Line data Source code
1 : /* SLP - Pattern matcher on SLP trees
2 : Copyright (C) 2020-2026 Free Software Foundation, Inc.
3 :
4 : This file is part of GCC.
5 :
6 : GCC is free software; you can redistribute it and/or modify it under
7 : the terms of the GNU General Public License as published by the Free
8 : Software Foundation; either version 3, or (at your option) any later
9 : version.
10 :
11 : GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 : WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 : FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 : for more details.
15 :
16 : You should have received a copy of the GNU General Public License
17 : along with GCC; see the file COPYING3. If not see
18 : <http://www.gnu.org/licenses/>. */
19 :
20 : #include "config.h"
21 : #include "system.h"
22 : #include "coretypes.h"
23 : #include "backend.h"
24 : #include "target.h"
25 : #include "rtl.h"
26 : #include "tree.h"
27 : #include "gimple.h"
28 : #include "tree-pass.h"
29 : #include "ssa.h"
30 : #include "optabs-tree.h"
31 : #include "insn-config.h"
32 : #include "recog.h" /* FIXME: for insn_data */
33 : #include "fold-const.h"
34 : #include "stor-layout.h"
35 : #include "gimple-iterator.h"
36 : #include "cfgloop.h"
37 : #include "tree-vectorizer.h"
38 : #include "langhooks.h"
39 : #include "gimple-walk.h"
40 : #include "dbgcnt.h"
41 : #include "tree-vector-builder.h"
42 : #include "vec-perm-indices.h"
43 : #include "gimple-fold.h"
44 : #include "internal-fn.h"
45 :
46 : /* SLP Pattern matching mechanism.
47 :
48 : This extension to the SLP vectorizer allows one to transform the generated SLP
49 : tree based on any pattern. The difference between this and the normal vect
50 : pattern matcher is that unlike the former, this matcher allows you to match
51 : with instructions that do not belong to the same SSA dominator graph.
52 :
53 : The only requirement that this pattern matcher has is that you are only
54 : only allowed to either match an entire group or none.
55 :
56 : The pattern matcher currently only allows you to perform replacements to
57 : internal functions.
58 :
59 : Once the patterns are matched it is one way, these cannot be undone. It is
60 : currently not supported to match patterns recursively.
61 :
62 : To add a new pattern, implement the vect_pattern class and add the type to
63 : slp_patterns.
64 :
65 : */
66 :
67 : /*******************************************************************************
68 : * vect_pattern class
69 : ******************************************************************************/
70 :
71 : /* Default implementation of recognize that performs matching, validation and
72 : replacement of nodes but that can be overriden if required. */
73 :
74 : static bool
75 4674 : vect_pattern_validate_optab (internal_fn ifn, slp_tree node)
76 : {
77 4674 : tree vectype = SLP_TREE_VECTYPE (node);
78 4674 : if (ifn == IFN_LAST || !vectype)
79 : return false;
80 :
81 4674 : if (dump_enabled_p ())
82 701 : dump_printf_loc (MSG_NOTE, vect_location,
83 : "Found %s pattern in SLP tree\n",
84 : internal_fn_name (ifn));
85 :
86 4674 : if (direct_internal_fn_supported_p (ifn, vectype, OPTIMIZE_FOR_SPEED))
87 : {
88 1096 : if (dump_enabled_p ())
89 17 : dump_printf_loc (MSG_NOTE, vect_location,
90 : "Target supports %s vectorization with mode %T\n",
91 : internal_fn_name (ifn), vectype);
92 : }
93 : else
94 : {
95 3578 : if (dump_enabled_p ())
96 : {
97 684 : if (!vectype)
98 : dump_printf_loc (MSG_NOTE, vect_location,
99 : "Target does not support vector type for %G\n",
100 : STMT_VINFO_STMT (SLP_TREE_REPRESENTATIVE (node)));
101 : else
102 684 : dump_printf_loc (MSG_NOTE, vect_location,
103 : "Target does not support %s for vector type "
104 : "%T\n", internal_fn_name (ifn), vectype);
105 : }
106 3578 : return false;
107 : }
108 : return true;
109 : }
110 :
111 : /*******************************************************************************
112 : * General helper types
113 : ******************************************************************************/
114 :
115 : /* The COMPLEX_OPERATION enum denotes the possible pair of operations that can
116 : be matched when looking for expressions that we are interested matching for
117 : complex numbers addition and mla. */
118 :
119 : typedef enum _complex_operation : unsigned {
120 : PLUS_PLUS,
121 : MINUS_PLUS,
122 : PLUS_MINUS,
123 : MULT_MULT,
124 : CMPLX_NONE
125 : } complex_operation_t;
126 :
127 : /*******************************************************************************
128 : * General helper functions
129 : ******************************************************************************/
130 :
131 : /* Helper function of linear_loads_p that checks to see if the load permutation
132 : is sequential and in monotonically increasing order of loads with no gaps.
