optimizing for speed, partition each SCC based on the containing
5861 : : cfg loop. The order we constructed above should ensure that, for natural
5862 : : cfg loops, we'll create sub-SCC partitions for outer loops before
5863 : : the corresponding sub-SCC partitions for inner loops. Similarly,
5864 : : when one sibling loop A dominates another sibling loop B, we should
5865 : : create a sub-SCC partition for A before a sub-SCC partition for B.
5866 : :
5867 : : As above, nothing depends for correctness on whether this achieves
5868 : : a natural nesting, but we should get better results when it does. */
5869 : 1106758 : m_partitions.reserve (m_vertices.length ());
5870 : 553379 : unsigned int next_partition_i = 0;
5871 : 553379 : hash_map<struct loop *, int> loop_partitions;
5872 : 553379 : unsigned int rpo_begin = 0;
5873 : 553379 : unsigned int num_partitioned_nodes = 0;
5874 : 5563177 : for (unsigned int rpo_end : scc_pos)
5875 : : {
5876 : 3903040 : loop_partitions.empty ();
5877 : : unsigned int partition_i = next_partition_i;
5878 : 7894208 : for (unsigned int rpo_i = rpo_begin; rpo_i < rpo_end; ++rpo_i)
5879 : : {
5880 : : /* Handle externals and constants optimistically throughout.
5881 : : But treat existing vectors as fixed since we do not handle
5882 : : permuting them. */
5883 : 3991168 : unsigned int node_i = m_partitioned_nodes[rpo_i];
5884 : 3991168 : auto &vertex = m_vertices[node_i];
5885 : 3991168 : if ((SLP_TREE_DEF_TYPE (vertex.node) == vect_external_def
5886 : 427500 : && !SLP_TREE_VEC_DEFS (vertex.node).exists ())
5887 : 3993485 : || SLP_TREE_DEF_TYPE (vertex.node) == vect_constant_def)
5888 : 1271871 : vertex.partition = -1;
5889 : : else
5890 : : {
5891 : 2719297 : bool existed;
5892 : 2719297 : if (m_optimize_size)
5893 : 28815 : existed = next_partition_i > partition_i;
5894 : : else
5895 : : {
5896 : 2690482 : struct loop *loop = containing_loop (vertex.node);
5897 : 2690482 : auto &entry = loop_partitions.get_or_insert (loop, &existed);
5898 : 2690482 : if (!existed)
5899 : 2603085 : entry = next_partition_i;
5900 : 2690482 : partition_i = entry;
5901 : : }
5902 : 2719297 : if (!existed)
5903 : : {
5904 : 2631804 : m_partitions.quick_push (slpg_partition_info ());
5905 : 2631804 : next_partition_i += 1;
5906 : : }
5907 : 2719297 : vertex.partition = partition_i;
5908 : 2719297 : num_partitioned_nodes += 1;
5909 : 2719297 : m_partitions[partition_i].node_end += 1;
5910 : : }
5911 : : }
5912 : 3903040 : rpo_begin = rpo_end;
5913 : : }
5914 : :
5915 : : /* Assign ranges of consecutive node indices to each partition,
5916 : : in partition order. Start with node_end being the same as
5917 : : node_begin so that the next loop can use it as a counter. */
5918 : 553379 : unsigned int node_begin = 0;
5919 : 4291941 : for (auto &partition : m_partitions)
5920 : : {
5921 : 2631804 : partition.node_begin = node_begin;
5922 : 2631804 : node_begin += partition.node_end;
5923 : 2631804 : partition.node_end = partition.node_begin;
5924 : : }
5925 : 553379 : gcc_assert (node_begin == num_partitioned_nodes);
5926 : :
5927 : : /* Finally build the list of nodes in partition order. */
5928 : 553379 : m_partitioned_nodes.truncate (num_partitioned_nodes);
5929 : 4544547 : for (unsigned int node_i = 0; node_i < m_vertices.length (); ++node_i)
5930 : : {
5931 : 3991168 : int partition_i = m_vertices[node_i].partition;
5932 : 3991168 : if (partition_i >= 0)
5933 : : {
5934 : 2719297 : unsigned int order_i = m_partitions[partition_i].node_end++;
5935 : 2719297 : m_partitioned_nodes[order_i] = node_i;
5936 : : }
5937 : : }
5938 : 553379 : }
5939 : :
5940 : : /* Look for edges from earlier partitions into node NODE_I and edges from
5941 : : node NODE_I into later partitions. Call:
5942 : :
5943 : : FN (ud, other_node_i)
5944 : :
5945 : : for each such use-to-def edge ud, where other_node_i is the node at the
5946 : : other end of the edge. */
5947 : :
5948 : : template<typename T>
5949 : : void
5950 : 3001941 : vect_optimize_slp_pass::for_each_partition_edge (unsigned int node_i, T fn)
5951 : : {
5952 : 3001941 : int partition_i = m_vertices[node_i].partition;
5953 : 3001941 : for (graph_edge *pred = m_slpg->vertices[node_i].pred;
5954 : 5057528 : pred; pred = pred->pred_next)
5955 : : {
5956 : 2055587 : int src_partition_i = m_vertices[pred->src].partition;
5957 : 2055587 : if (src_partition_i >= 0 && src_partition_i != partition_i)
5958 : 1880861 : fn (pred, pred->src);
5959 : : }
5960 : 3001941 : for (graph_edge *succ = m_slpg->vertices[node_i].succ;
5961 : 6442974 : succ; succ = succ->succ_next)
5962 : : {
5963 : 3441033 : int dest_partition_i = m_vertices[succ->dest].partition;
5964 : 3441033 : if (dest_partition_i >= 0 && dest_partition_i != partition_i)
5965 : 1887737 : fn (succ, succ->dest);
5966 : : }
5967 : 3001941 : }
5968 : :
5969 : : /* Return true if layout LAYOUT_I is compatible with the number of SLP lanes
5970 : : that NODE would operate on. This test is independent of NODE's actual
5971 : : operation. */
5972 : :
5973 : : bool
5974 : 1054970 : vect_optimize_slp_pass::is_compatible_layout (slp_tree node,
5975 : : unsigned int layout_i)
5976 : : {
5977 : 1054970 : if (layout_i == 0)
5978 : : return true;
5979 : :
5980 : 625094 : if (SLP_TREE_LANES (node) != m_perms[layout_i].length ())
5981 : 8379 : return false;
5982 : :
5983 : : return true;
5984 : : }
5985 : :
5986 : : /* Return the cost (in arbtirary units) of going from layout FROM_LAYOUT_I
5987 : : to layout TO_LAYOUT_I for a node like NODE. Return -1 if either of the
5988 : : layouts is incompatible with NODE or if the change is not possible for
5989 : : some other reason.
5990 : :
5991 : : The properties taken from NODE include the number of lanes and the
5992 : : vector type. The actual operation doesn't matter. */
5993 : :
5994 : : int
5995 : 452949 : vect_optimize_slp_pass::change_layout_cost (slp_tree node,
5996 : : unsigned int from_layout_i,
5997 : : unsigned int to_layout_i)
5998 : : {
5999 : 452949 : if (!is_compatible_layout (node, from_layout_i)
6000 : 452949 : || !is_compatible_layout (node, to_layout_i))
6001 : 662 : return -1;
6002 : :
6003 : 452287 : if (from_layout_i == to_layout_i)
6004 : : return 0;
6005 : :
6006 : 204561 : auto_vec<slp_tree, 1> children (1);
6007 : 204561 : children.quick_push (node);
6008 : 204561 : auto_lane_permutation_t perm (SLP_TREE_LANES (node));
6009 : 204561 : if (from_layout_i > 0)
6010 : 609188 : for (unsigned int i : m_perms[from_layout_i])
6011 : 279584 : perm.quick_push ({ 0, i });
6012 : : else
6013 : 319415 : for (unsigned int i = 0; i < SLP_TREE_LANES (node); ++i)
6014 : 224722 : perm.quick_push ({ 0, i });
6015 : 204561 : if (to_layout_i > 0)
6016 : 95124 : vect_slp_permute (m_perms[to_layout_i], perm, true);
6017 : 204561 : auto count = vectorizable_slp_permutation_1 (m_vinfo, nullptr, node, perm,
6018 : : children, false);
6019 : 204561 : if (count >= 0)
6020 : 200346 : return MAX (count, 1);
6021 : :
6022 : : /* ??? In principle we could try changing via layout 0, giving two
6023 : : layout changes rather than 1. Doing that would require
6024 : : corresponding support in get_result_with_layout. */
6025 : : return -1;
6026 : 204561 : }
6027 : :
6028 : : /* Return the costs of assigning layout LAYOUT_I to partition PARTITION_I. */
6029 : :
6030 : : inline slpg_partition_layout_costs &
6031 : 699428 : vect_optimize_slp_pass::partition_layout_costs (unsigned int partition_i,
6032 : : unsigned int layout_i)
6033 : : {
6034 : 1398856 : return m_partition_layout_costs[partition_i * m_perms.length () + layout_i];
6035 : : }
6036 : :
6037 : : /* Change PERM in one of two ways:
6038 : :
6039 : : - if IN_LAYOUT_I < 0, accept input operand I in the layout that has been
6040 : : chosen for child I of NODE.
6041 : :
6042 : : - if IN_LAYOUT >= 0, accept all inputs operands with that layout.
6043 : :
6044 : : In both cases, arrange for the output to have layout OUT_LAYOUT_I */
6045 : :
6046 : : void
6047 : 22588 : vect_optimize_slp_pass::
6048 : : change_vec_perm_layout (slp_tree node, lane_permutation_t &perm,
6049 : : int in_layout_i, unsigned int out_layout_i)
6050 : : {
6051 : 133380 : for (auto &entry : perm)
6052 : : {
6053 : 65616 : int this_in_layout_i = in_layout_i;
6054 : 65616 : if (this_in_layout_i < 0)
6055 : : {
6056 : 49455 : slp_tree in_node = SLP_TREE_CHILDREN (node)[entry.first];
6057 : 49455 : unsigned int in_partition_i = m_vertices[in_node->vertex].partition;
6058 : 49455 : if (in_partition_i == -1u)
6059 : 329 : continue;
6060 : 49126 : this_in_layout_i = m_partitions[in_partition_i].layout;
6061 : : }
6062 : 65287 : if (this_in_layout_i > 0)
6063 : 13347 : entry.second = m_perms[this_in_layout_i][entry.second];
6064 : : }
6065 : 22588 : if (out_layout_i > 0)
6066 : 4652 : vect_slp_permute (m_perms[out_layout_i], perm, true);
6067 : 22588 : }
6068 : :
6069 : : /* Check whether the target allows NODE to be rearranged so that the node's
6070 : : output has layout OUT_LAYOUT_I. Return the cost of the change if so,
6071 : : in the same arbitrary units as for change_layout_cost. Return -1 otherwise.
6072 : :
6073 : : If NODE is a VEC_PERM_EXPR and IN_LAYOUT_I < 0, also check whether
6074 : : NODE can adapt to the layout changes that have (perhaps provisionally)
6075 : : been chosen for NODE's children, so that no extra permutations are
6076 : : needed on either the input or the output of NODE.
6077 : :
6078 : : If NODE is a VEC_PERM_EXPR and IN_LAYOUT_I >= 0, instead assume
6079 : : that all inputs will be forced into layout IN_LAYOUT_I beforehand.
6080 : :
6081 : : IN_LAYOUT_I has no meaning for other types of node.
6082 : :
6083 : : Keeping the node as-is is always valid. If the target doesn't appear
6084 : : to support the node as-is, but might realistically support other layouts,
6085 : : then layout 0 instead has the cost of a worst-case permutation. On the
6086 : : one hand, this ensures that every node has at least one valid layout,
6087 : : avoiding what would otherwise be an awkward special case. On the other,
6088 : : it still encourages the pass to change an invalid pre-existing layout
6089 : : choice into a valid one. */
6090 : :
6091 : : int
6092 : 152884 : vect_optimize_slp_pass::internal_node_cost (slp_tree node, int in_layout_i,
6093 : : unsigned int out_layout_i)
6094 : : {
6095 : 152884 : const int fallback_cost = 1;
6096 : :
6097 : 152884 : if (SLP_TREE_CODE (node) == VEC_PERM_EXPR)
6098 : : {
6099 : 19816 : auto_lane_permutation_t tmp_perm;
6100 : 19816 : tmp_perm.safe_splice (SLP_TREE_LANE_PERMUTATION (node));
6101 : :
6102 : : /* Check that the child nodes support the chosen layout. Checking
6103 : : the first child is enough, since any second child would have the
6104 : : same shape. */
6105 : 19816 : auto first_child = SLP_TREE_CHILDREN (node)[0];
6106 : 19816 : if (in_layout_i > 0
6107 : 19816 : && !is_compatible_layout (first_child, in_layout_i))
6108 : : return -1;
6109 : :
6110 : 19157 : change_vec_perm_layout (node, tmp_perm, in_layout_i, out_layout_i);
6111 : 38314 : int count = vectorizable_slp_permutation_1 (m_vinfo, nullptr,
6112 : : node, tmp_perm,
6113 : 19157 : SLP_TREE_CHILDREN (node),
6114 : : false);
6115 : 19157 : if (count < 0)
6116 : : {
6117 : 1600 : if (in_layout_i == 0 && out_layout_i == 0)
6118 : : {
6119 : : /* Use the fallback cost if the node could in principle support
6120 : : some nonzero layout for both the inputs and the outputs.
6121 : : Otherwise assume that the node will be rejected later
6122 : : and rebuilt from scalars. */
6123 : 365 : if (SLP_TREE_LANES (node) == SLP_TREE_LANES (first_child))
6124 : : return fallback_cost;
6125 : 287 : return 0;
6126 : : }
6127 : : return -1;
6128 : : }
6129 : :
6130 : : /* We currently have no way of telling whether the new layout is cheaper
6131 : : or more expensive than the old one. But at least in principle,
6132 : : it should be worth making zero permutations (whole-vector shuffles)
6133 : : cheaper than real permutations, in case the pass is able to remove
6134 : : the latter. */
6135 : 17557 : return count == 0 ? 0 : 1;
6136 : 19816 : }
6137 : :
6138 : 133068 : stmt_vec_info rep = SLP_TREE_REPRESENTATIVE (node);
6139 : 133068 : if (rep
6140 : 132005 : && STMT_VINFO_DATA_REF (rep)
6141 : 51601 : && DR_IS_READ (STMT_VINFO_DATA_REF (rep))
6142 : 168664 : && SLP_TREE_LOAD_PERMUTATION (node).exists ())
6143 : : {
6144 : 31551 : auto_load_permutation_t tmp_perm;
6145 : 31551 : tmp_perm.safe_splice (SLP_TREE_LOAD_PERMUTATION (node));
6146 : 31551 : if (out_layout_i > 0)
6147 : 14343 : vect_slp_permute (m_perms[out_layout_i], tmp_perm, true);
6148 : :
6149 : 31551 : poly_uint64 vf = 1;
6150 : 31551 : if (auto loop_vinfo = dyn_cast<loop_vec_info> (m_vinfo))
6151 : 2197 : vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
6152 : 31551 : unsigned int n_perms;
6153 : 31551 : if (!vect_transform_slp_perm_load_1 (m_vinfo, node, tmp_perm, vNULL,
6154 : : nullptr, vf, true, false, &n_perms))
6155 : : {
6156 : 1987 : auto rep = SLP_TREE_REPRESENTATIVE (node);
6157 : 1987 : if (out_layout_i == 0)
6158 : : {
6159 : : /* Use the fallback cost if the load is an N-to-N permutation.
6160 : : Otherwise assume that the node will be rejected later
6161 : : and rebuilt from scalars. */
6162 : 1498 : if (STMT_VINFO_GROUPED_ACCESS (rep)
6163 : 2996 : && (DR_GROUP_SIZE (DR_GROUP_FIRST_ELEMENT (rep))
6164 : 1498 : == SLP_TREE_LANES (node)))
6165 : 915 : return fallback_cost;
6166 : : return 0;
6167 : : }
6168 : : return -1;
6169 : : }
6170 : :
6171 : : /* See the comment above the corresponding VEC_PERM_EXPR handling. */
6172 : 29564 : return n_perms == 0 ? 0 : 1;
6173 : 31551 : }
6174 : :
6175 : : return 0;
6176 : : }
6177 : :
6178 : : /* Decide which element layouts we should consider using. Calculate the
6179 : : weights associated with inserting layout changes on partition edges.
6180 : : Also mark partitions that cannot change layout, by setting their
6181 : : layout to zero. */
6182 : :
6183 : : void
6184 : 553379 : vect_optimize_slp_pass::start_choosing_layouts ()
6185 : : {
6186 : : /* Used to assign unique permutation indices. */
6187 : 553379 : using perm_hash = unbounded_hashmap_traits<
6188 : : vec_free_hash_base<int_hash_base<unsigned>>,
6189 : : int_hash<int, -1, -2>
6190 : : >;
6191 : 553379 : hash_map<vec<unsigned>, int, perm_hash> layout_ids;
6192 : :
6193 : : /* Layout 0 is "no change". */
6194 : 553379 : m_perms.safe_push (vNULL);
6195 : :
6196 : : /* Create layouts from existing permutations. */
6197 : 553379 : auto_load_permutation_t tmp_perm;
6198 : 4379434 : for (unsigned int node_i : m_partitioned_nodes)
6199 : : {
6200 : : /* Leafs also double as entries to the reverse graph. Allow the
6201 : : layout of those to be changed. */
6202 : 2719297 : auto &vertex = m_vertices[node_i];
6203 : 2719297 : auto &partition = m_partitions[vertex.partition];
6204 : 2719297 : if (!m_slpg->vertices[node_i].succ)
6205 : 602021 : partition.layout = 0;
6206 : :
6207 : : /* Loads and VEC_PERM_EXPRs are the only things generating permutes. */
6208 : 2719297 : slp_tree node = vertex.node;
6209 : 2719297 : stmt_vec_info dr_stmt = SLP_TREE_REPRESENTATIVE (node);
6210 : 2719297 : slp_tree child;
6211 : 2719297 : unsigned HOST_WIDE_INT imin, imax = 0;
6212 : 2719297 : bool any_permute = false;
6213 : 2719297 : tmp_perm.truncate (0);
6214 : 2719297 : if (SLP_TREE_LOAD_PERMUTATION (node).exists ())
6215 : : {
6216 : : /* If splitting out a SLP_TREE_LANE_PERMUTATION can make the node
6217 : : unpermuted, record a layout that reverses this permutation.
6218 : :
6219 : : We would need more work to cope with loads that are internally
6220 : : permuted and also have inputs (such as masks for
6221 : : IFN_MASK_LOADs). */
6222 : 302136 : gcc_assert (partition.layout == 0 && !m_slpg->vertices[node_i].succ);
6223 : 302136 : if (!STMT_VINFO_GROUPED_ACCESS (dr_stmt))
6224 : : {
6225 : 216351 : partition.layout = -1;
6226 : 2706141 : continue;
6227 : : }
6228 : 85785 : dr_stmt = DR_GROUP_FIRST_ELEMENT (dr_stmt);
6229 : 85785 : imin = DR_GROUP_SIZE (dr_stmt) + 1;
6230 : 85785 : tmp_perm.safe_splice (SLP_TREE_LOAD_PERMUTATION (node));
6231 : : }
6232 : 4723647 : else if (SLP_TREE_CODE (node) == VEC_PERM_EXPR
6233 : 125888 : && SLP_TREE_CHILDREN (node).length () == 1
6234 : 110675 : && (child = SLP_TREE_CHILDREN (node)[0])
6235 : 2527836 : && (TYPE_VECTOR_SUBPARTS (SLP_TREE_VECTYPE (child))
6236 : 110675 : .is_constant (&imin)))
6237 : : {
6238 : : /* If the child has the same vector size as this node,
6239 : : reversing the permutation can make the permutation a no-op.
6240 : : In other cases it can change a true permutation into a
6241 : : full-vector extract. */
6242 : 110675 : tmp_perm.reserve (SLP_TREE_LANES (node));
6243 : 319004 : for (unsigned j = 0; j < SLP_TREE_LANES (node); ++j)
6244 : 208329 : tmp_perm.quick_push (SLP_TREE_LANE_PERMUTATION (node)[j].second);
6245 : : }
6246 : : else
6247 : 2306486 : continue;
6248 : :
6249 : 581428 : for (unsigned j = 0; j < SLP_TREE_LANES (node); ++j)
6250 : : {
6251 : 384968 : unsigned idx = tmp_perm[j];
6252 : 384968 : imin = MIN (imin, idx);
6253 : 384968 : imax = MAX (imax, idx);
6254 : 384968 : if (idx - tmp_perm[0] != j)
6255 : 144815 : any_permute = true;
6256 : : }
6257 : : /* If the span doesn't match we'd disrupt VF computation, avoid
6258 : : that for now. */
6259 : 196460 : if (imax - imin + 1 != SLP_TREE_LANES (node))
6260 : 72565 : continue;
6261 : : /* If there's no permute no need to split one out. In this case
6262 : : we can consider turning a load into a permuted load, if that
6263 : : turns out to be cheaper than alternatives. */
6264 : 123895 : if (!any_permute)
6265 : : {
6266 : 110556 : partition.layout = -1;
6267 : 110556 : continue;
6268 : : }
6269 : :
6270 : : /* For now only handle true permutes, like
6271 : : vect_attempt_slp_rearrange_stmts did. This allows us to be lazy
6272 : : when permuting constants and invariants keeping the permute
6273 : : bijective. */
6274 : 13339 : auto_sbitmap load_index (SLP_TREE_LANES (node));
6275 : 13339 : bitmap_clear (load_index);
6276 : 54547 : for (unsigned j = 0; j < SLP_TREE_LANES (node); ++j)
6277 : 41208 : bitmap_set_bit (load_index, tmp_perm[j] - imin);
6278 : : unsigned j;
6279 : 53630 : for (j = 0; j < SLP_TREE_LANES (node); ++j)
6280 : 40474 : if (!bitmap_bit_p (load_index, j))
6281 : : break;
6282 : 13339 : if (j != SLP_TREE_LANES (node))
6283 : 183 : continue;
6284 : :
6285 : 13156 : vec<unsigned> perm = vNULL;
6286 : 13156 : perm.safe_grow (SLP_TREE_LANES (node), true);
6287 : 53320 : for (unsigned j = 0; j < SLP_TREE_LANES (node); ++j)
6288 : 40164 : perm[j] = tmp_perm[j] - imin;
6289 : :
6290 : 26312 : if (int (m_perms.length ()) >= param_vect_max_layout_candidates)
6291 : : {
6292 : : /* Continue to use existing layouts, but don't add any more. */
6293 : 0 : int *entry = layout_ids.get (perm);
6294 : 0 : partition.layout = entry ? *entry : 0;
6295 : 0 : perm.release ();
6296 : : }
6297 : : else
6298 : : {
6299 : 13156 : bool existed;
6300 : 13156 : int &layout_i = layout_ids.get_or_insert (perm, &existed);
6301 : 13156 : if (existed)
6302 : 3455 : perm.release ();
6303 : : else
6304 : : {
6305 : 9701 : layout_i = m_perms.length ();
6306 : 9701 : m_perms.safe_push (perm);
6307 : : }
6308 : 13156 : partition.layout = layout_i;
6309 : : }
6310 : 13339 : }
6311 : :
6312 : : /* Initially assume that every layout is possible and has zero cost
6313 : : in every partition. */
6314 : 553379 : m_partition_layout_costs.safe_grow_cleared (m_partitions.length ()
6315 : 1106758 : * m_perms.length ());
6316 : :
6317 : : /* We have to mark outgoing permutations facing non-associating-reduction
6318 : : graph entries that are not represented as to be materialized.
6319 : : slp_inst_kind_bb_reduc currently only covers associatable reductions. */
6320 : 2861597 : for (slp_instance instance : m_vinfo->slp_instances)
6321 : 1201460 : if (SLP_INSTANCE_KIND (instance) == slp_inst_kind_ctor)
6322 : : {
6323 : 1385 : unsigned int node_i = SLP_INSTANCE_TREE (instance)->vertex;
6324 : 1385 : m_partitions[m_vertices[node_i].partition].layout = 0;
6325 : : }
6326 : 1200075 : else if (SLP_INSTANCE_KIND (instance) == slp_inst_kind_reduc_chain)
6327 : : {
6328 : 343 : stmt_vec_info stmt_info
6329 : 343 : = SLP_TREE_REPRESENTATIVE (SLP_INSTANCE_TREE (instance));
6330 : 343 : stmt_vec_info reduc_info = info_for_reduction (m_vinfo, stmt_info);
6331 : 343 : if (needs_fold_left_reduction_p (TREE_TYPE
6332 : : (gimple_get_lhs (stmt_info->stmt)),
6333 : : STMT_VINFO_REDUC_CODE (reduc_info)))
6334 : : {
6335 : 71 : unsigned int node_i = SLP_INSTANCE_TREE (instance)->vertex;
6336 : 71 : m_partitions[m_vertices[node_i].partition].layout = 0;
6337 : : }
6338 : : }
6339 : :
6340 : : /* Check which layouts each node and partition can handle. Calculate the
6341 : : weights associated with inserting layout changes on edges. */
6342 : 4379434 : for (unsigned int node_i : m_partitioned_nodes)
6343 : : {
6344 : 2719297 : auto &vertex = m_vertices[node_i];
6345 : 2719297 : auto &partition = m_partitions[vertex.partition];
6346 : 2719297 : slp_tree node = vertex.node;
6347 : :
6348 : 2719297 : if (stmt_vec_info rep = SLP_TREE_REPRESENTATIVE (node))
6349 : : {
6350 : 2713958 : vertex.weight = vect_slp_node_weight (node);
6351 : :
6352 : : /* We do not handle stores with a permutation, so all
6353 : : incoming permutations must have been materialized.
6354 : :
6355 : : We also don't handle masked grouped loads, which lack a
6356 : : permutation vector. In this case the memory locations
6357 : : form an implicit second input to the loads, on top of the
6358 : : explicit mask input, and the memory input's layout cannot
6359 : : be changed.
6360 : :
6361 : : On the other hand, we do support permuting gather loads and
6362 : : masked gather loads, where each scalar load is independent
6363 : : of the others. This can be useful if the address/index input
6364 : : benefits from permutation. */
6365 : 2713958 : if (STMT_VINFO_DATA_REF (rep)
6366 : 1394842 : && STMT_VINFO_GROUPED_ACCESS (rep)
6367 : 3713705 : && !SLP_TREE_LOAD_PERMUTATION (node).exists ())
6368 : 913962 : partition.layout = 0;
6369 : :
6370 : : /* We cannot change the layout of an operation that is
6371 : : not independent on lanes. Note this is an explicit
6372 : : negative list since that's much shorter than the respective
6373 : : positive one but it's critical to keep maintaining it. */
6374 : 2713958 : if (is_gimple_call (STMT_VINFO_STMT (rep)))
6375 : 22639 : switch (gimple_call_combined_fn (STMT_VINFO_STMT (rep)))
6376 : : {
6377 : 438 : case CFN_COMPLEX_ADD_ROT90:
6378 : 438 : case CFN_COMPLEX_ADD_ROT270:
6379 : 438 : case CFN_COMPLEX_MUL:
6380 : 438 : case CFN_COMPLEX_MUL_CONJ:
6381 : 438 : case CFN_VEC_ADDSUB:
6382 : 438 : case CFN_VEC_FMADDSUB:
6383 : 438 : case CFN_VEC_FMSUBADD:
6384 : 438 : partition.layout = 0;
6385 : : default:;
6386 : : }
6387 : : }
6388 : :
6389 : 6009567 : auto process_edge = [&](graph_edge *ud, unsigned int other_node_i)
6390 : : {
6391 : 3290270 : auto &other_vertex = m_vertices[other_node_i];
6392 : :
6393 : : /* Count the number of edges from earlier partitions and the number
6394 : : of edges to later partitions. */
6395 : 3290270 : if (other_vertex.partition < vertex.partition)
6396 : 1645135 : partition.in_degree += 1;
6397 : : else
6398 : 1645135 : partition.out_degree += 1;
6399 : :
6400 : : /* If the current node uses the result of OTHER_NODE_I, accumulate
6401 : : the effects of that. */
6402 : 3290270 : if (ud->src == int (node_i))
6403 : : {
6404 : 1645135 : other_vertex.out_weight += vertex.weight;
6405 : 1645135 : other_vertex.out_degree += 1;
6406 : : }
6407 : 6009567 : };
6408 : 2719297 : for_each_partition_edge (node_i, process_edge);
6409 : : }
6410 : 553379 : }
6411 : :
6412 : : /* Return the incoming costs for node NODE_I, assuming that each input keeps
6413 : : its current (provisional) choice of layout. The inputs do not necessarily
6414 : : have the same layout as each other. */
6415 : :
6416 : : slpg_layout_cost
6417 : 3022 : vect_optimize_slp_pass::total_in_cost (unsigned int node_i)
6418 : : {
6419 : 3022 : auto &vertex = m_vertices[node_i];
6420 : 3022 : slpg_layout_cost cost;
6421 : 10816 : auto add_cost = [&](graph_edge *, unsigned int other_node_i)
6422 : : {
6423 : 7794 : auto &other_vertex = m_vertices[other_node_i];
6424 : 7794 : if (other_vertex.partition < vertex.partition)
6425 : : {
6426 : 4944 : auto &other_partition = m_partitions[other_vertex.partition];
6427 : 9888 : auto &other_costs = partition_layout_costs (other_vertex.partition,
6428 : 4944 : other_partition.layout);
6429 : 4944 : slpg_layout_cost this_cost = other_costs.in_cost;
6430 : 4944 : this_cost.add_serial_cost (other_costs.internal_cost);
6431 : 4944 : this_cost.split (other_partition.out_degree);
6432 : 4944 : cost.add_parallel_cost (this_cost);
6433 : : }
6434 : 10816 : };
6435 : 3022 : for_each_partition_edge (node_i, add_cost);
6436 : 3022 : return cost;
6437 : : }
6438 : :
6439 : : /* Return the cost of switching between layout LAYOUT1_I (at node NODE1_I)
6440 : : and layout LAYOUT2_I on cross-partition use-to-def edge UD. Return
6441 : : slpg_layout_cost::impossible () if the change isn't possible. */
6442 : :
6443 : : slpg_layout_cost
6444 : 452949 : vect_optimize_slp_pass::
6445 : : edge_layout_cost (graph_edge *ud, unsigned int node1_i, unsigned int layout1_i,
6446 : : unsigned int layout2_i)
6447 : : {
6448 : 452949 : auto &def_vertex = m_vertices[ud->dest];
6449 : 452949 : auto &use_vertex = m_vertices[ud->src];
6450 : 452949 : auto def_layout_i = ud->dest == int (node1_i) ? layout1_i : layout2_i;
6451 : 452949 : auto use_layout_i = ud->dest == int (node1_i) ? layout2_i : layout1_i;
6452 : 452949 : auto factor = change_layout_cost (def_vertex.node, def_layout_i,
6453 : : use_layout_i);
6454 : 452949 : if (factor < 0)
6455 : 4877 : return slpg_layout_cost::impossible ();
6456 : :
6457 : : /* We have a choice of putting the layout change at the site of the
6458 : : definition or at the site of the use. Prefer the former when
6459 : : optimizing for size or when the execution frequency of the
6460 : : definition is no greater than the combined execution frequencies of
6461 : : the uses. When putting the layout change at the site of the definition,
6462 : : divvy up the cost among all consumers. */
6463 : 448072 : if (m_optimize_size || def_vertex.weight <= def_vertex.out_weight)
6464 : : {
6465 : 433614 : slpg_layout_cost cost = { def_vertex.weight * factor, m_optimize_size };
6466 : 433614 : cost.split (def_vertex.out_degree);
6467 : 433614 : return cost;
6468 : : }
6469 : 14458 : return { use_vertex.weight * factor, m_optimize_size };
6470 : : }
6471 : :
6472 : : /* UD represents a use-def link between FROM_NODE_I and a node in a later
6473 : : partition; FROM_NODE_I could be the definition node or the use node.
6474 : : The node at the other end of the link wants to use layout TO_LAYOUT_I.
6475 : : Return the cost of any necessary fix-ups on edge UD, or return
6476 : : slpg_layout_cost::impossible () if the change isn't possible.
6477 : :
6478 : : At this point, FROM_NODE_I's partition has chosen the cheapest
6479 : : layout based on the information available so far, but this choice
6480 : : is only provisional. */
6481 : :
6482 : : slpg_layout_cost
6483 : 117799 : vect_optimize_slp_pass::forward_cost (graph_edge *ud, unsigned int from_node_i,
6484 : : unsigned int to_layout_i)
6485 : : {
6486 : 117799 : auto &from_vertex = m_vertices[from_node_i];
6487 : 117799 : unsigned int from_partition_i = from_vertex.partition;
6488 : 117799 : slpg_partition_info &from_partition = m_partitions[from_partition_i];
6489 : 117799 : gcc_assert (from_partition.layout >= 0);
6490 : :
6491 : : /* First calculate the cost on the assumption that FROM_PARTITION sticks
6492 : : with its current layout preference. */
6493 : 117799 : slpg_layout_cost cost = slpg_layout_cost::impossible ();
6494 : 117799 : auto edge_cost = edge_layout_cost (ud, from_node_i,
6495 : 117799 : from_partition.layout, to_layout_i);
6496 : 117799 : if (edge_cost.is_possible ())
6497 : : {
6498 : 230274 : auto &from_costs = partition_layout_costs (from_partition_i,
6499 : 115137 : from_partition.layout);
6500 : 115137 : cost = from_costs.in_cost;
6501 : 115137 : cost.add_serial_cost (from_costs.internal_cost);
6502 : 115137 : cost.split (from_partition.out_degree);
6503 : 115137 : cost.add_serial_cost (edge_cost);
6504 : : }
6505 : 2662 : else if (from_partition.layout == 0)
6506 : : /* We must allow the source partition to have layout 0 as a fallback,
6507 : : in case all other options turn out to be impossible. */
6508 : 2662 : return cost;
6509 : :
6510 : : /* Take the minimum of that cost and the cost that applies if
6511 : : FROM_PARTITION instead switches to TO_LAYOUT_I. */
6512 : 115137 : auto &direct_layout_costs = partition_layout_costs (from_partition_i,
6513 : : to_layout_i);
6514 : 115137 : if (direct_layout_costs.is_possible ())
6515 : : {
6516 : 107379 : slpg_layout_cost direct_cost = direct_layout_costs.in_cost;
6517 : 107379 : direct_cost.add_serial_cost (direct_layout_costs.internal_cost);
6518 : 107379 : direct_cost.split (from_partition.out_degree);
6519 : 107379 : if (!cost.is_possible ()
6520 : 107379 : || direct_cost.is_better_than (cost, m_optimize_size))
6521 : 33172 : cost = direct_cost;
6522 : : }
6523 : :
6524 : 115137 : return cost;
6525 : : }
6526 : :
6527 : : /* UD represents a use-def link between TO_NODE_I and a node in an earlier
6528 : : partition; TO_NODE_I could be the definition node or the use node.
6529 : : The node at the other end of the link wants to use layout FROM_LAYOUT_I;
6530 : : return the cost of any necessary fix-ups on edge UD, or
6531 : : slpg_layout_cost::impossible () if the choice cannot be made.
6532 : :
6533 : : At this point, TO_NODE_I's partition has a fixed choice of layout. */
6534 : :
6535 : : slpg_layout_cost
6536 : 112507 : vect_optimize_slp_pass::backward_cost (graph_edge *ud, unsigned int to_node_i,
6537 : : unsigned int from_layout_i)
6538 : : {
6539 : 112507 : auto &to_vertex = m_vertices[to_node_i];
6540 : 112507 : unsigned int to_partition_i = to_vertex.partition;
6541 : 112507 : slpg_partition_info &to_partition = m_partitions[to_partition_i];
6542 : 112507 : gcc_assert (to_partition.layout >= 0);
6543 : :
6544 : : /* If TO_NODE_I is a VEC_PERM_EXPR consumer, see whether it can be
6545 : : adjusted for this input having layout FROM_LAYOUT_I. Assume that
6546 : : any other inputs keep their current choice of layout. */
6547 : 112507 : auto &to_costs = partition_layout_costs (to_partition_i,
6548 : : to_partition.layout);
6549 : 112507 : if (ud->src == int (to_node_i)
6550 : 112363 : && SLP_TREE_CODE (to_vertex.node) == VEC_PERM_EXPR)
6551 : : {
6552 : 8633 : auto &from_partition = m_partitions[m_vertices[ud->dest].partition];
6553 : 8633 : auto old_layout = from_partition.layout;
6554 : 8633 : from_partition.layout = from_layout_i;
6555 : 17266 : int factor = internal_node_cost (to_vertex.node, -1,
6556 : 8633 : to_partition.layout);
6557 : 8633 : from_partition.layout = old_layout;
6558 : 8633 : if (factor >= 0)
6559 : : {
6560 : 7941 : slpg_layout_cost cost = to_costs.out_cost;
6561 : 15882 : cost.add_serial_cost ({ to_vertex.weight * factor,
6562 : 7941 : m_optimize_size });
6563 : 7941 : cost.split (to_partition.in_degree);
6564 : 7941 : return cost;
6565 : : }
6566 : : }
6567 : :
6568 : : /* Compute the cost if we insert any necessary layout change on edge UD. */
6569 : 104566 : auto edge_cost = edge_layout_cost (ud, to_node_i,
6570 : 104566 : to_partition.layout, from_layout_i);
6571 : 104566 : if (edge_cost.is_possible ())
6572 : : {
6573 : 104566 : slpg_layout_cost cost = to_costs.out_cost;
6574 : 104566 : cost.add_serial_cost (to_costs.internal_cost);
6575 : 104566 : cost.split (to_partition.in_degree);
6576 : 104566 : cost.add_serial_cost (edge_cost);
6577 : 104566 : return cost;
6578 : : }
6579 : :
6580 : 0 : return slpg_layout_cost::impossible ();
6581 : : }
6582 : :
6583 : : /* Make a forward pass through the partitions, accumulating input costs.