133 : */
134 :
135 : static inline complex_perm_kinds_t
136 2036 : is_linear_load_p (load_permutation_t loads)
137 : {
138 2102 : if (loads.length() == 0)
139 : return PERM_UNKNOWN;
140 :
141 2036 : unsigned load, i;
142 2036 : complex_perm_kinds_t candidates[4]
143 : = { PERM_ODDODD
144 : , PERM_EVENEVEN
145 : , PERM_EVENODD
146 : , PERM_ODDEVEN
147 : };
148 :
149 2036 : int valid_patterns = 4;
150 7567 : FOR_EACH_VEC_ELT (loads, i, load)
151 : {
152 5597 : unsigned adj_load = load % 2;
153 5597 : if (candidates[0] != PERM_UNKNOWN && adj_load != 1)
154 : {
155 1750 : candidates[0] = PERM_UNKNOWN;
156 1750 : valid_patterns--;
157 : }
158 5597 : if (candidates[1] != PERM_UNKNOWN && adj_load != 0)
159 : {
160 1101 : candidates[1] = PERM_UNKNOWN;
161 1101 : valid_patterns--;
162 : }
163 5597 : if (candidates[2] != PERM_UNKNOWN && load != i)
164 : {
165 2001 : candidates[2] = PERM_UNKNOWN;
166 2001 : valid_patterns--;
167 : }
168 5597 : if (candidates[3] != PERM_UNKNOWN
169 4511 : && load != (i % 2 == 0 ? i + 1 : i - 1))
170 : {
171 1322 : candidates[3] = PERM_UNKNOWN;
172 1322 : valid_patterns--;
173 : }
174 :
175 5597 : if (valid_patterns == 0)
176 : return PERM_UNKNOWN;
177 : }
178 :
179 3147 : for (i = 0; i < sizeof(candidates); i++)
180 5117 : if (candidates[i] != PERM_UNKNOWN)
181 : return candidates[i];
182 :
183 : return PERM_UNKNOWN;
184 : }
185 :
186 : /* Combine complex_perm_kinds A and B into a new permute kind that describes the
187 : resulting operation. */
188 :
189 : static inline complex_perm_kinds_t
190 15722 : vect_merge_perms (complex_perm_kinds_t a, complex_perm_kinds_t b)
191 : {
192 15722 : if (a == b)
193 : return a;
194 :
195 13326 : if (a == PERM_TOP)
196 : return b;
197 :
198 1809 : if (b == PERM_TOP)
199 : return a;
200 :
201 : return PERM_UNKNOWN;
202 : }
203 :
204 : /* Check to see if all loads rooted in ROOT are linear. Linearity is
205 : defined as having no gaps between values loaded. */
206 :
207 : static complex_perm_kinds_t
208 26235 : linear_loads_p (slp_tree_to_load_perm_map_t *perm_cache, slp_tree root)
209 : {
210 26235 : if (!root)
211 : return PERM_UNKNOWN;
212 :
213 26230 : unsigned i;
214 26230 : complex_perm_kinds_t *tmp;
215 :
216 26230 : if ((tmp = perm_cache->get (root)) != NULL)
217 6838 : return *tmp;
218 :
219 19392 : complex_perm_kinds_t retval = PERM_UNKNOWN;
220 19392 : perm_cache->put (root, retval);
221 :
222 : /* If it's a load node, then just read the load permute. */
223 19392 : if (SLP_TREE_DEF_TYPE (root) == vect_internal_def
224 17074 : && !SLP_TREE_PERMUTE_P (root)
225 14977 : && STMT_VINFO_DATA_REF (SLP_TREE_REPRESENTATIVE (root))
226 3231 : && DR_IS_READ (STMT_VINFO_DATA_REF (SLP_TREE_REPRESENTATIVE (root))))
227 : {
228 3231 : if (SLP_TREE_LOAD_PERMUTATION (root).exists ())
229 2036 : retval = is_linear_load_p (SLP_TREE_LOAD_PERMUTATION (root));
230 : else
231 1195 : retval = PERM_EVENODD;
232 3231 : perm_cache->put (root, retval);
233 3231 : return retval;
234 : }
235 16161 : else if (SLP_TREE_DEF_TYPE (root) != vect_internal_def)
236 : {
237 2318 : retval = PERM_TOP;
238 2318 : perm_cache->put (root, retval);
239 2318 : return retval;
240 : }
241 :
242 : complex_perm_kinds_t kind = PERM_TOP;
243 :
244 : slp_tree child;
245 16205 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (root), i, child)
246 : {
247 15722 : complex_perm_kinds_t res = linear_loads_p (perm_cache, child);
248 15722 : kind = vect_merge_perms (kind, res);
249 : /* Unknown and Top are not valid on blends as they produce no permute. */
250 15722 : retval = kind;
251 15722 : if (kind == PERM_UNKNOWN || kind == PERM_TOP)
252 : return retval;
253 : }
254 :
255 483 : retval = kind;
256 :
257 483 : perm_cache->put (root, retval);
258 483 : return retval;
259 : }
260 :
261 :
262 : /* This function attempts to make a node rooted in NODE is linear. If the node
263 : if already linear than the node itself is returned in RESULT.
264 :
265 : If the node is not linear then a new VEC_PERM_EXPR node is created with a
266 : lane permute that when applied will make the node linear. If such a
267 : permute cannot be created then FALSE is returned from the function.