6584 : : Make a tentative (provisional) choice of layout for each partition,
6585 : : ensuring that this choice still allows later partitions to keep
6586 : : their original layout. */
6587 : :
6588 : : void
6589 : 9178 : vect_optimize_slp_pass::forward_pass ()
6590 : : {
6591 : 90945 : for (unsigned int partition_i = 0; partition_i < m_partitions.length ();
6592 : : ++partition_i)
6593 : : {
6594 : 81767 : auto &partition = m_partitions[partition_i];
6595 : :
6596 : : /* If the partition consists of a single VEC_PERM_EXPR, precompute
6597 : : the incoming cost that would apply if every predecessor partition
6598 : : keeps its current layout. This is used within the loop below. */
6599 : 81767 : slpg_layout_cost in_cost;
6600 : 81767 : slp_tree single_node = nullptr;
6601 : 81767 : if (partition.node_end == partition.node_begin + 1)
6602 : : {
6603 : 81288 : unsigned int node_i = m_partitioned_nodes[partition.node_begin];
6604 : 81288 : single_node = m_vertices[node_i].node;
6605 : 81288 : if (SLP_TREE_CODE (single_node) == VEC_PERM_EXPR)
6606 : 3022 : in_cost = total_in_cost (node_i);
6607 : : }
6608 : :
6609 : : /* Go through the possible layouts. Decide which ones are valid
6610 : : for this partition and record which of the valid layouts has
6611 : : the lowest cost. */
6612 : 81767 : unsigned int min_layout_i = 0;
6613 : 81767 : slpg_layout_cost min_layout_cost = slpg_layout_cost::impossible ();
6614 : 252923 : for (unsigned int layout_i = 0; layout_i < m_perms.length (); ++layout_i)
6615 : : {
6616 : 171156 : auto &layout_costs = partition_layout_costs (partition_i, layout_i);
6617 : 171156 : if (!layout_costs.is_possible ())
6618 : 37276 : continue;
6619 : :
6620 : : /* If the recorded layout is already 0 then the layout cannot
6621 : : change. */
6622 : 171156 : if (partition.layout == 0 && layout_i != 0)
6623 : : {
6624 : 26952 : layout_costs.mark_impossible ();
6625 : 26952 : continue;
6626 : : }
6627 : :
6628 : 144204 : bool is_possible = true;
6629 : 282052 : for (unsigned int order_i = partition.node_begin;
6630 : 282052 : order_i < partition.node_end; ++order_i)
6631 : : {
6632 : 146132 : unsigned int node_i = m_partitioned_nodes[order_i];
6633 : 146132 : auto &vertex = m_vertices[node_i];
6634 : :
6635 : : /* Reject the layout if it is individually incompatible
6636 : : with any node in the partition. */
6637 : 146132 : if (!is_compatible_layout (vertex.node, layout_i))
6638 : : {
6639 : 7058 : is_possible = false;
6640 : 8284 : break;
6641 : : }
6642 : :
6643 : 372420 : auto add_cost = [&](graph_edge *ud, unsigned int other_node_i)
6644 : : {
6645 : 233346 : auto &other_vertex = m_vertices[other_node_i];
6646 : 233346 : if (other_vertex.partition < vertex.partition)
6647 : : {
6648 : : /* Accumulate the incoming costs from earlier
6649 : : partitions, plus the cost of any layout changes
6650 : : on UD itself. */
6651 : 117799 : auto cost = forward_cost (ud, other_node_i, layout_i);
6652 : 117799 : if (!cost.is_possible ())
6653 : 2662 : is_possible = false;
6654 : : else
6655 : 115137 : layout_costs.in_cost.add_parallel_cost (cost);
6656 : : }
6657 : : else
6658 : : /* Reject the layout if it would make layout 0 impossible
6659 : : for later partitions. This amounts to testing that the
6660 : : target supports reversing the layout change on edges
6661 : : to later partitions.
6662 : :
6663 : : In principle, it might be possible to push a layout
6664 : : change all the way down a graph, so that it never
6665 : : needs to be reversed and so that the target doesn't
6666 : : need to support the reverse operation. But it would
6667 : : be awkward to bail out if we hit a partition that
6668 : : does not support the new layout, especially since
6669 : : we are not dealing with a lattice. */
6670 : 115547 : is_possible &= edge_layout_cost (ud, other_node_i, 0,
6671 : 115547 : layout_i).is_possible ();
6672 : 372420 : };
6673 : 139074 : for_each_partition_edge (node_i, add_cost);
6674 : :
6675 : : /* Accumulate the cost of using LAYOUT_I within NODE,
6676 : : both for the inputs and the outputs. */
6677 : 139074 : int factor = internal_node_cost (vertex.node, layout_i,
6678 : : layout_i);
6679 : 139074 : if (factor < 0)
6680 : : {
6681 : 1226 : is_possible = false;
6682 : 1226 : break;
6683 : : }
6684 : 137848 : else if (factor)
6685 : 21137 : layout_costs.internal_cost.add_serial_cost
6686 : 21137 : ({ vertex.weight * factor, m_optimize_size });
6687 : : }
6688 : 144204 : if (!is_possible)
6689 : : {
6690 : 10324 : layout_costs.mark_impossible ();
6691 : 10324 : continue;
6692 : : }
6693 : :
6694 : : /* Combine the incoming and partition-internal costs. */
6695 : 133880 : slpg_layout_cost combined_cost = layout_costs.in_cost;
6696 : 133880 : combined_cost.add_serial_cost (layout_costs.internal_cost);
6697 : :
6698 : : /* If this partition consists of a single VEC_PERM_EXPR, see
6699 : : if the VEC_PERM_EXPR can be changed to support output layout
6700 : : LAYOUT_I while keeping all the provisional choices of input
6701 : : layout. */
6702 : 133880 : if (single_node
6703 : 132959 : && SLP_TREE_CODE (single_node) == VEC_PERM_EXPR)
6704 : : {
6705 : 5177 : int factor = internal_node_cost (single_node, -1, layout_i);
6706 : 5177 : if (factor >= 0)
6707 : : {
6708 : 4712 : auto weight = m_vertices[single_node->vertex].weight;
6709 : 4712 : slpg_layout_cost internal_cost
6710 : 4712 : = { weight * factor, m_optimize_size };
6711 : :
6712 : 4712 : slpg_layout_cost alt_cost = in_cost;
6713 : 4712 : alt_cost.add_serial_cost (internal_cost);
6714 : 4712 : if (alt_cost.is_better_than (combined_cost, m_optimize_size))
6715 : : {
6716 : 1381 : combined_cost = alt_cost;
6717 : 1381 : layout_costs.in_cost = in_cost;
6718 : 1381 : layout_costs.internal_cost = internal_cost;
6719 : : }
6720 : : }
6721 : : }
6722 : :
6723 : : /* Record the layout with the lowest cost. Prefer layout 0 in
6724 : : the event of a tie between it and another layout. */
6725 : 133880 : if (!min_layout_cost.is_possible ()
6726 : 52113 : || combined_cost.is_better_than (min_layout_cost,
6727 : 52113 : m_optimize_size))
6728 : : {
6729 : 95611 : min_layout_i = layout_i;
6730 : 95611 : min_layout_cost = combined_cost;
6731 : : }
6732 : : }
6733 : :
6734 : : /* This loop's handling of earlier partitions should ensure that
6735 : : choosing the original layout for the current partition is no
6736 : : less valid than it was in the original graph, even with the
6737 : : provisional layout choices for those earlier partitions. */
6738 : 81767 : gcc_assert (min_layout_cost.is_possible ());
6739 : 81767 : partition.layout = min_layout_i;
6740 : : }
6741 : 9178 : }
6742 : :
6743 : : /* Make a backward pass through the partitions, accumulating output costs.
6744 : : Make a final choice of layout for each partition. */
6745 : :
6746 : : void
6747 : 9178 : vect_optimize_slp_pass::backward_pass ()
6748 : : {
6749 : 100123 : for (unsigned int partition_i = m_partitions.length (); partition_i-- > 0;)
6750 : : {
6751 : 81767 : auto &partition = m_partitions[partition_i];
6752 : :
6753 : 81767 : unsigned int min_layout_i = 0;
6754 : 81767 : slpg_layout_cost min_layout_cost = slpg_layout_cost::impossible ();
6755 : 252923 : for (unsigned int layout_i = 0; layout_i < m_perms.length (); ++layout_i)
6756 : : {
6757 : 171156 : auto &layout_costs = partition_layout_costs (partition_i, layout_i);
6758 : 171156 : if (!layout_costs.is_possible ())
6759 : 37276 : continue;
6760 : :
6761 : : /* Accumulate the costs from successor partitions. */
6762 : 133880 : bool is_possible = true;
6763 : 269659 : for (unsigned int order_i = partition.node_begin;
6764 : 269659 : order_i < partition.node_end; ++order_i)
6765 : : {
6766 : 135779 : unsigned int node_i = m_partitioned_nodes[order_i];
6767 : 135779 : auto &vertex = m_vertices[node_i];
6768 : 363323 : auto add_cost = [&](graph_edge *ud, unsigned int other_node_i)
6769 : : {
6770 : 227544 : auto &other_vertex = m_vertices[other_node_i];
6771 : 227544 : auto &other_partition = m_partitions[other_vertex.partition];
6772 : 227544 : if (other_vertex.partition > vertex.partition)
6773 : : {
6774 : : /* Accumulate the incoming costs from later
6775 : : partitions, plus the cost of any layout changes
6776 : : on UD itself. */
6777 : 112507 : auto cost = backward_cost (ud, other_node_i, layout_i);
6778 : 112507 : if (!cost.is_possible ())
6779 : 0 : is_possible = false;
6780 : : else
6781 : 112507 : layout_costs.out_cost.add_parallel_cost (cost);
6782 : : }
6783 : : else
6784 : : /* Make sure that earlier partitions can (if necessary
6785 : : or beneficial) keep the layout that they chose in
6786 : : the forward pass. This ensures that there is at
6787 : : least one valid choice of layout. */
6788 : 115037 : is_possible &= edge_layout_cost (ud, other_node_i,
6789 : 115037 : other_partition.layout,
6790 : 115037 : layout_i).is_possible ();
6791 : 363323 : };
6792 : 135779 : for_each_partition_edge (node_i, add_cost);
6793 : : }
6794 : 133880 : if (!is_possible)
6795 : : {
6796 : 0 : layout_costs.mark_impossible ();
6797 : 0 : continue;
6798 : : }
6799 : :
6800 : : /* Locally combine the costs from the forward and backward passes.
6801 : : (This combined cost is not passed on, since that would lead
6802 : : to double counting.) */
6803 : 133880 : slpg_layout_cost combined_cost = layout_costs.in_cost;
6804 : 133880 : combined_cost.add_serial_cost (layout_costs.internal_cost);
6805 : 133880 : combined_cost.add_serial_cost (layout_costs.out_cost);
6806 : :
6807 : : /* Record the layout with the lowest cost. Prefer layout 0 in
6808 : : the event of a tie between it and another layout. */
6809 : 133880 : if (!min_layout_cost.is_possible ()
6810 : 52113 : || combined_cost.is_better_than (min_layout_cost,
6811 : 52113 : m_optimize_size))
6812 : : {
6813 : 93184 : min_layout_i = layout_i;
6814 : 93184 : min_layout_cost = combined_cost;
6815 : : }
6816 : : }
6817 : :
6818 : 81767 : gcc_assert (min_layout_cost.is_possible ());
6819 : 81767 : partition.layout = min_layout_i;
6820 : : }
6821 : 9178 : }
6822 : :
6823 : : /* Return a node that applies layout TO_LAYOUT_I to the original form of NODE.
6824 : : NODE already has the layout that was selected for its partition. */
6825 : :
6826 : : slp_tree
6827 : 92351 : vect_optimize_slp_pass::get_result_with_layout (slp_tree node,
6828 : : unsigned int to_layout_i)
6829 : : {
6830 : 92351 : unsigned int result_i = node->vertex * m_perms.length () + to_layout_i;
6831 : 92351 : slp_tree result = m_node_layouts[result_i];
6832 : 92351 : if (result)
6833 : : return result;
6834 : :
6835 : 91976 : if (SLP_TREE_DEF_TYPE (node) == vect_constant_def
6836 : 91976 : || (SLP_TREE_DEF_TYPE (node) == vect_external_def
6837 : : /* We can't permute vector defs in place. */
6838 : 15759 : && SLP_TREE_VEC_DEFS (node).is_empty ()))
6839 : : {
6840 : : /* If the vector is uniform or unchanged, there's nothing to do. */
6841 : 31372 : if (to_layout_i == 0 || vect_slp_tree_uniform_p (node))
6842 : : result = node;
6843 : : else
6844 : : {
6845 : 1509 : auto scalar_ops = SLP_TREE_SCALAR_OPS (node).copy ();
6846 : 1509 : result = vect_create_new_slp_node (scalar_ops);
6847 : 1509 : vect_slp_permute (m_perms[to_layout_i], scalar_ops, true);
6848 : : }
6849 : : }
6850 : : else
6851 : : {
6852 : 60604 : unsigned int partition_i = m_vertices[node->vertex].partition;
6853 : 60604 : unsigned int from_layout_i = m_partitions[partition_i].layout;
6854 : 60604 : if (from_layout_i == to_layout_i)
6855 : 60318 : return node;
6856 : :
6857 : : /* If NODE is itself a VEC_PERM_EXPR, try to create a parallel
6858 : : permutation instead of a serial one. Leave the new permutation
6859 : : in TMP_PERM on success. */
6860 : 286 : auto_lane_permutation_t tmp_perm;
6861 : 286 : unsigned int num_inputs = 1;
6862 : 286 : if (SLP_TREE_CODE (node) == VEC_PERM_EXPR)
6863 : : {
6864 : 9 : tmp_perm.safe_splice (SLP_TREE_LANE_PERMUTATION (node));
6865 : 9 : if (from_layout_i != 0)
6866 : 9 : vect_slp_permute (m_perms[from_layout_i], tmp_perm, false);
6867 : 9 : if (to_layout_i != 0)
6868 : 4 : vect_slp_permute (m_perms[to_layout_i], tmp_perm, true);
6869 : 9 : if (vectorizable_slp_permutation_1 (m_vinfo, nullptr, node,
6870 : : tmp_perm,
6871 : 9 : SLP_TREE_CHILDREN (node),
6872 : : false) >= 0)
6873 : 9 : num_inputs = SLP_TREE_CHILDREN (node).length ();
6874 : : else
6875 : 0 : tmp_perm.truncate (0);
6876 : : }
6877 : :
6878 : 286 : if (dump_enabled_p ())
6879 : : {
6880 : 61 : if (tmp_perm.length () > 0)
6881 : 6 : dump_printf_loc (MSG_NOTE, vect_location,
6882 : : "duplicating permutation node %p with"
6883 : : " layout %d\n",
6884 : : (void *) node, to_layout_i);
6885 : : else
6886 : 55 : dump_printf_loc (MSG_NOTE, vect_location,
6887 : : "inserting permutation node in place of %p\n",
6888 : : (void *) node);
6889 : : }
6890 : :
6891 : 286 : unsigned int num_lanes = SLP_TREE_LANES (node);
6892 : 286 : result = vect_create_new_slp_node (num_inputs, VEC_PERM_EXPR);
6893 : 286 : if (SLP_TREE_SCALAR_STMTS (node).length ())
6894 : : {
6895 : 285 : auto &stmts = SLP_TREE_SCALAR_STMTS (result);
6896 : 285 : stmts.safe_splice (SLP_TREE_SCALAR_STMTS (node));
6897 : 285 : if (from_layout_i != 0)
6898 : 237 : vect_slp_permute (m_perms[from_layout_i], stmts, false);
6899 : 285 : if (to_layout_i != 0)
6900 : 52 : vect_slp_permute (m_perms[to_layout_i], stmts, true);
6901 : : }
6902 : 286 : SLP_TREE_REPRESENTATIVE (result) = SLP_TREE_REPRESENTATIVE (node);
6903 : 286 : SLP_TREE_LANES (result) = num_lanes;
6904 : 286 : SLP_TREE_VECTYPE (result) = SLP_TREE_VECTYPE (node);
6905 : 286 : result->vertex = -1;
6906 : :
6907 : 286 : auto &lane_perm = SLP_TREE_LANE_PERMUTATION (result);
6908 : 286 : if (tmp_perm.length ())
6909 : : {
6910 : 9 : lane_perm.safe_splice (tmp_perm);
6911 : 9 : SLP_TREE_CHILDREN (result).safe_splice (SLP_TREE_CHILDREN (node));
6912 : : }
6913 : : else
6914 : : {
6915 : 277 : lane_perm.create (num_lanes);
6916 : 889 : for (unsigned j = 0; j < num_lanes; ++j)
6917 : 612 : lane_perm.quick_push ({ 0, j });
6918 : 277 : if (from_layout_i != 0)
6919 : 228 : vect_slp_permute (m_perms[from_layout_i], lane_perm, false);
6920 : 277 : if (to_layout_i != 0)
6921 : 49 : vect_slp_permute (m_perms[to_layout_i], lane_perm, true);
6922 : 277 : SLP_TREE_CHILDREN (result).safe_push (node);
6923 : : }
6924 : 1148 : for (slp_tree child : SLP_TREE_CHILDREN (result))
6925 : 290 : child->refcnt++;
6926 : 286 : }
6927 : 31658 : m_node_layouts[result_i] = result;
6928 : 31658 : return result;
6929 : : }
6930 : :
6931 : : /* Apply the chosen vector layouts to the SLP graph. */
6932 : :
6933 : : void
6934 : 9178 : vect_optimize_slp_pass::materialize ()
6935 : : {
6936 : : /* We no longer need the costs, so avoid having two O(N * P) arrays
6937 : : live at the same time. */
6938 : 9178 : m_partition_layout_costs.release ();
6939 : 27534 : m_node_layouts.safe_grow_cleared (m_vertices.length () * m_perms.length ());
6940 : :
6941 : 18356 : auto_sbitmap fully_folded (m_vertices.length ());
6942 : 9178 : bitmap_clear (fully_folded);
6943 : 110287 : for (unsigned int node_i : m_partitioned_nodes)
6944 : : {
6945 : 82753 : auto &vertex = m_vertices[node_i];
6946 : 82753 : slp_tree node = vertex.node;
6947 : 82753 : int layout_i = m_partitions[vertex.partition].layout;
6948 : 82753 : gcc_assert (layout_i >= 0);
6949 : :
6950 : : /* Rearrange the scalar statements to match the chosen layout. */
6951 : 82753 : if (layout_i > 0)
6952 : 11625 : vect_slp_permute (m_perms[layout_i],
6953 : 11625 : SLP_TREE_SCALAR_STMTS (node), true);
6954 : :
6955 : : /* Update load and lane permutations. */
6956 : 82753 : if (SLP_TREE_CODE (node) == VEC_PERM_EXPR)
6957 : : {
6958 : : /* First try to absorb the input vector layouts. If that fails,
6959 : : force the inputs to have layout LAYOUT_I too. We checked that
6960 : : that was possible before deciding to use nonzero output layouts.
6961 : : (Note that at this stage we don't really have any guarantee that
6962 : : the target supports the original VEC_PERM_EXPR.) */
6963 : 3066 : auto &perm = SLP_TREE_LANE_PERMUTATION (node);
6964 : 3066 : auto_lane_permutation_t tmp_perm;
6965 : 3066 : tmp_perm.safe_splice (perm);
6966 : 3066 : change_vec_perm_layout (node, tmp_perm, -1, layout_i);
6967 : 3066 : if (vectorizable_slp_permutation_1 (m_vinfo, nullptr, node,
6968 : : tmp_perm,
6969 : 3066 : SLP_TREE_CHILDREN (node),
6970 : : false) >= 0)
6971 : : {
6972 : 2701 : if (dump_enabled_p ()
6973 : 3533 : && !std::equal (tmp_perm.begin (), tmp_perm.end (),
6974 : : perm.begin ()))
6975 : 54 : dump_printf_loc (MSG_NOTE, vect_location,
6976 : : "absorbing input layouts into %p\n",
6977 : : (void *) node);
6978 : 8103 : std::copy (tmp_perm.begin (), tmp_perm.end (), perm.begin ());
6979 : 2701 : bitmap_set_bit (fully_folded, node_i);
6980 : : }
6981 : : else
6982 : : {
6983 : : /* Not MSG_MISSED because it would make no sense to users. */
6984 : 365 : if (dump_enabled_p ())
6985 : 46 : dump_printf_loc (MSG_NOTE, vect_location,
6986 : : "failed to absorb input layouts into %p\n",
6987 : : (void *) node);
6988 : 365 : change_vec_perm_layout (nullptr, perm, layout_i, layout_i);
6989 : : }
6990 : 3066 : }
6991 : : else
6992 : : {
6993 : 79687 : gcc_assert (!SLP_TREE_LANE_PERMUTATION (node).exists ());
6994 : 79687 : auto &load_perm = SLP_TREE_LOAD_PERMUTATION (node);
6995 : 79687 : if (layout_i > 0)
6996 : : /* ??? When we handle non-bijective permutes the idea
6997 : : is that we can force the load-permutation to be
6998 : : { min, min + 1, min + 2, ... max }. But then the
6999 : : scalar defs might no longer match the lane content
7000 : : which means wrong-code with live lane vectorization.
7001 : : So we possibly have to have NULL entries for those. */
7002 : 11532 : vect_slp_permute (m_perms[layout_i], load_perm, true);
7003 : : }
7004 : : }
7005 : :
7006 : : /* Do this before any nodes disappear, since it involves a walk
7007 : : over the leaves. */
7008 : 9178 : remove_redundant_permutations ();
7009 : :
7010 : : /* Replace each child with a correctly laid-out version. */
7011 : 110287 : for (unsigned int node_i : m_partitioned_nodes)
7012 : : {
7013 : : /* Skip nodes that have already been handled above. */
7014 : 82753 : if (bitmap_bit_p (fully_folded, node_i))
7015 : 2701 : continue;
7016 : :
7017 : 80052 : auto &vertex = m_vertices[node_i];
7018 : 80052 : int in_layout_i = m_partitions[vertex.partition].layout;
7019 : 80052 : gcc_assert (in_layout_i >= 0);
7020 : :
7021 : : unsigned j;
7022 : : slp_tree child;
7023 : 233034 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (vertex.node), j, child)
7024 : : {
7025 : 92523 : if (!child)
7026 : 172 : continue;
7027 : :
7028 : 92351 : slp_tree new_child = get_result_with_layout (child, in_layout_i);
7029 : 92351 : if (new_child != child)
7030 : : {
7031 : 1893 : vect_free_slp_tree (child);
7032 : 1893 : SLP_TREE_CHILDREN (vertex.node)[j] = new_child;
7033 : 1893 : new_child->refcnt += 1;
7034 : : }
7035 : : }
7036 : : }
7037 : 9178 : }
7038 : :
7039 : : /* Elide load permutations that are not necessary. Such permutations might
7040 : : be pre-existing, rather than created by the layout optimizations. */
7041 : :
7042 : : void
7043 : 553379 : vect_optimize_slp_pass::remove_redundant_permutations ()
7044 : : {
7045 : 3601208 : for (unsigned int node_i : m_leafs)
7046 : : {
7047 : 1941071 : slp_tree node = m_vertices[node_i].node;
7048 : 1941071 : if (!SLP_TREE_LOAD_PERMUTATION (node).exists ())
7049 : 1638935 : continue;
7050 : :
7051 : : /* In basic block vectorization we allow any subchain of an interleaving
7052 : : chain.
7053 : : FORNOW: not in loop SLP because of realignment complications. */
7054 : 302136 : if (is_a <bb_vec_info> (m_vinfo))
7055 : : {
7056 : 142483 : bool subchain_p = true;
7057 : : stmt_vec_info next_load_info = NULL;
7058 : : stmt_vec_info load_info;
7059 : : unsigned j;
7060 : 142483 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), j, load_info)
7061 : : {
7062 : 113559 : if (j != 0
7063 : 113559 : && (next_load_info != load_info
7064 : 52151 : || ! load_info
7065 : 52151 : || DR_GROUP_GAP (load_info) != 1))
7066 : : {
7067 : : subchain_p = false;
7068 : : break;
7069 : : }
7070 : 96136 : next_load_info = DR_GROUP_NEXT_ELEMENT (load_info);
7071 : : }
7072 : 46347 : if (subchain_p)
7073 : : {
7074 : 28924 : SLP_TREE_LOAD_PERMUTATION (node).release ();
7075 : 28924 : continue;
7076 : : }
7077 : : }
7078 : : else
7079 : : {
7080 : 255789 : loop_vec_info loop_vinfo = as_a<loop_vec_info> (m_vinfo);
7081 : 255789 : stmt_vec_info load_info;
7082 : 255789 : bool this_load_permuted = false;
7083 : 255789 : unsigned j;
7084 : 757890 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), j, load_info)
7085 : 259628 : if (SLP_TREE_LOAD_PERMUTATION (node)[j] != j)
7086 : : {
7087 : : this_load_permuted = true;
7088 : : break;
7089 : : }
7090 : : /* When this isn't a grouped access we know it's single element
7091 : : and contiguous. */
7092 : 255789 : if (!STMT_VINFO_GROUPED_ACCESS (SLP_TREE_SCALAR_STMTS (node)[0]))
7093 : : {
7094 : 216351 : if (!this_load_permuted
7095 : 216351 : && (known_eq (LOOP_VINFO_VECT_FACTOR (loop_vinfo), 1U)
7096 : 215816 : || SLP_TREE_LANES (node) == 1))
7097 : 215816 : SLP_TREE_LOAD_PERMUTATION (node).release ();
7098 : 216351 : continue;
7099 : : }
7100 : 39438 : stmt_vec_info first_stmt_info
7101 : 39438 : = DR_GROUP_FIRST_ELEMENT (SLP_TREE_SCALAR_STMTS (node)[0]);
7102 : 39776 : if (!this_load_permuted
7103 : : /* The load requires permutation when unrolling exposes
7104 : : a gap either because the group is larger than the SLP
7105 : : group-size or because there is a gap between the groups. */
7106 : 39438 : && (known_eq (LOOP_VINFO_VECT_FACTOR (loop_vinfo), 1U)
7107 : 26421 : || ((SLP_TREE_LANES (node) == DR_GROUP_SIZE (first_stmt_info))
7108 : 102 : && DR_GROUP_GAP (first_stmt_info) == 0)))
7109 : : {
7110 : 338 : SLP_TREE_LOAD_PERMUTATION (node).release ();
7111 : 338 : continue;
7112 : : }
7113 : : }
7114 : : }
7115 : 553379 : }
7116 : :
7117 : : /* Print the partition graph and layout information to the dump file. */
7118 : :
7119 : : void
7120 : 607 : vect_optimize_slp_pass::dump ()
7121 : : {
7122 : 607 : dump_printf_loc (MSG_NOTE, vect_location,
7123 : : "SLP optimize permutations:\n");
7124 : 1230 : for (unsigned int layout_i = 1; layout_i < m_perms.length (); ++layout_i)
7125 : : {
7126 : 623 : dump_printf_loc (MSG_NOTE, vect_location, " %d: { ", layout_i);
7127 : 623 : const char *sep = "";
7128 : 5329 : for (unsigned int idx : m_perms[layout_i])
7129 : : {
7130 : 3460 : dump_printf (MSG_NOTE, "%s%d", sep, idx);
7131 : 3460 : sep = ", ";
7132 : : }
7133 : 623 : dump_printf (MSG_NOTE, " }\n");
7134 : : }
7135 : 607 : dump_printf_loc (MSG_NOTE, vect_location,
7136 : : "SLP optimize partitions:\n");
7137 : 5176 : for (unsigned int partition_i = 0; partition_i < m_partitions.length ();
7138 : : ++partition_i)
7139 : : {
7140 : 4569 : auto &partition = m_partitions[partition_i];
7141 : 4569 : dump_printf_loc (MSG_NOTE, vect_location, " -------------\n");
7142 : 4569 : dump_printf_loc (MSG_NOTE, vect_location,
7143 : : " partition %d (layout %d):\n",
7144 : : partition_i, partition.layout);
7145 : 4569 : dump_printf_loc (MSG_NOTE, vect_location, " nodes:\n");
7146 : 9338 : for (unsigned int order_i = partition.node_begin;
7147 : 9338 : order_i < partition.node_end; ++order_i)
7148 : : {
7149 : 4769 : auto &vertex = m_vertices[m_partitioned_nodes[order_i]];
7150 : 9538 : dump_printf_loc (MSG_NOTE, vect_location, " - %p:\n",
7151 : 4769 : (void *) vertex.node);
7152 : 4769 : dump_printf_loc (MSG_NOTE, vect_location,
7153 : : " weight: %f\n",
7154 : : vertex.weight.to_double ());
7155 : 4769 : if (vertex.out_degree)
7156 : 3754 : dump_printf_loc (MSG_NOTE, vect_location,
7157 : : " out weight: %f (degree %d)\n",
7158 : : vertex.out_weight.to_double (),
7159 : : vertex.out_degree);
7160 : 4769 : if (SLP_TREE_CODE (vertex.node) == VEC_PERM_EXPR)
7161 : 462 : dump_printf_loc (MSG_NOTE, vect_location,
7162 : : " op: VEC_PERM_EXPR\n");
7163 : 4307 : else if (auto rep = SLP_TREE_REPRESENTATIVE (vertex.node))
7164 : 4289 : dump_printf_loc (MSG_NOTE, vect_location,
7165 : : " op template: %G", rep->stmt);
7166 : : }
7167 : 4569 : dump_printf_loc (MSG_NOTE, vect_location, " edges:\n");
7168 : 9338 : for (unsigned int order_i = partition.node_begin;
7169 : 9338 : order_i < partition.node_end; ++order_i)
7170 : : {
7171 : 4769 : unsigned int node_i = m_partitioned_nodes[order_i];
7172 : 4769 : auto &vertex = m_vertices[node_i];
7173 : 14413 : auto print_edge = [&](graph_edge *, unsigned int other_node_i)
7174 : : {
7175 : 9644 : auto &other_vertex = m_vertices[other_node_i];
7176 : 9644 : if (other_vertex.partition < vertex.partition)
7177 : 4822 : dump_printf_loc (MSG_NOTE, vect_location,
7178 : : " - %p [%d] --> %p\n",
7179 : 4822 : (void *) other_vertex.node,
7180 : : other_vertex.partition,
7181 : 4822 : (void *) vertex.node);
7182 : : else
7183 : 4822 : dump_printf_loc (MSG_NOTE, vect_location,
7184 : : " - %p --> [%d] %p\n",
7185 : 4822 : (void *) vertex.node,
7186 : : other_vertex.partition,
7187 : 4822 : (void *) other_vertex.node);
7188 : 14413 : };
7189 : 4769 : for_each_partition_edge (node_i, print_edge);
7190 : : }
7191 : :
7192 : 13960 : for (unsigned int layout_i = 0; layout_i < m_perms.length (); ++layout_i)
7193 : : {
7194 : 9391 : auto &layout_costs = partition_layout_costs (partition_i, layout_i);
7195 : 9391 : if (layout_costs.is_possible ())
7196 : : {
7197 : 7570 : dump_printf_loc (MSG_NOTE, vect_location,
7198 : : " layout %d:%s\n", layout_i,
7199 : 7570 : partition.layout == int (layout_i)
7200 : : ? " (*)" : "");
7201 : 7570 : slpg_layout_cost combined_cost = layout_costs.in_cost;
7202 : 7570 : combined_cost.add_serial_cost (layout_costs.internal_cost);
7203 : 7570 : combined_cost.add_serial_cost (layout_costs.out_cost);
7204 : : #define TEMPLATE "{depth: %f, total: %f}"
7205 : 7570 : dump_printf_loc (MSG_NOTE, vect_location,
7206 : : " " TEMPLATE "\n",
7207 : : layout_costs.in_cost.depth.to_double (),
7208 : : layout_costs.in_cost.total.to_double ());
7209 : 7570 : dump_printf_loc (MSG_NOTE, vect_location,
7210 : : " + " TEMPLATE "\n",
7211 : : layout_costs.internal_cost.depth.to_double (),
7212 : : layout_costs.internal_cost.total.to_double ());
7213 : 7570 : dump_printf_loc (MSG_NOTE, vect_location,
7214 : : " + " TEMPLATE "\n",
7215 : : layout_costs.out_cost.depth.to_double (),
7216 : : layout_costs.out_cost.total.to_double ());
7217 : 7570 : dump_printf_loc (MSG_NOTE, vect_location,
7218 : : " = " TEMPLATE "\n",
7219 : : combined_cost.depth.to_double (),
7220 : : combined_cost.total.to_double ());
7221 : : #undef TEMPLATE
7222 : : }
7223 : : else
7224 : 1821 : dump_printf_loc (MSG_NOTE, vect_location,
7225 : : " layout %d: rejected\n", layout_i);
7226 : : }
7227 : : }
7228 : 607 : }
7229 : :
7230 : : /* Masked load lanes discovery. */
7231 : :
7232 : : void
7233 : 553379 : vect_optimize_slp_pass::decide_masked_load_lanes ()
7234 : : {
7235 : 5651586 : for (auto v : m_vertices)
7236 : : {
7237 : 3991449 : slp_tree node = v.node;
7238 : 3991449 : if (SLP_TREE_DEF_TYPE (node) != vect_internal_def
7239 : 2717243 : || SLP_TREE_CODE (node) == VEC_PERM_EXPR)
7240 : 1400375 : continue;
7241 : 2591074 : stmt_vec_info stmt_info = SLP_TREE_REPRESENTATIVE (node);
7242 : 1283562 : if (! STMT_VINFO_GROUPED_ACCESS (stmt_info)
7243 : : /* The mask has to be uniform. */
7244 : 888979 : || STMT_VINFO_SLP_VECT_ONLY (stmt_info)
7245 : 888833 : || ! is_a <gcall *> (STMT_VINFO_STMT (stmt_info))
7246 : 2591125 : || ! gimple_call_internal_p (STMT_VINFO_STMT (stmt_info),
7247 : : IFN_MASK_LOAD))
7248 : 2591071 : continue;
7249 : 3 : stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info);
7250 : 6 : if (STMT_VINFO_STRIDED_P (stmt_info)
7251 : 3 : || compare_step_with_zero (m_vinfo, stmt_info) <= 0
7252 : 3 : || vect_load_lanes_supported (SLP_TREE_VECTYPE (node),
7253 : 0 : DR_GROUP_SIZE (stmt_info),
7254 : : true) == IFN_LAST)
7255 : 3 : continue;
7256 : :
7257 : : /* Uniform masks need to be suitably represented. */
7258 : 0 : slp_tree mask = SLP_TREE_CHILDREN (node)[0];
7259 : 0 : if (SLP_TREE_CODE (mask) != VEC_PERM_EXPR
7260 : 0 : || SLP_TREE_CHILDREN (mask).length () != 1)
7261 : 0 : continue;
7262 : 0 : bool match = true;
7263 : 0 : for (auto perm : SLP_TREE_LANE_PERMUTATION (mask))
7264 : 0 : if (perm.first != 0 || perm.second != 0)
7265 : : {
7266 : : match = false;
7267 : : break;
7268 : : }
7269 : 0 : if (!match)
7270 : 0 : continue;
7271 : :
7272 : : /* Now see if the consumer side matches. */
7273 : 0 : for (graph_edge *pred = m_slpg->vertices[node->vertex].pred;
7274 : 0 : pred; pred = pred->pred_next)
7275 : : {
7276 : 0 : slp_tree pred_node = m_vertices[pred->src].node;
7277 : : /* All consumers should be a permute with a single outgoing lane. */
7278 : 0 : if (SLP_TREE_CODE (pred_node) != VEC_PERM_EXPR
7279 : 0 : || SLP_TREE_LANES (pred_node) != 1)
7280 : : {
7281 : : match = false;
7282 : : break;
7283 : : }
7284 : 0 : gcc_assert (SLP_TREE_CHILDREN (pred_node).length () == 1);
7285 : : }
7286 : 0 : if (!match)
7287 : 0 : continue;
7288 : : /* Now we can mark the nodes as to use load lanes. */
7289 : 0 : node->ldst_lanes = true;
7290 : 0 : for (graph_edge *pred = m_slpg->vertices[node->vertex].pred;
7291 : 0 : pred; pred = pred->pred_next)
7292 : 0 : m_vertices[pred->src].node->ldst_lanes = true;
7293 : : /* The catch is we have to massage the mask. We have arranged
7294 : : analyzed uniform masks to be represented by a splat VEC_PERM
7295 : : which we can now simply elide as we cannot easily re-do SLP
7296 : : discovery here. */
7297 : 0 : slp_tree new_mask = SLP_TREE_CHILDREN (mask)[0];
7298 : 0 : SLP_TREE_REF_COUNT (new_mask)++;
7299 : 0 : SLP_TREE_CHILDREN (node)[0] = new_mask;
7300 : 0 : vect_free_slp_tree (mask);
7301 : : }
7302 : 553379 : }
7303 : :
7304 : : /* Main entry point for the SLP graph optimization pass. */
7305 : :
7306 : : void
7307 : 553379 : vect_optimize_slp_pass::run ()
7308 : : {
7309 : 553379 : build_graph ();
7310 : 553379 : create_partitions ();
7311 : 553379 : start_choosing_layouts ();
7312 : 553379 : if (m_perms.length () > 1)
7313 : : {
7314 : 9178 : forward_pass ();
7315 : 9178 : backward_pass ();
7316 : 9178 : if (dump_enabled_p ())
7317 : 607 : dump ();
7318 : 9178 : materialize ();
7319 : 37235 : while (!m_perms.is_empty ())
7320 : 18879 : m_perms.pop ().release ();
7321 : : }
7322 : : else
7323 : 544201 : remove_redundant_permutations ();
7324 : 553379 : free_graph (m_slpg);
7325 : 553379 : build_graph ();
7326 : 553379 : decide_masked_load_lanes ();
7327 : 553379 : free_graph (m_slpg);
7328 : 553379 : }
7329 : :
7330 : : /* Apply CSE to NODE and its children using BST_MAP. */
7331 : :
7332 : : static void
7333 : 4285154 : vect_cse_slp_nodes (scalar_stmts_to_slp_tree_map_t *bst_map, slp_tree& node)
7334 : : {
7335 : 4285154 : bool put_p = false;
7336 : 4285154 : if (SLP_TREE_DEF_TYPE (node) == vect_internal_def
7337 : : /* Besides some VEC_PERM_EXPR, two-operator nodes also
7338 : : lack scalar stmts and thus CSE doesn't work via bst_map. Ideally
7339 : : we'd have sth that works for all internal and external nodes. */
7340 : 4285154 : && !SLP_TREE_SCALAR_STMTS (node).is_empty ())
7341 : : {
7342 : 2990775 : slp_tree *leader = bst_map->get (SLP_TREE_SCALAR_STMTS (node));
7343 : 2990775 : if (leader)
7344 : : {
7345 : : /* We've visited this node already. */
7346 : 295537 : if (!*leader || *leader == node)
7347 : : return;
7348 : :
7349 : 2111 : if (dump_enabled_p ())
7350 : 940 : dump_printf_loc (MSG_NOTE, vect_location,
7351 : : "re-using SLP tree %p for %p\n",
7352 : : (void *)*leader, (void *)node);
7353 : 2111 : vect_free_slp_tree (node);
7354 : 2111 : (*leader)->refcnt += 1;
7355 : 2111 : node = *leader;
7356 : 2111 : return;
7357 : : }
7358 : :
7359 : : /* Avoid creating a cycle by populating the map only after recursion. */
7360 : 2695238 : bst_map->put (SLP_TREE_SCALAR_STMTS (node).copy (), nullptr);
7361 : 2695238 : node->refcnt += 1;
7362 : 2695238 : put_p = true;
7363 : : /* And recurse. */
7364 : : }
7365 : :
7366 : 12006186 : for (slp_tree &child : SLP_TREE_CHILDREN (node))
7367 : 3469127 : if (child)
7368 : 3083694 : vect_cse_slp_nodes (bst_map, child);
7369 : :
7370 : : /* Now record the node for CSE in other siblings. */
7371 : 3989617 : if (put_p)
7372 : 2695238 : *bst_map->get (SLP_TREE_SCALAR_STMTS (node)) = node;
7373 : : }
7374 : :
7375 : : /* Optimize the SLP graph of VINFO. */
7376 : :
7377 : : void
7378 : 832804 : vect_optimize_slp (vec_info *vinfo)
7379 : : {
7380 : 832804 : if (vinfo->slp_instances.is_empty ())
7381 : : return;
7382 : 553379 : vect_optimize_slp_pass (vinfo).run ();
7383 : :
7384 : : /* Apply CSE again to nodes after permute optimization. */
7385 : 553379 : scalar_stmts_to_slp_tree_map_t *bst_map
7386 : 553379 : = new scalar_stmts_to_slp_tree_map_t ();
7387 : :
7388 : 2861597 : for (auto inst : vinfo->slp_instances)
7389 : 1201460 : vect_cse_slp_nodes (bst_map, SLP_INSTANCE_TREE (inst));
7390 : :
7391 : 553379 : release_scalar_stmts_to_slp_tree_map (bst_map);
7392 : : }
7393 : :
7394 : : /* Gather loads reachable from the individual SLP graph entries. */
7395 : :
7396 : : void
7397 : 832804 : vect_gather_slp_loads (vec_info *vinfo)
7398 : : {
7399 : 832804 : unsigned i;
7400 : 832804 : slp_instance instance;
7401 : 2034264 : FOR_EACH_VEC_ELT (vinfo->slp_instances, i, instance)
7402 : : {
7403 : 1201460 : hash_set<slp_tree> visited;
7404 : 1201460 : vect_gather_slp_loads (SLP_INSTANCE_LOADS (instance),
7405 : : SLP_INSTANCE_TREE (instance), visited);
7406 : 1201460 : }
7407 : 832804 : }
7408 : :
7409 : : /* For NODE update VF based on the number of lanes and the vector types
7410 : : used. */
7411 : :
7412 : : static void
7413 : 2736350 : vect_update_slp_vf_for_node (slp_tree node, poly_uint64 &vf,
7414 : : hash_set<slp_tree> &visited)
7415 : : {
7416 : 2736350 : if (!node || SLP_TREE_DEF_TYPE (node) != vect_internal_def)
7417 : 1095878 : return;
7418 : 1885458 : if (visited.add (node))
7419 : : return;
7420 : :
7421 : 6635801 : for (slp_tree child : SLP_TREE_CHILDREN (node))
7422 : 2297253 : vect_update_slp_vf_for_node (child, vf, visited);
7423 : :
7424 : : /* We do not visit SLP nodes for constants or externals - those neither
7425 : : have a vector type set yet (vectorizable_* does this) nor do they
7426 : : have max_nunits set. Instead we rely on internal nodes max_nunit
7427 : : to cover constant/external operands.
7428 : : Note that when we stop using fixed size vectors externs and constants
7429 : : shouldn't influence the (minimum) vectorization factor, instead
7430 : : vectorizable_* should honor the vectorization factor when trying to
7431 : : assign vector types to constants and externals and cause iteration
7432 : : to a higher vectorization factor when required. */
7433 : 1640472 : poly_uint64 node_vf
7434 : 1640472 : = calculate_unrolling_factor (node->max_nunits, SLP_TREE_LANES (node));
7435 : 1640472 : vf = force_common_multiple (vf, node_vf);
7436 : :
7437 : : /* For permute nodes that are fed from externs or constants we have to
7438 : : consider their number of lanes as well. Likewise for store-lanes. */
7439 : 1640472 : if (SLP_TREE_CODE (node) == VEC_PERM_EXPR
7440 : 1519695 : || node->ldst_lanes)
7441 : 495527 : for (slp_tree child : SLP_TREE_CHILDREN (node))
7442 : 133196 : if (SLP_TREE_DEF_TYPE (child) != vect_internal_def)
7443 : : {
7444 : 2759 : poly_uint64 child_vf
7445 : 2759 : = calculate_unrolling_factor (node->max_nunits,
7446 : : SLP_TREE_LANES (child));
7447 : 2759 : vf = force_common_multiple (vf, child_vf);
7448 : : }
7449 : : }
7450 : :
7451 : : /* For each possible SLP instance decide whether to SLP it and calculate overall
7452 : : unrolling factor needed to SLP the loop. Return TRUE if decided to SLP at
7453 : : least one instance. */
7454 : :
7455 : : bool
7456 : 241907 : vect_make_slp_decision (loop_vec_info loop_vinfo)
7457 : : {
7458 : 241907 : unsigned int i;
7459 : 241907 : poly_uint64 unrolling_factor = 1;
7460 : 241907 : const vec<slp_instance> &slp_instances
7461 : : = LOOP_VINFO_SLP_INSTANCES (loop_vinfo);
7462 : 241907 : slp_instance instance;
7463 : 241907 : int decided_to_slp = 0;
7464 : :
7465 : 241907 : DUMP_VECT_SCOPE ("vect_make_slp_decision");
7466 : :
7467 : 241907 : hash_set<slp_tree> visited;
7468 : 922911 : FOR_EACH_VEC_ELT (slp_instances, i, instance)
7469 : : {
7470 : : /* FORNOW: SLP if you can. */
7471 : : /* All unroll factors have the form:
7472 : :
7473 : : GET_MODE_SIZE (vinfo->vector_mode) * X
7474 : :
7475 : : for some rational X, so they must have a common multiple. */
7476 : 439097 : vect_update_slp_vf_for_node (SLP_INSTANCE_TREE (instance),
7477 : : unrolling_factor, visited);
7478 : :
7479 : : /* Mark all the stmts that belong to INSTANCE as PURE_SLP stmts. Later we
7480 : : call vect_detect_hybrid_slp () to find stmts that need hybrid SLP and
7481 : : loop-based vectorization. Such stmts will be marked as HYBRID. */
7482 : 439097 : vect_mark_slp_stmts (loop_vinfo, SLP_INSTANCE_TREE (instance));
7483 : 439097 : decided_to_slp++;
7484 : : }
7485 : :
7486 : 241907 : LOOP_VINFO_SLP_UNROLLING_FACTOR (loop_vinfo) = unrolling_factor;
7487 : :
7488 : 241907 : if (decided_to_slp && dump_enabled_p ())
7489 : : {
7490 : 19578 : dump_printf_loc (MSG_NOTE, vect_location,
7491 : : "Decided to SLP %d instances. Unrolling factor ",
7492 : : decided_to_slp);
7493 : 19578 : dump_dec (MSG_NOTE, unrolling_factor);
7494 : 19578 : dump_printf (MSG_NOTE, "\n");
7495 : : }
7496 : :
7497 : 241907 : return (decided_to_slp > 0);
7498 : 241907 : }
7499 : :
7500 : : /* Private data for vect_detect_hybrid_slp. */
7501 : : struct vdhs_data
7502 : : {
7503 : : loop_vec_info loop_vinfo;
7504 : : vec<stmt_vec_info> *worklist;
7505 : : };
7506 : :
7507 : : /* Walker for walk_gimple_op. */
7508 : :
7509 : : static tree
7510 : 57897 : vect_detect_hybrid_slp (tree *tp, int *, void *data)
7511 : : {
7512 : 57897 : walk_stmt_info *wi = (walk_stmt_info *)data;
7513 : 57897 : vdhs_data *dat = (vdhs_data *)wi->info;
7514 : :
7515 : 57897 : if (wi->is_lhs)
7516 : : return NULL_TREE;
7517 : :
7518 : 37634 : stmt_vec_info def_stmt_info = dat->loop_vinfo->lookup_def (*tp);
7519 : 37634 : if (!def_stmt_info)
7520 : : return NULL_TREE;
7521 : 16487 : def_stmt_info = vect_stmt_to_vectorize (def_stmt_info);
7522 : 16487 : if (PURE_SLP_STMT (def_stmt_info))
7523 : : {
7524 : 5164 : if (dump_enabled_p ())
7525 : 610 : dump_printf_loc (MSG_NOTE, vect_location, "marking hybrid: %G",
7526 : : def_stmt_info->stmt);
7527 : 5164 : STMT_SLP_TYPE (def_stmt_info) = hybrid;
7528 : 5164 : dat->worklist->safe_push (def_stmt_info);
7529 : : }
7530 : :
7531 : : return NULL_TREE;
7532 : : }
7533 : :
7534 : : /* Look if STMT_INFO is consumed by SLP indirectly and mark it pure_slp
7535 : : if so, otherwise pushing it to WORKLIST. */
7536 : :
7537 : : static void
7538 : 76330 : maybe_push_to_hybrid_worklist (vec_info *vinfo,
7539 : : vec<stmt_vec_info> &worklist,
7540 : : stmt_vec_info stmt_info)
7541 : : {
7542 : 76330 : if (dump_enabled_p ())
7543 : 3679 : dump_printf_loc (MSG_NOTE, vect_location,
7544 : : "Processing hybrid candidate : %G", stmt_info->stmt);
7545 : 76330 : stmt_vec_info orig_info = vect_orig_stmt (stmt_info);
7546 : 76330 : imm_use_iterator iter2;
7547 : 76330 : ssa_op_iter iter1;
7548 : 76330 : use_operand_p use_p;
7549 : 76330 : def_operand_p def_p;
7550 : 76330 : bool any_def = false;
7551 : 156050 : FOR_EACH_PHI_OR_STMT_DEF (def_p, orig_info->stmt, iter1, SSA_OP_DEF)
7552 : : {
7553 : 11213 : any_def = true;
7554 : 14904 : FOR_EACH_IMM_USE_FAST (use_p, iter2, DEF_FROM_PTR (def_p))
7555 : : {
7556 : 11514 : if (is_gimple_debug (USE_STMT (use_p)))
7557 : 193 : continue;
7558 : 11321 : stmt_vec_info use_info = vinfo->lookup_stmt (USE_STMT (use_p));
7559 : : /* An out-of loop use means this is a loop_vect sink. */
7560 : 11321 : if (!use_info)
7561 : : {
7562 : 1214 : if (dump_enabled_p ())
7563 : 168 : dump_printf_loc (MSG_NOTE, vect_location,
7564 : : "Found loop_vect sink: %G", stmt_info->stmt);
7565 : 1214 : worklist.safe_push (stmt_info);
7566 : 10741 : return;
7567 : : }
7568 : 12038 : else if (!STMT_SLP_TYPE (vect_stmt_to_vectorize (use_info)))
7569 : : {
7570 : 6609 : if (dump_enabled_p ())
7571 : 952 : dump_printf_loc (MSG_NOTE, vect_location,
7572 : : "Found loop_vect use: %G", use_info->stmt);
7573 : 6609 : worklist.safe_push (stmt_info);
7574 : 6609 : return;
7575 : : }
7576 : : }
7577 : : }
7578 : : /* No def means this is a loop_vect sink. Gimple conditionals also don't have a
7579 : : def but shouldn't be considered sinks. */
7580 : 68507 : if (!any_def && STMT_VINFO_DEF_TYPE (stmt_info) != vect_condition_def)
7581 : : {
7582 : 1704 : if (dump_enabled_p ())
7583 : 192 : dump_printf_loc (MSG_NOTE, vect_location,
7584 : : "Found loop_vect sink: %G", stmt_info->stmt);
7585 : 1704 : worklist.safe_push (stmt_info);
7586 : 1704 : return;
7587 : : }
7588 : 66803 : if (dump_enabled_p ())
7589 : 2367 : dump_printf_loc (MSG_NOTE, vect_location,
7590 : : "Marked SLP consumed stmt pure: %G", stmt_info->stmt);
7591 : 66803 : STMT_SLP_TYPE (stmt_info) = pure_slp;
7592 : : }
7593 : :
7594 : : /* Find stmts that must be both vectorized and SLPed. */
7595 : :
7596 : : void
7597 : 236902 : vect_detect_hybrid_slp (loop_vec_info loop_vinfo)
7598 : : {
7599 : 236902 : DUMP_VECT_SCOPE ("vect_detect_hybrid_slp");
7600 : :
7601 : : /* All stmts participating in SLP are marked pure_slp, all other
7602 : : stmts are loop_vect.
7603 : : First collect all loop_vect stmts into a worklist.
7604 : : SLP patterns cause not all original scalar stmts to appear in
7605 : : SLP_TREE_SCALAR_STMTS and thus not all of them are marked pure_slp.
7606 : : Rectify this here and do a backward walk over the IL only considering
7607 : : stmts as loop_vect when they are used by a loop_vect stmt and otherwise
7608 : : mark them as pure_slp. */
7609 : 236902 : auto_vec<stmt_vec_info> worklist;
7610 : 779408 : for (int i = LOOP_VINFO_LOOP (loop_vinfo)->num_nodes - 1; i >= 0; --i)
7611 : : {
7612 : 542506 : basic_block bb = LOOP_VINFO_BBS (loop_vinfo)[i];
7613 : 1192483 : for (gphi_iterator gsi = gsi_start_phis (bb); !gsi_end_p (gsi);
7614 : 649977 : gsi_next (&gsi))
7615 : : {
7616 : 649977 : gphi *phi = gsi.phi ();
7617 : 649977 : stmt_vec_info stmt_info = loop_vinfo->lookup_stmt (phi);
7618 : 649977 : if (!STMT_SLP_TYPE (stmt_info) && STMT_VINFO_RELEVANT (stmt_info))
7619 : 1826 : maybe_push_to_hybrid_worklist (loop_vinfo,
7620 : : worklist, stmt_info);
7621 : : }
7622 : 542506 : for (gimple_stmt_iterator gsi = gsi_last_bb (bb); !gsi_end_p (gsi);
7623 : 6738950 : gsi_prev (&gsi))
7624 : : {
7625 : 3098222 : gimple *stmt = gsi_stmt (gsi);
7626 : 3098222 : if (is_gimple_debug (stmt))
7627 : 644831 : continue;
7628 : 2453391 : stmt_vec_info stmt_info = loop_vinfo->lookup_stmt (stmt);
7629 : 2453391 : if (STMT_VINFO_IN_PATTERN_P (stmt_info))
7630 : : {
7631 : 206591 : for (gimple_stmt_iterator gsi2
7632 : 206591 : = gsi_start (STMT_VINFO_PATTERN_DEF_SEQ (stmt_info));
7633 : 433617 : !gsi_end_p (gsi2); gsi_next (&gsi2))
7634 : : {
7635 : 227026 : stmt_vec_info patt_info
7636 : 227026 : = loop_vinfo->lookup_stmt (gsi_stmt (gsi2));
7637 : 227026 : if (!STMT_SLP_TYPE (patt_info)
7638 : 83384 : && STMT_VINFO_RELEVANT (patt_info))
7639 : 2837 : maybe_push_to_hybrid_worklist (loop_vinfo,
7640 : : worklist, patt_info);
7641 : : }
7642 : 206591 : stmt_info = STMT_VINFO_RELATED_STMT (stmt_info);
7643 : : }
7644 : 2453391 : if (!STMT_SLP_TYPE (stmt_info) && STMT_VINFO_RELEVANT (stmt_info))
7645 : 71667 : maybe_push_to_hybrid_worklist (loop_vinfo,
7646 : : worklist, stmt_info);
7647 : : }
7648 : : }
7649 : :
7650 : : /* Now we have a worklist of non-SLP stmts, follow use->def chains and
7651 : : mark any SLP vectorized stmt as hybrid.
7652 : : ??? We're visiting def stmts N times (once for each non-SLP and
7653 : : once for each hybrid-SLP use). */
7654 : 236902 : walk_stmt_info wi;
7655 : 236902 : vdhs_data dat;
7656 : 236902 : dat.worklist = &worklist;
7657 : 236902 : dat.loop_vinfo = loop_vinfo;
7658 : 236902 : memset (&wi, 0, sizeof (wi));
7659 : 236902 : wi.info = (void *)&dat;
7660 : 253759 : while (!worklist.is_empty ())
7661 : : {
7662 : 14691 : stmt_vec_info stmt_info = worklist.pop ();
7663 : : /* Since SSA operands are not set up for pattern stmts we need
7664 : : to use walk_gimple_op. */
7665 : 14691 : wi.is_lhs = 0;
7666 : 14691 : walk_gimple_op (stmt_info->stmt, vect_detect_hybrid_slp, &wi);
7667 : : /* For gather/scatter make sure to walk the offset operand, that
7668 : : can be a scaling and conversion away. */
7669 : 14691 : gather_scatter_info gs_info;
7670 : 14691 : if (STMT_VINFO_GATHER_SCATTER_P (stmt_info)
7671 : 14691 : && vect_check_gather_scatter (stmt_info, loop_vinfo, &gs_info))
7672 : : {
7673 : 880 : int dummy;
7674 : 880 : vect_detect_hybrid_slp (&gs_info.offset, &dummy, &wi);
7675 : : }
7676 : : }
7677 : 236902 : }
7678 : :
7679 : :
7680 : : /* Initialize a bb_vec_info struct for the statements in BBS basic blocks. */
7681 : :
7682 : 2297833 : _bb_vec_info::_bb_vec_info (vec<basic_block> _bbs, vec_info_shared *shared)
7683 : : : vec_info (vec_info::bb, shared),
7684 : 2297833 : roots (vNULL)
7685 : : {
7686 : : /* The region we are operating on. bbs[0] is the entry, excluding
7687 : : its PHI nodes. In the future we might want to track an explicit
7688 : : entry edge to cover bbs[0] PHI nodes and have a region entry
7689 : : insert location. */
7690 : 2297833 : bbs = _bbs.address ();
7691 : 2297833 : nbbs = _bbs.length ();
7692 : :
7693 : 17109235 : for (unsigned i = 0; i < nbbs; ++i)
7694 : : {
7695 : 14811402 : if (i != 0)
7696 : 19476122 : for (gphi_iterator si = gsi_start_phis (bbs[i]); !gsi_end_p (si);
7697 : 6962553 : gsi_next (&si))
7698 : : {
7699 : 6962553 : gphi *phi = si.phi ();
7700 : 6962553 : gimple_set_uid (phi, 0);
7701 : 6962553 : add_stmt (phi);
7702 : : }
7703 : 29622804 : for (gimple_stmt_iterator gsi = gsi_start_bb (bbs[i]);
7704 : 116192481 : !gsi_end_p (gsi); gsi_next (&gsi))
7705 : : {
7706 : 101381079 : gimple *stmt = gsi_stmt (gsi);
7707 : 101381079 : gimple_set_uid (stmt, 0);
7708 : 101381079 : if (is_gimple_debug (stmt))
7709 : 59128200 : continue;
7710 : 42252879 : add_stmt (stmt);
7711 : : }
7712 : : }
7713 : 2297833 : }
7714 : :
7715 : :
7716 : : /* Free BB_VINFO struct, as well as all the stmt_vec_info structs of all the
7717 : : stmts in the basic block. */
7718 : :
7719 : 2297833 : _bb_vec_info::~_bb_vec_info ()
7720 : : {
7721 : : /* Reset region marker. */
7722 : 17109235 : for (unsigned i = 0; i < nbbs; ++i)
7723 : : {
7724 : 14811402 : if (i != 0)
7725 : 19491216 : for (gphi_iterator si = gsi_start_phis (bbs[i]); !gsi_end_p (si);
7726 : 6977647 : gsi_next (&si))
7727 : : {
7728 : 6977647 : gphi *phi = si.phi ();
7729 : 6977647 : gimple_set_uid (phi, -1);
7730 : : }
7731 : 29622804 : for (gimple_stmt_iterator gsi = gsi_start_bb (bbs[i]);
7732 : 116135093 : !gsi_end_p (gsi); gsi_next (&gsi))
7733 : : {
7734 : 101323691 : gimple *stmt = gsi_stmt (gsi);
7735 : 101323691 : gimple_set_uid (stmt, -1);
7736 : : }
7737 : : }
7738 : :
7739 : 3422793 : for (unsigned i = 0; i < roots.length (); ++i)
7740 : : {
7741 : 1124960 : roots[i].stmts.release ();
7742 : 1124960 : roots[i].roots.release ();
7743 : 1124960 : roots[i].remain.release ();
7744 : : }
7745 : 2297833 : roots.release ();
7746 : 2297833 : }
7747 : :
7748 : : /* Subroutine of vect_slp_analyze_node_operations. Handle the root of NODE,
7749 : : given then that child nodes have already been processed, and that
7750 : : their def types currently match their SLP node's def type. */
7751 : :
7752 : : static bool
7753 : 2212864 : vect_slp_analyze_node_operations_1 (vec_info *vinfo, slp_tree node,
7754 : : slp_instance node_instance,
7755 : : stmt_vector_for_cost *cost_vec)
7756 : : {
7757 : 2212864 : stmt_vec_info stmt_info = SLP_TREE_REPRESENTATIVE (node);
7758 : :
7759 : : /* Calculate the number of vector statements to be created for the scalar
7760 : : stmts in this node. It is the number of scalar elements in one scalar
7761 : : iteration (DR_GROUP_SIZE) multiplied by VF divided by the number of
7762 : : elements in a vector. For single-defuse-cycle, lane-reducing op, and
7763 : : PHI statement that starts reduction comprised of only lane-reducing ops,
7764 : : the number is more than effective vector statements actually required. */
7765 : 2212864 : SLP_TREE_NUMBER_OF_VEC_STMTS (node) = vect_get_num_copies (vinfo, node);
7766 : :
7767 : : /* Handle purely internal nodes. */
7768 : 2212864 : if (SLP_TREE_CODE (node) == VEC_PERM_EXPR)
7769 : : {
7770 : 109558 : if (!vectorizable_slp_permutation (vinfo, NULL, node, cost_vec))
7771 : : return false;
7772 : :
7773 : : stmt_vec_info slp_stmt_info;
7774 : : unsigned int i;
7775 : 281242 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, slp_stmt_info)
7776 : : {
7777 : 173003 : if (slp_stmt_info
7778 : 166920 : && STMT_VINFO_LIVE_P (slp_stmt_info)
7779 : 173147 : && !vectorizable_live_operation (vinfo, slp_stmt_info, node,
7780 : : node_instance, i,
7781 : : false, cost_vec))
7782 : : return false;
7783 : : }
7784 : : return true;
7785 : : }
7786 : :
7787 : 2103306 : bool dummy;
7788 : 2103306 : return vect_analyze_stmt (vinfo, stmt_info, &dummy,
7789 : : node, node_instance, cost_vec);
7790 : : }
7791 : :
7792 : : /* Verify if we can externalize a set of internal defs. */
7793 : :
7794 : : static bool
7795 : 401493 : vect_slp_can_convert_to_external (const vec<stmt_vec_info> &stmts)
7796 : : {
7797 : 401493 : basic_block bb = NULL;
7798 : 1975981 : for (stmt_vec_info stmt : stmts)
7799 : 893604 : if (!stmt)
7800 : : return false;
7801 : : /* Constant generation uses get_later_stmt which can only handle
7802 : : defs from the same BB. */
7803 : 893604 : else if (!bb)
7804 : 401493 : bb = gimple_bb (stmt->stmt);
7805 : 492111 : else if (gimple_bb (stmt->stmt) != bb)
7806 : : return false;
7807 : : return true;
7808 : : }
7809 : :
7810 : : /* Try to build NODE from scalars, returning true on success.
7811 : : NODE_INSTANCE is the SLP instance that contains NODE. */
7812 : :
7813 : : static bool
7814 : 355213 : vect_slp_convert_to_external (vec_info *vinfo, slp_tree node,
7815 : : slp_instance node_instance)
7816 : : {
7817 : 355213 : stmt_vec_info stmt_info;
7818 : 355213 : unsigned int i;
7819 : :
7820 : 355213 : if (!is_a <bb_vec_info> (vinfo)
7821 : 75593 : || node == SLP_INSTANCE_TREE (node_instance)
7822 : 23683 : || !SLP_TREE_SCALAR_STMTS (node).exists ()
7823 : 23629 : || vect_contains_pattern_stmt_p (SLP_TREE_SCALAR_STMTS (node))
7824 : : /* Force the mask use to be built from scalars instead. */
7825 : 21631 : || VECTOR_BOOLEAN_TYPE_P (SLP_TREE_VECTYPE (node))
7826 : 376709 : || !vect_slp_can_convert_to_external (SLP_TREE_SCALAR_STMTS (node)))
7827 : : return false;
7828 : :
7829 : 18984 : if (dump_enabled_p ())
7830 : 80 : dump_printf_loc (MSG_NOTE, vect_location,
7831 : : "Building vector operands of %p from scalars instead\n",
7832 : : (void *) node);
7833 : :
7834 : : /* Don't remove and free the child nodes here, since they could be
7835 : : referenced by other structures. The analysis and scheduling phases
7836 : : (need to) ignore child nodes of anything that isn't vect_internal_def. */
7837 : 18984 : unsigned int group_size = SLP_TREE_LANES (node);
7838 : 18984 : SLP_TREE_DEF_TYPE (node) = vect_external_def;
7839 : : /* Invariants get their vector type from the uses. */
7840 : 18984 : SLP_TREE_VECTYPE (node) = NULL_TREE;
7841 : 18984 : SLP_TREE_SCALAR_OPS (node).safe_grow (group_size, true);
7842 : 18984 : SLP_TREE_LOAD_PERMUTATION (node).release ();
7843 : 67307 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
7844 : : {
7845 : 48323 : tree lhs = gimple_get_lhs (vect_orig_stmt (stmt_info)->stmt);
7846 : 48323 : SLP_TREE_SCALAR_OPS (node)[i] = lhs;
7847 : : }
7848 : : return true;
7849 : : }
7850 : :
7851 : : /* Return true if all elements of the slice are the same. */
7852 : : bool
7853 : 464820 : vect_scalar_ops_slice::all_same_p () const
7854 : : {
7855 : 507052 : for (unsigned int i = 1; i < length; ++i)
7856 : 341775 : if (!operand_equal_p (op (0), op (i)))
7857 : : return false;
7858 : : return true;
7859 : : }
7860 : :
7861 : : hashval_t
7862 : 636222 : vect_scalar_ops_slice_hash::hash (const value_type &s)
7863 : : {
7864 : 636222 : hashval_t hash = 0;
7865 : 1997396 : for (unsigned i = 0; i < s.length; ++i)
7866 : 1361174 : hash = iterative_hash_expr (s.op (i), hash);
7867 : 636222 : return hash;
7868 : : }
7869 : :
7870 : : bool
7871 : 349988 : vect_scalar_ops_slice_hash::equal (const value_type &s1,
7872 : : const compare_type &s2)
7873 : : {
7874 : 349988 : if (s1.length != s2.length)
7875 : : return false;
7876 : 546158 : for (unsigned i = 0; i < s1.length; ++i)
7877 : 468206 : if (!operand_equal_p (s1.op (i), s2.op (i)))
7878 : : return false;
7879 : : return true;
7880 : : }
7881 : :
7882 : : /* Compute the prologue cost for invariant or constant operands represented
7883 : : by NODE. */
7884 : :
7885 : : static void
7886 : 933747 : vect_prologue_cost_for_slp (slp_tree node,
7887 : : stmt_vector_for_cost *cost_vec)
7888 : : {
7889 : : /* There's a special case of an existing vector, that costs nothing. */
7890 : 933747 : if (SLP_TREE_SCALAR_OPS (node).length () == 0
7891 : 933747 : && !SLP_TREE_VEC_DEFS (node).is_empty ())
7892 : 1755 : return;
7893 : : /* Without looking at the actual initializer a vector of
7894 : : constants can be implemented as load from the constant pool.
7895 : : When all elements are the same we can use a splat. */
7896 : 931992 : tree vectype = SLP_TREE_VECTYPE (node);
7897 : 931992 : unsigned group_size = SLP_TREE_SCALAR_OPS (node).length ();
7898 : 931992 : unsigned HOST_WIDE_INT const_nunits;
7899 : 931992 : unsigned nelt_limit;
7900 : 931992 : auto ops = &SLP_TREE_SCALAR_OPS (node);
7901 : 931992 : auto_vec<unsigned int> starts (SLP_TREE_NUMBER_OF_VEC_STMTS (node));
7902 : 931992 : if (TYPE_VECTOR_SUBPARTS (vectype).is_constant (&const_nunits)
7903 : 931992 : && ! multiple_p (const_nunits, group_size))
7904 : : {
7905 : 107267 : nelt_limit = const_nunits;
7906 : 107267 : hash_set<vect_scalar_ops_slice_hash> vector_ops;
7907 : 425023 : for (unsigned int i = 0; i < SLP_TREE_NUMBER_OF_VEC_STMTS (node); ++i)
7908 : 317756 : if (!vector_ops.add ({ ops, i * nelt_limit, nelt_limit }))
7909 : 239804 : starts.quick_push (i * nelt_limit);
7910 : 107267 : }
7911 : : else
7912 : : {
7913 : : /* If either the vector has variable length or the vectors
7914 : : are composed of repeated whole groups we only need to
7915 : : cost construction once. All vectors will be the same. */
7916 : 824725 : nelt_limit = group_size;
7917 : 824725 : starts.quick_push (0);
7918 : : }
7919 : : /* ??? We're just tracking whether vectors in a single node are the same.
7920 : : Ideally we'd do something more global. */
7921 : 931992 : bool passed = false;
7922 : 3860505 : for (unsigned int start : starts)
7923 : : {
7924 : 1064529 : vect_cost_for_stmt kind;
7925 : 1064529 : if (SLP_TREE_DEF_TYPE (node) == vect_constant_def)
7926 : : kind = vector_load;
7927 : 464820 : else if (vect_scalar_ops_slice { ops, start, nelt_limit }.all_same_p ())
7928 : : kind = scalar_to_vec;
7929 : : else
7930 : 299543 : kind = vec_construct;
7931 : : /* The target cost hook has no idea which part of the SLP node
7932 : : we are costing so avoid passing it down more than once. Pass
7933 : : it to the first vec_construct or scalar_to_vec part since for those
7934 : : the x86 backend tries to account for GPR to XMM register moves. */
7935 : 1064529 : record_stmt_cost (cost_vec, 1, kind,
7936 : 1064529 : (kind != vector_load && !passed) ? node : nullptr,
7937 : : vectype, 0, vect_prologue);
7938 : 1064529 : if (kind != vector_load)
7939 : 464820 : passed = true;
7940 : : }
7941 : 931992 : }
7942 : :
7943 : : /* Analyze statements contained in SLP tree NODE after recursively analyzing
7944 : : the subtree. NODE_INSTANCE contains NODE and VINFO contains INSTANCE.
7945 : :
7946 : : Return true if the operations are supported. */
7947 : :
7948 : : static bool
7949 : 4016344 : vect_slp_analyze_node_operations (vec_info *vinfo, slp_tree node,
7950 : : slp_instance node_instance,
7951 : : hash_set<slp_tree> &visited_set,
7952 : : vec<slp_tree> &visited_vec,
7953 : : stmt_vector_for_cost *cost_vec)
7954 : : {
7955 : 4016344 : int i, j;
7956 : 4016344 : slp_tree child;
7957 : :
7958 : : /* Assume we can code-generate all invariants. */
7959 : 4016344 : if (!node
7960 : 3792435 : || SLP_TREE_DEF_TYPE (node) == vect_constant_def
7961 : 3127662 : || SLP_TREE_DEF_TYPE (node) == vect_external_def)
7962 : : return true;
7963 : :
7964 : 2666938 : if (SLP_TREE_DEF_TYPE (node) == vect_uninitialized_def)
7965 : : {
7966 : 18 : if (dump_enabled_p ())
7967 : 0 : dump_printf_loc (MSG_NOTE, vect_location,
7968 : : "Failed cyclic SLP reference in %p\n", (void *) node);
7969 : 18 : return false;
7970 : : }
7971 : 2666920 : gcc_assert (SLP_TREE_DEF_TYPE (node) == vect_internal_def);
7972 : :
7973 : : /* If we already analyzed the exact same set of scalar stmts we're done.
7974 : : We share the generated vector stmts for those. */
7975 : 2666920 : if (visited_set.add (node))
7976 : : return true;
7977 : 2419565 : visited_vec.safe_push (node);
7978 : :
7979 : 2419565 : bool res = true;
7980 : 2419565 : unsigned visited_rec_start = visited_vec.length ();
7981 : 2419565 : unsigned cost_vec_rec_start = cost_vec->length ();
7982 : 2419565 : bool seen_non_constant_child = false;
7983 : 5179516 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
7984 : : {
7985 : 2966501 : res = vect_slp_analyze_node_operations (vinfo, child, node_instance,
7986 : : visited_set, visited_vec,
7987 : : cost_vec);
7988 : 2966501 : if (!res)
7989 : : break;
7990 : 2759951 : if (child && SLP_TREE_DEF_TYPE (child) != vect_constant_def)
7991 : 2759951 : seen_non_constant_child = true;
7992 : : }
7993 : : /* We're having difficulties scheduling nodes with just constant
7994 : : operands and no scalar stmts since we then cannot compute a stmt
7995 : : insertion place. */
7996 : 2419565 : if (res
7997 : 2419565 : && !seen_non_constant_child
7998 : 2419565 : && SLP_TREE_SCALAR_STMTS (node).is_empty ())
7999 : : {
8000 : 151 : if (dump_enabled_p ())
8001 : 6 : dump_printf_loc (MSG_NOTE, vect_location,
8002 : : "Cannot vectorize all-constant op node %p\n",
8003 : : (void *) node);
8004 : : res = false;
8005 : : }
8006 : :
8007 : 2419414 : if (res)
8008 : 2212864 : res = vect_slp_analyze_node_operations_1 (vinfo, node, node_instance,
8009 : : cost_vec);
8010 : : /* If analysis failed we have to pop all recursive visited nodes
8011 : : plus ourselves. */
8012 : 2419565 : if (!res)
8013 : : {
8014 : 1979392 : while (visited_vec.length () >= visited_rec_start)
8015 : 634483 : visited_set.remove (visited_vec.pop ());
8016 : 355213 : cost_vec->truncate (cost_vec_rec_start);
8017 : : }
8018 : :
8019 : : /* When the node can be vectorized cost invariant nodes it references.
8020 : : This is not done in DFS order to allow the refering node
8021 : : vectorizable_* calls to nail down the invariant nodes vector type
8022 : : and possibly unshare it if it needs a different vector type than
8023 : : other referrers. */
8024 : 2419565 : if (res)
8025 : 4541098 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), j, child)
8026 : 2476746 : if (child
8027 : 2282754 : && (SLP_TREE_DEF_TYPE (child) == vect_constant_def
8028 : 2282754 : || SLP_TREE_DEF_TYPE (child) == vect_external_def)
8029 : : /* Perform usual caching, note code-generation still
8030 : : code-gens these nodes multiple times but we expect
8031 : : to CSE them later. */
8032 : 3468880 : && !visited_set.add (child))
8033 : : {
8034 : 967963 : visited_vec.safe_push (child);
8035 : : /* ??? After auditing more code paths make a "default"
8036 : : and push the vector type from NODE to all children
8037 : : if it is not already set. */
8038 : : /* Compute the number of vectors to be generated. */
8039 : 967963 : tree vector_type = SLP_TREE_VECTYPE (child);
8040 : 967963 : if (!vector_type)
8041 : : {
8042 : : /* Masked loads can have an undefined (default SSA definition)
8043 : : else operand. We do not need to cost it. */
8044 : 34216 : vec<tree> ops = SLP_TREE_SCALAR_OPS (child);
8045 : 35387 : if ((STMT_VINFO_TYPE (SLP_TREE_REPRESENTATIVE (node))
8046 : : == load_vec_info_type)
8047 : 35387 : && ((ops.length ()
8048 : 1171 : && TREE_CODE (ops[0]) == SSA_NAME
8049 : 0 : && SSA_NAME_IS_DEFAULT_DEF (ops[0])
8050 : 0 : && VAR_P (SSA_NAME_VAR (ops[0])))
8051 : 1171 : || SLP_TREE_DEF_TYPE (child) == vect_constant_def))
8052 : 1171 : continue;
8053 : :
8054 : : /* For shifts with a scalar argument we don't need
8055 : : to cost or code-generate anything.