268 :
269 : Here linearity is defined as having a sequential, monotically increasing
270 : load position inside the load permute generated by the loads reachable from
271 : NODE. */
272 :
273 : static slp_tree
274 0 : vect_build_swap_evenodd_node (slp_tree node)
275 : {
276 : /* Attempt to linearise the permute. */
277 0 : vec<std::pair<unsigned, unsigned> > zipped;
278 0 : zipped.create (SLP_TREE_LANES (node));
279 :
280 0 : for (unsigned x = 0; x < SLP_TREE_LANES (node); x+=2)
281 : {
282 0 : zipped.quick_push (std::make_pair (0, x+1));
283 0 : zipped.quick_push (std::make_pair (0, x));
284 : }
285 :
286 : /* Create the new permute node and store it instead. */
287 0 : slp_tree vnode = vect_create_new_slp_node (1, VEC_PERM_EXPR);
288 0 : SLP_TREE_LANE_PERMUTATION (vnode) = zipped;
289 0 : SLP_TREE_VECTYPE (vnode) = SLP_TREE_VECTYPE (node);
290 0 : SLP_TREE_CHILDREN (vnode).quick_push (node);
291 0 : SLP_TREE_REF_COUNT (vnode) = 1;
292 0 : SLP_TREE_LANES (vnode) = SLP_TREE_LANES (node);
293 0 : SLP_TREE_REPRESENTATIVE (vnode) = SLP_TREE_REPRESENTATIVE (node);
294 0 : SLP_TREE_REF_COUNT (node)++;
295 0 : return vnode;
296 : }
297 :
298 : /* Checks to see of the expression represented by NODE is a gimple assign with
299 : code CODE. */
300 :
301 : static inline bool
302 10535691 : vect_match_expression_p (slp_tree node, code_helper code)
303 : {
304 10535691 : if (!node
305 9764446 : || SLP_TREE_PERMUTE_P (node)
306 9715370 : || !SLP_TREE_REPRESENTATIVE (node))
307 : return false;
308 :
309 7388462 : gimple* expr = STMT_VINFO_STMT (SLP_TREE_REPRESENTATIVE (node));
310 7388462 : if (is_gimple_assign (expr)
311 6325466 : && code.is_tree_code ()
312 13707917 : && gimple_assign_rhs_code (expr) == (tree_code) code)
313 : return true;
314 6935521 : if (is_a <gcall *> (expr)
315 62451 : && !code.is_tree_code ()
316 6935569 : && gimple_call_combined_fn (expr) == (combined_fn) code)
317 : return true;
318 :
319 : return false;
320 : }
321 :
322 : /* Check if the given lane permute in PERMUTES matches an alternating sequence
323 : of {even odd even odd ...}. This to account for unrolled loops. Further
324 : mode there resulting permute must be linear. */
325 :
326 : static inline bool
327 6338 : vect_check_evenodd_blend (lane_permutation_t &permutes,
328 : unsigned even, unsigned odd)
329 : {
330 6659 : if (permutes.length () == 0
331 6149 : || permutes.length () % 2 != 0)
332 : return false;
333 :
334 6127 : unsigned val[2] = {even, odd};
335 6127 : unsigned seed = 0;
336 21412 : for (unsigned i = 0; i < permutes.length (); i++)
337 15584 : if (permutes[i].first != val[i % 2]
338 15584 : || permutes[i].second != seed++)
339 : return false;
340 :
341 : return true;
342 : }
343 :
344 : /* This function will match the two gimple expressions representing NODE1 and
345 : NODE2 in parallel and returns the pair operation that represents the two
346 : expressions in the two statements.
347 :
348 : If match is successful then the corresponding complex_operation is
349 : returned and the arguments to the two matched operations are returned in OPS.
350 :
351 : If TWO_OPERANDS it is expected that the LANES of the parent VEC_PERM select
352 : from the two nodes alternatingly.
353 :
354 : If unsuccessful then CMPLX_NONE is returned and OPS is untouched.
355 :
356 : e.g. the following gimple statements
357 :
358 : stmt 0 _39 = _37 + _12;
359 : stmt 1 _6 = _38 - _36;
360 :
361 : will return PLUS_MINUS along with OPS containing {_37, _12, _38, _36}.
362 : */
363 :
364 : static complex_operation_t
365 1343055 : vect_detect_pair_op (slp_tree node1, slp_tree node2, lane_permutation_t &lanes,
366 : bool two_operands = true, vec<slp_tree> *ops = NULL)
367 : {
368 1343055 : complex_operation_t result = CMPLX_NONE;
369 :
370 1343055 : if (vect_match_expression_p (node1, MINUS_EXPR)
371 42306 : && vect_match_expression_p (node2, PLUS_EXPR)
372 1346274 : && (!two_operands || vect_check_evenodd_blend (lanes, 0, 1)))
373 : result = MINUS_PLUS;
374 1340118 : else if (vect_match_expression_p (node1, PLUS_EXPR)
375 135975 : && vect_match_expression_p (node2, MINUS_EXPR)
376 1343237 : && (!two_operands || vect_check_evenodd_blend (lanes, 0, 1)))
377 : result = PLUS_MINUS;
378 1337227 : else if (vect_match_expression_p (node1, PLUS_EXPR)
379 1337227 : && vect_match_expression_p (node2, PLUS_EXPR))
380 : result = PLUS_PLUS;
381 1333789 : else if (vect_match_expression_p (node1, MULT_EXPR)
382 1333789 : && vect_match_expression_p (node2, MULT_EXPR))
383 3658 : result = MULT_MULT;
384 :
385 1343055 : if (result != CMPLX_NONE && ops != NULL)
386 : {
387 12895 : if (two_operands)
388 : {
389 12895 : auto l0node = SLP_TREE_CHILDREN (node1);
390 12895 : auto l1node = SLP_TREE_CHILDREN (node2);
391 :
392 : /* Check if the tree is connected as we expect it. */
393 19042 : if (!((l0node[0] == l1node[0] && l0node[1] == l1node[1])
394 6850 : || (l0node[0] == l1node[1] && l0node[1] == l1node[0])))
395 1343055 : return CMPLX_NONE;
396 : }
397 6069 : ops->safe_push (node1);
398 6069 : ops->safe_push (node2);
399 : }
400 : return result;
401 : }
402 :
403 : /* Overload of vect_detect_pair_op that matches against the representative
404 : statements in the children of NODE. It is expected that NODE has exactly
405 : two children and when TWO_OPERANDS then NODE must be a VEC_PERM. */
406 :
407 : static complex_operation_t
408 4733471 : vect_detect_pair_op (slp_tree node, bool two_operands = true,
409 : vec<slp_tree> *ops = NULL)
410 : {
411 4733471 : if (!two_operands && SLP_TREE_PERMUTE_P (node))
412 : return CMPLX_NONE;
413 :
414 4733471 : if (SLP_TREE_CHILDREN (node).length () != 2)
415 : return CMPLX_NONE;
416 :
417 1343055 : vec<slp_tree> children = SLP_TREE_CHILDREN (node);
418 1343055 : lane_permutation_t &lanes = SLP_TREE_LANE_PERMUTATION (node);
419 :
420 1343055 : return vect_detect_pair_op (children[0], children[1], lanes, two_operands,
421 1343055 : ops);
422 : }
423 :
424 : /*******************************************************************************
425 : * complex_pattern class
426 : ******************************************************************************/
427 :
428 : /* SLP Complex Numbers pattern matching.
429 :
430 : As an example, the following simple loop:
431 :
432 : double a[restrict N]; double b[restrict N]; double c[restrict N];
433 :
434 : for (int i=0; i < N; i+=2)
435 : {
436 : c[i] = a[i] - b[i+1];
437 : c[i+1] = a[i+1] + b[i];
438 : }
439 :
440 : which represents a complex addition on with a rotation of 90* around the
441 : argand plane. i.e. if `a` and `b` were complex numbers then this would be the
442 : same as `a + (b * I)`.
443 :
444 : Here the expressions for `c[i]` and `c[i+1]` are independent but have to be
445 : both recognized in order for the pattern to work. As an SLP tree this is
446 : represented as
447 :
448 : +--------------------------------+
449 : | stmt 0 *_9 = _10; |
450 : | stmt 1 *_15 = _16; |
451 : +--------------------------------+
452 : |
453 : |
454 : v
455 : +--------------------------------+
456 : | stmt 0 _10 = _4 - _8; |
457 : | stmt 1 _16 = _12 + _14; |
458 : | lane permutation { 0[0] 1[1] } |
459 : +--------------------------------+
460 : | |
461 : | |
462 : | |
463 : +-----+ | | +-----+
464 : | | | | | |
465 : +-----| { } |<-----+ +----->| { } --------+
466 : | | | +------------------| | |
467 : | +-----+ | +-----+ |
468 : | | | |
469 : | | | |
470 : | +------|------------------+ |
471 : | | | |
472 : v v v v
473 : +--------------------------+ +--------------------------------+
474 : | stmt 0 _8 = *_7; | | stmt 0 _4 = *_3; |
475 : | stmt 1 _14 = *_13; | | stmt 1 _12 = *_11; |
476 : | load permutation { 1 0 } | | load permutation { 0 1 } |
477 : +--------------------------+ +--------------------------------+
478 :
479 : The pattern matcher allows you to replace both statements 0 and 1 or none at
480 : all. Because this operation is a two operands operation the actual nodes
481 : being replaced are those in the { } nodes. The actual scalar statements
482 : themselves are not replaced or used during the matching but instead the
483 : SLP_TREE_REPRESENTATIVE statements are inspected. You are also allowed to
484 : replace and match on any number of nodes.
485 :
486 : Because the pattern matcher matches on the representative statement for the
487 : SLP node the case of two_operators it allows you to match the children of the
488 : node. This is done using the method `recognize ()`.
489 :
490 : */
491 :
492 : /* The complex_pattern class contains common code for pattern matchers that work
493 : on complex numbers. These provide functionality to allow de-construction and
494 : validation of sequences depicting/transforming REAL and IMAG pairs. */
495 :
496 : class complex_pattern : public vect_pattern
497 : {
498 : protected:
499 : auto_vec<slp_tree> m_workset;
500 20 : complex_pattern (slp_tree *node, vec<slp_tree> *m_ops, internal_fn ifn)
501 40 : : vect_pattern (node, m_ops, ifn)
502 : {
503 20 : this->m_workset.safe_push (*node);
504 20 : }
505 :
506 : public:
507 : void build (vec_info *) override;
508 :
509 : static internal_fn
510 : matches (complex_operation_t op, slp_tree_to_load_perm_map_t *, slp_tree *,
511 : vec<slp_tree> *);
512 : };
513 :
514 : /* Create a replacement pattern statement for each node in m_node and inserts
515 : the new statement into m_node as the new representative statement. The old
516 : statement is marked as being in a pattern defined by the new statement. The
517 : statement is created as call to internal function IFN with m_num_args
518 : arguments.