8056 : : ??? Represent this more explicitely. */
8057 : 33045 : gcc_assert ((STMT_VINFO_TYPE (SLP_TREE_REPRESENTATIVE (node))
8058 : : == shift_vec_info_type)
8059 : : && j == 1);
8060 : 33045 : continue;
8061 : 33045 : }
8062 : :
8063 : 933747 : SLP_TREE_NUMBER_OF_VEC_STMTS (child)
8064 : 933747 : = vect_get_num_copies (vinfo, child);
8065 : : /* And cost them. */
8066 : 933747 : vect_prologue_cost_for_slp (child, cost_vec);
8067 : : }
8068 : :
8069 : : /* If this node or any of its children can't be vectorized, try pruning
8070 : : the tree here rather than felling the whole thing. */
8071 : 355213 : if (!res && vect_slp_convert_to_external (vinfo, node, node_instance))
8072 : : {
8073 : : /* We'll need to revisit this for invariant costing and number
8074 : : of vectorized stmt setting. */
8075 : : res = true;
8076 : : }
8077 : :
8078 : : return res;
8079 : : }
8080 : :
8081 : : /* Given a definition DEF, analyze if it will have any live scalar use after
8082 : : performing SLP vectorization whose information is represented by BB_VINFO,
8083 : : and record result into hash map SCALAR_USE_MAP as cache for later fast
8084 : : check. If recursion DEPTH exceeds a limit, stop analysis and make a
8085 : : conservative assumption. Return 0 if no scalar use, 1 if there is, -1
8086 : : means recursion is limited. */
8087 : :
8088 : : static int
8089 : 541044 : vec_slp_has_scalar_use (bb_vec_info bb_vinfo, tree def,
8090 : : hash_map<tree, int> &scalar_use_map,
8091 : : int depth = 0)
8092 : : {
8093 : 541044 : const int depth_limit = 2;
8094 : 541044 : imm_use_iterator use_iter;
8095 : 541044 : gimple *use_stmt;
8096 : :
8097 : 541044 : if (int *res = scalar_use_map.get (def))
8098 : 15244 : return *res;
8099 : :
8100 : 525800 : int scalar_use = 1;
8101 : :
8102 : 1212483 : FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, def)
8103 : : {
8104 : 793489 : if (is_gimple_debug (use_stmt))
8105 : 196570 : continue;
8106 : :
8107 : 596919 : stmt_vec_info use_stmt_info = bb_vinfo->lookup_stmt (use_stmt);
8108 : :
8109 : 596919 : if (!use_stmt_info)
8110 : : break;
8111 : :
8112 : 598327 : if (PURE_SLP_STMT (vect_stmt_to_vectorize (use_stmt_info)))
8113 : 488399 : continue;
8114 : :
8115 : : /* Do not step forward when encounter PHI statement, since it may
8116 : : involve cyclic reference and cause infinite recursive invocation. */
8117 : 102365 : if (gimple_code (use_stmt) == GIMPLE_PHI)
8118 : : break;
8119 : :
8120 : : /* When pattern recognition is involved, a statement whose definition is
8121 : : consumed in some pattern, may not be included in the final replacement
8122 : : pattern statements, so would be skipped when building SLP graph.
8123 : :
8124 : : * Original
8125 : : char a_c = *(char *) a;
8126 : : char b_c = *(char *) b;
8127 : : unsigned short a_s = (unsigned short) a_c;
8128 : : int a_i = (int) a_s;
8129 : : int b_i = (int) b_c;
8130 : : int r_i = a_i - b_i;
8131 : :
8132 : : * After pattern replacement
8133 : : a_s = (unsigned short) a_c;
8134 : : a_i = (int) a_s;
8135 : :
8136 : : patt_b_s = (unsigned short) b_c; // b_i = (int) b_c
8137 : : patt_b_i = (int) patt_b_s; // b_i = (int) b_c
8138 : :
8139 : : patt_r_s = widen_minus(a_c, b_c); // r_i = a_i - b_i
8140 : : patt_r_i = (int) patt_r_s; // r_i = a_i - b_i
8141 : :
8142 : : The definitions of a_i(original statement) and b_i(pattern statement)
8143 : : are related to, but actually not part of widen_minus pattern.
8144 : : Vectorizing the pattern does not cause these definition statements to
8145 : : be marked as PURE_SLP. For this case, we need to recursively check
8146 : : whether their uses are all absorbed into vectorized code. But there
8147 : : is an exception that some use may participate in an vectorized
8148 : : operation via an external SLP node containing that use as an element.
8149 : : The parameter "scalar_use_map" tags such kind of SSA as having scalar
8150 : : use in advance. */
8151 : 82758 : tree lhs = gimple_get_lhs (use_stmt);
8152 : :
8153 : 82758 : if (!lhs || TREE_CODE (lhs) != SSA_NAME)
8154 : : break;
8155 : :
8156 : 49935 : if (depth_limit && depth >= depth_limit)
8157 : 9244 : return -1;
8158 : :
8159 : 40691 : if ((scalar_use = vec_slp_has_scalar_use (bb_vinfo, lhs, scalar_use_map,
8160 : : depth + 1)))
8161 : : break;
8162 : 525800 : }
8163 : :
8164 : 516556 : if (end_imm_use_stmt_p (&use_iter))
8165 : 418994 : scalar_use = 0;
8166 : :
8167 : : /* If recursion is limited, do not cache result for non-root defs. */
8168 : 516556 : if (!depth || scalar_use >= 0)
8169 : : {
8170 : 507312 : bool added = scalar_use_map.put (def, scalar_use);
8171 : 507312 : gcc_assert (!added);
8172 : : }
8173 : :
8174 : 516556 : return scalar_use;
8175 : : }
8176 : :
8177 : : /* Mark lanes of NODE that are live outside of the basic-block vectorized
8178 : : region and that can be vectorized using vectorizable_live_operation
8179 : : with STMT_VINFO_LIVE_P. Not handled live operations will cause the
8180 : : scalar code computing it to be retained. */
8181 : :
8182 : : static void
8183 : 883827 : vect_bb_slp_mark_live_stmts (bb_vec_info bb_vinfo, slp_tree node,
8184 : : slp_instance instance,
8185 : : stmt_vector_for_cost *cost_vec,
8186 : : hash_map<tree, int> &scalar_use_map,
8187 : : hash_set<stmt_vec_info> &svisited,
8188 : : hash_set<slp_tree> &visited)
8189 : : {
8190 : 883827 : if (visited.add (node))
8191 : 26307 : return;
8192 : :
8193 : 857520 : unsigned i;
8194 : 857520 : stmt_vec_info stmt_info;
8195 : 857520 : stmt_vec_info last_stmt = vect_find_last_scalar_stmt_in_slp (node);
8196 : 3065065 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
8197 : : {
8198 : 2207545 : if (!stmt_info || svisited.contains (stmt_info))
8199 : 23901 : continue;
8200 : 2192065 : stmt_vec_info orig_stmt_info = vect_orig_stmt (stmt_info);
8201 : 2192065 : if (STMT_VINFO_IN_PATTERN_P (orig_stmt_info)
8202 : 11201 : && STMT_VINFO_RELATED_STMT (orig_stmt_info) != stmt_info)
8203 : : /* Only the pattern root stmt computes the original scalar value. */
8204 : 8421 : continue;
8205 : 2183644 : bool mark_visited = true;
8206 : 2183644 : gimple *orig_stmt = orig_stmt_info->stmt;
8207 : 2183644 : ssa_op_iter op_iter;
8208 : 2183644 : def_operand_p def_p;
8209 : 4867641 : FOR_EACH_PHI_OR_STMT_DEF (def_p, orig_stmt, op_iter, SSA_OP_DEF)
8210 : : {
8211 : 500353 : if (vec_slp_has_scalar_use (bb_vinfo, DEF_FROM_PTR (def_p),
8212 : : scalar_use_map))
8213 : : {
8214 : 82827 : STMT_VINFO_LIVE_P (stmt_info) = true;
8215 : 82827 : if (vectorizable_live_operation (bb_vinfo, stmt_info, node,
8216 : : instance, i, false, cost_vec))
8217 : : /* ??? So we know we can vectorize the live stmt from one SLP
8218 : : node. If we cannot do so from all or none consistently
8219 : : we'd have to record which SLP node (and lane) we want to
8220 : : use for the live operation. So make sure we can
8221 : : code-generate from all nodes. */
8222 : : mark_visited = false;
8223 : : else
8224 : 0 : STMT_VINFO_LIVE_P (stmt_info) = false;
8225 : : }
8226 : :
8227 : : /* We have to verify whether we can insert the lane extract
8228 : : before all uses. The following is a conservative approximation.
8229 : : We cannot put this into vectorizable_live_operation because
8230 : : iterating over all use stmts from inside a FOR_EACH_IMM_USE_STMT
8231 : : doesn't work.
8232 : : Note that while the fact that we emit code for loads at the
8233 : : first load should make this a non-problem leafs we construct
8234 : : from scalars are vectorized after the last scalar def.
8235 : : ??? If we'd actually compute the insert location during
8236 : : analysis we could use sth less conservative than the last
8237 : : scalar stmt in the node for the dominance check. */
8238 : : /* ??? What remains is "live" uses in vector CTORs in the same
8239 : : SLP graph which is where those uses can end up code-generated
8240 : : right after their definition instead of close to their original
8241 : : use. But that would restrict us to code-generate lane-extracts
8242 : : from the latest stmt in a node. So we compensate for this
8243 : : during code-generation, simply not replacing uses for those
8244 : : hopefully rare cases. */
8245 : 500353 : imm_use_iterator use_iter;
8246 : 500353 : gimple *use_stmt;
8247 : 500353 : stmt_vec_info use_stmt_info;
8248 : :
8249 : 500353 : if (STMT_VINFO_LIVE_P (stmt_info))
8250 : 452904 : FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, DEF_FROM_PTR (def_p))
8251 : 370077 : if (!is_gimple_debug (use_stmt)
8252 : 282636 : && (!(use_stmt_info = bb_vinfo->lookup_stmt (use_stmt))
8253 : 271782 : || !PURE_SLP_STMT (vect_stmt_to_vectorize (use_stmt_info)))
8254 : 531600 : && !vect_stmt_dominates_stmt_p (last_stmt->stmt, use_stmt))
8255 : : {
8256 : 12159 : if (dump_enabled_p ())
8257 : 284 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
8258 : : "Cannot determine insertion place for "
8259 : : "lane extract\n");
8260 : 12159 : STMT_VINFO_LIVE_P (stmt_info) = false;
8261 : 12159 : mark_visited = true;
8262 : 82827 : }
8263 : : }
8264 : 2183644 : if (mark_visited)
8265 : 2111871 : svisited.add (stmt_info);
8266 : : }
8267 : :
8268 : : slp_tree child;
8269 : 2430793 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
8270 : 833275 : if (child && SLP_TREE_DEF_TYPE (child) == vect_internal_def)
8271 : 225179 : vect_bb_slp_mark_live_stmts (bb_vinfo, child, instance, cost_vec,
8272 : : scalar_use_map, svisited, visited);
8273 : : }
8274 : :
8275 : : /* Traverse all slp instances of BB_VINFO, and mark lanes of every node that
8276 : : are live outside of the basic-block vectorized region and that can be
8277 : : vectorized using vectorizable_live_operation with STMT_VINFO_LIVE_P. */
8278 : :
8279 : : static void
8280 : 314290 : vect_bb_slp_mark_live_stmts (bb_vec_info bb_vinfo)
8281 : : {
8282 : 314290 : if (bb_vinfo->slp_instances.is_empty ())
8283 : 34824 : return;
8284 : :
8285 : 279466 : hash_set<stmt_vec_info> svisited;
8286 : 279466 : hash_set<slp_tree> visited;
8287 : 279466 : hash_map<tree, int> scalar_use_map;
8288 : 279466 : auto_vec<slp_tree> worklist;
8289 : :
8290 : 1497046 : for (slp_instance instance : bb_vinfo->slp_instances)
8291 : : {
8292 : 658648 : if (SLP_INSTANCE_KIND (instance) == slp_inst_kind_bb_reduc)
8293 : 41790 : for (tree op : SLP_INSTANCE_REMAIN_DEFS (instance))
8294 : 13703 : if (TREE_CODE (op) == SSA_NAME)
8295 : 12385 : scalar_use_map.put (op, 1);
8296 : 658648 : if (!visited.add (SLP_INSTANCE_TREE (instance)))
8297 : 657235 : worklist.safe_push (SLP_INSTANCE_TREE (instance));
8298 : : }
8299 : :
8300 : 1464701 : do
8301 : : {
8302 : 1464701 : slp_tree node = worklist.pop ();
8303 : :
8304 : 1464701 : if (SLP_TREE_DEF_TYPE (node) == vect_external_def)
8305 : : {
8306 : 1466515 : for (tree op : SLP_TREE_SCALAR_OPS (node))
8307 : 646157 : if (TREE_CODE (op) == SSA_NAME)
8308 : 419360 : scalar_use_map.put (op, 1);
8309 : : }
8310 : : else
8311 : : {
8312 : 3504032 : for (slp_tree child : SLP_TREE_CHILDREN (node))
8313 : 833251 : if (child && !visited.add (child))
8314 : 807466 : worklist.safe_push (child);
8315 : : }
8316 : : }
8317 : 2929402 : while (!worklist.is_empty ());
8318 : :
8319 : 279466 : visited.empty ();
8320 : :
8321 : 1497046 : for (slp_instance instance : bb_vinfo->slp_instances)
8322 : : {
8323 : 658648 : vect_location = instance->location ();
8324 : 658648 : vect_bb_slp_mark_live_stmts (bb_vinfo, SLP_INSTANCE_TREE (instance),
8325 : : instance, &instance->cost_vec,
8326 : : scalar_use_map, svisited, visited);
8327 : : }
8328 : 279466 : }
8329 : :
8330 : : /* Determine whether we can vectorize the reduction epilogue for INSTANCE. */
8331 : :
8332 : : static bool
8333 : 59987 : vectorizable_bb_reduc_epilogue (slp_instance instance,
8334 : : stmt_vector_for_cost *cost_vec)
8335 : : {
8336 : 59987 : gassign *stmt = as_a <gassign *> (instance->root_stmts[0]->stmt);
8337 : 59987 : enum tree_code reduc_code = gimple_assign_rhs_code (stmt);
8338 : 59987 : if (reduc_code == MINUS_EXPR)
8339 : 0 : reduc_code = PLUS_EXPR;
8340 : 59987 : internal_fn reduc_fn;
8341 : 59987 : tree vectype = SLP_TREE_VECTYPE (SLP_INSTANCE_TREE (instance));
8342 : 59987 : if (!vectype
8343 : 59894 : || !reduction_fn_for_scalar_code (reduc_code, &reduc_fn)
8344 : 59894 : || reduc_fn == IFN_LAST
8345 : 59894 : || !direct_internal_fn_supported_p (reduc_fn, vectype, OPTIMIZE_FOR_BOTH)
8346 : 85908 : || !useless_type_conversion_p (TREE_TYPE (gimple_assign_lhs (stmt)),
8347 : 25921 : TREE_TYPE (vectype)))
8348 : : {
8349 : 44279 : if (dump_enabled_p ())
8350 : 254 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
8351 : : "not vectorized: basic block reduction epilogue "
8352 : : "operation unsupported.\n");
8353 : 44279 : return false;
8354 : : }
8355 : :
8356 : : /* There's no way to cost a horizontal vector reduction via REDUC_FN so
8357 : : cost log2 vector operations plus shuffles and one extraction. */
8358 : 15708 : unsigned steps = floor_log2 (vect_nunits_for_cost (vectype));
8359 : 15708 : record_stmt_cost (cost_vec, steps, vector_stmt, instance->root_stmts[0],
8360 : : vectype, 0, vect_body);
8361 : 15708 : record_stmt_cost (cost_vec, steps, vec_perm, instance->root_stmts[0],
8362 : : vectype, 0, vect_body);
8363 : 15708 : record_stmt_cost (cost_vec, 1, vec_to_scalar, instance->root_stmts[0],
8364 : : vectype, 0, vect_body);
8365 : :
8366 : : /* Since we replace all stmts of a possibly longer scalar reduction
8367 : : chain account for the extra scalar stmts for that. */
8368 : 15708 : record_stmt_cost (cost_vec, instance->remain_defs.length (), scalar_stmt,
8369 : 15708 : instance->root_stmts[0], 0, vect_body);
8370 : 15708 : return true;
8371 : : }
8372 : :
8373 : : /* Prune from ROOTS all stmts that are computed as part of lanes of NODE
8374 : : and recurse to children. */
8375 : :
8376 : : static void
8377 : 121310 : vect_slp_prune_covered_roots (slp_tree node, hash_set<stmt_vec_info> &roots,
8378 : : hash_set<slp_tree> &visited)
8379 : : {
8380 : 121310 : if (SLP_TREE_DEF_TYPE (node) != vect_internal_def
8381 : 121310 : || visited.add (node))
8382 : 54237 : return;
8383 : :
8384 : : stmt_vec_info stmt;
8385 : : unsigned i;
8386 : 218943 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt)
8387 : 151870 : if (stmt)
8388 : 154858 : roots.remove (vect_orig_stmt (stmt));
8389 : :
8390 : : slp_tree child;
8391 : 158378 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
8392 : 91305 : if (child)
8393 : 85636 : vect_slp_prune_covered_roots (child, roots, visited);
8394 : : }
8395 : :
8396 : : /* Analyze statements in SLP instances of VINFO. Return true if the
8397 : : operations are supported. */
8398 : :
8399 : : bool
8400 : 545311 : vect_slp_analyze_operations (vec_info *vinfo)
8401 : : {
8402 : 545311 : slp_instance instance;
8403 : 545311 : int i;
8404 : :
8405 : 545311 : DUMP_VECT_SCOPE ("vect_slp_analyze_operations");
8406 : :
8407 : 545311 : hash_set<slp_tree> visited;
8408 : 1481323 : for (i = 0; vinfo->slp_instances.iterate (i, &instance); )
8409 : : {
8410 : 1049843 : auto_vec<slp_tree> visited_vec;
8411 : 1049843 : stmt_vector_for_cost cost_vec;
8412 : 1049843 : cost_vec.create (2);
8413 : 1049843 : if (is_a <bb_vec_info> (vinfo))
8414 : 756273 : vect_location = instance->location ();
8415 : 1049843 : if (!vect_slp_analyze_node_operations (vinfo,
8416 : : SLP_INSTANCE_TREE (instance),
8417 : : instance, visited, visited_vec,
8418 : : &cost_vec)
8419 : : /* CTOR instances require vectorized defs for the SLP tree root. */
8420 : 920146 : || (SLP_INSTANCE_KIND (instance) == slp_inst_kind_ctor
8421 : 639 : && (SLP_TREE_DEF_TYPE (SLP_INSTANCE_TREE (instance))
8422 : : != vect_internal_def
8423 : : /* Make sure we vectorized with the expected type. */
8424 : 635 : || !useless_type_conversion_p
8425 : 635 : (TREE_TYPE (TREE_TYPE (gimple_assign_rhs1
8426 : : (instance->root_stmts[0]->stmt))),
8427 : 635 : TREE_TYPE (SLP_TREE_VECTYPE
8428 : : (SLP_INSTANCE_TREE (instance))))))
8429 : : /* Check we can vectorize the reduction. */
8430 : 920139 : || (SLP_INSTANCE_KIND (instance) == slp_inst_kind_bb_reduc
8431 : 59987 : && !vectorizable_bb_reduc_epilogue (instance, &cost_vec))
8432 : : /* Check we can vectorize the gcond. */
8433 : 1925703 : || (SLP_INSTANCE_KIND (instance) == slp_inst_kind_gcond
8434 : 37797 : && !vectorizable_early_exit (vinfo,
8435 : 37797 : SLP_INSTANCE_ROOT_STMTS (instance)[0],
8436 : : NULL, NULL,
8437 : : SLP_INSTANCE_TREE (instance),
8438 : : &cost_vec)))
8439 : : {
8440 : 210027 : cost_vec.release ();
8441 : 210027 : slp_tree node = SLP_INSTANCE_TREE (instance);
8442 : 210027 : stmt_vec_info stmt_info;
8443 : 210027 : if (!SLP_INSTANCE_ROOT_STMTS (instance).is_empty ())
8444 : 142014 : stmt_info = SLP_INSTANCE_ROOT_STMTS (instance)[0];
8445 : : else
8446 : 68013 : stmt_info = SLP_TREE_SCALAR_STMTS (node)[0];
8447 : 210027 : if (is_a <loop_vec_info> (vinfo))
8448 : : {
8449 : 113831 : if (dump_enabled_p ())
8450 : 6552 : dump_printf_loc (MSG_NOTE, vect_location,
8451 : : "unsupported SLP instance starting from: %G",
8452 : : stmt_info->stmt);
8453 : 113831 : return false;
8454 : : }
8455 : 96196 : if (dump_enabled_p ())
8456 : 321 : dump_printf_loc (MSG_NOTE, vect_location,
8457 : : "removing SLP instance operations starting from: %G",
8458 : : stmt_info->stmt);
8459 : 96196 : vect_free_slp_instance (instance);
8460 : 96196 : vinfo->slp_instances.ordered_remove (i);
8461 : 1420070 : while (!visited_vec.is_empty ())
8462 : 291788 : visited.remove (visited_vec.pop ());
8463 : : }
8464 : : else
8465 : : {
8466 : 839816 : i++;
8467 : 839816 : if (loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo))
8468 : : {
8469 : 179739 : add_stmt_costs (loop_vinfo->vector_costs, &cost_vec);
8470 : 179739 : cost_vec.release ();
8471 : : }
8472 : : else
8473 : : /* For BB vectorization remember the SLP graph entry
8474 : : cost for later. */
8475 : 660077 : instance->cost_vec = cost_vec;
8476 : : }
8477 : 1049843 : }
8478 : :
8479 : : /* Now look for SLP instances with a root that are covered by other
8480 : : instances and remove them. */
8481 : 431480 : hash_set<stmt_vec_info> roots;
8482 : 1688706 : for (i = 0; vinfo->slp_instances.iterate (i, &instance); ++i)
8483 : 843741 : if (!SLP_INSTANCE_ROOT_STMTS (instance).is_empty ())
8484 : 17995 : roots.add (SLP_INSTANCE_ROOT_STMTS (instance)[0]);
8485 : 431480 : if (!roots.is_empty ())
8486 : : {
8487 : 9654 : visited.empty ();
8488 : 45328 : for (i = 0; vinfo->slp_instances.iterate (i, &instance); ++i)
8489 : 35674 : vect_slp_prune_covered_roots (SLP_INSTANCE_TREE (instance), roots,
8490 : : visited);
8491 : 45328 : for (i = 0; vinfo->slp_instances.iterate (i, &instance); )
8492 : 35674 : if (!SLP_INSTANCE_ROOT_STMTS (instance).is_empty ()
8493 : 17995 : && !roots.contains (SLP_INSTANCE_ROOT_STMTS (instance)[0]))
8494 : : {
8495 : 1429 : stmt_vec_info root = SLP_INSTANCE_ROOT_STMTS (instance)[0];
8496 : 1429 : if (dump_enabled_p ())
8497 : 20 : dump_printf_loc (MSG_NOTE, vect_location,
8498 : : "removing SLP instance operations starting "
8499 : : "from: %G", root->stmt);
8500 : 1429 : vect_free_slp_instance (instance);
8501 : 1429 : vinfo->slp_instances.ordered_remove (i);
8502 : : }
8503 : : else
8504 : 34245 : ++i;
8505 : : }
8506 : :
8507 : : /* Compute vectorizable live stmts. */
8508 : 431480 : if (bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (vinfo))
8509 : 314290 : vect_bb_slp_mark_live_stmts (bb_vinfo);
8510 : :
8511 : 862960 : return !vinfo->slp_instances.is_empty ();
8512 : 976791 : }
8513 : :
8514 : : /* Get the SLP instance leader from INSTANCE_LEADER thereby transitively
8515 : : closing the eventual chain. */
8516 : :
8517 : : static slp_instance
8518 : 699907 : get_ultimate_leader (slp_instance instance,
8519 : : hash_map<slp_instance, slp_instance> &instance_leader)
8520 : : {
8521 : 699907 : auto_vec<slp_instance *, 8> chain;
8522 : 699907 : slp_instance *tem;
8523 : 747379 : while (*(tem = instance_leader.get (instance)) != instance)
8524 : : {
8525 : 47472 : chain.safe_push (tem);
8526 : 47472 : instance = *tem;
8527 : : }
8528 : 747379 : while (!chain.is_empty ())
8529 : 47472 : *chain.pop () = instance;
8530 : 699907 : return instance;
8531 : 699907 : }
8532 : :
8533 : : namespace {
8534 : : /* Subroutine of vect_bb_partition_graph_r. Map KEY to INSTANCE in
8535 : : KEY_TO_INSTANCE, making INSTANCE the leader of any previous mapping
8536 : : for KEY. Return true if KEY was already in KEY_TO_INSTANCE.
8537 : :
8538 : : INSTANCE_LEADER is as for get_ultimate_leader. */
8539 : :
8540 : : template<typename T>
8541 : : bool
8542 : 3157425 : vect_map_to_instance (slp_instance instance, T key,
8543 : : hash_map<T, slp_instance> &key_to_instance,
8544 : : hash_map<slp_instance, slp_instance> &instance_leader)
8545 : : {
8546 : : bool existed_p;
8547 : 3157425 : slp_instance &key_instance = key_to_instance.get_or_insert (key, &existed_p);
8548 : 3157425 : if (!existed_p)
8549 : : ;
8550 : 115314 : else if (key_instance != instance)
8551 : : {
8552 : : /* If we're running into a previously marked key make us the
8553 : : leader of the current ultimate leader. This keeps the
8554 : : leader chain acyclic and works even when the current instance
8555 : : connects two previously independent graph parts. */
8556 : 41259 : slp_instance key_leader
8557 : 41259 : = get_ultimate_leader (key_instance, instance_leader);
8558 : 41259 : if (key_leader != instance)
8559 : 11320 : instance_leader.put (key_leader, instance);
8560 : : }
8561 : 3157425 : key_instance = instance;
8562 : 3157425 : return existed_p;
8563 : : }
8564 : : }
8565 : :
8566 : : /* Worker of vect_bb_partition_graph, recurse on NODE. */
8567 : :
8568 : : static void
8569 : 883827 : vect_bb_partition_graph_r (bb_vec_info bb_vinfo,
8570 : : slp_instance instance, slp_tree node,
8571 : : hash_map<stmt_vec_info, slp_instance> &stmt_to_instance,
8572 : : hash_map<slp_tree, slp_instance> &node_to_instance,
8573 : : hash_map<slp_instance, slp_instance> &instance_leader)
8574 : : {
8575 : 883827 : stmt_vec_info stmt_info;
8576 : 883827 : unsigned i;
8577 : :
8578 : 3157425 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
8579 : 2273598 : if (stmt_info)
8580 : 2273598 : vect_map_to_instance (instance, stmt_info, stmt_to_instance,
8581 : : instance_leader);
8582 : :
8583 : 883827 : if (vect_map_to_instance (instance, node, node_to_instance,
8584 : : instance_leader))
8585 : 883827 : return;
8586 : :
8587 : : slp_tree child;
8588 : 1690795 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
8589 : 833275 : if (child && SLP_TREE_DEF_TYPE (child) == vect_internal_def)
8590 : 225179 : vect_bb_partition_graph_r (bb_vinfo, instance, child, stmt_to_instance,
8591 : : node_to_instance, instance_leader);
8592 : : }
8593 : :
8594 : : /* Partition the SLP graph into pieces that can be costed independently. */
8595 : :
8596 : : static void
8597 : 279466 : vect_bb_partition_graph (bb_vec_info bb_vinfo)
8598 : : {
8599 : 279466 : DUMP_VECT_SCOPE ("vect_bb_partition_graph");
8600 : :
8601 : : /* First walk the SLP graph assigning each involved scalar stmt a
8602 : : corresponding SLP graph entry and upon visiting a previously
8603 : : marked stmt, make the stmts leader the current SLP graph entry. */
8604 : 279466 : hash_map<stmt_vec_info, slp_instance> stmt_to_instance;
8605 : 279466 : hash_map<slp_tree, slp_instance> node_to_instance;
8606 : 279466 : hash_map<slp_instance, slp_instance> instance_leader;
8607 : 279466 : slp_instance instance;
8608 : 938114 : for (unsigned i = 0; bb_vinfo->slp_instances.iterate (i, &instance); ++i)
8609 : : {
8610 : 658648 : instance_leader.put (instance, instance);
8611 : 658648 : vect_bb_partition_graph_r (bb_vinfo,
8612 : : instance, SLP_INSTANCE_TREE (instance),
8613 : : stmt_to_instance, node_to_instance,
8614 : : instance_leader);
8615 : : }
8616 : :
8617 : : /* Then collect entries to each independent subgraph. */
8618 : 1217580 : for (unsigned i = 0; bb_vinfo->slp_instances.iterate (i, &instance); ++i)
8619 : : {
8620 : 658648 : slp_instance leader = get_ultimate_leader (instance, instance_leader);
8621 : 658648 : leader->subgraph_entries.safe_push (instance);
8622 : 658648 : if (dump_enabled_p ()
8623 : 658648 : && leader != instance)
8624 : 67 : dump_printf_loc (MSG_NOTE, vect_location,
8625 : : "instance %p is leader of %p\n",
8626 : : (void *) leader, (void *) instance);
8627 : : }
8628 : 279466 : }
8629 : :
8630 : : /* Compute the set of scalar stmts participating in internal and external
8631 : : nodes. */
8632 : :
8633 : : static void
8634 : 1478706 : vect_slp_gather_vectorized_scalar_stmts (vec_info *vinfo, slp_tree node,
8635 : : hash_set<slp_tree> &visited,
8636 : : hash_set<stmt_vec_info> &vstmts,
8637 : : hash_set<stmt_vec_info> &estmts)
8638 : : {
8639 : 1478706 : int i;
8640 : 1478706 : stmt_vec_info stmt_info;
8641 : 1478706 : slp_tree child;
8642 : :
8643 : 1478706 : if (visited.add (node))
8644 : 26131 : return;
8645 : :
8646 : 1452575 : if (SLP_TREE_DEF_TYPE (node) == vect_internal_def)
8647 : : {
8648 : 3008713 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
8649 : 2159853 : if (stmt_info)
8650 : 2159853 : vstmts.add (stmt_info);
8651 : :
8652 : 3009365 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
8653 : 823680 : if (child)
8654 : 823680 : vect_slp_gather_vectorized_scalar_stmts (vinfo, child, visited,
8655 : : vstmts, estmts);
8656 : : }
8657 : : else
8658 : 3374117 : for (tree def : SLP_TREE_SCALAR_OPS (node))
8659 : : {
8660 : 1563954 : stmt_vec_info def_stmt = vinfo->lookup_def (def);
8661 : 1563954 : if (def_stmt)
8662 : 251528 : estmts.add (def_stmt);
8663 : : }
8664 : : }
8665 : :
8666 : :
8667 : : /* Compute the scalar cost of the SLP node NODE and its children
8668 : : and return it. Do not account defs that are marked in LIFE and
8669 : : update LIFE according to uses of NODE. */
8670 : :
8671 : : static void
8672 : 874415 : vect_bb_slp_scalar_cost (vec_info *vinfo,
8673 : : slp_tree node, vec<bool, va_heap> *life,
8674 : : stmt_vector_for_cost *cost_vec,
8675 : : hash_set<stmt_vec_info> &vectorized_scalar_stmts,
8676 : : hash_set<slp_tree> &visited)
8677 : : {
8678 : 874415 : unsigned i;
8679 : 874415 : stmt_vec_info stmt_info;
8680 : 874415 : slp_tree child;
8681 : :
8682 : 874415 : if (visited.add (node))
8683 : 25538 : return;
8684 : :
8685 : 3008764 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
8686 : : {
8687 : 2159887 : ssa_op_iter op_iter;
8688 : 2159887 : def_operand_p def_p;
8689 : :
8690 : 4319774 : if (!stmt_info || (*life)[i])
8691 : 78005 : continue;
8692 : :
8693 : 2139781 : stmt_vec_info orig_stmt_info = vect_orig_stmt (stmt_info);
8694 : 2139781 : gimple *orig_stmt = orig_stmt_info->stmt;
8695 : :
8696 : : /* If there is a non-vectorized use of the defs then the scalar
8697 : : stmt is kept live in which case we do not account it or any
8698 : : required defs in the SLP children in the scalar cost. This
8699 : : way we make the vectorization more costly when compared to
8700 : : the scalar cost. */
8701 : 2139781 : if (!STMT_VINFO_LIVE_P (stmt_info))
8702 : : {
8703 : 2073144 : auto_vec<gimple *, 8> worklist;
8704 : 2073144 : hash_set<gimple *> *worklist_visited = NULL;
8705 : 2073144 : worklist.quick_push (orig_stmt);
8706 : 2077967 : do
8707 : : {
8708 : 2077967 : gimple *work_stmt = worklist.pop ();
8709 : 4559158 : FOR_EACH_PHI_OR_STMT_DEF (def_p, work_stmt, op_iter, SSA_OP_DEF)
8710 : : {
8711 : 416108 : imm_use_iterator use_iter;
8712 : 416108 : gimple *use_stmt;
8713 : 1042456 : FOR_EACH_IMM_USE_STMT (use_stmt, use_iter,
8714 : : DEF_FROM_PTR (def_p))
8715 : 639232 : if (!is_gimple_debug (use_stmt))
8716 : : {
8717 : 458362 : stmt_vec_info use_stmt_info
8718 : 458362 : = vinfo->lookup_stmt (use_stmt);
8719 : 458362 : if (!use_stmt_info
8720 : 458362 : || !vectorized_scalar_stmts.contains (use_stmt_info))
8721 : : {
8722 : 17777 : if (use_stmt_info
8723 : 16025 : && STMT_VINFO_IN_PATTERN_P (use_stmt_info))
8724 : : {
8725 : : /* For stmts participating in patterns we have
8726 : : to check its uses recursively. */
8727 : 4893 : if (!worklist_visited)
8728 : 3785 : worklist_visited = new hash_set<gimple *> ();
8729 : 4893 : if (!worklist_visited->add (use_stmt))
8730 : 4893 : worklist.safe_push (use_stmt);
8731 : 4893 : continue;
8732 : : }
8733 : 12884 : (*life)[i] = true;
8734 : 12884 : goto next_lane;
8735 : : }
8736 : 416108 : }
8737 : : }
8738 : : }
8739 : 4130166 : while (!worklist.is_empty ());
8740 : 2060260 : next_lane:
8741 : 2073144 : if (worklist_visited)
8742 : 3785 : delete worklist_visited;
8743 : 2073144 : if ((*life)[i])
8744 : 12884 : continue;
8745 : 2073144 : }
8746 : :
8747 : : /* Count scalar stmts only once. */
8748 : 2126897 : if (gimple_visited_p (orig_stmt))
8749 : 22043 : continue;
8750 : 2104854 : gimple_set_visited (orig_stmt, true);
8751 : :
8752 : 2104854 : vect_cost_for_stmt kind;
8753 : 2104854 : if (STMT_VINFO_DATA_REF (orig_stmt_info))
8754 : : {
8755 : 1928734 : data_reference_p dr = STMT_VINFO_DATA_REF (orig_stmt_info);
8756 : 1928734 : tree base = get_base_address (DR_REF (dr));
8757 : : /* When the scalar access is to a non-global not address-taken
8758 : : decl that is not BLKmode assume we can access it with a single
8759 : : non-load/store instruction. */
8760 : 1928734 : if (DECL_P (base)
8761 : 1549838 : && !is_global_var (base)
8762 : 1484154 : && !TREE_ADDRESSABLE (base)
8763 : 2621312 : && DECL_MODE (base) != BLKmode)
8764 : : kind = scalar_stmt;
8765 : 1671428 : else if (DR_IS_READ (STMT_VINFO_DATA_REF (orig_stmt_info)))
8766 : : kind = scalar_load;
8767 : : else
8768 : 1455009 : kind = scalar_store;
8769 : : }
8770 : 176120 : else if (vect_nop_conversion_p (orig_stmt_info))
8771 : 14607 : continue;