519 :
520 : Futhermore the new pattern is also added to the vectorization information
521 : structure VINFO and the old statement STMT_INFO is marked as unused while
522 : the new statement is marked as used and the number of SLP uses of the new
523 : statement is incremented.
524 :
525 : The newly created SLP nodes are marked as SLP only and will be dissolved
526 : if SLP is aborted.
527 :
528 : The newly created gimple call is returned and the BB remains unchanged.
529 :
530 : This default method is designed to only match against simple operands where
531 : all the input and output types are the same.
532 : */
533 :
534 : void
535 20 : complex_pattern::build (vec_info *vinfo)
536 : {
537 20 : stmt_vec_info stmt_info;
538 :
539 20 : auto_vec<tree> args;
540 20 : args.create (this->m_num_args);
541 20 : args.quick_grow_cleared (this->m_num_args);
542 20 : slp_tree node;
543 20 : unsigned ix;
544 20 : stmt_vec_info call_stmt_info;
545 20 : gcall *call_stmt = NULL;
546 :
547 : /* Now modify the nodes themselves. */
548 60 : FOR_EACH_VEC_ELT (this->m_workset, ix, node)
549 : {
550 : /* Calculate the location of the statement in NODE to replace. */
551 20 : stmt_info = SLP_TREE_REPRESENTATIVE (node);
552 20 : stmt_vec_info reduc_def
553 20 : = STMT_VINFO_REDUC_DEF (vect_orig_stmt (stmt_info));
554 20 : gimple* old_stmt = STMT_VINFO_STMT (stmt_info);
555 20 : tree lhs_old_stmt = gimple_get_lhs (old_stmt);
556 20 : tree type = TREE_TYPE (lhs_old_stmt);
557 :
558 : /* Create the argument set for use by gimple_build_call_internal_vec. */
559 70 : for (unsigned i = 0; i < this->m_num_args; i++)
560 50 : args[i] = lhs_old_stmt;
561 :
562 : /* Create the new pattern statements. */
563 20 : call_stmt = gimple_build_call_internal_vec (this->m_ifn, args);
564 20 : tree var = make_temp_ssa_name (type, call_stmt, "slp_patt");
565 20 : gimple_call_set_lhs (call_stmt, var);
566 20 : gimple_set_location (call_stmt, gimple_location (old_stmt));
567 20 : gimple_call_set_nothrow (call_stmt, true);
568 :
569 : /* Adjust the book-keeping for the new and old statements for use during
570 : SLP. This is required to get the right VF and statement during SLP
571 : analysis. These changes are created after relevancy has been set for
572 : the nodes as such we need to manually update them. Any changes will be
573 : undone if SLP is cancelled. */
574 20 : call_stmt_info
575 20 : = vinfo->add_pattern_stmt (call_stmt, stmt_info);
576 :
577 : /* Make sure to mark the representative statement pure_slp and
578 : relevant and transfer reduction info. */
579 20 : STMT_VINFO_RELEVANT (call_stmt_info) = vect_used_in_scope;
580 20 : STMT_SLP_TYPE (call_stmt_info) = pure_slp;
581 20 : STMT_VINFO_REDUC_DEF (call_stmt_info) = reduc_def;
582 :
583 20 : gimple_set_bb (call_stmt, gimple_bb (stmt_info->stmt));
584 20 : STMT_VINFO_VECTYPE (call_stmt_info) = SLP_TREE_VECTYPE (node);
585 20 : STMT_VINFO_SLP_VECT_ONLY_PATTERN (call_stmt_info) = true;
586 :
587 : /* Since we are replacing all the statements in the group with the same
588 : thing it doesn't really matter. So just set it every time a new stmt
589 : is created. */
590 20 : SLP_TREE_REPRESENTATIVE (node) = call_stmt_info;
591 20 : SLP_TREE_LANE_PERMUTATION (node).release ();
592 20 : SLP_TREE_CODE (node) = CALL_EXPR;
593 : }
594 20 : }
595 :
596 : /*******************************************************************************
597 : * complex_add_pattern class
598 : ******************************************************************************/
599 :
600 : class complex_add_pattern : public complex_pattern
601 : {
602 : protected:
603 0 : complex_add_pattern (slp_tree *node, vec<slp_tree> *m_ops, internal_fn ifn)
604 0 : : complex_pattern (node, m_ops, ifn)
605 : {
606 0 : this->m_num_args = 2;
607 : }
608 :
609 : public:
610 : void build (vec_info *) final override;
611 : static internal_fn
612 : matches (complex_operation_t op, slp_tree_to_load_perm_map_t *,
613 : slp_compat_nodes_map_t *, slp_tree *, vec<slp_tree> *);
614 :
615 : static vect_pattern*
616 : recognize (slp_tree_to_load_perm_map_t *, slp_compat_nodes_map_t *,
617 : slp_tree *);
618 :
619 : static vect_pattern*
620 0 : mkInstance (slp_tree *node, vec<slp_tree> *m_ops, internal_fn ifn)
621 : {
622 0 : return new complex_add_pattern (node, m_ops, ifn);
623 : }
624 : };
625 :
626 : /* Perform a replacement of the detected complex add pattern with the new
627 : instruction sequences. */
628 :
629 : void
630 0 : complex_add_pattern::build (vec_info *vinfo)
631 : {
632 0 : SLP_TREE_CHILDREN (*this->m_node).reserve_exact (2);
633 :
634 0 : slp_tree node = this->m_ops[0];
635 0 : vec<slp_tree> children = SLP_TREE_CHILDREN (node);
636 :
637 : /* First re-arrange the children. */
638 0 : SLP_TREE_CHILDREN (*this->m_node)[0] = children[0];
639 0 : SLP_TREE_CHILDREN (*this->m_node)[1] =
640 0 : vect_build_swap_evenodd_node (children[1]);
641 :
642 0 : SLP_TREE_REF_COUNT (SLP_TREE_CHILDREN (*this->m_node)[0])++;
643 0 : SLP_TREE_REF_COUNT (SLP_TREE_CHILDREN (*this->m_node)[1])++;
644 0 : vect_free_slp_tree (this->m_ops[0]);
645 0 : vect_free_slp_tree (this->m_ops[1]);
646 :
647 0 : complex_pattern::build (vinfo);
648 0 : }
649 :
650 : /* Pattern matcher for trying to match complex addition pattern in SLP tree.
651 :
652 : If no match is found then IFN is set to IFN_LAST.
653 : This function matches the patterns shaped as:
654 :
655 : c[i] = a[i] - b[i+1];
656 : c[i+1] = a[i+1] + b[i];
657 :
658 : If a match occurred then TRUE is returned, else FALSE. The initial match is
659 : expected to be in OP1 and the initial match operands in args0. */
660 :
661 : internal_fn
662 4722760 : complex_add_pattern::matches (complex_operation_t op,
663 : slp_tree_to_load_perm_map_t *perm_cache,
664 : slp_compat_nodes_map_t * /* compat_cache */,
665 : slp_tree *node, vec<slp_tree> *ops)
666 : {
667 4722760 : internal_fn ifn = IFN_LAST;
668 :
669 : /* Find the two components. Rotation in the complex plane will modify
670 : the operations:
671 :
672 : * Rotation 0: + +
673 : * Rotation 90: - +
674 : * Rotation 180: - -
675 : * Rotation 270: + -
676 :
677 : Rotation 0 and 180 can be handled by normal SIMD code, so we don't need
678 : to care about them here. */
679 4722760 : if (op == MINUS_PLUS)
680 : ifn = IFN_COMPLEX_ADD_ROT90;
681 4719845 : else if (op == PLUS_MINUS)
682 : ifn = IFN_COMPLEX_ADD_ROT270;
683 : else
684 : return ifn;
685 :
686 : /* verify that there is a permute, otherwise this isn't a pattern we
687 : we support. */
688 5787 : gcc_assert (ops->length () == 2);
689 :
690 5787 : vec<slp_tree> children = SLP_TREE_CHILDREN ((*ops)[0]);
691 :
692 : /* First node must be unpermuted. */
693 5787 : if (linear_loads_p (perm_cache, children[0]) != PERM_EVENODD)
694 : return IFN_LAST;
695 :
696 : /* Second node must be permuted. */
697 492 : if (linear_loads_p (perm_cache, children[1]) != PERM_ODDEVEN)
698 : return IFN_LAST;
699 :
700 338 : if (!vect_pattern_validate_optab (ifn, *node))
701 : return IFN_LAST;
702 :
703 : return ifn;
704 : }
705 :
706 : /* Attempt to recognize a complex add pattern. */
707 :
708 : vect_pattern*
709 0 : complex_add_pattern::recognize (slp_tree_to_load_perm_map_t *perm_cache,
710 : slp_compat_nodes_map_t *compat_cache,
711 : slp_tree *node)
712 : {
713 0 : auto_vec<slp_tree> ops;
714 0 : complex_operation_t op
715 0 : = vect_detect_pair_op (*node, true, &ops);
716 0 : internal_fn ifn
717 0 : = complex_add_pattern::matches (op, perm_cache, compat_cache, node, &ops);
718 0 : if (ifn == IFN_LAST)
719 : return NULL;
720 :
721 0 : return new complex_add_pattern (node, &ops, ifn);
722 0 : }
723 :
724 : /*******************************************************************************
725 : * complex_mul_pattern
726 : ******************************************************************************/
727 :
728 : /* Helper function to check if PERM is KIND or PERM_TOP. */
729 :
730 : static inline bool
731 533 : is_eq_or_top (slp_tree_to_load_perm_map_t *perm_cache,
732 : slp_tree op1, complex_perm_kinds_t kind1,
733 : slp_tree op2, complex_perm_kinds_t kind2)
734 : {
735 533 : complex_perm_kinds_t perm1 = linear_loads_p (perm_cache, op1);
736 533 : if (perm1 != kind1 && perm1 != PERM_TOP)
737 : return false;
738 :
739 173 : complex_perm_kinds_t perm2 = linear_loads_p (perm_cache, op2);
740 173 : if (perm2 != kind2 && perm2 != PERM_TOP)
741 : return false;
742 :
743 : return true;
744 : }
745 :
746 : enum _conj_status { CONJ_NONE, CONJ_FST, CONJ_SND };