8772 : : /* For single-argument PHIs assume coalescing which means zero cost
8773 : : for the scalar and the vector PHIs. This avoids artificially
8774 : : favoring the vector path (but may pessimize it in some cases). */
8775 : 161513 : else if (is_a <gphi *> (orig_stmt_info->stmt)
8776 : 161513 : && gimple_phi_num_args
8777 : 87064 : (as_a <gphi *> (orig_stmt_info->stmt)) == 1)
8778 : 8365 : continue;
8779 : : else
8780 : : kind = scalar_stmt;
8781 : 2081882 : record_stmt_cost (cost_vec, 1, kind, orig_stmt_info,
8782 : : SLP_TREE_VECTYPE (node), 0, vect_body);
8783 : : }
8784 : :
8785 : 1697754 : auto_vec<bool, 20> subtree_life;
8786 : 2405700 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
8787 : : {
8788 : 823704 : if (child && SLP_TREE_DEF_TYPE (child) == vect_internal_def)
8789 : : {
8790 : : /* Do not directly pass LIFE to the recursive call, copy it to
8791 : : confine changes in the callee to the current child/subtree. */
8792 : 219389 : if (SLP_TREE_CODE (node) == VEC_PERM_EXPR)
8793 : : {
8794 : 3130 : subtree_life.safe_grow_cleared (SLP_TREE_LANES (child), true);
8795 : 11096 : for (unsigned j = 0;
8796 : 11096 : j < SLP_TREE_LANE_PERMUTATION (node).length (); ++j)
8797 : : {
8798 : 7966 : auto perm = SLP_TREE_LANE_PERMUTATION (node)[j];
8799 : 7966 : if (perm.first == i)
8800 : 4170 : subtree_life[perm.second] = (*life)[j];
8801 : : }
8802 : : }
8803 : : else
8804 : : {
8805 : 216259 : gcc_assert (SLP_TREE_LANES (node) == SLP_TREE_LANES (child));
8806 : 216259 : subtree_life.safe_splice (*life);
8807 : : }
8808 : 219389 : vect_bb_slp_scalar_cost (vinfo, child, &subtree_life, cost_vec,
8809 : : vectorized_scalar_stmts, visited);
8810 : 219389 : subtree_life.truncate (0);
8811 : : }
8812 : : }
8813 : : }
8814 : :
8815 : : /* Comparator for the loop-index sorted cost vectors. */
8816 : :
8817 : : static int
8818 : 16898777 : li_cost_vec_cmp (const void *a_, const void *b_)
8819 : : {
8820 : 16898777 : auto *a = (const std::pair<unsigned, stmt_info_for_cost *> *)a_;
8821 : 16898777 : auto *b = (const std::pair<unsigned, stmt_info_for_cost *> *)b_;
8822 : 16898777 : if (a->first < b->first)
8823 : : return -1;
8824 : 16315182 : else if (a->first == b->first)
8825 : 15803271 : return 0;
8826 : : return 1;
8827 : : }
8828 : :
8829 : : /* Check if vectorization of the basic block is profitable for the
8830 : : subgraph denoted by SLP_INSTANCES. */
8831 : :
8832 : : static bool
8833 : 643837 : vect_bb_vectorization_profitable_p (bb_vec_info bb_vinfo,
8834 : : vec<slp_instance> slp_instances,
8835 : : loop_p orig_loop)
8836 : : {
8837 : 643837 : slp_instance instance;
8838 : 643837 : int i;
8839 : 643837 : unsigned int vec_inside_cost = 0, vec_outside_cost = 0, scalar_cost = 0;
8840 : 643837 : unsigned int vec_prologue_cost = 0, vec_epilogue_cost = 0;
8841 : :
8842 : 643837 : if (dump_enabled_p ())
8843 : : {
8844 : 82 : dump_printf_loc (MSG_NOTE, vect_location, "Costing subgraph: \n");
8845 : 82 : hash_set<slp_tree> visited;
8846 : 331 : FOR_EACH_VEC_ELT (slp_instances, i, instance)
8847 : 85 : vect_print_slp_graph (MSG_NOTE, vect_location,
8848 : : SLP_INSTANCE_TREE (instance), visited);
8849 : 82 : }
8850 : :
8851 : : /* Compute the set of scalar stmts we know will go away 'locally' when
8852 : : vectorizing. This used to be tracked with just PURE_SLP_STMT but that's
8853 : : not accurate for nodes promoted extern late or for scalar stmts that
8854 : : are used both in extern defs and in vectorized defs. */
8855 : 643837 : hash_set<stmt_vec_info> vectorized_scalar_stmts;
8856 : 643837 : hash_set<stmt_vec_info> scalar_stmts_in_externs;
8857 : 643837 : hash_set<slp_tree> visited;
8858 : 1298863 : FOR_EACH_VEC_ELT (slp_instances, i, instance)
8859 : : {
8860 : 655026 : vect_slp_gather_vectorized_scalar_stmts (bb_vinfo,
8861 : : SLP_INSTANCE_TREE (instance),
8862 : : visited,
8863 : : vectorized_scalar_stmts,
8864 : : scalar_stmts_in_externs);
8865 : 718921 : for (stmt_vec_info rstmt : SLP_INSTANCE_ROOT_STMTS (instance))
8866 : 34507 : vectorized_scalar_stmts.add (rstmt);
8867 : : }
8868 : : /* Scalar stmts used as defs in external nodes need to be preseved, so
8869 : : remove them from vectorized_scalar_stmts. */
8870 : 868150 : for (stmt_vec_info stmt : scalar_stmts_in_externs)
8871 : 224313 : vectorized_scalar_stmts.remove (stmt);
8872 : :
8873 : : /* Calculate scalar cost and sum the cost for the vector stmts
8874 : : previously collected. */
8875 : 643837 : stmt_vector_for_cost scalar_costs = vNULL;
8876 : 643837 : stmt_vector_for_cost vector_costs = vNULL;
8877 : 643837 : visited.empty ();
8878 : 1298863 : FOR_EACH_VEC_ELT (slp_instances, i, instance)
8879 : : {
8880 : 655026 : auto_vec<bool, 20> life;
8881 : 655026 : life.safe_grow_cleared (SLP_TREE_LANES (SLP_INSTANCE_TREE (instance)),
8882 : : true);
8883 : 655026 : if (!SLP_INSTANCE_ROOT_STMTS (instance).is_empty ())
8884 : 29388 : record_stmt_cost (&scalar_costs,
8885 : 14694 : SLP_INSTANCE_ROOT_STMTS (instance).length (),
8886 : : scalar_stmt,
8887 : 14694 : SLP_INSTANCE_ROOT_STMTS (instance)[0], 0, vect_body);
8888 : 655026 : vect_bb_slp_scalar_cost (bb_vinfo,
8889 : : SLP_INSTANCE_TREE (instance),
8890 : : &life, &scalar_costs, vectorized_scalar_stmts,
8891 : : visited);
8892 : 655026 : vector_costs.safe_splice (instance->cost_vec);
8893 : 655026 : instance->cost_vec.release ();
8894 : 655026 : }
8895 : :
8896 : 643837 : if (dump_enabled_p ())
8897 : 82 : dump_printf_loc (MSG_NOTE, vect_location, "Cost model analysis: \n");
8898 : :
8899 : : /* When costing non-loop vectorization we need to consider each covered
8900 : : loop independently and make sure vectorization is profitable. For
8901 : : now we assume a loop may be not entered or executed an arbitrary
8902 : : number of iterations (??? static information can provide more
8903 : : precise info here) which means we can simply cost each containing
8904 : : loops stmts separately. */
8905 : :
8906 : : /* First produce cost vectors sorted by loop index. */
8907 : 643837 : auto_vec<std::pair<unsigned, stmt_info_for_cost *> >
8908 : 643837 : li_scalar_costs (scalar_costs.length ());
8909 : 643837 : auto_vec<std::pair<unsigned, stmt_info_for_cost *> >
8910 : 643837 : li_vector_costs (vector_costs.length ());
8911 : 643837 : stmt_info_for_cost *cost;
8912 : 2740413 : FOR_EACH_VEC_ELT (scalar_costs, i, cost)
8913 : : {
8914 : 2096576 : unsigned l = gimple_bb (cost->stmt_info->stmt)->loop_father->num;
8915 : 2096576 : li_scalar_costs.quick_push (std::make_pair (l, cost));
8916 : : }
8917 : : /* Use a random used loop as fallback in case the first vector_costs
8918 : : entry does not have a stmt_info associated with it. */
8919 : 643837 : unsigned l = li_scalar_costs[0].first;
8920 : 2384369 : FOR_EACH_VEC_ELT (vector_costs, i, cost)
8921 : : {
8922 : : /* We inherit from the previous COST, invariants, externals and
8923 : : extracts immediately follow the cost for the related stmt. */
8924 : 1740532 : if (cost->stmt_info)
8925 : 1020644 : l = gimple_bb (cost->stmt_info->stmt)->loop_father->num;
8926 : 1740532 : li_vector_costs.quick_push (std::make_pair (l, cost));
8927 : : }
8928 : 643837 : li_scalar_costs.qsort (li_cost_vec_cmp);
8929 : 643837 : li_vector_costs.qsort (li_cost_vec_cmp);
8930 : :
8931 : : /* Now cost the portions individually. */
8932 : : unsigned vi = 0;
8933 : : unsigned si = 0;
8934 : 1115170 : bool profitable = true;
8935 : 1115170 : while (si < li_scalar_costs.length ()
8936 : 1763096 : && vi < li_vector_costs.length ())
8937 : : {
8938 : 647926 : unsigned sl = li_scalar_costs[si].first;
8939 : 647926 : unsigned vl = li_vector_costs[vi].first;
8940 : 647926 : if (sl != vl)
8941 : : {
8942 : 135 : if (dump_enabled_p ())
8943 : 0 : dump_printf_loc (MSG_NOTE, vect_location,
8944 : : "Scalar %d and vector %d loop part do not "
8945 : : "match up, skipping scalar part\n", sl, vl);
8946 : : /* Skip the scalar part, assuming zero cost on the vector side. */
8947 : 333 : do
8948 : : {
8949 : 333 : si++;
8950 : : }
8951 : 333 : while (si < li_scalar_costs.length ()
8952 : 470 : && li_scalar_costs[si].first == sl);
8953 : 135 : continue;
8954 : : }
8955 : :
8956 : 647791 : class vector_costs *scalar_target_cost_data = init_cost (bb_vinfo, true);
8957 : 2082861 : do
8958 : : {
8959 : 2082861 : add_stmt_cost (scalar_target_cost_data, li_scalar_costs[si].second);
8960 : 2082861 : si++;
8961 : : }
8962 : 2082861 : while (si < li_scalar_costs.length ()
8963 : 4172898 : && li_scalar_costs[si].first == sl);
8964 : 647791 : scalar_target_cost_data->finish_cost (nullptr);
8965 : 647791 : scalar_cost = scalar_target_cost_data->body_cost ();
8966 : :
8967 : : /* Complete the target-specific vector cost calculation. */
8968 : 647791 : class vector_costs *vect_target_cost_data = init_cost (bb_vinfo, false);
8969 : 1721548 : do
8970 : : {
8971 : 1721548 : add_stmt_cost (vect_target_cost_data, li_vector_costs[vi].second);
8972 : 1721548 : vi++;
8973 : : }
8974 : 1721548 : while (vi < li_vector_costs.length ()
8975 : 3450815 : && li_vector_costs[vi].first == vl);
8976 : 647791 : vect_target_cost_data->finish_cost (scalar_target_cost_data);
8977 : 647791 : vec_prologue_cost = vect_target_cost_data->prologue_cost ();
8978 : 647791 : vec_inside_cost = vect_target_cost_data->body_cost ();
8979 : 647791 : vec_epilogue_cost = vect_target_cost_data->epilogue_cost ();
8980 : 647791 : delete scalar_target_cost_data;
8981 : 647791 : delete vect_target_cost_data;
8982 : :
8983 : 647791 : vec_outside_cost = vec_prologue_cost + vec_epilogue_cost;
8984 : :
8985 : 647791 : if (dump_enabled_p ())
8986 : : {
8987 : 88 : dump_printf_loc (MSG_NOTE, vect_location,
8988 : : "Cost model analysis for part in loop %d:\n", sl);
8989 : 88 : dump_printf (MSG_NOTE, " Vector cost: %d\n",
8990 : : vec_inside_cost + vec_outside_cost);
8991 : 88 : dump_printf (MSG_NOTE, " Scalar cost: %d\n", scalar_cost);
8992 : : }
8993 : :
8994 : : /* Vectorization is profitable if its cost is more than the cost of scalar
8995 : : version. Note that we err on the vector side for equal cost because
8996 : : the cost estimate is otherwise quite pessimistic (constant uses are
8997 : : free on the scalar side but cost a load on the vector side for
8998 : : example). */
8999 : 647791 : if (vec_outside_cost + vec_inside_cost > scalar_cost)
9000 : : {
9001 : : profitable = false;
9002 : : break;
9003 : : }
9004 : : }
9005 : 1111081 : if (profitable && vi < li_vector_costs.length ())
9006 : : {
9007 : 575 : if (dump_enabled_p ())
9008 : 0 : dump_printf_loc (MSG_NOTE, vect_location,
9009 : : "Excess vector cost for part in loop %d:\n",
9010 : 0 : li_vector_costs[vi].first);
9011 : : profitable = false;
9012 : : }
9013 : :
9014 : : /* Unset visited flag. This is delayed when the subgraph is profitable
9015 : : and we process the loop for remaining unvectorized if-converted code. */
9016 : 643837 : if (!orig_loop || !profitable)
9017 : 2738956 : FOR_EACH_VEC_ELT (scalar_costs, i, cost)
9018 : 2095246 : gimple_set_visited (cost->stmt_info->stmt, false);
9019 : :
9020 : 643837 : scalar_costs.release ();
9021 : 643837 : vector_costs.release ();
9022 : :
9023 : 643837 : return profitable;
9024 : 643837 : }
9025 : :
9026 : : /* qsort comparator for lane defs. */
9027 : :
9028 : : static int
9029 : 40 : vld_cmp (const void *a_, const void *b_)
9030 : : {
9031 : 40 : auto *a = (const std::pair<unsigned, tree> *)a_;
9032 : 40 : auto *b = (const std::pair<unsigned, tree> *)b_;
9033 : 40 : return a->first - b->first;
9034 : : }
9035 : :
9036 : : /* Return true if USE_STMT is a vector lane insert into VEC and set
9037 : : *THIS_LANE to the lane number that is set. */
9038 : :
9039 : : static bool
9040 : 240 : vect_slp_is_lane_insert (gimple *use_stmt, tree vec, unsigned *this_lane)
9041 : : {
9042 : 240 : gassign *use_ass = dyn_cast <gassign *> (use_stmt);
9043 : 91 : if (!use_ass
9044 : 91 : || gimple_assign_rhs_code (use_ass) != BIT_INSERT_EXPR
9045 : 22 : || (vec
9046 : 22 : ? gimple_assign_rhs1 (use_ass) != vec
9047 : 24 : : ((vec = gimple_assign_rhs1 (use_ass)), false))
9048 : 46 : || !useless_type_conversion_p (TREE_TYPE (TREE_TYPE (vec)),
9049 : 46 : TREE_TYPE (gimple_assign_rhs2 (use_ass)))
9050 : 46 : || !constant_multiple_p
9051 : 46 : (tree_to_poly_uint64 (gimple_assign_rhs3 (use_ass)),
9052 : 92 : tree_to_poly_uint64 (TYPE_SIZE (TREE_TYPE (TREE_TYPE (vec)))),
9053 : : this_lane))
9054 : 194 : return false;
9055 : : return true;
9056 : : }
9057 : :
9058 : : /* Find any vectorizable constructors and add them to the grouped_store
9059 : : array. */
9060 : :
9061 : : static void
9062 : 2297833 : vect_slp_check_for_roots (bb_vec_info bb_vinfo)
9063 : : {
9064 : 17109235 : for (unsigned i = 0; i < bb_vinfo->nbbs; ++i)
9065 : 29622804 : for (gimple_stmt_iterator gsi = gsi_start_bb (bb_vinfo->bbs[i]);
9066 : 116192481 : !gsi_end_p (gsi); gsi_next (&gsi))
9067 : : {
9068 : 101381079 : gassign *assign = dyn_cast<gassign *> (gsi_stmt (gsi));
9069 : : /* This can be used to start SLP discovery for early breaks for BB early breaks
9070 : : when we get that far. */
9071 : 101381079 : if (!assign)
9072 : 146835886 : continue;
9073 : :
9074 : 28903269 : tree rhs = gimple_assign_rhs1 (assign);
9075 : 28903269 : enum tree_code code = gimple_assign_rhs_code (assign);
9076 : 28903269 : use_operand_p use_p;
9077 : 28903269 : gimple *use_stmt;
9078 : 28903269 : if (code == CONSTRUCTOR)
9079 : : {
9080 : 1617913 : if (!VECTOR_TYPE_P (TREE_TYPE (rhs))
9081 : 49505 : || maybe_ne (TYPE_VECTOR_SUBPARTS (TREE_TYPE (rhs)),
9082 : 76492 : CONSTRUCTOR_NELTS (rhs))
9083 : 35224 : || VECTOR_TYPE_P (TREE_TYPE (CONSTRUCTOR_ELT (rhs, 0)->value))
9084 : 1653136 : || uniform_vector_p (rhs))
9085 : 1610249 : continue;
9086 : :
9087 : : unsigned j;
9088 : : tree val;
9089 : 45634 : FOR_EACH_CONSTRUCTOR_VALUE (CONSTRUCTOR_ELTS (rhs), j, val)
9090 : 37970 : if (TREE_CODE (val) != SSA_NAME
9091 : 37970 : || !bb_vinfo->lookup_def (val))
9092 : : break;
9093 : 19716 : if (j != CONSTRUCTOR_NELTS (rhs))
9094 : 2194 : continue;
9095 : :
9096 : 7664 : vec<stmt_vec_info> roots = vNULL;
9097 : 7664 : roots.safe_push (bb_vinfo->lookup_stmt (assign));
9098 : 7664 : vec<stmt_vec_info> stmts;
9099 : 7664 : stmts.create (CONSTRUCTOR_NELTS (rhs));
9100 : 49830 : FOR_EACH_CONSTRUCTOR_VALUE (CONSTRUCTOR_ELTS (rhs), j, val)
9101 : 34502 : stmts.quick_push
9102 : 34502 : (vect_stmt_to_vectorize (bb_vinfo->lookup_def (val)));
9103 : 7664 : bb_vinfo->roots.safe_push (slp_root (slp_inst_kind_ctor,
9104 : 7664 : stmts, roots));
9105 : : }
9106 : 27285356 : else if (code == BIT_INSERT_EXPR
9107 : 813 : && VECTOR_TYPE_P (TREE_TYPE (rhs))
9108 : 550 : && TYPE_VECTOR_SUBPARTS (TREE_TYPE (rhs)).is_constant ()
9109 : 550 : && TYPE_VECTOR_SUBPARTS (TREE_TYPE (rhs)).to_constant () > 1
9110 : 550 : && integer_zerop (gimple_assign_rhs3 (assign))
9111 : 312 : && useless_type_conversion_p
9112 : 312 : (TREE_TYPE (TREE_TYPE (rhs)),
9113 : 312 : TREE_TYPE (gimple_assign_rhs2 (assign)))
9114 : 27285920 : && bb_vinfo->lookup_def (gimple_assign_rhs2 (assign)))
9115 : : {
9116 : : /* We start to match on insert to lane zero but since the
9117 : : inserts need not be ordered we'd have to search both
9118 : : the def and the use chains. */
9119 : 211 : tree vectype = TREE_TYPE (rhs);
9120 : 211 : unsigned nlanes = TYPE_VECTOR_SUBPARTS (vectype).to_constant ();
9121 : 211 : auto_vec<std::pair<unsigned, tree> > lane_defs (nlanes);
9122 : 211 : auto_sbitmap lanes (nlanes);
9123 : 211 : bitmap_clear (lanes);
9124 : 211 : bitmap_set_bit (lanes, 0);
9125 : 211 : tree def = gimple_assign_lhs (assign);
9126 : 211 : lane_defs.quick_push
9127 : 211 : (std::make_pair (0, gimple_assign_rhs2 (assign)));
9128 : 211 : unsigned lanes_found = 1;
9129 : : /* Start with the use chains, the last stmt will be the root. */
9130 : 211 : stmt_vec_info last = bb_vinfo->lookup_stmt (assign);
9131 : 211 : vec<stmt_vec_info> roots = vNULL;
9132 : 211 : roots.safe_push (last);
9133 : 213 : do
9134 : : {
9135 : 213 : use_operand_p use_p;
9136 : 213 : gimple *use_stmt;
9137 : 213 : if (!single_imm_use (def, &use_p, &use_stmt))
9138 : : break;
9139 : 207 : unsigned this_lane;
9140 : 207 : if (!bb_vinfo->lookup_stmt (use_stmt)
9141 : 207 : || !vect_slp_is_lane_insert (use_stmt, def, &this_lane)
9142 : 229 : || !bb_vinfo->lookup_def (gimple_assign_rhs2 (use_stmt)))
9143 : : break;
9144 : 22 : if (bitmap_bit_p (lanes, this_lane))
9145 : : break;
9146 : 2 : lanes_found++;
9147 : 2 : bitmap_set_bit (lanes, this_lane);
9148 : 2 : gassign *use_ass = as_a <gassign *> (use_stmt);
9149 : 2 : lane_defs.quick_push (std::make_pair
9150 : 2 : (this_lane, gimple_assign_rhs2 (use_ass)));
9151 : 2 : last = bb_vinfo->lookup_stmt (use_ass);
9152 : 2 : roots.safe_push (last);
9153 : 2 : def = gimple_assign_lhs (use_ass);
9154 : : }
9155 : 2 : while (lanes_found < nlanes);
9156 : 211 : if (roots.length () > 1)
9157 : 2 : std::swap(roots[0], roots[roots.length () - 1]);
9158 : 211 : if (lanes_found < nlanes)
9159 : : {
9160 : : /* Now search the def chain. */
9161 : 211 : def = gimple_assign_rhs1 (assign);
9162 : 213 : do
9163 : : {
9164 : 213 : if (TREE_CODE (def) != SSA_NAME
9165 : 213 : || !has_single_use (def))
9166 : : break;
9167 : 56 : gimple *def_stmt = SSA_NAME_DEF_STMT (def);
9168 : 56 : unsigned this_lane;
9169 : 56 : if (!bb_vinfo->lookup_stmt (def_stmt)
9170 : 33 : || !vect_slp_is_lane_insert (def_stmt,
9171 : : NULL_TREE, &this_lane)
9172 : 80 : || !bb_vinfo->lookup_def (gimple_assign_rhs2 (def_stmt)))
9173 : : break;
9174 : 24 : if (bitmap_bit_p (lanes, this_lane))
9175 : : break;
9176 : 4 : lanes_found++;
9177 : 4 : bitmap_set_bit (lanes, this_lane);
9178 : 8 : lane_defs.quick_push (std::make_pair
9179 : 4 : (this_lane,
9180 : 4 : gimple_assign_rhs2 (def_stmt)));
9181 : 4 : roots.safe_push (bb_vinfo->lookup_stmt (def_stmt));
9182 : 4 : def = gimple_assign_rhs1 (def_stmt);
9183 : : }
9184 : 4 : while (lanes_found < nlanes);
9185 : : }
9186 : 211 : if (lanes_found == nlanes)
9187 : : {
9188 : : /* Sort lane_defs after the lane index and register the root. */
9189 : 2 : lane_defs.qsort (vld_cmp);
9190 : 2 : vec<stmt_vec_info> stmts;
9191 : 2 : stmts.create (nlanes);
9192 : 10 : for (unsigned i = 0; i < nlanes; ++i)
9193 : 8 : stmts.quick_push (bb_vinfo->lookup_def (lane_defs[i].second));
9194 : 2 : bb_vinfo->roots.safe_push (slp_root (slp_inst_kind_ctor,
9195 : 2 : stmts, roots));
9196 : : }
9197 : : else
9198 : 209 : roots.release ();
9199 : 211 : }
9200 : 27285145 : else if (!VECTOR_TYPE_P (TREE_TYPE (rhs))
9201 : 26427477 : && (associative_tree_code (code) || code == MINUS_EXPR)
9202 : : /* ??? This pessimizes a two-element reduction. PR54400.
9203 : : ??? In-order reduction could be handled if we only
9204 : : traverse one operand chain in vect_slp_linearize_chain. */
9205 : 30828830 : && !needs_fold_left_reduction_p (TREE_TYPE (rhs), code)
9206 : : /* Ops with constants at the tail can be stripped here. */
9207 : 5294392 : && TREE_CODE (rhs) == SSA_NAME
9208 : 5242971 : && TREE_CODE (gimple_assign_rhs2 (assign)) == SSA_NAME
9209 : : /* Should be the chain end. */
9210 : 29412067 : && (!single_imm_use (gimple_assign_lhs (assign),
9211 : : &use_p, &use_stmt)
9212 : 1638845 : || !is_gimple_assign (use_stmt)
9213 : 1092914 : || (gimple_assign_rhs_code (use_stmt) != code
9214 : 795497 : && ((code != PLUS_EXPR && code != MINUS_EXPR)
9215 : 431826 : || (gimple_assign_rhs_code (use_stmt)
9216 : 431826 : != (code == PLUS_EXPR ? MINUS_EXPR : PLUS_EXPR))))))
9217 : : {
9218 : : /* We start the match at the end of a possible association
9219 : : chain. */
9220 : 1750707 : auto_vec<chain_op_t> chain;
9221 : 1750707 : auto_vec<std::pair<tree_code, gimple *> > worklist;
9222 : 1750707 : auto_vec<gimple *> chain_stmts;
9223 : 1750707 : gimple *code_stmt = NULL, *alt_code_stmt = NULL;
9224 : 1750707 : if (code == MINUS_EXPR)
9225 : 280808 : code = PLUS_EXPR;
9226 : 1750707 : internal_fn reduc_fn;
9227 : 2020724 : if (!reduction_fn_for_scalar_code (code, &reduc_fn)
9228 : 1750707 : || reduc_fn == IFN_LAST)
9229 : 270017 : continue;
9230 : 1480690 : vect_slp_linearize_chain (bb_vinfo, worklist, chain, code, assign,
9231 : : /* ??? */
9232 : : code_stmt, alt_code_stmt, &chain_stmts);
9233 : 2961380 : if (chain.length () > 1)
9234 : : {
9235 : : /* Sort the chain according to def_type and operation. */
9236 : 1480690 : chain.sort (dt_sort_cmp, bb_vinfo);
9237 : : /* ??? Now we'd want to strip externals and constants
9238 : : but record those to be handled in the epilogue. */
9239 : : /* ??? For now do not allow mixing ops or externs/constants. */
9240 : 1480690 : bool invalid = false;
9241 : 1480690 : unsigned remain_cnt = 0;
9242 : 1480690 : unsigned last_idx = 0;
9243 : 4479308 : for (unsigned i = 0; i < chain.length (); ++i)
9244 : : {
9245 : 3299359 : if (chain[i].code != code)
9246 : : {
9247 : : invalid = true;
9248 : : break;
9249 : : }
9250 : 2998618 : if (chain[i].dt != vect_internal_def
9251 : : /* Avoid stmts where the def is not the LHS, like
9252 : : ASMs. */
9253 : 5785338 : || (gimple_get_lhs (bb_vinfo->lookup_def
9254 : 2786720 : (chain[i].op)->stmt)
9255 : 2786720 : != chain[i].op))
9256 : 214790 : remain_cnt++;
9257 : : else
9258 : : last_idx = i;
9259 : : }
9260 : : /* Make sure to have an even number of lanes as we later do
9261 : : all-or-nothing discovery, not trying to split further. */
9262 : 1480690 : if ((chain.length () - remain_cnt) & 1)
9263 : 171782 : remain_cnt++;
9264 : 1480690 : if (!invalid && chain.length () - remain_cnt > 1)
9265 : : {
9266 : 1117294 : vec<stmt_vec_info> stmts;
9267 : 1117294 : vec<tree> remain = vNULL;
9268 : 1117294 : stmts.create (chain.length ());
9269 : 1117294 : if (remain_cnt > 0)
9270 : 99738 : remain.create (remain_cnt);
9271 : 3593349 : for (unsigned i = 0; i < chain.length (); ++i)
9272 : : {
9273 : 2476055 : stmt_vec_info stmt_info;
9274 : 2476055 : if (chain[i].dt == vect_internal_def
9275 : 2443699 : && ((stmt_info = bb_vinfo->lookup_def (chain[i].op)),
9276 : 2443699 : gimple_get_lhs (stmt_info->stmt) == chain[i].op)
9277 : 4919670 : && (i != last_idx
9278 : 1117294 : || (stmts.length () & 1)))
9279 : 2365814 : stmts.quick_push (stmt_info);
9280 : : else
9281 : 110241 : remain.quick_push (chain[i].op);
9282 : : }
9283 : 1117294 : vec<stmt_vec_info> roots;
9284 : 1117294 : roots.create (chain_stmts.length ());
9285 : 2476055 : for (unsigned i = 0; i < chain_stmts.length (); ++i)
9286 : 1358761 : roots.quick_push (bb_vinfo->lookup_stmt (chain_stmts[i]));
9287 : 1117294 : bb_vinfo->roots.safe_push (slp_root (slp_inst_kind_bb_reduc,
9288 : 1117294 : stmts, roots, remain));
9289 : : }
9290 : : }
9291 : 1750707 : }
9292 : : }
9293 : 2297833 : }
9294 : :
9295 : : /* Walk the grouped store chains and replace entries with their
9296 : : pattern variant if any. */
9297 : :
9298 : : static void
9299 : 595902 : vect_fixup_store_groups_with_patterns (vec_info *vinfo)
9300 : : {
9301 : 595902 : stmt_vec_info first_element;
9302 : 595902 : unsigned i;
9303 : :
9304 : 1425572 : FOR_EACH_VEC_ELT (vinfo->grouped_stores, i, first_element)
9305 : : {
9306 : : /* We also have CTORs in this array. */
9307 : 829670 : if (!STMT_VINFO_GROUPED_ACCESS (first_element))
9308 : 0 : continue;
9309 : 829670 : if (STMT_VINFO_IN_PATTERN_P (first_element))
9310 : : {
9311 : 238 : stmt_vec_info orig = first_element;
9312 : 238 : first_element = STMT_VINFO_RELATED_STMT (first_element);
9313 : 238 : DR_GROUP_FIRST_ELEMENT (first_element) = first_element;
9314 : 238 : DR_GROUP_SIZE (first_element) = DR_GROUP_SIZE (orig);
9315 : 238 : DR_GROUP_GAP (first_element) = DR_GROUP_GAP (orig);
9316 : 238 : DR_GROUP_NEXT_ELEMENT (first_element) = DR_GROUP_NEXT_ELEMENT (orig);
9317 : 238 : vinfo->grouped_stores[i] = first_element;
9318 : : }
9319 : 829670 : stmt_vec_info prev = first_element;
9320 : 2305995 : while (DR_GROUP_NEXT_ELEMENT (prev))
9321 : : {
9322 : 1476325 : stmt_vec_info elt = DR_GROUP_NEXT_ELEMENT (prev);
9323 : 1476325 : if (STMT_VINFO_IN_PATTERN_P (elt))
9324 : : {
9325 : 855 : stmt_vec_info orig = elt;
9326 : 855 : elt = STMT_VINFO_RELATED_STMT (elt);
9327 : 855 : DR_GROUP_NEXT_ELEMENT (prev) = elt;
9328 : 855 : DR_GROUP_GAP (elt) = DR_GROUP_GAP (orig);
9329 : 855 : DR_GROUP_NEXT_ELEMENT (elt) = DR_GROUP_NEXT_ELEMENT (orig);
9330 : : }
9331 : 1476325 : DR_GROUP_FIRST_ELEMENT (elt) = first_element;
9332 : 1476325 : prev = elt;
9333 : : }
9334 : : }
9335 : 595902 : }
9336 : :
9337 : : /* Check if the region described by BB_VINFO can be vectorized, returning
9338 : : true if so. When returning false, set FATAL to true if the same failure
9339 : : would prevent vectorization at other vector sizes, false if it is still
9340 : : worth trying other sizes. N_STMTS is the number of statements in the
9341 : : region. */
9342 : :
9343 : : static bool
9344 : 2297833 : vect_slp_analyze_bb_1 (bb_vec_info bb_vinfo, int n_stmts, bool &fatal,
9345 : : vec<int> *dataref_groups)
9346 : : {
9347 : 2297833 : DUMP_VECT_SCOPE ("vect_slp_analyze_bb");
9348 : :
9349 : 2297833 : slp_instance instance;
9350 : 2297833 : int i;
9351 : 2297833 : poly_uint64 min_vf = 2;
9352 : :
9353 : : /* The first group of checks is independent of the vector size. */
9354 : 2297833 : fatal = true;
9355 : :
9356 : : /* Analyze the data references. */
9357 : :
9358 : 2297833 : if (!vect_analyze_data_refs (bb_vinfo, &min_vf, NULL))
9359 : : {
9360 : 0 : if (dump_enabled_p ())
9361 : 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9362 : : "not vectorized: unhandled data-ref in basic "
9363 : : "block.\n");
9364 : 0 : return false;
9365 : : }
9366 : :
9367 : 2297833 : if (!vect_analyze_data_ref_accesses (bb_vinfo, dataref_groups))
9368 : : {
9369 : 0 : if (dump_enabled_p ())
9370 : 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9371 : : "not vectorized: unhandled data access in "
9372 : : "basic block.\n");
9373 : 0 : return false;
9374 : : }
9375 : :
9376 : 2297833 : vect_slp_check_for_roots (bb_vinfo);
9377 : :
9378 : : /* If there are no grouped stores and no constructors in the region
9379 : : there is no need to continue with pattern recog as vect_analyze_slp
9380 : : will fail anyway. */
9381 : 2297833 : if (bb_vinfo->grouped_stores.is_empty ()
9382 : 1958644 : && bb_vinfo->roots.is_empty ())
9383 : : {
9384 : 1701931 : if (dump_enabled_p ())
9385 : 1000 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9386 : : "not vectorized: no grouped stores in "
9387 : : "basic block.\n");
9388 : 1701931 : return false;
9389 : : }
9390 : :
9391 : : /* While the rest of the analysis below depends on it in some way. */
9392 : 595902 : fatal = false;
9393 : :
9394 : 595902 : vect_pattern_recog (bb_vinfo);
9395 : :
9396 : : /* Update store groups from pattern processing. */
9397 : 595902 : vect_fixup_store_groups_with_patterns (bb_vinfo);
9398 : :
9399 : : /* Check the SLP opportunities in the basic block, analyze and build SLP
9400 : : trees. */
9401 : 595902 : if (!vect_analyze_slp (bb_vinfo, n_stmts, false))
9402 : : {
9403 : 0 : if (dump_enabled_p ())
9404 : : {
9405 : 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9406 : : "Failed to SLP the basic block.\n");
9407 : 0 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9408 : : "not vectorized: failed to find SLP opportunities "
9409 : : "in basic block.\n");
9410 : : }
9411 : 0 : return false;
9412 : : }
9413 : :
9414 : : /* Optimize permutations. */
9415 : 595902 : vect_optimize_slp (bb_vinfo);
9416 : :
9417 : : /* Gather the loads reachable from the SLP graph entries. */
9418 : 595902 : vect_gather_slp_loads (bb_vinfo);
9419 : :
9420 : 595902 : vect_record_base_alignments (bb_vinfo);
9421 : :
9422 : : /* Analyze and verify the alignment of data references and the
9423 : : dependence in the SLP instances. */
9424 : 1358265 : for (i = 0; BB_VINFO_SLP_INSTANCES (bb_vinfo).iterate (i, &instance); )
9425 : : {
9426 : 762363 : vect_location = instance->location ();
9427 : 762363 : if (! vect_slp_analyze_instance_alignment (bb_vinfo, instance)
9428 : 762363 : || ! vect_slp_analyze_instance_dependence (bb_vinfo, instance))
9429 : : {
9430 : 6090 : slp_tree node = SLP_INSTANCE_TREE (instance);
9431 : 6090 : stmt_vec_info stmt_info = SLP_TREE_SCALAR_STMTS (node)[0];
9432 : 6090 : if (dump_enabled_p ())
9433 : 6 : dump_printf_loc (MSG_NOTE, vect_location,
9434 : : "removing SLP instance operations starting from: %G",
9435 : : stmt_info->stmt);
9436 : 6090 : vect_free_slp_instance (instance);
9437 : 6090 : BB_VINFO_SLP_INSTANCES (bb_vinfo).ordered_remove (i);
9438 : 6090 : continue;
9439 : 6090 : }
9440 : :
9441 : : /* Mark all the statements that we want to vectorize as pure SLP and
9442 : : relevant. */
9443 : 756273 : vect_mark_slp_stmts (bb_vinfo, SLP_INSTANCE_TREE (instance));
9444 : 756273 : vect_mark_slp_stmts_relevant (SLP_INSTANCE_TREE (instance));
9445 : 756273 : unsigned j;
9446 : 756273 : stmt_vec_info root;
9447 : : /* Likewise consider instance root stmts as vectorized. */
9448 : 1654788 : FOR_EACH_VEC_ELT (SLP_INSTANCE_ROOT_STMTS (instance), j, root)
9449 : 142242 : STMT_SLP_TYPE (root) = pure_slp;
9450 : :
9451 : 756273 : i++;
9452 : : }
9453 : 2332657 : if (! BB_VINFO_SLP_INSTANCES (bb_vinfo).length ())
9454 : : return false;
9455 : :
9456 : 314290 : if (!vect_slp_analyze_operations (bb_vinfo))
9457 : : {
9458 : 34824 : if (dump_enabled_p ())
9459 : 93 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9460 : : "not vectorized: bad operation in basic block.\n");
9461 : 34824 : return false;
9462 : : }
9463 : :
9464 : 279466 : vect_bb_partition_graph (bb_vinfo);
9465 : :
9466 : 279466 : return true;
9467 : : }
9468 : :
9469 : : /* Subroutine of vect_slp_bb. Try to vectorize the statements for all
9470 : : basic blocks in BBS, returning true on success.