747 :
748 : static inline bool
749 369 : compatible_complex_nodes_p (slp_compat_nodes_map_t *compat_cache,
750 : slp_tree a, int *pa, slp_tree b, int *pb)
751 : {
752 369 : bool *tmp;
753 369 : std::pair<slp_tree, slp_tree> key = std::make_pair(a, b);
754 369 : if ((tmp = compat_cache->get (key)) != NULL)
755 27 : return *tmp;
756 :
757 342 : compat_cache->put (key, false);
758 :
759 406 : if (SLP_TREE_CHILDREN (a).length () != SLP_TREE_CHILDREN (b).length ())
760 : return false;
761 :
762 340 : if (SLP_TREE_DEF_TYPE (a) != SLP_TREE_DEF_TYPE (b))
763 : return false;
764 :
765 : /* Only internal nodes can be loads, as such we can't check further if they
766 : are externals. */
767 340 : if (SLP_TREE_DEF_TYPE (a) != vect_internal_def)
768 : {
769 188 : for (unsigned i = 0; i < SLP_TREE_SCALAR_OPS (a).length (); i++)
770 : {
771 130 : tree op1 = SLP_TREE_SCALAR_OPS (a)[pa[i % 2]];
772 130 : tree op2 = SLP_TREE_SCALAR_OPS (b)[pb[i % 2]];
773 130 : if (!operand_equal_p (op1, op2, 0))
774 : return false;
775 : }
776 :
777 58 : compat_cache->put (key, true);
778 58 : return true;
779 : }
780 :
781 280 : auto a_stmt = STMT_VINFO_STMT (SLP_TREE_REPRESENTATIVE (a));
782 280 : auto b_stmt = STMT_VINFO_STMT (SLP_TREE_REPRESENTATIVE (b));
783 :
784 280 : if (gimple_code (a_stmt) != gimple_code (b_stmt))
785 : return false;
786 :
787 : /* code, children, type, externals, loads, constants */
788 280 : if (gimple_num_args (a_stmt) != gimple_num_args (b_stmt))
789 : return false;
790 :
791 : /* At this point, a and b are known to be the same gimple operations. */
792 280 : if (is_gimple_call (a_stmt))
793 : {
794 0 : if (!compatible_calls_p (dyn_cast <gcall *> (a_stmt),
795 : dyn_cast <gcall *> (b_stmt), false))
796 : return false;
797 : }
798 280 : else if (!is_gimple_assign (a_stmt))
799 : return false;
800 : else
801 : {
802 280 : tree_code acode = gimple_assign_rhs_code (a_stmt);
803 280 : tree_code bcode = gimple_assign_rhs_code (b_stmt);
804 280 : if ((acode == REALPART_EXPR || acode == IMAGPART_EXPR)
805 171 : && (bcode == REALPART_EXPR || bcode == IMAGPART_EXPR))
806 : return true;
807 :
808 109 : if (acode != bcode)
809 : return false;
810 : }
811 :
812 109 : if (!STMT_VINFO_DATA_REF (SLP_TREE_REPRESENTATIVE (a))
813 78 : || !STMT_VINFO_DATA_REF (SLP_TREE_REPRESENTATIVE (b)))
814 : {
815 92 : for (unsigned i = 0; i < gimple_num_args (a_stmt); i++)
816 : {
817 61 : tree t1 = gimple_arg (a_stmt, i);
818 61 : tree t2 = gimple_arg (b_stmt, i);
819 61 : if (TREE_CODE (t1) != TREE_CODE (t2))
820 : return false;
821 :
822 : /* If SSA name then we will need to inspect the children
823 : so we can punt here. */
824 61 : if (TREE_CODE (t1) == SSA_NAME)
825 43 : continue;
826 :
827 18 : if (!operand_equal_p (t1, t2, 0))
828 : return false;
829 : }
830 : }
831 : else
832 : {
833 78 : auto dr1 = STMT_VINFO_DATA_REF (SLP_TREE_REPRESENTATIVE (a));
834 78 : auto dr2 = STMT_VINFO_DATA_REF (SLP_TREE_REPRESENTATIVE (b));
835 : /* Don't check the last dimension as that's checked by the lineary
836 : checks. This check is also much stricter than what we need
837 : because it doesn't consider loading from adjacent elements
838 : in the same struct as loading from the same base object.