9471 : : The region has N_STMTS statements and has the datarefs given by DATAREFS. */
9472 : :
9473 : : static bool
9474 : 2004560 : vect_slp_region (vec<basic_block> bbs, vec<data_reference_p> datarefs,
9475 : : vec<int> *dataref_groups, unsigned int n_stmts,
9476 : : loop_p orig_loop)
9477 : : {
9478 : 2004560 : bb_vec_info bb_vinfo;
9479 : 2004560 : auto_vector_modes vector_modes;
9480 : :
9481 : : /* Autodetect first vector size we try. */
9482 : 2004560 : machine_mode next_vector_mode = VOIDmode;
9483 : 2004560 : targetm.vectorize.autovectorize_vector_modes (&vector_modes, false);
9484 : 2004560 : unsigned int mode_i = 0;
9485 : :
9486 : 2004560 : vec_info_shared shared;
9487 : :
9488 : 2004560 : machine_mode autodetected_vector_mode = VOIDmode;
9489 : 2591106 : while (1)
9490 : : {
9491 : 2297833 : bool vectorized = false;
9492 : 2297833 : bool fatal = false;
9493 : 2297833 : bb_vinfo = new _bb_vec_info (bbs, &shared);
9494 : :
9495 : 2297833 : bool first_time_p = shared.datarefs.is_empty ();
9496 : 2297833 : BB_VINFO_DATAREFS (bb_vinfo) = datarefs;
9497 : 2297833 : if (first_time_p)
9498 : 2027028 : bb_vinfo->shared->save_datarefs ();
9499 : : else
9500 : 270805 : bb_vinfo->shared->check_datarefs ();
9501 : 2297833 : bb_vinfo->vector_mode = next_vector_mode;
9502 : :
9503 : 2297833 : if (vect_slp_analyze_bb_1 (bb_vinfo, n_stmts, fatal, dataref_groups))
9504 : : {
9505 : 279466 : if (dump_enabled_p ())
9506 : : {
9507 : 1430 : dump_printf_loc (MSG_NOTE, vect_location,
9508 : : "***** Analysis succeeded with vector mode"
9509 : 715 : " %s\n", GET_MODE_NAME (bb_vinfo->vector_mode));
9510 : 715 : dump_printf_loc (MSG_NOTE, vect_location, "SLPing BB part\n");
9511 : : }
9512 : :
9513 : 279466 : bb_vinfo->shared->check_datarefs ();
9514 : :
9515 : 279466 : bool force_clear = false;
9516 : 279466 : auto_vec<slp_instance> profitable_subgraphs;
9517 : 1497046 : for (slp_instance instance : BB_VINFO_SLP_INSTANCES (bb_vinfo))
9518 : : {
9519 : 658648 : if (instance->subgraph_entries.is_empty ())
9520 : 199808 : continue;
9521 : :
9522 : 647328 : dump_user_location_t saved_vect_location = vect_location;
9523 : 647328 : vect_location = instance->location ();
9524 : 647328 : if (!unlimited_cost_model (NULL)
9525 : 1291165 : && !vect_bb_vectorization_profitable_p
9526 : 643837 : (bb_vinfo, instance->subgraph_entries, orig_loop))
9527 : : {
9528 : 177168 : if (dump_enabled_p ())
9529 : 14 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9530 : : "not vectorized: vectorization is not "
9531 : : "profitable.\n");
9532 : 177168 : vect_location = saved_vect_location;
9533 : 177168 : continue;
9534 : : }
9535 : :
9536 : 470160 : vect_location = saved_vect_location;
9537 : 470160 : if (!dbg_cnt (vect_slp))
9538 : : {
9539 : 0 : force_clear = true;
9540 : 0 : continue;
9541 : : }
9542 : :
9543 : 470160 : profitable_subgraphs.safe_push (instance);
9544 : : }
9545 : :
9546 : : /* When we're vectorizing an if-converted loop body make sure
9547 : : we vectorized all if-converted code. */
9548 : 455535 : if ((!profitable_subgraphs.is_empty () || force_clear) && orig_loop)
9549 : : {
9550 : 125 : gcc_assert (bb_vinfo->nbbs == 1);
9551 : 250 : for (gimple_stmt_iterator gsi = gsi_start_bb (bb_vinfo->bbs[0]);
9552 : 5244 : !gsi_end_p (gsi); gsi_next (&gsi))
9553 : : {
9554 : : /* The costing above left us with DCEable vectorized scalar
9555 : : stmts having the visited flag set on profitable
9556 : : subgraphs. Do the delayed clearing of the flag here. */
9557 : 5119 : if (gimple_visited_p (gsi_stmt (gsi)))
9558 : : {
9559 : 1304 : gimple_set_visited (gsi_stmt (gsi), false);
9560 : 1304 : continue;
9561 : : }
9562 : 3815 : if (flag_vect_cost_model == VECT_COST_MODEL_UNLIMITED)
9563 : 868 : continue;
9564 : :
9565 : 7538 : if (gassign *ass = dyn_cast <gassign *> (gsi_stmt (gsi)))
9566 : 3100 : if (gimple_assign_rhs_code (ass) == COND_EXPR)
9567 : : {
9568 : 142 : if (!profitable_subgraphs.is_empty ()
9569 : 48 : && dump_enabled_p ())
9570 : 0 : dump_printf_loc (MSG_NOTE, vect_location,
9571 : : "not profitable because of "
9572 : : "unprofitable if-converted scalar "
9573 : : "code\n");
9574 : 94 : profitable_subgraphs.truncate (0);
9575 : : }
9576 : : }
9577 : : }
9578 : :
9579 : : /* Finally schedule the profitable subgraphs. */
9580 : 1101694 : for (slp_instance instance : profitable_subgraphs)
9581 : : {
9582 : 470090 : if (!vectorized && dump_enabled_p ())
9583 : 702 : dump_printf_loc (MSG_NOTE, vect_location,
9584 : : "Basic block will be vectorized "
9585 : : "using SLP\n");
9586 : 470090 : vectorized = true;
9587 : :
9588 : : /* Dump before scheduling as store vectorization will remove
9589 : : the original stores and mess with the instance tree
9590 : : so querying its location will eventually ICE. */
9591 : 470090 : if (flag_checking)
9592 : 1886684 : for (slp_instance sub : instance->subgraph_entries)
9593 : 476414 : gcc_assert (SLP_TREE_VECTYPE (SLP_INSTANCE_TREE (sub)));
9594 : 470090 : unsigned HOST_WIDE_INT bytes;
9595 : 470090 : if (dump_enabled_p ())
9596 : 3327 : for (slp_instance sub : instance->subgraph_entries)
9597 : : {
9598 : 882 : tree vtype = SLP_TREE_VECTYPE (SLP_INSTANCE_TREE (sub));
9599 : 1764 : if (GET_MODE_SIZE (TYPE_MODE (vtype)).is_constant (&bytes))
9600 : 882 : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS,
9601 : 882 : sub->location (),
9602 : : "basic block part vectorized using %wu "
9603 : : "byte vectors\n", bytes);
9604 : : else
9605 : : dump_printf_loc (MSG_OPTIMIZED_LOCATIONS,
9606 : : sub->location (),
9607 : : "basic block part vectorized using "
9608 : : "variable length vectors\n");
9609 : : }
9610 : :
9611 : 470090 : dump_user_location_t saved_vect_location = vect_location;
9612 : 470090 : vect_location = instance->location ();
9613 : :
9614 : 470090 : vect_schedule_slp (bb_vinfo, instance->subgraph_entries);
9615 : :
9616 : 470090 : vect_location = saved_vect_location;
9617 : : }
9618 : :
9619 : :
9620 : : /* Generate the invariant statements. */
9621 : 279466 : if (!gimple_seq_empty_p (bb_vinfo->inv_pattern_def_seq))
9622 : : {
9623 : 45 : if (dump_enabled_p ())
9624 : 0 : dump_printf_loc (MSG_NOTE, vect_location,
9625 : : "------>generating invariant statements\n");
9626 : :
9627 : 45 : bb_vinfo->insert_seq_on_entry (NULL,
9628 : : bb_vinfo->inv_pattern_def_seq);
9629 : : }
9630 : 279466 : }
9631 : : else
9632 : : {
9633 : 2018367 : if (dump_enabled_p ())
9634 : 1359 : dump_printf_loc (MSG_NOTE, vect_location,
9635 : : "***** Analysis failed with vector mode %s\n",
9636 : 1359 : GET_MODE_NAME (bb_vinfo->vector_mode));
9637 : : }
9638 : :
9639 : 2297833 : if (mode_i == 0)
9640 : 2004560 : autodetected_vector_mode = bb_vinfo->vector_mode;
9641 : :
9642 : 2297833 : if (!fatal)
9643 : 3232476 : while (mode_i < vector_modes.length ()
9644 : 1742584 : && vect_chooses_same_modes_p (bb_vinfo, vector_modes[mode_i]))
9645 : : {
9646 : 338741 : if (dump_enabled_p ())
9647 : 1646 : dump_printf_loc (MSG_NOTE, vect_location,
9648 : : "***** The result for vector mode %s would"
9649 : : " be the same\n",
9650 : 823 : GET_MODE_NAME (vector_modes[mode_i]));
9651 : 338741 : mode_i += 1;
9652 : : }
9653 : :
9654 : 2297833 : delete bb_vinfo;
9655 : :
9656 : 2297833 : if (mode_i < vector_modes.length ()
9657 : 2133000 : && VECTOR_MODE_P (autodetected_vector_mode)
9658 : 1994138 : && (related_vector_mode (vector_modes[mode_i],
9659 : : GET_MODE_INNER (autodetected_vector_mode))
9660 : 997069 : == autodetected_vector_mode)
9661 : 4430833 : && (related_vector_mode (autodetected_vector_mode,
9662 : 527392 : GET_MODE_INNER (vector_modes[mode_i]))
9663 : 1054784 : == vector_modes[mode_i]))
9664 : : {
9665 : 527392 : if (dump_enabled_p ())
9666 : 232 : dump_printf_loc (MSG_NOTE, vect_location,
9667 : : "***** Skipping vector mode %s, which would"
9668 : : " repeat the analysis for %s\n",
9669 : 232 : GET_MODE_NAME (vector_modes[mode_i]),
9670 : 232 : GET_MODE_NAME (autodetected_vector_mode));
9671 : 527392 : mode_i += 1;
9672 : : }
9673 : :
9674 : 2297833 : if (vectorized
9675 : 2121812 : || mode_i == vector_modes.length ()
9676 : 1957073 : || autodetected_vector_mode == VOIDmode
9677 : : /* If vect_slp_analyze_bb_1 signaled that analysis for all
9678 : : vector sizes will fail do not bother iterating. */
9679 : 3118975 : || fatal)
9680 : 4009120 : return vectorized;
9681 : :
9682 : : /* Try the next biggest vector size. */
9683 : 293273 : next_vector_mode = vector_modes[mode_i++];
9684 : 293273 : if (dump_enabled_p ())
9685 : 255 : dump_printf_loc (MSG_NOTE, vect_location,
9686 : : "***** Re-trying analysis with vector mode %s\n",
9687 : 255 : GET_MODE_NAME (next_vector_mode));
9688 : 293273 : }
9689 : 2004560 : }
9690 : :
9691 : :
9692 : : /* Main entry for the BB vectorizer. Analyze and transform BBS, returns
9693 : : true if anything in the basic-block was vectorized. */
9694 : :
9695 : : static bool
9696 : 2004560 : vect_slp_bbs (const vec<basic_block> &bbs, loop_p orig_loop)
9697 : : {
9698 : 2004560 : vec<data_reference_p> datarefs = vNULL;
9699 : 2004560 : auto_vec<int> dataref_groups;
9700 : 2004560 : int insns = 0;
9701 : 2004560 : int current_group = 0;
9702 : :
9703 : 12284787 : for (unsigned i = 0; i < bbs.length (); i++)
9704 : : {
9705 : 10280227 : basic_block bb = bbs[i];
9706 : 78536450 : for (gimple_stmt_iterator gsi = gsi_after_labels (bb); !gsi_end_p (gsi);
9707 : 68256223 : gsi_next (&gsi))
9708 : : {
9709 : 68256223 : gimple *stmt = gsi_stmt (gsi);
9710 : 68256223 : if (is_gimple_debug (stmt))
9711 : 40199152 : continue;
9712 : :
9713 : 28057071 : insns++;
9714 : :
9715 : 28057071 : if (gimple_location (stmt) != UNKNOWN_LOCATION)
9716 : 25522306 : vect_location = stmt;
9717 : :
9718 : 28057071 : if (!vect_find_stmt_data_reference (NULL, stmt, &datarefs,
9719 : 28057071 : &dataref_groups, current_group))
9720 : 4883430 : ++current_group;
9721 : : }
9722 : : /* New BBs always start a new DR group. */
9723 : 10280227 : ++current_group;
9724 : : }
9725 : :
9726 : 2004560 : return vect_slp_region (bbs, datarefs, &dataref_groups, insns, orig_loop);
9727 : 2004560 : }
9728 : :
9729 : : /* Special entry for the BB vectorizer. Analyze and transform a single
9730 : : if-converted BB with ORIG_LOOPs body being the not if-converted
9731 : : representation. Returns true if anything in the basic-block was
9732 : : vectorized. */
9733 : :
9734 : : bool
9735 : 16421 : vect_slp_if_converted_bb (basic_block bb, loop_p orig_loop)
9736 : : {
9737 : 16421 : auto_vec<basic_block> bbs;
9738 : 16421 : bbs.safe_push (bb);
9739 : 16421 : return vect_slp_bbs (bbs, orig_loop);
9740 : 16421 : }
9741 : :
9742 : : /* Main entry for the BB vectorizer. Analyze and transform BB, returns
9743 : : true if anything in the basic-block was vectorized. */
9744 : :
9745 : : bool
9746 : 872849 : vect_slp_function (function *fun)
9747 : : {
9748 : 872849 : bool r = false;
9749 : 872849 : int *rpo = XNEWVEC (int, n_basic_blocks_for_fn (fun));
9750 : 872849 : auto_bitmap exit_bbs;
9751 : 872849 : bitmap_set_bit (exit_bbs, EXIT_BLOCK);
9752 : 872849 : edge entry = single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (fun));
9753 : 872849 : unsigned n = rev_post_order_and_mark_dfs_back_seme (fun, entry, exit_bbs,
9754 : 872849 : true, rpo, NULL);
9755 : :
9756 : : /* For the moment split the function into pieces to avoid making
9757 : : the iteration on the vector mode moot. Split at points we know
9758 : : to not handle well which is CFG merges (SLP discovery doesn't
9759 : : handle non-loop-header PHIs) and loop exits. Since pattern
9760 : : recog requires reverse iteration to visit uses before defs
9761 : : simply chop RPO into pieces. */
9762 : 872849 : auto_vec<basic_block> bbs;
9763 : 11162242 : for (unsigned i = 0; i < n; i++)
9764 : : {
9765 : 10289393 : basic_block bb = BASIC_BLOCK_FOR_FN (fun, rpo[i]);
9766 : 10289393 : bool split = false;
9767 : :
9768 : : /* Split when a BB is not dominated by the first block. */
9769 : 19355147 : if (!bbs.is_empty ()
9770 : 9065754 : && !dominated_by_p (CDI_DOMINATORS, bb, bbs[0]))
9771 : : {
9772 : 779333 : if (dump_enabled_p ())
9773 : 105 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9774 : : "splitting region at dominance boundary bb%d\n",
9775 : : bb->index);
9776 : : split = true;
9777 : : }
9778 : : /* Split when the loop determined by the first block
9779 : : is exited. This is because we eventually insert
9780 : : invariants at region begin. */
9781 : 17796481 : else if (!bbs.is_empty ()
9782 : 8286421 : && bbs[0]->loop_father != bb->loop_father
9783 : 2074347 : && !flow_loop_nested_p (bbs[0]->loop_father, bb->loop_father))
9784 : : {
9785 : 4918 : if (dump_enabled_p ())
9786 : 6 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9787 : : "splitting region at loop %d exit at bb%d\n",
9788 : 3 : bbs[0]->loop_father->num, bb->index);
9789 : : split = true;
9790 : : }
9791 : 9505142 : else if (!bbs.is_empty ()
9792 : 8281503 : && bb->loop_father->header == bb
9793 : 435959 : && bb->loop_father->dont_vectorize)
9794 : : {
9795 : 5836 : if (dump_enabled_p ())
9796 : 66 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9797 : : "splitting region at dont-vectorize loop %d "
9798 : : "entry at bb%d\n",
9799 : : bb->loop_father->num, bb->index);
9800 : : split = true;
9801 : : }
9802 : :
9803 : 11079480 : if (split && !bbs.is_empty ())
9804 : : {
9805 : 790087 : r |= vect_slp_bbs (bbs, NULL);
9806 : 790087 : bbs.truncate (0);
9807 : : }
9808 : :
9809 : 10289393 : if (bbs.is_empty ())
9810 : : {
9811 : : /* We need to be able to insert at the head of the region which
9812 : : we cannot for region starting with a returns-twice call. */
9813 : 2013726 : if (gcall *first = safe_dyn_cast <gcall *> (first_stmt (bb)))
9814 : 380920 : if (gimple_call_flags (first) & ECF_RETURNS_TWICE)
9815 : : {
9816 : 287 : if (dump_enabled_p ())
9817 : 2 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9818 : : "skipping bb%d as start of region as it "
9819 : : "starts with returns-twice call\n",
9820 : : bb->index);
9821 : 25587 : continue;
9822 : : }
9823 : : /* If the loop this BB belongs to is marked as not to be vectorized
9824 : : honor that also for BB vectorization. */
9825 : 2013439 : if (bb->loop_father->dont_vectorize)
9826 : 25300 : continue;
9827 : : }
9828 : :
9829 : 10263806 : bbs.safe_push (bb);
9830 : :
9831 : : /* When we have a stmt ending this block and defining a
9832 : : value we have to insert on edges when inserting after it for
9833 : : a vector containing its definition. Avoid this for now. */
9834 : 20527612 : if (gimple *last = *gsi_last_bb (bb))
9835 : 8156996 : if (gimple_get_lhs (last)
9836 : 8156996 : && is_ctrl_altering_stmt (last))
9837 : : {
9838 : 325210 : if (dump_enabled_p ())
9839 : 2 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
9840 : : "splitting region at control altering "
9841 : : "definition %G", last);
9842 : 325210 : r |= vect_slp_bbs (bbs, NULL);
9843 : 325210 : bbs.truncate (0);
9844 : : }
9845 : : }
9846 : :
9847 : 872849 : if (!bbs.is_empty ())
9848 : 872842 : r |= vect_slp_bbs (bbs, NULL);
9849 : :
9850 : 872849 : free (rpo);
9851 : :
9852 : 872849 : return r;
9853 : 872849 : }
9854 : :
9855 : : /* Build a variable-length vector in which the elements in ELTS are repeated
9856 : : to a fill NRESULTS vectors of type VECTOR_TYPE. Store the vectors in
9857 : : RESULTS and add any new instructions to SEQ.
9858 : :
9859 : : The approach we use is:
9860 : :
9861 : : (1) Find a vector mode VM with integer elements of mode IM.
9862 : :
9863 : : (2) Replace ELTS[0:NELTS] with ELTS'[0:NELTS'], where each element of
9864 : : ELTS' has mode IM. This involves creating NELTS' VIEW_CONVERT_EXPRs
9865 : : from small vectors to IM.
9866 : :
9867 : : (3) Duplicate each ELTS'[I] into a vector of mode VM.
9868 : :
9869 : : (4) Use a tree of interleaving VEC_PERM_EXPRs to create VMs with the
9870 : : correct byte contents.
9871 : :
9872 : : (5) Use VIEW_CONVERT_EXPR to cast the final VMs to the required type.
9873 : :
9874 : : We try to find the largest IM for which this sequence works, in order
9875 : : to cut down on the number of interleaves. */
9876 : :
9877 : : void
9878 : 0 : duplicate_and_interleave (vec_info *vinfo, gimple_seq *seq, tree vector_type,
9879 : : const vec<tree> &elts, unsigned int nresults,
9880 : : vec<tree> &results)
9881 : : {
9882 : 0 : unsigned int nelts = elts.length ();
9883 : 0 : tree element_type = TREE_TYPE (vector_type);
9884 : :
9885 : : /* (1) Find a vector mode VM with integer elements of mode IM. */
9886 : 0 : unsigned int nvectors = 1;
9887 : 0 : tree new_vector_type;
9888 : 0 : tree permutes[2];
9889 : 0 : if (!can_duplicate_and_interleave_p (vinfo, nelts, element_type,
9890 : : &nvectors, &new_vector_type,
9891 : : permutes))
9892 : 0 : gcc_unreachable ();
9893 : :
9894 : : /* Get a vector type that holds ELTS[0:NELTS/NELTS']. */
9895 : 0 : unsigned int partial_nelts = nelts / nvectors;
9896 : 0 : tree partial_vector_type = build_vector_type (element_type, partial_nelts);
9897 : :
9898 : 0 : tree_vector_builder partial_elts;
9899 : 0 : auto_vec<tree, 32> pieces (nvectors * 2);
9900 : 0 : pieces.quick_grow_cleared (nvectors * 2);
9901 : 0 : for (unsigned int i = 0; i < nvectors; ++i)
9902 : : {
9903 : : /* (2) Replace ELTS[0:NELTS] with ELTS'[0:NELTS'], where each element of
9904 : : ELTS' has mode IM. */
9905 : 0 : partial_elts.new_vector (partial_vector_type, partial_nelts, 1);
9906 : 0 : for (unsigned int j = 0; j < partial_nelts; ++j)
9907 : 0 : partial_elts.quick_push (elts[i * partial_nelts + j]);
9908 : 0 : tree t = gimple_build_vector (seq, &partial_elts);
9909 : 0 : t = gimple_build (seq, VIEW_CONVERT_EXPR,
9910 : 0 : TREE_TYPE (new_vector_type), t);
9911 : :
9912 : : /* (3) Duplicate each ELTS'[I] into a vector of mode VM. */
9913 : 0 : pieces[i] = gimple_build_vector_from_val (seq, new_vector_type, t);
9914 : : }
9915 : :
9916 : : /* (4) Use a tree of VEC_PERM_EXPRs to create a single VM with the
9917 : : correct byte contents.
9918 : :
9919 : : Conceptually, we need to repeat the following operation log2(nvectors)
9920 : : times, where hi_start = nvectors / 2:
9921 : :
9922 : : out[i * 2] = VEC_PERM_EXPR (in[i], in[i + hi_start], lo_permute);
9923 : : out[i * 2 + 1] = VEC_PERM_EXPR (in[i], in[i + hi_start], hi_permute);
9924 : :
9925 : : However, if each input repeats every N elements and the VF is
9926 : : a multiple of N * 2, the HI result is the same as the LO result.
9927 : : This will be true for the first N1 iterations of the outer loop,
9928 : : followed by N2 iterations for which both the LO and HI results
9929 : : are needed. I.e.:
9930 : :
9931 : : N1 + N2 = log2(nvectors)
9932 : :
9933 : : Each "N1 iteration" doubles the number of redundant vectors and the
9934 : : effect of the process as a whole is to have a sequence of nvectors/2**N1
9935 : : vectors that repeats 2**N1 times. Rather than generate these redundant
9936 : : vectors, we halve the number of vectors for each N1 iteration. */
9937 : : unsigned int in_start = 0;
9938 : : unsigned int out_start = nvectors;
9939 : : unsigned int new_nvectors = nvectors;
9940 : 0 : for (unsigned int in_repeat = 1; in_repeat < nvectors; in_repeat *= 2)
9941 : : {
9942 : 0 : unsigned int hi_start = new_nvectors / 2;
9943 : 0 : unsigned int out_i = 0;
9944 : 0 : for (unsigned int in_i = 0; in_i < new_nvectors; ++in_i)
9945 : : {
9946 : 0 : if ((in_i & 1) != 0
9947 : 0 : && multiple_p (TYPE_VECTOR_SUBPARTS (new_vector_type),
9948 : : 2 * in_repeat))
9949 : 0 : continue;
9950 : :
9951 : 0 : tree output = make_ssa_name (new_vector_type);
9952 : 0 : tree input1 = pieces[in_start + (in_i / 2)];
9953 : 0 : tree input2 = pieces[in_start + (in_i / 2) + hi_start];
9954 : 0 : gassign *stmt = gimple_build_assign (output, VEC_PERM_EXPR,
9955 : : input1, input2,
9956 : : permutes[in_i & 1]);
9957 : 0 : gimple_seq_add_stmt (seq, stmt);
9958 : 0 : pieces[out_start + out_i] = output;
9959 : 0 : out_i += 1;
9960 : : }
9961 : 0 : std::swap (in_start, out_start);
9962 : 0 : new_nvectors = out_i;
9963 : : }
9964 : :
9965 : : /* (5) Use VIEW_CONVERT_EXPR to cast the final VM to the required type. */
9966 : 0 : results.reserve (nresults);
9967 : 0 : for (unsigned int i = 0; i < nresults; ++i)
9968 : 0 : if (i < new_nvectors)
9969 : 0 : results.quick_push (gimple_build (seq, VIEW_CONVERT_EXPR, vector_type,
9970 : 0 : pieces[in_start + i]));
9971 : : else
9972 : 0 : results.quick_push (results[i - new_nvectors]);
9973 : 0 : }
9974 : :
9975 : :
9976 : : /* For constant and loop invariant defs in OP_NODE this function creates
9977 : : vector defs that will be used in the vectorized stmts and stores them
9978 : : to SLP_TREE_VEC_DEFS of OP_NODE. */
9979 : :
9980 : : static void
9981 : 471079 : vect_create_constant_vectors (vec_info *vinfo, slp_tree op_node)
9982 : : {
9983 : 471079 : unsigned HOST_WIDE_INT nunits;
9984 : 471079 : tree vec_cst;
9985 : 471079 : unsigned j, number_of_places_left_in_vector;
9986 : 471079 : tree vector_type;
9987 : 471079 : tree vop;
9988 : 471079 : int group_size = op_node->ops.length ();
9989 : 471079 : unsigned int vec_num, i;
9990 : 471079 : unsigned number_of_copies = 1;
9991 : 471079 : bool constant_p;
9992 : 471079 : gimple_seq ctor_seq = NULL;
9993 : 471079 : auto_vec<tree, 16> permute_results;
9994 : :
9995 : : /* We always want SLP_TREE_VECTYPE (op_node) here correctly set. */
9996 : 471079 : vector_type = SLP_TREE_VECTYPE (op_node);
9997 : :
9998 : 471079 : unsigned int number_of_vectors = SLP_TREE_NUMBER_OF_VEC_STMTS (op_node);
9999 : 471079 : SLP_TREE_VEC_DEFS (op_node).create (number_of_vectors);
10000 : 471079 : auto_vec<tree> voprnds (number_of_vectors);
10001 : :
10002 : : /* NUMBER_OF_COPIES is the number of times we need to use the same values in
10003 : : created vectors. It is greater than 1 if unrolling is performed.
10004 : :
10005 : : For example, we have two scalar operands, s1 and s2 (e.g., group of
10006 : : strided accesses of size two), while NUNITS is four (i.e., four scalars
10007 : : of this type can be packed in a vector). The output vector will contain
10008 : : two copies of each scalar operand: {s1, s2, s1, s2}. (NUMBER_OF_COPIES
10009 : : will be 2).
10010 : :
10011 : : If GROUP_SIZE > NUNITS, the scalars will be split into several vectors
10012 : : containing the operands.
10013 : :
10014 : : For example, NUNITS is four as before, and the group size is 8
10015 : : (s1, s2, ..., s8). We will create two vectors {s1, s2, s3, s4} and
10016 : : {s5, s6, s7, s8}. */
10017 : :
10018 : : /* When using duplicate_and_interleave, we just need one element for
10019 : : each scalar statement. */
10020 : 471079 : if (!TYPE_VECTOR_SUBPARTS (vector_type).is_constant (&nunits))
10021 : : nunits = group_size;
10022 : :
10023 : 471079 : number_of_copies = nunits * number_of_vectors / group_size;
10024 : :
10025 : 471079 : number_of_places_left_in_vector = nunits;
10026 : 471079 : constant_p = true;
10027 : 471079 : tree uniform_elt = NULL_TREE;
10028 : 471079 : tree_vector_builder elts (vector_type, nunits, 1);
10029 : 471079 : elts.quick_grow (nunits);
10030 : 471079 : stmt_vec_info insert_after = NULL;
10031 : 1309434 : for (j = 0; j < number_of_copies; j++)
10032 : : {
10033 : 838355 : tree op;
10034 : 3290860 : for (i = group_size - 1; op_node->ops.iterate (i, &op); i--)
10035 : : {
10036 : : /* Create 'vect_ = {op0,op1,...,opn}'. */
10037 : 1614150 : tree orig_op = op;
10038 : 1614150 : if (number_of_places_left_in_vector == nunits)
10039 : : uniform_elt = op;
10040 : 1028521 : else if (uniform_elt && operand_equal_p (uniform_elt, op))
10041 : 619594 : op = elts[number_of_places_left_in_vector];
10042 : : else
10043 : : uniform_elt = NULL_TREE;
10044 : 1614150 : number_of_places_left_in_vector--;
10045 : 1614150 : if (!types_compatible_p (TREE_TYPE (vector_type), TREE_TYPE (op)))
10046 : : {
10047 : 274420 : if (CONSTANT_CLASS_P (op))
10048 : : {
10049 : 96359 : if (VECTOR_BOOLEAN_TYPE_P (vector_type))
10050 : : {
10051 : : /* Can't use VIEW_CONVERT_EXPR for booleans because
10052 : : of possibly different sizes of scalar value and
10053 : : vector element. */
10054 : 76 : if (integer_zerop (op))
10055 : 76 : op = build_int_cst (TREE_TYPE (vector_type), 0);
10056 : 0 : else if (integer_onep (op))
10057 : 0 : op = build_all_ones_cst (TREE_TYPE (vector_type));
10058 : : else
10059 : 0 : gcc_unreachable ();
10060 : : }
10061 : : else
10062 : 96283 : op = fold_unary (VIEW_CONVERT_EXPR,
10063 : : TREE_TYPE (vector_type), op);
10064 : 96359 : gcc_assert (op && CONSTANT_CLASS_P (op));
10065 : : }
10066 : : else
10067 : : {
10068 : 178061 : tree new_temp = make_ssa_name (TREE_TYPE (vector_type));
10069 : 178061 : gimple *init_stmt;
10070 : 178061 : if (VECTOR_BOOLEAN_TYPE_P (vector_type))
10071 : : {
10072 : 391 : tree true_val
10073 : 391 : = build_all_ones_cst (TREE_TYPE (vector_type));
10074 : 391 : tree false_val
10075 : 391 : = build_zero_cst (TREE_TYPE (vector_type));
10076 : 391 : gcc_assert (INTEGRAL_TYPE_P (TREE_TYPE (op)));
10077 : 391 : init_stmt = gimple_build_assign (new_temp, COND_EXPR,
10078 : : op, true_val,
10079 : : false_val);
10080 : : }
10081 : : else
10082 : : {
10083 : 177670 : op = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (vector_type),
10084 : : op);
10085 : 177670 : init_stmt
10086 : 177670 : = gimple_build_assign (new_temp, VIEW_CONVERT_EXPR,
10087 : : op);
10088 : : }
10089 : 178061 : gimple_seq_add_stmt (&ctor_seq, init_stmt);
10090 : 178061 : op = new_temp;
10091 : : }
10092 : : }
10093 : 1614150 : elts[number_of_places_left_in_vector] = op;
10094 : 1614150 : if (!CONSTANT_CLASS_P (op))
10095 : 326739 : constant_p = false;
10096 : : /* For BB vectorization we have to compute an insert location
10097 : : when a def is inside the analyzed region since we cannot
10098 : : simply insert at the BB start in this case. */
10099 : 1614150 : stmt_vec_info opdef;
10100 : 1614150 : if (TREE_CODE (orig_op) == SSA_NAME
10101 : 204177 : && !SSA_NAME_IS_DEFAULT_DEF (orig_op)
10102 : 181036 : && is_a <bb_vec_info> (vinfo)
10103 : 1743503 : && (opdef = vinfo->lookup_def (orig_op)))
10104 : : {
10105 : 82994 : if (!insert_after)
10106 : : insert_after = opdef;
10107 : : else
10108 : 46322 : insert_after = get_later_stmt (insert_after, opdef);
10109 : : }
10110 : :
10111 : 1614150 : if (number_of_places_left_in_vector == 0)
10112 : : {
10113 : 585629 : auto type_nunits = TYPE_VECTOR_SUBPARTS (vector_type);
10114 : 585629 : if (uniform_elt)
10115 : 606450 : vec_cst = gimple_build_vector_from_val (&ctor_seq, vector_type,
10116 : 303225 : elts[0]);
10117 : 564808 : else if (constant_p
10118 : 564808 : ? multiple_p (type_nunits, nunits)
10119 : 109274 : : known_eq (type_nunits, nunits))
10120 : 282404 : vec_cst = gimple_build_vector (&ctor_seq, &elts);
10121 : : else
10122 : : {
10123 : 0 : if (permute_results.is_empty ())
10124 : 0 : duplicate_and_interleave (vinfo, &ctor_seq, vector_type,
10125 : : elts, number_of_vectors,
10126 : : permute_results);
10127 : 0 : vec_cst = permute_results[number_of_vectors - j - 1];
10128 : : }
10129 : 585629 : if (!gimple_seq_empty_p (ctor_seq))
10130 : : {
10131 : 144407 : if (insert_after)
10132 : : {
10133 : 36672 : gimple_stmt_iterator gsi;
10134 : 36672 : if (gimple_code (insert_after->stmt) == GIMPLE_PHI)
10135 : : {
10136 : 3227 : gsi = gsi_after_labels (gimple_bb (insert_after->stmt));
10137 : 3227 : gsi_insert_seq_before (&gsi, ctor_seq,
10138 : : GSI_CONTINUE_LINKING);
10139 : : }
10140 : 33445 : else if (!stmt_ends_bb_p (insert_after->stmt))
10141 : : {
10142 : 33445 : gsi = gsi_for_stmt (insert_after->stmt);
10143 : 33445 : gsi_insert_seq_after (&gsi, ctor_seq,
10144 : : GSI_CONTINUE_LINKING);
10145 : : }
10146 : : else
10147 : : {
10148 : : /* When we want to insert after a def where the
10149 : : defining stmt throws then insert on the fallthru
10150 : : edge. */
10151 : 0 : edge e = find_fallthru_edge
10152 : 0 : (gimple_bb (insert_after->stmt)->succs);
10153 : 0 : basic_block new_bb
10154 : 0 : = gsi_insert_seq_on_edge_immediate (e, ctor_seq);
10155 : 0 : gcc_assert (!new_bb);
10156 : : }
10157 : : }
10158 : : else
10159 : 107735 : vinfo->insert_seq_on_entry (NULL, ctor_seq);
10160 : 144407 : ctor_seq = NULL;
10161 : : }
10162 : 585629 : voprnds.quick_push (vec_cst);
10163 : 585629 : insert_after = NULL;
10164 : 585629 : number_of_places_left_in_vector = nunits;
10165 : 585629 : constant_p = true;
10166 : 585629 : elts.new_vector (vector_type, nunits, 1);
10167 : 585629 : elts.quick_grow (nunits);
10168 : : }
10169 : : }
10170 : : }
10171 : :
10172 : : /* Since the vectors are created in the reverse order, we should invert
10173 : : them. */
10174 : 471079 : vec_num = voprnds.length ();
10175 : 1056708 : for (j = vec_num; j != 0; j--)
10176 : : {
10177 : 585629 : vop = voprnds[j - 1];
10178 : 585629 : SLP_TREE_VEC_DEFS (op_node).quick_push (vop);
10179 : : }
10180 : :
10181 : : /* In case that VF is greater than the unrolling factor needed for the SLP
10182 : : group of stmts, NUMBER_OF_VECTORS to be created is greater than
10183 : : NUMBER_OF_SCALARS/NUNITS or NUNITS/NUMBER_OF_SCALARS, and hence we have
10184 : : to replicate the vectors. */
10185 : 471079 : while (number_of_vectors > SLP_TREE_VEC_DEFS (op_node).length ())
10186 : 471079 : for (i = 0; SLP_TREE_VEC_DEFS (op_node).iterate (i, &vop) && i < vec_num;
10187 : : i++)
10188 : 0 : SLP_TREE_VEC_DEFS (op_node).quick_push (vop);
10189 : 471079 : }
10190 : :
10191 : : /* Get the Ith vectorized definition from SLP_NODE. */
10192 : :
10193 : : tree
10194 : 129254 : vect_get_slp_vect_def (slp_tree slp_node, unsigned i)
10195 : : {
10196 : 129254 : return SLP_TREE_VEC_DEFS (slp_node)[i];
10197 : : }
10198 : :
10199 : : /* Get the vectorized definitions of SLP_NODE in *VEC_DEFS. */
10200 : :
10201 : : void
10202 : 867148 : vect_get_slp_defs (slp_tree slp_node, vec<tree> *vec_defs)
10203 : : {
10204 : 867148 : vec_defs->create (SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node));
10205 : 867148 : vec_defs->splice (SLP_TREE_VEC_DEFS (slp_node));
10206 : 867148 : }
10207 : :
10208 : : /* Get N vectorized definitions for SLP_NODE. */
10209 : :
10210 : : void
10211 : 2160 : vect_get_slp_defs (vec_info *,
10212 : : slp_tree slp_node, vec<vec<tree> > *vec_oprnds, unsigned n)
10213 : : {
10214 : 2160 : if (n == -1U)
10215 : 2160 : n = SLP_TREE_CHILDREN (slp_node).length ();
10216 : :
10217 : 7851 : for (unsigned i = 0; i < n; ++i)
10218 : : {
10219 : 5691 : slp_tree child = SLP_TREE_CHILDREN (slp_node)[i];
10220 : 5691 : vec<tree> vec_defs = vNULL;
10221 : 5691 : vect_get_slp_defs (child, &vec_defs);
10222 : 5691 : vec_oprnds->quick_push (vec_defs);
10223 : : }
10224 : 2160 : }
10225 : :
10226 : : /* A subroutine of vect_transform_slp_perm_load with two extra arguments:
10227 : : - PERM gives the permutation that the caller wants to use for NODE,
10228 : : which might be different from SLP_LOAD_PERMUTATION.
10229 : : - DUMP_P controls whether the function dumps information. */
10230 : :
10231 : : static bool
10232 : 127026 : vect_transform_slp_perm_load_1 (vec_info *vinfo, slp_tree node,
10233 : : load_permutation_t &perm,
10234 : : const vec<tree> &dr_chain,
10235 : : gimple_stmt_iterator *gsi, poly_uint64 vf,
10236 : : bool analyze_only, bool dump_p,
10237 : : unsigned *n_perms, unsigned int *n_loads,
10238 : : bool dce_chain)
10239 : : {
10240 : 127026 : stmt_vec_info stmt_info = SLP_TREE_SCALAR_STMTS (node)[0];
10241 : 127026 : int vec_index = 0;
10242 : 127026 : tree vectype = SLP_TREE_VECTYPE (node);
10243 : 127026 : unsigned int group_size = SLP_TREE_SCALAR_STMTS (node).length ();
10244 : 127026 : unsigned int mask_element;
10245 : 127026 : unsigned dr_group_size;
10246 : 127026 : machine_mode mode;
10247 : :
10248 : 127026 : if (!STMT_VINFO_GROUPED_ACCESS (stmt_info))
10249 : : dr_group_size = 1;
10250 : : else
10251 : : {
10252 : 125417 : stmt_info = DR_GROUP_FIRST_ELEMENT (stmt_info);
10253 : 125417 : dr_group_size = DR_GROUP_SIZE (stmt_info);
10254 : : }
10255 : :
10256 : 127026 : mode = TYPE_MODE (vectype);
10257 : 127026 : poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
10258 : 127026 : unsigned int nstmts = SLP_TREE_NUMBER_OF_VEC_STMTS (node);
10259 : :
10260 : : /* Initialize the vect stmts of NODE to properly insert the generated
10261 : : stmts later. */
10262 : 127026 : if (! analyze_only)
10263 : 32289 : for (unsigned i = SLP_TREE_VEC_DEFS (node).length (); i < nstmts; i++)
10264 : 12791 : SLP_TREE_VEC_DEFS (node).quick_push (NULL_TREE);
10265 : :
10266 : : /* Generate permutation masks for every NODE. Number of masks for each NODE
10267 : : is equal to GROUP_SIZE.
10268 : : E.g., we have a group of three nodes with three loads from the same
10269 : : location in each node, and the vector size is 4. I.e., we have a
10270 : : a0b0c0a1b1c1... sequence and we need to create the following vectors:
10271 : : for a's: a0a0a0a1 a1a1a2a2 a2a3a3a3
10272 : : for b's: b0b0b0b1 b1b1b2b2 b2b3b3b3
10273 : : ...
10274 : :
10275 : : The masks for a's should be: {0,0,0,3} {3,3,6,6} {6,9,9,9}.
10276 : : The last mask is illegal since we assume two operands for permute
10277 : : operation, and the mask element values can't be outside that range.
10278 : : Hence, the last mask must be converted into {2,5,5,5}.
10279 : : For the first two permutations we need the first and the second input
10280 : : vectors: {a0,b0,c0,a1} and {b1,c1,a2,b2}, and for the last permutation
10281 : : we need the second and the third vectors: {b1,c1,a2,b2} and
10282 : : {c2,a3,b3,c3}. */
10283 : :
10284 : 127026 : int vect_stmts_counter = 0;
10285 : 127026 : unsigned int index = 0;
10286 : 127026 : int first_vec_index = -1;
10287 : 127026 : int second_vec_index = -1;
10288 : 127026 : bool noop_p = true;
10289 : 127026 : *n_perms = 0;
10290 : :
10291 : 127026 : vec_perm_builder mask;
10292 : 127026 : unsigned int nelts_to_build;
10293 : 127026 : unsigned int nvectors_per_build;
10294 : 127026 : unsigned int in_nlanes;
10295 : 127026 : bool repeating_p = (group_size == dr_group_size
10296 : 159827 : && multiple_p (nunits, group_size));
10297 : 127026 : if (repeating_p)
10298 : : {
10299 : : /* A single vector contains a whole number of copies of the node, so:
10300 : : (a) all permutes can use the same mask; and
10301 : : (b) the permutes only need a single vector input. */
10302 : 22724 : mask.new_vector (nunits, group_size, 3);
10303 : 22724 : nelts_to_build = mask.encoded_nelts ();
10304 : : /* It's possible to obtain zero nstmts during analyze_only, so make
10305 : : it at least one to ensure the later computation for n_perms
10306 : : proceed. */
10307 : 22724 : nvectors_per_build = nstmts > 0 ? nstmts : 1;
10308 : 22724 : in_nlanes = dr_group_size * 3;
10309 : : }
10310 : : else
10311 : : {
10312 : : /* We need to construct a separate mask for each vector statement. */
10313 : 104302 : unsigned HOST_WIDE_INT const_nunits, const_vf;
10314 : 104302 : if (!nunits.is_constant (&const_nunits)
10315 : 104302 : || !vf.is_constant (&const_vf))
10316 : : return false;
10317 : 104302 : mask.new_vector (const_nunits, const_nunits, 1);
10318 : 104302 : nelts_to_build = const_vf * group_size;
10319 : 104302 : nvectors_per_build = 1;
10320 : 104302 : in_nlanes = const_vf * dr_group_size;
10321 : : }
10322 : 127026 : auto_sbitmap used_in_lanes (in_nlanes);
10323 : 127026 : bitmap_clear (used_in_lanes);
10324 : 127026 : auto_bitmap used_defs;
10325 : :
10326 : 127026 : unsigned int count = mask.encoded_nelts ();
10327 : 127026 : mask.quick_grow (count);
10328 : 127026 : vec_perm_indices indices;
10329 : :
10330 : 732188 : for (unsigned int j = 0; j < nelts_to_build; j++)
10331 : : {
10332 : 610291 : unsigned int iter_num = j / group_size;
10333 : 610291 : unsigned int stmt_num = j % group_size;
10334 : 610291 : unsigned int i = (iter_num * dr_group_size + perm[stmt_num]);
10335 : 610291 : bitmap_set_bit (used_in_lanes, i);
10336 : 610291 : if (repeating_p)
10337 : : {
10338 : : first_vec_index = 0;
10339 : : mask_element = i;
10340 : : }
10341 : : else
10342 : : {
10343 : : /* Enforced before the loop when !repeating_p. */
10344 : 453187 : unsigned int const_nunits = nunits.to_constant ();
10345 : 453187 : vec_index = i / const_nunits;
10346 : 453187 : mask_element = i % const_nunits;
10347 : 453187 : if (vec_index == first_vec_index
10348 : 453187 : || first_vec_index == -1)
10349 : : {
10350 : : first_vec_index = vec_index;
10351 : : }
10352 : 164614 : else if (vec_index == second_vec_index
10353 : 164614 : || second_vec_index == -1)
10354 : : {
10355 : 162777 : second_vec_index = vec_index;
10356 : 162777 : mask_element += const_nunits;
10357 : : }
10358 : : else
10359 : : {
10360 : 1837 : if (dump_p)
10361 : 153 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
10362 : : "permutation requires at "
10363 : : "least three vectors %G",
10364 : : stmt_info->stmt);
10365 : 1837 : gcc_assert (analyze_only);
10366 : : return false;
10367 : : }
10368 : :
10369 : 451350 : gcc_assert (mask_element < 2 * const_nunits);
10370 : : }
10371 : :
10372 : 608454 : if (mask_element != index)
10373 : 411249 : noop_p = false;
10374 : 608454 : mask[index++] = mask_element;
10375 : :
10376 : 608454 : if (index == count)
10377 : : {
10378 : 173052 : if (!noop_p)
10379 : : {
10380 : 204987 : indices.new_vector (mask, second_vec_index == -1 ? 1 : 2, nunits);
10381 : 120378 : if (!can_vec_perm_const_p (mode, mode, indices))
10382 : : {
10383 : 3292 : if (dump_p)
10384 : : {
10385 : 81 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
10386 : : "unsupported vect permute { ");
10387 : 721 : for (i = 0; i < count; ++i)
10388 : : {
10389 : 640 : dump_dec (MSG_MISSED_OPTIMIZATION, mask[i]);
10390 : 640 : dump_printf (MSG_MISSED_OPTIMIZATION, " ");
10391 : : }
10392 : 81 : dump_printf (MSG_MISSED_OPTIMIZATION, "}\n");
10393 : : }
10394 : 3292 : gcc_assert (analyze_only);
10395 : : return false;
10396 : : }
10397 : :
10398 : 117086 : tree mask_vec = NULL_TREE;
10399 : 117086 : if (!analyze_only)
10400 : 11440 : mask_vec = vect_gen_perm_mask_checked (vectype, indices);
10401 : :
10402 : 117086 : if (second_vec_index == -1)
10403 : 33772 : second_vec_index = first_vec_index;
10404 : :
10405 : 234776 : for (unsigned int ri = 0; ri < nvectors_per_build; ++ri)
10406 : : {
10407 : 117690 : ++*n_perms;
10408 : 117690 : if (analyze_only)
10409 : 106122 : continue;
10410 : : /* Generate the permute statement if necessary. */
10411 : 11568 : tree first_vec = dr_chain[first_vec_index + ri];
10412 : 11568 : tree second_vec = dr_chain[second_vec_index + ri];
10413 : 11568 : gassign *stmt = as_a<gassign *> (stmt_info->stmt);
10414 : 11568 : tree perm_dest
10415 : 11568 : = vect_create_destination_var (gimple_assign_lhs (stmt),
10416 : : vectype);
10417 : 11568 : perm_dest = make_ssa_name (perm_dest);
10418 : 11568 : gimple *perm_stmt
10419 : 11568 : = gimple_build_assign (perm_dest, VEC_PERM_EXPR, first_vec,
10420 : : second_vec, mask_vec);
10421 : 11568 : vect_finish_stmt_generation (vinfo, stmt_info, perm_stmt,
10422 : : gsi);
10423 : 11568 : if (dce_chain)
10424 : : {
10425 : 10945 : bitmap_set_bit (used_defs, first_vec_index + ri);
10426 : 10945 : bitmap_set_bit (used_defs, second_vec_index + ri);
10427 : : }
10428 : :
10429 : : /* Store the vector statement in NODE. */
10430 : 11568 : SLP_TREE_VEC_DEFS (node)[vect_stmts_counter++] = perm_dest;
10431 : : }
10432 : : }
10433 : 52674 : else if (!analyze_only)
10434 : : {
10435 : 2446 : for (unsigned int ri = 0; ri < nvectors_per_build; ++ri)
10436 : : {
10437 : 1223 : tree first_vec = dr_chain[first_vec_index + ri];
10438 : : /* If mask was NULL_TREE generate the requested
10439 : : identity transform. */
10440 : 1223 : if (dce_chain)
10441 : 1222 : bitmap_set_bit (used_defs, first_vec_index + ri);
10442 : :
10443 : : /* Store the vector statement in NODE. */
10444 : 1223 : SLP_TREE_VEC_DEFS (node)[vect_stmts_counter++] = first_vec;
10445 : : }
10446 : : }
10447 : :
10448 : : index = 0;
10449 : : first_vec_index = -1;
10450 : : second_vec_index = -1;
10451 : : noop_p = true;
10452 : : }
10453 : : }
10454 : :
10455 : 121897 : if (n_loads)
10456 : : {
10457 : 2371 : if (repeating_p)
10458 : 192 : *n_loads = SLP_TREE_NUMBER_OF_VEC_STMTS (node);
10459 : : else
10460 : : {
10461 : : /* Enforced above when !repeating_p. */
10462 : 2179 : unsigned int const_nunits = nunits.to_constant ();
10463 : 2179 : *n_loads = 0;
10464 : 2179 : bool load_seen = false;
10465 : 26289 : for (unsigned i = 0; i < in_nlanes; ++i)
10466 : : {
10467 : 24110 : if (i % const_nunits == 0)
10468 : : {
10469 : 4328 : if (load_seen)
10470 : 2149 : *n_loads += 1;
10471 : : load_seen = false;
10472 : : }
10473 : 24110 : if (bitmap_bit_p (used_in_lanes, i))
10474 : 12086 : load_seen = true;
10475 : : }
10476 : 2179 : if (load_seen)
10477 : 2179 : *n_loads += 1;
10478 : : }
10479 : : }
10480 : :
10481 : 121897 : if (dce_chain)
10482 : 180612 : for (unsigned i = 0; i < dr_chain.length (); ++i)
10483 : 44434 : if (!bitmap_bit_p (used_defs, i))
10484 : : {
10485 : 26111 : tree def = dr_chain[i];
10486 : 26186 : do
10487 : : {
10488 : 26186 : gimple *stmt = SSA_NAME_DEF_STMT (def);
10489 : 26186 : if (is_gimple_assign (stmt)
10490 : 26186 : && (gimple_assign_rhs_code (stmt) == VIEW_CONVERT_EXPR
10491 : 26186 : || gimple_assign_rhs_code (stmt) == CONSTRUCTOR))
10492 : 513 : def = single_ssa_tree_operand (stmt, SSA_OP_USE);
10493 : : else
10494 : : def = NULL;
10495 : 26186 : gimple_stmt_iterator rgsi = gsi_for_stmt (stmt);
10496 : 26186 : gsi_remove (&rgsi, true);
10497 : 26186 : release_defs (stmt);
10498 : : }
10499 : 26186 : while (def);
10500 : : }
10501 : :
10502 : : return true;
10503 : 127026 : }
10504 : :
10505 : : /* Generate vector permute statements from a list of loads in DR_CHAIN.
10506 : : If ANALYZE_ONLY is TRUE, only check that it is possible to create valid
10507 : : permute statements for the SLP node NODE. Store the number of vector
10508 : : permute instructions in *N_PERMS and the number of vector load
10509 : : instructions in *N_LOADS. If DCE_CHAIN is true, remove all definitions
10510 : : that were not needed. */
10511 : :
10512 : : bool
10513 : 95475 : vect_transform_slp_perm_load (vec_info *vinfo,
10514 : : slp_tree node, const vec<tree> &dr_chain,
10515 : : gimple_stmt_iterator *gsi, poly_uint64 vf,
10516 : : bool analyze_only, unsigned *n_perms,
10517 : : unsigned int *n_loads, bool dce_chain)
10518 : : {
10519 : 95475 : return vect_transform_slp_perm_load_1 (vinfo, node,
10520 : 95475 : SLP_TREE_LOAD_PERMUTATION (node),
10521 : : dr_chain, gsi, vf, analyze_only,
10522 : : dump_enabled_p (), n_perms, n_loads,
10523 : 95475 : dce_chain);
10524 : : }
10525 : :
10526 : : /* Produce the next vector result for SLP permutation NODE by adding a vector
10527 : : statement at GSI. If MASK_VEC is nonnull, add:
10528 : :
10529 : : <new SSA name> = VEC_PERM_EXPR <FIRST_DEF, SECOND_DEF, MASK_VEC>
10530 : :
10531 : : otherwise add:
10532 : :
10533 : : <new SSA name> = FIRST_DEF. */
10534 : :
10535 : : static void
10536 : 28948 : vect_add_slp_permutation (vec_info *vinfo, gimple_stmt_iterator *gsi,
10537 : : slp_tree node, tree first_def, tree second_def,
10538 : : tree mask_vec, poly_uint64 identity_offset)
10539 : : {
10540 : 28948 : tree vectype = SLP_TREE_VECTYPE (node);
10541 : :
10542 : : /* ??? We SLP match existing vector element extracts but
10543 : : allow punning which we need to re-instantiate at uses
10544 : : but have no good way of explicitly representing. */
10545 : 28948 : if (operand_equal_p (TYPE_SIZE (TREE_TYPE (first_def)), TYPE_SIZE (vectype))
10546 : 28948 : && !types_compatible_p (TREE_TYPE (first_def), vectype))
10547 : : {
10548 : 26 : gassign *conv_stmt
10549 : 26 : = gimple_build_assign (make_ssa_name (vectype),
10550 : : build1 (VIEW_CONVERT_EXPR, vectype, first_def));
10551 : 26 : vect_finish_stmt_generation (vinfo, NULL, conv_stmt, gsi);
10552 : 26 : first_def = gimple_assign_lhs (conv_stmt);
10553 : : }
10554 : 28948 : gassign *perm_stmt;
10555 : 28948 : tree perm_dest = make_ssa_name (vectype);
10556 : 28948 : if (mask_vec)
10557 : : {
10558 : 26001 : if (operand_equal_p (TYPE_SIZE (TREE_TYPE (first_def)),
10559 : 26001 : TYPE_SIZE (vectype))
10560 : 26001 : && !types_compatible_p (TREE_TYPE (second_def), vectype))
10561 : : {
10562 : 9 : gassign *conv_stmt
10563 : 9 : = gimple_build_assign (make_ssa_name (vectype),
10564 : : build1 (VIEW_CONVERT_EXPR,
10565 : : vectype, second_def));
10566 : 9 : vect_finish_stmt_generation (vinfo, NULL, conv_stmt, gsi);
10567 : 9 : second_def = gimple_assign_lhs (conv_stmt);
10568 : : }
10569 : 26001 : perm_stmt = gimple_build_assign (perm_dest, VEC_PERM_EXPR,
10570 : : first_def, second_def,
10571 : : mask_vec);
10572 : : }
10573 : 2947 : else if (!types_compatible_p (TREE_TYPE (first_def), vectype))
10574 : : {
10575 : : /* For identity permutes we still need to handle the case
10576 : : of offsetted extracts or concats. */
10577 : 196 : unsigned HOST_WIDE_INT c;
10578 : 196 : auto first_def_nunits
10579 : 196 : = TYPE_VECTOR_SUBPARTS (TREE_TYPE (first_def));
10580 : 196 : if (known_le (TYPE_VECTOR_SUBPARTS (vectype), first_def_nunits))
10581 : : {
10582 : 192 : unsigned HOST_WIDE_INT elsz
10583 : 192 : = tree_to_uhwi (TYPE_SIZE (TREE_TYPE (TREE_TYPE (first_def))));
10584 : 384 : tree lowpart = build3 (BIT_FIELD_REF, vectype, first_def,
10585 : 192 : TYPE_SIZE (vectype),
10586 : 192 : bitsize_int (identity_offset * elsz));
10587 : 192 : perm_stmt = gimple_build_assign (perm_dest, lowpart);
10588 : : }
10589 : 4 : else if (constant_multiple_p (TYPE_VECTOR_SUBPARTS (vectype),
10590 : 4 : first_def_nunits, &c) && c == 2)
10591 : : {
10592 : 4 : tree ctor = build_constructor_va (vectype, 2, NULL_TREE, first_def,
10593 : : NULL_TREE, second_def);
10594 : 4 : perm_stmt = gimple_build_assign (perm_dest, ctor);
10595 : : }
10596 : : else
10597 : 0 : gcc_unreachable ();
10598 : : }
10599 : : else
10600 : : {
10601 : : /* We need a copy here in case the def was external. */
10602 : 2751 : perm_stmt = gimple_build_assign (perm_dest, first_def);
10603 : : }
10604 : 28948 : vect_finish_stmt_generation (vinfo, NULL, perm_stmt, gsi);
10605 : : /* Store the vector statement in NODE. */
10606 : 28948 : node->push_vec_def (perm_stmt);
10607 : 28948 : }
10608 : :
10609 : : /* Subroutine of vectorizable_slp_permutation. Check whether the target
10610 : : can perform permutation PERM on the (1 or 2) input nodes in CHILDREN.
10611 : : If GSI is nonnull, emit the permutation there.
10612 : :
10613 : : When GSI is null, the only purpose of NODE is to give properties
10614 : : of the result, such as the vector type and number of SLP lanes.
10615 : : The node does not need to be a VEC_PERM_EXPR.
10616 : :
10617 : : If the target supports the operation, return the number of individual
10618 : : VEC_PERM_EXPRs needed, otherwise return -1. Print information to the
10619 : : dump file if DUMP_P is true. */
10620 : :
10621 : : static int
10622 : 351717 : vectorizable_slp_permutation_1 (vec_info *vinfo, gimple_stmt_iterator *gsi,
10623 : : slp_tree node, lane_permutation_t &perm,
10624 : : vec<slp_tree> &children, bool dump_p)
10625 : : {
10626 : 351717 : tree vectype = SLP_TREE_VECTYPE (node);
10627 : :
10628 : : /* ??? We currently only support all same vector input types
10629 : : while the SLP IL should really do a concat + select and thus accept
10630 : : arbitrary mismatches. */
10631 : 351717 : slp_tree child;
10632 : 351717 : unsigned i;
10633 : 351717 : poly_uint64 nunits = TYPE_VECTOR_SUBPARTS (vectype);
10634 : 351717 : bool repeating_p = multiple_p (nunits, SLP_TREE_LANES (node));
10635 : : /* True if we're permuting a single input of 2N vectors down
10636 : : to N vectors. This case doesn't generalize beyond 2 since
10637 : : VEC_PERM_EXPR only takes 2 inputs. */
10638 : 351717 : bool pack_p = false;
10639 : : /* If we're permuting inputs of N vectors each into X*N outputs,
10640 : : this is the value of X, otherwise it is 1. */
10641 : 351717 : unsigned int unpack_factor = 1;
10642 : 351717 : tree op_vectype = NULL_TREE;
10643 : 353385 : FOR_EACH_VEC_ELT (children, i, child)
10644 : 353174 : if (SLP_TREE_VECTYPE (child))
10645 : : {
10646 : : op_vectype = SLP_TREE_VECTYPE (child);
10647 : : break;
10648 : : }
10649 : 351717 : if (!op_vectype)
10650 : 211 : op_vectype = vectype;
10651 : 757668 : FOR_EACH_VEC_ELT (children, i, child)
10652 : : {
10653 : 405957 : if ((SLP_TREE_DEF_TYPE (child) != vect_internal_def
10654 : 11168 : && !vect_maybe_update_slp_op_vectype (child, op_vectype))
10655 : 405957 : || !types_compatible_p (SLP_TREE_VECTYPE (child), op_vectype)
10656 : 811914 : || !types_compatible_p (TREE_TYPE (vectype), TREE_TYPE (op_vectype)))
10657 : : {
10658 : 6 : if (dump_p)
10659 : 6 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
10660 : : "Unsupported vector types in lane permutation\n");
10661 : 6 : return -1;
10662 : : }
10663 : 405951 : auto op_nunits = TYPE_VECTOR_SUBPARTS (op_vectype);
10664 : 405951 : unsigned int this_unpack_factor;
10665 : : /* Detect permutations of external, pre-existing vectors. The external
10666 : : node's SLP_TREE_LANES stores the total number of units in the vector,
10667 : : or zero if the vector has variable length.
10668 : :
10669 : : We are expected to keep the original VEC_PERM_EXPR for such cases.
10670 : : There is no repetition to model. */
10671 : 405951 : if (SLP_TREE_DEF_TYPE (child) == vect_external_def
10672 : 405951 : && SLP_TREE_SCALAR_OPS (child).is_empty ())
10673 : : repeating_p = false;
10674 : : /* Check whether the input has twice as many lanes per vector. */
10675 : 397370 : else if (children.length () == 1
10676 : 397370 : && known_eq (SLP_TREE_LANES (child) * nunits,
10677 : : SLP_TREE_LANES (node) * op_nunits * 2))
10678 : : pack_p = true;
10679 : : /* Check whether the output has N times as many lanes per vector. */
10680 : 405951 : else if (constant_multiple_p (SLP_TREE_LANES (node) * op_nunits,
10681 : 350487 : SLP_TREE_LANES (child) * nunits,
10682 : : &this_unpack_factor)
10683 : 311483 : && (i == 0 || unpack_factor == this_unpack_factor))
10684 : : unpack_factor = this_unpack_factor;
10685 : : else
10686 : : repeating_p = false;
10687 : : }
10688 : :
10689 : 703422 : gcc_assert (perm.length () == SLP_TREE_LANES (node));
10690 : :
10691 : : /* Load-lanes permute. This permute only acts as a forwarder to
10692 : : select the correct vector def of the load-lanes load which
10693 : : has the permuted vectors in its vector defs like
10694 : : { v0, w0, r0, v1, w1, r1 ... } for a ld3. All costs are
10695 : : accounted for in the costing for the actual load so we
10696 : : return zero here. */
10697 : 351711 : if (node->ldst_lanes)
10698 : : {
10699 : 0 : gcc_assert (children.length () == 1);
10700 : 0 : if (!gsi)
10701 : : /* This is a trivial op always supported. */
10702 : : return 0;
10703 : 0 : slp_tree child = children[0];
10704 : 0 : unsigned vec_idx = (SLP_TREE_LANE_PERMUTATION (node)[0].second
10705 : 0 : / SLP_TREE_LANES (node));
10706 : 0 : unsigned vec_num = SLP_TREE_LANES (child) / SLP_TREE_LANES (node);
10707 : 0 : for (unsigned i = 0; i < SLP_TREE_NUMBER_OF_VEC_STMTS (node); ++i)
10708 : : {
10709 : 0 : tree def = SLP_TREE_VEC_DEFS (child)[i * vec_num + vec_idx];
10710 : 0 : node->push_vec_def (def);
10711 : : }
10712 : : return 0;
10713 : : }
10714 : :
10715 : : /* Set REPEATING_P to true if the permutations are cylical wrt UNPACK_FACTOR
10716 : : and if we can generate the vectors in a vector-length agnostic way.
10717 : : This requires UNPACK_STEP == NUNITS / UNPACK_FACTOR to be known at
10718 : : compile time.
10719 : :
10720 : : The significance of UNPACK_STEP is that, when PACK_P is false,
10721 : : output vector I operates on a window of UNPACK_STEP elements from each
10722 : : input, starting at lane UNPACK_STEP * (I % UNPACK_FACTOR). For example,
10723 : : when UNPACK_FACTOR is 2, the first output vector operates on lanes
10724 : : [0, NUNITS / 2 - 1] of each input vector and the second output vector
10725 : : operates on lanes [NUNITS / 2, NUNITS - 1] of each input vector.
10726 : :
10727 : : When REPEATING_P is true, NOUTPUTS holds the total number of outputs
10728 : : that we actually need to generate. */
10729 : 351711 : uint64_t noutputs = 0;
10730 : 351711 : poly_uint64 unpack_step = 0;
10731 : 351711 : loop_vec_info linfo = dyn_cast <loop_vec_info> (vinfo);
10732 : 115784 : if (!linfo
10733 : 394024 : || !multiple_p (nunits, unpack_factor, &unpack_step)
10734 : 114716 : || !constant_multiple_p (LOOP_VINFO_VECT_FACTOR (linfo)
10735 : 114716 : * SLP_TREE_LANES (node), nunits, &noutputs))
10736 : : repeating_p = false;
10737 : :
10738 : : /* We can handle the conditions described for REPEATING_P above for
10739 : : both variable- and constant-length vectors. The fallback requires
10740 : : us to generate every element of every permute vector explicitly,
10741 : : which is only possible for constant-length permute vectors.
10742 : :
10743 : : Set:
10744 : :
10745 : : - NPATTERNS and NELTS_PER_PATTERN to the encoding of the permute
10746 : : mask vectors that we want to build.
10747 : :
10748 : : - NCOPIES to the number of copies of PERM that we need in order
10749 : : to build the necessary permute mask vectors. */
10750 : 114716 : uint64_t npatterns;
10751 : 114716 : unsigned nelts_per_pattern;
10752 : 114716 : uint64_t ncopies;
10753 : 114716 : if (repeating_p)
10754 : : {
10755 : : /* We need permute mask vectors that have the form:
10756 : :
10757 : : { X1, ..., Xn, X1 + n, ..., Xn + n, X1 + 2n, ..., Xn + 2n, ... }
10758 : :
10759 : : In other words, the original n-element permute in PERM is
10760 : : "unrolled" to fill a full vector. The stepped vector encoding
10761 : : that we use for permutes requires 3n elements. */
10762 : 72403 : npatterns = SLP_TREE_LANES (node);
10763 : 72403 : nelts_per_pattern = ncopies = 3;
10764 : : }
10765 : : else
10766 : : {
10767 : : /* Calculate every element of every permute mask vector explicitly,
10768 : : instead of relying on the pattern described above. */
10769 : 279308 : if (!nunits.is_constant (&npatterns)
10770 : 279308 : || !TYPE_VECTOR_SUBPARTS (op_vectype).is_constant ())
10771 : : {
10772 : : if (dump_p)
10773 : : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
10774 : : "unsupported permutation %p on variable-length"
10775 : : " vectors\n", (void *) node);
10776 : : return -1;
10777 : : }
10778 : 279308 : nelts_per_pattern = ncopies = 1;
10779 : 279308 : if (linfo && !LOOP_VINFO_VECT_FACTOR (linfo).is_constant (&ncopies))
10780 : : {
10781 : : if (dump_p)
10782 : : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
10783 : : "unsupported permutation %p for variable VF\n",
10784 : : (void *) node);
10785 : : return -1;
10786 : : }
10787 : : pack_p = false;
10788 : : unpack_factor = 1;
10789 : : }
10790 : 351711 : unsigned olanes = unpack_factor * ncopies * SLP_TREE_LANES (node);
10791 : 351711 : gcc_assert (repeating_p || multiple_p (olanes, nunits));
10792 : :
10793 : : /* Compute the { { SLP operand, vector index}, lane } permutation sequence
10794 : : from the { SLP operand, scalar lane } permutation as recorded in the
10795 : : SLP node as intermediate step. This part should already work
10796 : : with SLP children with arbitrary number of lanes. */
10797 : 351711 : auto_vec<std::pair<std::pair<unsigned, unsigned>, poly_uint64>> vperm;
10798 : 351711 : auto_vec<poly_uint64> active_lane;
10799 : 351711 : vperm.create (olanes);
10800 : 351711 : active_lane.safe_grow_cleared (children.length (), true);
10801 : 711232 : for (unsigned int ui = 0; ui < unpack_factor; ++ui)
10802 : : {
10803 : 1562184 : for (unsigned j = 0; j < children.length (); ++j)
10804 : 421571 : active_lane[j] = ui * unpack_step;
10805 : 1000043 : for (unsigned i = 0; i < ncopies; ++i)
10806 : : {
10807 : 4031380 : for (unsigned pi = 0; pi < perm.length (); ++pi)
10808 : : {
10809 : 1375168 : std::pair<unsigned, unsigned> p = perm[pi];
10810 : 1375168 : tree vtype = SLP_TREE_VECTYPE (children[p.first]);
10811 : 1375168 : if (repeating_p)
10812 : 404526 : vperm.quick_push ({{p.first, 0},
10813 : 404526 : p.second + active_lane[p.first]});
10814 : : else
10815 : : {
10816 : : /* We checked above that the vectors are constant-length. */
10817 : 970642 : unsigned vnunits = TYPE_VECTOR_SUBPARTS (vtype)
10818 : 970642 : .to_constant ();
10819 : 970642 : unsigned lane = active_lane[p.first].to_constant ();
10820 : 970642 : unsigned vi = (lane + p.second) / vnunits;
10821 : 970642 : unsigned vl = (lane + p.second) % vnunits;
10822 : 970642 : vperm.quick_push ({{p.first, vi}, vl});
10823 : : }
10824 : : }
10825 : : /* Advance to the next group. */
10826 : 1392171 : for (unsigned j = 0; j < children.length (); ++j)
10827 : 751649 : active_lane[j] += SLP_TREE_LANES (children[j]);
10828 : : }
10829 : : }
10830 : :
10831 : 351711 : if (dump_p)
10832 : : {
10833 : 9023 : dump_printf_loc (MSG_NOTE, vect_location,
10834 : : "vectorizing permutation %p", (void *)node);
10835 : 32227 : for (unsigned i = 0; i < perm.length (); ++i)
10836 : 23204 : dump_printf (MSG_NOTE, " op%u[%u]", perm[i].first, perm[i].second);
10837 : 9023 : if (repeating_p)
10838 : 7746 : dump_printf (MSG_NOTE, " (repeat %d)\n", SLP_TREE_LANES (node));
10839 : 9023 : dump_printf (MSG_NOTE, "\n");
10840 : 9023 : dump_printf_loc (MSG_NOTE, vect_location, "as");
10841 : 90792 : for (unsigned i = 0; i < vperm.length (); ++i)
10842 : : {
10843 : 81769 : if (i != 0
10844 : 81769 : && (repeating_p
10845 : 56913 : ? multiple_p (i, npatterns)
10846 : 60334 : : multiple_p (i, TYPE_VECTOR_SUBPARTS (vectype))))
10847 : 24219 : dump_printf (MSG_NOTE, ",");
10848 : 81769 : dump_printf (MSG_NOTE, " vops%u[%u][",
10849 : 81769 : vperm[i].first.first, vperm[i].first.second);
10850 : 81769 : dump_dec (MSG_NOTE, vperm[i].second);
10851 : 81769 : dump_printf (MSG_NOTE, "]");
10852 : : }
10853 : 9023 : dump_printf (MSG_NOTE, "\n");
10854 : : }
10855 : :
10856 : : /* We can only handle two-vector permutes, everything else should
10857 : : be lowered on the SLP level. The following is closely inspired
10858 : : by vect_transform_slp_perm_load and is supposed to eventually
10859 : : replace it.
10860 : : ??? As intermediate step do code-gen in the SLP tree representation
10861 : : somehow? */
10862 : 351711 : std::pair<unsigned, unsigned> first_vec = std::make_pair (-1U, -1U);
10863 : 351711 : std::pair<unsigned, unsigned> second_vec = std::make_pair (-1U, -1U);
10864 : 351711 : unsigned int index = 0;
10865 : 351711 : poly_uint64 mask_element;
10866 : 351711 : vec_perm_builder mask;
10867 : 351711 : mask.new_vector (nunits, npatterns, nelts_per_pattern);
10868 : 351711 : unsigned int count = mask.encoded_nelts ();
10869 : 351711 : mask.quick_grow (count);
10870 : 351711 : vec_perm_indices indices;
10871 : 351711 : unsigned nperms = 0;
10872 : : /* When REPEATING_P is true, we only have UNPACK_FACTOR unique permute
10873 : : vectors to check during analysis, but we need to generate NOUTPUTS
10874 : : vectors during transformation. */
10875 : 351711 : unsigned total_nelts = olanes;
10876 : 351711 : if (repeating_p && gsi)
10877 : 9363 : total_nelts = (total_nelts / unpack_factor) * noutputs;
10878 : 1731227 : for (unsigned i = 0; i < total_nelts; ++i)
10879 : : {
10880 : : /* VI is the input vector index when generating code for REPEATING_P. */
10881 : 1387009 : unsigned vi = i / olanes * (pack_p ? 2 : 1);
10882 : 1387009 : unsigned ei = i % olanes;
10883 : 1387009 : mask_element = vperm[ei].second;
10884 : 1387009 : if (pack_p)
10885 : : {
10886 : : /* In this case, we have N outputs and the single child provides 2N
10887 : : inputs. Output X permutes inputs 2X and 2X+1.
10888 : :
10889 : : The mask indices are taken directly from the SLP permutation node.
10890 : : Index X selects from the first vector if (X / NUNITS) % 2 == 0;
10891 : : X selects from the second vector otherwise. These conditions
10892 : : are only known at compile time for constant-length vectors. */
10893 : : first_vec = std::make_pair (0, 0);
10894 : : second_vec = std::make_pair (0, 1);
10895 : : }
10896 : 1209709 : else if (first_vec.first == -1U
10897 : 1209709 : || first_vec == vperm[ei].first)
10898 : 1025524 : first_vec = vperm[ei].first;
10899 : 184185 : else if (second_vec.first == -1U
10900 : 184185 : || second_vec == vperm[ei].first)
10901 : : {
10902 : 183553 : second_vec = vperm[ei].first;
10903 : 183553 : mask_element += nunits;
10904 : : }
10905 : : else
10906 : : {
10907 : 632 : if (dump_p)
10908 : 24 : dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
10909 : : "permutation requires at "
10910 : : "least three vectors\n");
10911 : 632 : gcc_assert (!gsi);
10912 : : return -1;
10913 : : }
10914 : :
10915 : 1386377 : mask[index++] = mask_element;
10916 : :
10917 : 1386377 : if (index == count)
10918 : : {
10919 : 677926 : indices.new_vector (mask, second_vec.first == -1U ? 1 : 2,
10920 : : TYPE_VECTOR_SUBPARTS (op_vectype));
10921 : 536636 : bool identity_p = (indices.series_p (0, 1, mask[0], 1)
10922 : 835323 : && constant_multiple_p (mask[0], nunits));
10923 : 536636 : machine_mode vmode = TYPE_MODE (vectype);
10924 : 536636 : machine_mode op_vmode = TYPE_MODE (op_vectype);
10925 : 536636 : unsigned HOST_WIDE_INT c;
10926 : 536636 : if ((!identity_p
10927 : 432170 : && !can_vec_perm_const_p (vmode, op_vmode, indices))
10928 : 536636 : || (identity_p
10929 : 104466 : && !known_le (nunits,
10930 : : TYPE_VECTOR_SUBPARTS (op_vectype))
10931 : 6869 : && (!constant_multiple_p (nunits,
10932 : 8 : TYPE_VECTOR_SUBPARTS (op_vectype),
10933 : 8 : &c) || c != 2)))
10934 : : {
10935 : 6861 : if (dump_p)
10936 : : {
10937 : 164 : dump_printf_loc (MSG_MISSED_OPTIMIZATION,
10938 : : vect_location,
10939 : : "unsupported vect permute { ");
10940 : 1734 : for (i = 0; i < count; ++i)
10941 : : {
10942 : 1570 : dump_dec (MSG_MISSED_OPTIMIZATION, mask[i]);
10943 : 1570 : dump_printf (MSG_MISSED_OPTIMIZATION, " ");
10944 : : }
10945 : 164 : dump_printf (MSG_MISSED_OPTIMIZATION, "}\n");
10946 : : }
10947 : 6861 : gcc_assert (!gsi);
10948 : 7493 : return -1;
10949 : : }
10950 : :
10951 : 529775 : if (!identity_p)
10952 : 425309 : nperms++;
10953 : 529775 : if (gsi)
10954 : : {
10955 : 28948 : if (second_vec.first == -1U)
10956 : 6494 : second_vec = first_vec;
10957 : :
10958 : 28948 : slp_tree
10959 : 28948 : first_node = children[first_vec.first],
10960 : 28948 : second_node = children[second_vec.first];
10961 : :
10962 : 28948 : tree mask_vec = NULL_TREE;
10963 : 28948 : if (!identity_p)
10964 : 26001 : mask_vec = vect_gen_perm_mask_checked (vectype, indices);
10965 : :
10966 : 28948 : tree first_def
10967 : 28948 : = vect_get_slp_vect_def (first_node, first_vec.second + vi);
10968 : 28948 : tree second_def
10969 : 28948 : = vect_get_slp_vect_def (second_node, second_vec.second + vi);
10970 : 28948 : vect_add_slp_permutation (vinfo, gsi, node, first_def,
10971 : 28948 : second_def, mask_vec, mask[0]);
10972 : : }
10973 : :
10974 : : index = 0;
10975 : : first_vec = std::make_pair (-1U, -1U);
10976 : : second_vec = std::make_pair (-1U, -1U);
10977 : : }
10978 : : }
10979 : :
10980 : 344218 : return nperms;
10981 : 351711 : }
10982 : :
10983 : : /* Vectorize the SLP permutations in NODE as specified
10984 : : in SLP_TREE_LANE_PERMUTATION which is a vector of pairs of SLP
10985 : : child number and lane number.
10986 : : Interleaving of two two-lane two-child SLP subtrees (not supported):
10987 : : [ { 0, 0 }, { 1, 0 }, { 0, 1 }, { 1, 1 } ]
10988 : : A blend of two four-lane two-child SLP subtrees:
10989 : : [ { 0, 0 }, { 1, 1 }, { 0, 2 }, { 1, 3 } ]
10990 : : Highpart of a four-lane one-child SLP subtree (not supported):
10991 : : [ { 0, 2 }, { 0, 3 } ]
10992 : : Where currently only a subset is supported by code generating below. */
10993 : :
10994 : : static bool
10995 : 124924 : vectorizable_slp_permutation (vec_info *vinfo, gimple_stmt_iterator *gsi,
10996 : : slp_tree node, stmt_vector_for_cost *cost_vec)
10997 : : {
10998 : 124924 : tree vectype = SLP_TREE_VECTYPE (node);
10999 : 124924 : lane_permutation_t &perm = SLP_TREE_LANE_PERMUTATION (node);
11000 : 124924 : int nperms = vectorizable_slp_permutation_1 (vinfo, gsi, node, perm,
11001 : 124924 : SLP_TREE_CHILDREN (node),
11002 : : dump_enabled_p ());
11003 : 124924 : if (nperms < 0)
11004 : : return false;
11005 : :
11006 : 123605 : if (!gsi)
11007 : 108239 : record_stmt_cost (cost_vec, nperms, vec_perm, node, vectype, 0, vect_body);
11008 : :
11009 : : return true;
11010 : : }
11011 : :
11012 : : /* Vectorize SLP NODE. */
11013 : :
11014 : : static void
11015 : 1394407 : vect_schedule_slp_node (vec_info *vinfo,
11016 : : slp_tree node, slp_instance instance)
11017 : : {
11018 : 1394407 : gimple_stmt_iterator si;
11019 : 1394407 : int i;
11020 : 1394407 : slp_tree child;
11021 : :
11022 : : /* Vectorize externals and constants. */
11023 : 1394407 : if (SLP_TREE_DEF_TYPE (node) == vect_constant_def
11024 : 1394407 : || SLP_TREE_DEF_TYPE (node) == vect_external_def)
11025 : : {
11026 : : /* ??? vectorizable_shift can end up using a scalar operand which is
11027 : : currently denoted as !SLP_TREE_VECTYPE. No need to vectorize the
11028 : : node in this case. */
11029 : 477304 : if (!SLP_TREE_VECTYPE (node))
11030 : 477304 : return;
11031 : :
11032 : : /* There are two reasons vector defs might already exist. The first
11033 : : is that we are vectorizing an existing vector def. The second is
11034 : : when performing BB vectorization shared constant/external nodes
11035 : : are not split apart during partitioning so during the code-gen
11036 : : DFS walk we can end up visiting them twice. */
11037 : 471751 : if (! SLP_TREE_VEC_DEFS (node).exists ())
11038 : 471079 : vect_create_constant_vectors (vinfo, node);
11039 : 471751 : return;
11040 : : }
11041 : :
11042 : 917103 : gcc_assert (SLP_TREE_VEC_DEFS (node).is_empty ());
11043 : :
11044 : 917103 : stmt_vec_info stmt_info = SLP_TREE_REPRESENTATIVE (node);
11045 : :
11046 : 917103 : gcc_assert (SLP_TREE_NUMBER_OF_VEC_STMTS (node) != 0);
11047 : 917103 : SLP_TREE_VEC_DEFS (node).create (SLP_TREE_NUMBER_OF_VEC_STMTS (node));
11048 : :
11049 : 917103 : if (SLP_TREE_CODE (node) != VEC_PERM_EXPR
11050 : 901737 : && STMT_VINFO_DATA_REF (stmt_info))
11051 : : {
11052 : : /* Vectorized loads go before the first scalar load to make it
11053 : : ready early, vectorized stores go before the last scalar
11054 : : stmt which is where all uses are ready. */
11055 : 683544 : stmt_vec_info last_stmt_info = NULL;
11056 : 683544 : if (DR_IS_READ (STMT_VINFO_DATA_REF (stmt_info)))
11057 : 154485 : last_stmt_info = vect_find_first_scalar_stmt_in_slp (node);
11058 : : else /* DR_IS_WRITE */
11059 : 529059 : last_stmt_info = vect_find_last_scalar_stmt_in_slp (node);
11060 : 683544 : si = gsi_for_stmt (last_stmt_info->stmt);
11061 : 683544 : }
11062 : 233559 : else if (SLP_TREE_CODE (node) != VEC_PERM_EXPR
11063 : 218193 : && (STMT_VINFO_TYPE (stmt_info) == cycle_phi_info_type
11064 : : || STMT_VINFO_TYPE (stmt_info) == induc_vec_info_type
11065 : : || STMT_VINFO_TYPE (stmt_info) == phi_info_type))
11066 : : {
11067 : : /* For PHI node vectorization we do not use the insertion iterator. */
11068 : 50963 : si = gsi_none ();
11069 : : }
11070 : : else
11071 : : {
11072 : : /* Emit other stmts after the children vectorized defs which is
11073 : : earliest possible. */
11074 : : gimple *last_stmt = NULL;
11075 : : bool seen_vector_def = false;
11076 : 506308 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
11077 : 323712 : if (SLP_TREE_DEF_TYPE (child) == vect_internal_def)
11078 : : {
11079 : : /* For fold-left reductions we are retaining the scalar
11080 : : reduction PHI but we still have SLP_TREE_NUM_VEC_STMTS
11081 : : set so the representation isn't perfect. Resort to the
11082 : : last scalar def here. */
11083 : 260807 : if (SLP_TREE_VEC_DEFS (child).is_empty ())
11084 : : {
11085 : 890 : gcc_assert (STMT_VINFO_TYPE (SLP_TREE_REPRESENTATIVE (child))
11086 : : == cycle_phi_info_type);
11087 : 890 : gphi *phi = as_a <gphi *>
11088 : 890 : (vect_find_last_scalar_stmt_in_slp (child)->stmt);
11089 : 890 : if (!last_stmt
11090 : 890 : || vect_stmt_dominates_stmt_p (last_stmt, phi))
11091 : : last_stmt = phi;
11092 : : }
11093 : : /* We are emitting all vectorized stmts in the same place and
11094 : : the last one is the last.
11095 : : ??? Unless we have a load permutation applied and that
11096 : : figures to re-use an earlier generated load. */
11097 : : unsigned j;
11098 : : tree vdef;
11099 : 600430 : FOR_EACH_VEC_ELT (SLP_TREE_VEC_DEFS (child), j, vdef)
11100 : : {
11101 : 339623 : gimple *vstmt = SSA_NAME_DEF_STMT (vdef);
11102 : 339623 : if (!last_stmt
11103 : 339623 : || vect_stmt_dominates_stmt_p (last_stmt, vstmt))
11104 : : last_stmt = vstmt;
11105 : : }
11106 : : }
11107 : 62905 : else if (!SLP_TREE_VECTYPE (child))
11108 : : {
11109 : : /* For externals we use unvectorized at all scalar defs. */
11110 : : unsigned j;
11111 : : tree def;
11112 : 12107 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_OPS (child), j, def)
11113 : 7136 : if (TREE_CODE (def) == SSA_NAME
11114 : 7136 : && !SSA_NAME_IS_DEFAULT_DEF (def))
11115 : : {
11116 : 168 : gimple *stmt = SSA_NAME_DEF_STMT (def);
11117 : 168 : if (!last_stmt
11118 : 168 : || vect_stmt_dominates_stmt_p (last_stmt, stmt))
11119 : : last_stmt = stmt;
11120 : : }
11121 : : }
11122 : : else
11123 : : {
11124 : : /* For externals we have to look at all defs since their
11125 : : insertion place is decided per vector. But beware
11126 : : of pre-existing vectors where we need to make sure
11127 : : we do not insert before the region boundary. */
11128 : 57934 : if (SLP_TREE_SCALAR_OPS (child).is_empty ()
11129 : 548 : && !vinfo->lookup_def (SLP_TREE_VEC_DEFS (child)[0]))
11130 : : seen_vector_def = true;
11131 : : else
11132 : : {
11133 : : unsigned j;
11134 : : tree vdef;
11135 : 457647 : FOR_EACH_VEC_ELT (SLP_TREE_VEC_DEFS (child), j, vdef)
11136 : 76116 : if (TREE_CODE (vdef) == SSA_NAME
11137 : 76116 : && !SSA_NAME_IS_DEFAULT_DEF (vdef))
11138 : : {
11139 : 16626 : gimple *vstmt = SSA_NAME_DEF_STMT (vdef);
11140 : 16626 : if (!last_stmt
11141 : 16626 : || vect_stmt_dominates_stmt_p (last_stmt, vstmt))
11142 : : last_stmt = vstmt;
11143 : : }
11144 : : }
11145 : : }
11146 : : /* This can happen when all children are pre-existing vectors or
11147 : : constants. */
11148 : 182596 : if (!last_stmt)
11149 : 1713 : last_stmt = vect_find_first_scalar_stmt_in_slp (node)->stmt;
11150 : 1713 : if (!last_stmt)
11151 : : {
11152 : 0 : gcc_assert (seen_vector_def);
11153 : 0 : si = gsi_after_labels (vinfo->bbs[0]);
11154 : : }
11155 : 182596 : else if (is_ctrl_altering_stmt (last_stmt))
11156 : : {
11157 : : /* We split regions to vectorize at control altering stmts
11158 : : with a definition so this must be an external which
11159 : : we can insert at the start of the region. */
11160 : 3 : si = gsi_after_labels (vinfo->bbs[0]);
11161 : : }
11162 : 182593 : else if (is_a <bb_vec_info> (vinfo)
11163 : 13755 : && SLP_TREE_CODE (node) != VEC_PERM_EXPR
11164 : 12701 : && gimple_bb (last_stmt) != gimple_bb (stmt_info->stmt)
11165 : 183526 : && gimple_could_trap_p (stmt_info->stmt))
11166 : : {
11167 : : /* We've constrained possibly trapping operations to all come
11168 : : from the same basic-block, if vectorized defs would allow earlier
11169 : : scheduling still force vectorized stmts to the original block.
11170 : : This is only necessary for BB vectorization since for loop vect
11171 : : all operations are in a single BB and scalar stmt based
11172 : : placement doesn't play well with epilogue vectorization. */
11173 : 76 : gcc_assert (dominated_by_p (CDI_DOMINATORS,
11174 : : gimple_bb (stmt_info->stmt),
11175 : : gimple_bb (last_stmt)));
11176 : 76 : si = gsi_after_labels (gimple_bb (stmt_info->stmt));
11177 : : }
11178 : 182517 : else if (is_a <gphi *> (last_stmt))
11179 : 13747 : si = gsi_after_labels (gimple_bb (last_stmt));
11180 : : else
11181 : : {
11182 : 168770 : si = gsi_for_stmt (last_stmt);
11183 : 168770 : gsi_next (&si);
11184 : :
11185 : : /* Avoid scheduling internal defs outside of the loop when
11186 : : we might have only implicitly tracked loop mask/len defs. */
11187 : 168770 : if (auto loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
11188 : 38 : if (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)
11189 : 155257 : || LOOP_VINFO_FULLY_WITH_LENGTH_P (loop_vinfo))
11190 : : {
11191 : 38 : gimple_stmt_iterator si2
11192 : 38 : = gsi_after_labels (LOOP_VINFO_LOOP (loop_vinfo)->header);
11193 : 38 : if ((gsi_end_p (si2)
11194 : 0 : && (LOOP_VINFO_LOOP (loop_vinfo)->header
11195 : 0 : != gimple_bb (last_stmt))
11196 : 0 : && dominated_by_p (CDI_DOMINATORS,
11197 : : LOOP_VINFO_LOOP (loop_vinfo)->header,
11198 : 0 : gimple_bb (last_stmt)))
11199 : 38 : || (!gsi_end_p (si2)
11200 : 38 : && last_stmt != *si2
11201 : 37 : && vect_stmt_dominates_stmt_p (last_stmt, *si2)))
11202 : 1 : si = si2;
11203 : : }
11204 : : }
11205 : : }
11206 : :
11207 : : /* Handle purely internal nodes. */
11208 : 917103 : if (SLP_TREE_CODE (node) == VEC_PERM_EXPR)
11209 : : {
11210 : 15366 : if (dump_enabled_p ())
11211 : 3061 : dump_printf_loc (MSG_NOTE, vect_location,
11212 : : "------>vectorizing SLP permutation node\n");
11213 : : /* ??? the transform kind is stored to STMT_VINFO_TYPE which might
11214 : : be shared with different SLP nodes (but usually it's the same
11215 : : operation apart from the case the stmt is only there for denoting
11216 : : the actual scalar lane defs ...). So do not call vect_transform_stmt
11217 : : but open-code it here (partly). */
11218 : 15366 : bool done = vectorizable_slp_permutation (vinfo, &si, node, NULL);
11219 : 15366 : gcc_assert (done);
11220 : : stmt_vec_info slp_stmt_info;
11221 : : unsigned int i;
11222 : 952309 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, slp_stmt_info)
11223 : 22861 : if (slp_stmt_info && STMT_VINFO_LIVE_P (slp_stmt_info))
11224 : : {
11225 : 564 : done = vectorizable_live_operation (vinfo, slp_stmt_info, node,
11226 : : instance, i, true, NULL);
11227 : 564 : gcc_assert (done);
11228 : : }
11229 : : }
11230 : : else
11231 : : {
11232 : 901737 : if (dump_enabled_p ())
11233 : 62737 : dump_printf_loc (MSG_NOTE, vect_location,
11234 : : "------>vectorizing SLP node starting from: %G",
11235 : : stmt_info->stmt);
11236 : 901737 : vect_transform_stmt (vinfo, stmt_info, &si, node, instance);
11237 : : }
11238 : : }
11239 : :
11240 : : /* Replace scalar calls from SLP node NODE with setting of their lhs to zero.
11241 : : For loop vectorization this is done in vectorizable_call, but for SLP
11242 : : it needs to be deferred until end of vect_schedule_slp, because multiple
11243 : : SLP instances may refer to the same scalar stmt. */
11244 : :
11245 : : static void
11246 : 548707 : vect_remove_slp_scalar_calls (vec_info *vinfo,
11247 : : slp_tree node, hash_set<slp_tree> &visited)
11248 : : {
11249 : 548707 : gimple *new_stmt;
11250 : 548707 : gimple_stmt_iterator gsi;
11251 : 548707 : int i;
11252 : 548707 : slp_tree child;
11253 : 548707 : tree lhs;
11254 : 548707 : stmt_vec_info stmt_info;
11255 : :
11256 : 548707 : if (!node || SLP_TREE_DEF_TYPE (node) != vect_internal_def)
11257 : 171865 : return;
11258 : :
11259 : 416819 : if (visited.add (node))
11260 : : return;
11261 : :
11262 : 841151 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
11263 : 464309 : vect_remove_slp_scalar_calls (vinfo, child, visited);
11264 : :
11265 : 1188618 : FOR_EACH_VEC_ELT (SLP_TREE_SCALAR_STMTS (node), i, stmt_info)
11266 : : {
11267 : 438871 : if (!stmt_info)
11268 : 3611 : continue;
11269 : 435260 : gcall *stmt = dyn_cast <gcall *> (stmt_info->stmt);
11270 : 4799 : if (!stmt || gimple_bb (stmt) == NULL)
11271 : 430467 : continue;
11272 : 4793 : if (is_pattern_stmt_p (stmt_info)
11273 : 4793 : || !PURE_SLP_STMT (stmt_info))
11274 : 732 : continue;
11275 : 4061 : lhs = gimple_call_lhs (stmt);
11276 : 4061 : if (lhs)
11277 : 3619 : new_stmt = gimple_build_assign (lhs, build_zero_cst (TREE_TYPE (lhs)));
11278 : : else
11279 : : {
11280 : 442 : new_stmt = gimple_build_nop ();
11281 : 442 : unlink_stmt_vdef (stmt_info->stmt);
11282 : : }
11283 : 4061 : gsi = gsi_for_stmt (stmt);
11284 : 4061 : vinfo->replace_stmt (&gsi, stmt_info, new_stmt);
11285 : 4061 : if (lhs)
11286 : 3619 : SSA_NAME_DEF_STMT (lhs) = new_stmt;
11287 : : }
11288 : : }
11289 : :
11290 : : static void
11291 : 84398 : vect_remove_slp_scalar_calls (vec_info *vinfo, slp_tree node)
11292 : : {
11293 : 84398 : hash_set<slp_tree> visited;
11294 : 84398 : vect_remove_slp_scalar_calls (vinfo, node, visited);
11295 : 84398 : }
11296 : :
11297 : : /* Vectorize the instance root. */
11298 : :
11299 : : void
11300 : 5608 : vectorize_slp_instance_root_stmt (vec_info *vinfo, slp_tree node, slp_instance instance)
11301 : : {
11302 : 5608 : gassign *rstmt = NULL;
11303 : :
11304 : 5608 : if (instance->kind == slp_inst_kind_ctor)
11305 : : {
11306 : 248 : if (SLP_TREE_NUMBER_OF_VEC_STMTS (node) == 1)
11307 : : {
11308 : 217 : tree vect_lhs = SLP_TREE_VEC_DEFS (node)[0];
11309 : 217 : tree root_lhs = gimple_get_lhs (instance->root_stmts[0]->stmt);
11310 : 217 : if (!useless_type_conversion_p (TREE_TYPE (root_lhs),
11311 : 217 : TREE_TYPE (vect_lhs)))
11312 : 0 : vect_lhs = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (root_lhs),
11313 : : vect_lhs);
11314 : 217 : rstmt = gimple_build_assign (root_lhs, vect_lhs);
11315 : : }
11316 : 31 : else if (SLP_TREE_NUMBER_OF_VEC_STMTS (node) > 1)
11317 : : {
11318 : 31 : int nelts = SLP_TREE_NUMBER_OF_VEC_STMTS (node);
11319 : 31 : tree child_def;
11320 : 31 : int j;
11321 : 31 : vec<constructor_elt, va_gc> *v;
11322 : 31 : vec_alloc (v, nelts);
11323 : :
11324 : : /* A CTOR can handle V16HI composition from VNx8HI so we
11325 : : do not need to convert vector elements if the types
11326 : : do not match. */
11327 : 125 : FOR_EACH_VEC_ELT (SLP_TREE_VEC_DEFS (node), j, child_def)
11328 : 94 : CONSTRUCTOR_APPEND_ELT (v, NULL_TREE, child_def);
11329 : 31 : tree lhs = gimple_get_lhs (instance->root_stmts[0]->stmt);
11330 : 31 : tree rtype
11331 : 31 : = TREE_TYPE (gimple_assign_rhs1 (instance->root_stmts[0]->stmt));
11332 : 31 : tree r_constructor = build_constructor (rtype, v);
11333 : 31 : rstmt = gimple_build_assign (lhs, r_constructor);
11334 : : }
11335 : : }
11336 : 5360 : else if (instance->kind == slp_inst_kind_bb_reduc)
11337 : : {
11338 : : /* Largely inspired by reduction chain epilogue handling in
11339 : : vect_create_epilog_for_reduction. */
11340 : 4055 : vec<tree> vec_defs = vNULL;
11341 : 4055 : vect_get_slp_defs (node, &vec_defs);
11342 : 4055 : enum tree_code reduc_code
11343 : 4055 : = gimple_assign_rhs_code (instance->root_stmts[0]->stmt);
11344 : : /* ??? We actually have to reflect signs somewhere. */
11345 : 4055 : if (reduc_code == MINUS_EXPR)
11346 : 0 : reduc_code = PLUS_EXPR;
11347 : 4055 : gimple_seq epilogue = NULL;
11348 : : /* We may end up with more than one vector result, reduce them
11349 : : to one vector. */
11350 : 4055 : tree vec_def = vec_defs[0];
11351 : 4055 : tree vectype = TREE_TYPE (vec_def);
11352 : 4055 : tree compute_vectype = vectype;
11353 : 4055 : bool pun_for_overflow_p = (ANY_INTEGRAL_TYPE_P (vectype)
11354 : 3989 : && TYPE_OVERFLOW_UNDEFINED (vectype)
11355 : 6957 : && operation_can_overflow (reduc_code));
11356 : 2780 : if (pun_for_overflow_p)
11357 : : {
11358 : 2780 : compute_vectype = unsigned_type_for (vectype);
11359 : 2780 : vec_def = gimple_build (&epilogue, VIEW_CONVERT_EXPR,
11360 : : compute_vectype, vec_def);
11361 : : }
11362 : 6424 : for (unsigned i = 1; i < vec_defs.length (); ++i)
11363 : : {
11364 : 2369 : tree def = vec_defs[i];
11365 : 2369 : if (pun_for_overflow_p)
11366 : 2273 : def = gimple_build (&epilogue, VIEW_CONVERT_EXPR,
11367 : : compute_vectype, def);
11368 : 2369 : vec_def = gimple_build (&epilogue, reduc_code, compute_vectype,
11369 : : vec_def, def);
11370 : : }
11371 : 4055 : vec_defs.release ();
11372 : : /* ??? Support other schemes than direct internal fn. */
11373 : 4055 : internal_fn reduc_fn;
11374 : 4055 : if (!reduction_fn_for_scalar_code (reduc_code, &reduc_fn)
11375 : 4055 : || reduc_fn == IFN_LAST)
11376 : 0 : gcc_unreachable ();
11377 : 4055 : tree scalar_def = gimple_build (&epilogue, as_combined_fn (reduc_fn),
11378 : 4055 : TREE_TYPE (compute_vectype), vec_def);
11379 : 4055 : if (!SLP_INSTANCE_REMAIN_DEFS (instance).is_empty ())
11380 : : {
11381 : 2643 : tree rem_def = NULL_TREE;
11382 : 12063 : for (auto def : SLP_INSTANCE_REMAIN_DEFS (instance))
11383 : : {
11384 : 9420 : def = gimple_convert (&epilogue, TREE_TYPE (scalar_def), def);
11385 : 9420 : if (!rem_def)
11386 : : rem_def = def;
11387 : : else
11388 : 6777 : rem_def = gimple_build (&epilogue, reduc_code,
11389 : 6777 : TREE_TYPE (scalar_def),
11390 : : rem_def, def);
11391 : : }
11392 : 2643 : scalar_def = gimple_build (&epilogue, reduc_code,
11393 : 2643 : TREE_TYPE (scalar_def),
11394 : : scalar_def, rem_def);
11395 : : }
11396 : 4055 : scalar_def = gimple_convert (&epilogue,
11397 : 4055 : TREE_TYPE (vectype), scalar_def);
11398 : 4055 : gimple_stmt_iterator rgsi = gsi_for_stmt (instance->root_stmts[0]->stmt);
11399 : 4055 : gsi_insert_seq_before (&rgsi, epilogue, GSI_SAME_STMT);
11400 : 4055 : gimple_assign_set_rhs_from_tree (&rgsi, scalar_def);
11401 : 4055 : update_stmt (gsi_stmt (rgsi));
11402 : 4055 : return;
11403 : : }
11404 : 1305 : else if (instance->kind == slp_inst_kind_gcond)
11405 : : {
11406 : : /* Only support a single root for now as we can't codegen CFG yet and so we
11407 : : can't support lane > 1 at this time. */
11408 : 1305 : gcc_assert (instance->root_stmts.length () == 1);
11409 : 1305 : auto root_stmt_info = instance->root_stmts[0];
11410 : 1305 : auto last_stmt = STMT_VINFO_STMT (vect_orig_stmt (root_stmt_info));
11411 : 1305 : gimple_stmt_iterator rgsi = gsi_for_stmt (last_stmt);
11412 : 1305 : gimple *vec_stmt = NULL;
11413 : 1305 : gcc_assert (!SLP_TREE_VEC_DEFS (node).is_empty ());
11414 : 1305 : bool res = vectorizable_early_exit (vinfo, root_stmt_info, &rgsi,
11415 : : &vec_stmt, node, NULL);
11416 : 1305 : gcc_assert (res);
11417 : 1305 : return;
11418 : : }
11419 : : else
11420 : 0 : gcc_unreachable ();
11421 : :
11422 : 248 : gcc_assert (rstmt);
11423 : :
11424 : 248 : gimple_stmt_iterator rgsi = gsi_for_stmt (instance->root_stmts[0]->stmt);
11425 : 248 : gsi_replace (&rgsi, rstmt, true);
11426 : : }
11427 : :
11428 : : struct slp_scc_info
11429 : : {
11430 : : bool on_stack;
11431 : : int dfs;
11432 : : int lowlink;
11433 : : };
11434 : :
11435 : : /* Schedule the SLP INSTANCE doing a DFS walk and collecting SCCs. */
11436 : :
11437 : : static void
11438 : 1394407 : vect_schedule_scc (vec_info *vinfo, slp_tree node, slp_instance instance,
11439 : : hash_map<slp_tree, slp_scc_info> &scc_info,
11440 : : int &maxdfs, vec<slp_tree> &stack)
11441 : : {
11442 : 1394407 : bool existed_p;
11443 : 1394407 : slp_scc_info *info = &scc_info.get_or_insert (node, &existed_p);
11444 : 1394407 : gcc_assert (!existed_p);
11445 : 1394407 : info->dfs = maxdfs;
11446 : 1394407 : info->lowlink = maxdfs;
11447 : 1394407 : maxdfs++;
11448 : :
11449 : : /* Leaf. */
11450 : 1394407 : if (SLP_TREE_DEF_TYPE (node) != vect_internal_def)
11451 : : {
11452 : 477304 : info->on_stack = false;
11453 : 477304 : vect_schedule_slp_node (vinfo, node, instance);
11454 : 981981 : return;
11455 : : }
11456 : :
11457 : 917103 : info->on_stack = true;
11458 : 917103 : stack.safe_push (node);
11459 : :
11460 : 917103 : unsigned i;
11461 : 917103 : slp_tree child;
11462 : : /* DFS recurse. */
11463 : 1880282 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (node), i, child)
11464 : : {
11465 : 963179 : if (!child)
11466 : 52051 : continue;
11467 : 911128 : slp_scc_info *child_info = scc_info.get (child);
11468 : 911128 : if (!child_info)
11469 : : {
11470 : 833759 : vect_schedule_scc (vinfo, child, instance, scc_info, maxdfs, stack);
11471 : : /* Recursion might have re-allocated the node. */
11472 : 833759 : info = scc_info.get (node);
11473 : 833759 : child_info = scc_info.get (child);
11474 : 833759 : info->lowlink = MIN (info->lowlink, child_info->lowlink);
11475 : : }
11476 : 77369 : else if (child_info->on_stack)
11477 : 23345 : info->lowlink = MIN (info->lowlink, child_info->dfs);
11478 : : }
11479 : 917103 : if (info->lowlink != info->dfs)
11480 : : return;
11481 : :
11482 : 889730 : auto_vec<slp_tree, 4> phis_to_fixup;
11483 : :
11484 : : /* Singleton. */
11485 : 889730 : if (stack.last () == node)
11486 : : {
11487 : 867449 : stack.pop ();
11488 : 867449 : info->on_stack = false;
11489 : 867449 : vect_schedule_slp_node (vinfo, node, instance);
11490 : 867449 : if (SLP_TREE_CODE (node) != VEC_PERM_EXPR
11491 : 867449 : && is_a <gphi *> (SLP_TREE_REPRESENTATIVE (node)->stmt))
11492 : 28633 : phis_to_fixup.quick_push (node);
11493 : : }
11494 : : else
11495 : : {
11496 : : /* SCC. */
11497 : 22281 : int last_idx = stack.length () - 1;
11498 : 49654 : while (stack[last_idx] != node)
11499 : 27373 : last_idx--;
11500 : : /* We can break the cycle at PHIs who have at least one child
11501 : : code generated. Then we could re-start the DFS walk until
11502 : : all nodes in the SCC are covered (we might have new entries
11503 : : for only back-reachable nodes). But it's simpler to just
11504 : : iterate and schedule those that are ready. */
11505 : 22281 : unsigned todo = stack.length () - last_idx;
11506 : 22434 : do
11507 : : {
11508 : 95218 : for (int idx = stack.length () - 1; idx >= last_idx; --idx)
11509 : : {
11510 : 50350 : slp_tree entry = stack[idx];
11511 : 50350 : if (!entry)
11512 : 378 : continue;
11513 : 49972 : bool phi = (SLP_TREE_CODE (entry) != VEC_PERM_EXPR
11514 : 49972 : && is_a <gphi *> (SLP_TREE_REPRESENTATIVE (entry)->stmt));
11515 : 49972 : bool ready = !phi;
11516 : 125446 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (entry), i, child)
11517 : 98471 : if (!child)
11518 : : {
11519 : 21668 : gcc_assert (phi);
11520 : : ready = true;
11521 : : break;
11522 : : }
11523 : 76803 : else if (scc_info.get (child)->on_stack)
11524 : : {
11525 : 22000 : if (!phi)
11526 : : {
11527 : : ready = false;
11528 : : break;
11529 : : }
11530 : : }
11531 : : else
11532 : : {
11533 : 54803 : if (phi)
11534 : : {
11535 : : ready = true;
11536 : : break;
11537 : : }
11538 : : }
11539 : 28304 : if (ready)
11540 : : {
11541 : 49654 : vect_schedule_slp_node (vinfo, entry, instance);
11542 : 49654 : scc_info.get (entry)->on_stack = false;
11543 : 49654 : stack[idx] = NULL;
11544 : 49654 : todo--;
11545 : 49654 : if (phi)
11546 : 22703 : phis_to_fixup.safe_push (entry);
11547 : : }
11548 : : }
11549 : : }
11550 : 22434 : while (todo != 0);
11551 : :
11552 : : /* Pop the SCC. */
11553 : 22281 : stack.truncate (last_idx);
11554 : : }
11555 : :
11556 : : /* Now fixup the backedge def of the vectorized PHIs in this SCC. */
11557 : : slp_tree phi_node;
11558 : 1830796 : FOR_EACH_VEC_ELT (phis_to_fixup, i, phi_node)
11559 : : {
11560 : 51336 : gphi *phi = as_a <gphi *> (SLP_TREE_REPRESENTATIVE (phi_node)->stmt);
11561 : 51336 : edge_iterator ei;
11562 : 51336 : edge e;
11563 : 159278 : FOR_EACH_EDGE (e, ei, gimple_bb (phi)->preds)
11564 : : {
11565 : 107942 : unsigned dest_idx = e->dest_idx;
11566 : 107942 : child = SLP_TREE_CHILDREN (phi_node)[dest_idx];
11567 : 107942 : if (!child || SLP_TREE_DEF_TYPE (child) != vect_internal_def)
11568 : 61535 : continue;
11569 : 46407 : unsigned n = SLP_TREE_VEC_DEFS (phi_node).length ();
11570 : : /* Simply fill all args. */
11571 : 46407 : if (STMT_VINFO_DEF_TYPE (SLP_TREE_REPRESENTATIVE (phi_node))
11572 : : != vect_first_order_recurrence)
11573 : 95685 : for (unsigned i = 0; i < n; ++i)
11574 : : {
11575 : 49310 : tree phidef = SLP_TREE_VEC_DEFS (phi_node)[i];
11576 : 49310 : gphi *phi = as_a <gphi *> (SSA_NAME_DEF_STMT (phidef));
11577 : 49310 : add_phi_arg (phi, vect_get_slp_vect_def (child, i),
11578 : : e, gimple_phi_arg_location (phi, dest_idx));
11579 : : }
11580 : : else
11581 : : {
11582 : : /* Unless it is a first order recurrence which needs
11583 : : args filled in for both the PHI node and the permutes. */
11584 : 32 : gimple *perm
11585 : 32 : = SSA_NAME_DEF_STMT (SLP_TREE_VEC_DEFS (phi_node)[0]);
11586 : 32 : gimple *rphi = SSA_NAME_DEF_STMT (gimple_assign_rhs1 (perm));
11587 : 32 : add_phi_arg (as_a <gphi *> (rphi),
11588 : : vect_get_slp_vect_def (child, n - 1),
11589 : : e, gimple_phi_arg_location (phi, dest_idx));
11590 : 84 : for (unsigned i = 0; i < n; ++i)
11591 : : {
11592 : 52 : gimple *perm
11593 : 52 : = SSA_NAME_DEF_STMT (SLP_TREE_VEC_DEFS (phi_node)[i]);
11594 : 52 : if (i > 0)
11595 : 20 : gimple_assign_set_rhs1 (perm,
11596 : : vect_get_slp_vect_def (child, i - 1));
11597 : 52 : gimple_assign_set_rhs2 (perm,
11598 : : vect_get_slp_vect_def (child, i));
11599 : 52 : update_stmt (perm);
11600 : : }
11601 : : }
11602 : : }
11603 : : }
11604 : 889730 : }
11605 : :
11606 : : /* Generate vector code for SLP_INSTANCES in the loop/basic block. */
11607 : :
11608 : : void
11609 : 524801 : vect_schedule_slp (vec_info *vinfo, const vec<slp_instance> &slp_instances)
11610 : : {
11611 : 524801 : slp_instance instance;
11612 : 524801 : unsigned int i;
11613 : :
11614 : 524801 : hash_map<slp_tree, slp_scc_info> scc_info;
11615 : 524801 : int maxdfs = 0;
11616 : 1085613 : FOR_EACH_VEC_ELT (slp_instances, i, instance)
11617 : : {
11618 : 560812 : slp_tree node = SLP_INSTANCE_TREE (instance);
11619 : 560812 : if (dump_enabled_p ())
11620 : : {
11621 : 15260 : dump_printf_loc (MSG_NOTE, vect_location,
11622 : : "Vectorizing SLP tree:\n");
11623 : : /* ??? Dump all? */
11624 : 15260 : if (!SLP_INSTANCE_ROOT_STMTS (instance).is_empty ())
11625 : 425 : dump_printf_loc (MSG_NOTE, vect_location, "Root stmt: %G",
11626 : 425 : SLP_INSTANCE_ROOT_STMTS (instance)[0]->stmt);
11627 : 15260 : vect_print_slp_graph (MSG_NOTE, vect_location,
11628 : : SLP_INSTANCE_TREE (instance));
11629 : : }
11630 : : /* Schedule the tree of INSTANCE, scheduling SCCs in a way to
11631 : : have a PHI be the node breaking the cycle. */
11632 : 560812 : auto_vec<slp_tree> stack;
11633 : 560812 : if (!scc_info.get (node))
11634 : 560648 : vect_schedule_scc (vinfo, node, instance, scc_info, maxdfs, stack);
11635 : :
11636 : 560812 : if (!SLP_INSTANCE_ROOT_STMTS (instance).is_empty ())
11637 : 5608 : vectorize_slp_instance_root_stmt (vinfo, node, instance);
11638 : :
11639 : 560812 : if (dump_enabled_p ())
11640 : 15260 : dump_printf_loc (MSG_NOTE, vect_location,
11641 : : "vectorizing stmts using SLP.\n");
11642 : 560812 : }
11643 : :
11644 : 1610414 : FOR_EACH_VEC_ELT (slp_instances, i, instance)
11645 : : {
11646 : 560812 : slp_tree root = SLP_INSTANCE_TREE (instance);
11647 : 560812 : stmt_vec_info store_info;
11648 : 560812 : unsigned int j;
11649 : :
11650 : : /* Remove scalar call stmts. Do not do this for basic-block
11651 : : vectorization as not all uses may be vectorized.
11652 : : ??? Why should this be necessary? DCE should be able to
11653 : : remove the stmts itself.
11654 : : ??? For BB vectorization we can as well remove scalar
11655 : : stmts starting from the SLP tree root if they have no
11656 : : uses. */
11657 : 560812 : if (is_a <loop_vec_info> (vinfo))
11658 : 84398 : vect_remove_slp_scalar_calls (vinfo, root);
11659 : :
11660 : : /* Remove vectorized stores original scalar stmts. */
11661 : 2503723 : for (j = 0; SLP_TREE_SCALAR_STMTS (root).iterate (j, &store_info); j++)
11662 : : {
11663 : 1413852 : if (!STMT_VINFO_DATA_REF (store_info)
11664 : 1386015 : || !DR_IS_WRITE (STMT_VINFO_DATA_REF (store_info)))
11665 : : break;
11666 : :
11667 : 1382099 : store_info = vect_orig_stmt (store_info);
11668 : : /* Free the attached stmt_vec_info and remove the stmt. */
11669 : 1382099 : vinfo->remove_stmt (store_info);
11670 : :
11671 : : /* Invalidate SLP_TREE_REPRESENTATIVE in case we released it
11672 : : to not crash in vect_free_slp_tree later. */
11673 : 1382099 : if (SLP_TREE_REPRESENTATIVE (root) == store_info)
11674 : 528824 : SLP_TREE_REPRESENTATIVE (root) = NULL;
11675 : : }
11676 : : }
11677 : 524801 : }
|