839 : But for now, I'll play it safe. */
840 78 : if (!same_data_refs (dr1, dr2, 1))
841 : return false;
842 : }
843 :
844 148 : for (unsigned i = 0; i < SLP_TREE_CHILDREN (a).length (); i++)
845 : {
846 61 : if (!compatible_complex_nodes_p (compat_cache,
847 61 : SLP_TREE_CHILDREN (a)[i], pa,
848 61 : SLP_TREE_CHILDREN (b)[i], pb))
849 : return false;
850 : }
851 :
852 87 : compat_cache->put (key, true);
853 87 : return true;
854 : }
855 :
856 :
857 : /* Check to see if the oprands to two multiplies, 2 each in LEFT_OP and
858 : RIGHT_OP match a complex multiplication or complex multiply-and-accumulate
859 : or complex multiply-and-subtract pattern. Do this using the permute cache
860 : PERM_CACHE and the combination compatibility list COMPAT_CACHE. If
861 : the operation is successful the macthing operands are returned in OPS and
862 : _STATUS indicates if the operation matched includes a conjugate of one of the
863 : operands. If the operation succeeds True is returned, otherwise False and
864 : the values in ops are meaningless. */
865 : static inline bool
866 1506 : vect_validate_multiplication (slp_tree_to_load_perm_map_t *perm_cache,
867 : slp_compat_nodes_map_t *compat_cache,
868 : const vec<slp_tree> &left_op,
869 : const vec<slp_tree> &right_op,
870 : bool subtract, vec<slp_tree> &ops,
871 : enum _conj_status *_status)
872 : {
873 1506 : enum _conj_status stats = CONJ_NONE;
874 :
875 : /* The complex operations can occur in two layouts and two permute sequences
876 : so declare them and re-use them. */
877 1506 : int styles[][4] = { { 0, 2, 1, 3} /* {L1, R1} + {L2, R2}. */
878 : , { 0, 3, 1, 2} /* {L1, R2} + {L2, R1}. */
879 : };
880 :
881 : /* Now for the corresponding permutes that go with these values. */
882 1506 : complex_perm_kinds_t perms[][4]
883 : = { { PERM_EVENEVEN, PERM_ODDODD, PERM_EVENODD, PERM_ODDEVEN }
884 : , { PERM_EVENODD, PERM_ODDEVEN, PERM_EVENEVEN, PERM_ODDODD }
885 : };
886 :
887 : /* These permutes are used during comparisons of externals on which
888 : we require strict equality. */
889 1506 : int cq[][4][2]
890 : = { { { 0, 0 }, { 1, 1 }, { 0, 1 }, { 1, 0 } }
891 : , { { 0, 1 }, { 1, 0 }, { 0, 0 }, { 1, 1 } }
892 : };
893 :
894 : /* Default to style and perm 0, most operations use this one. */
895 1506 : int style = 0;
896 1506 : int perm = subtract ? 1 : 0;
897 :
898 : /* Check if we have a negate operation, if so absorb the node and continue
899 : looking. */
900 1506 : bool neg0 = vect_match_expression_p (right_op[0], NEGATE_EXPR);
901 1506 : bool neg1 = vect_match_expression_p (right_op[1], NEGATE_EXPR);
902 :
903 : /* Create the combined inputs after remapping and flattening. */
904 1506 : ops.create (4);
905 1506 : ops.safe_splice (left_op);
906 1506 : ops.safe_splice (right_op);
907 :
908 : /* Determine which style we're looking at. We only have different ones
909 : whenever a conjugate is involved. */
910 1506 : if (neg0 && neg1)
911 : ;
912 1506 : else if (neg0)
913 : {
914 0 : ops[2] = SLP_TREE_CHILDREN (right_op[0])[0];
915 0 : stats = CONJ_FST;
916 0 : if (subtract)
917 0 : perm = 0;
918 : }
919 1506 : else if (neg1)
920 : {
921 10 : ops[3] = SLP_TREE_CHILDREN (right_op[1])[0];
922 10 : stats = CONJ_SND;
923 10 : perm = 1;
924 : }
925 :
926 1506 : *_status = stats;
927 :
928 : /* Extract out the elements to check. */
929 1506 : slp_tree op0 = ops[styles[style][0]];
930 1506 : slp_tree op1 = ops[styles[style][1]];
931 1506 : slp_tree op2 = ops[styles[style][2]];
932 1506 : slp_tree op3 = ops[styles[style][3]];
933 :
934 : /* Do cheapest test first. If failed no need to analyze further. */
935 1506 : if (linear_loads_p (perm_cache, op0) != perms[perm][0]
936 589 : || linear_loads_p (perm_cache, op1) != perms[perm][1]
937 2039 : || !is_eq_or_top (perm_cache, op2, perms[perm][2], op3, perms[perm][3]))
938 1333 : return false;
939 :
940 173 : return compatible_complex_nodes_p (compat_cache, op0, cq[perm][0], op1,
941 173 : cq[perm][1])
942 308 : && compatible_complex_nodes_p (compat_cache, op2, cq[perm][2], op3,
943 135 : cq[perm][3]);
944 : }
945 :
946 : /* This function combines two nodes containing only even and only odd lanes
947 : together into a single node which contains the nodes in even/odd order
948 : by using a lane permute.
949 :
950 : The lanes in EVEN and ODD are duplicated 2 times inside the vectors.
951 : So for a lanes = 4 EVEN contains {EVEN1, EVEN1, EVEN2, EVEN2}.
952 :
953 : The tree REPRESENTATION is taken from the supplied REP along with the
954 : vectype which must be the same between all three nodes.
955 : */
956 :
957 : static slp_tree
958 20 : vect_build_combine_node (slp_tree even, slp_tree odd, slp_tree rep)
959 : {
960 20 : vec<std::pair<unsigned, unsigned> > perm;
961 20 : perm.create (SLP_TREE_LANES (rep));
962 :
963 40 : for (unsigned x = 0; x < SLP_TREE_LANES (rep); x+=2)
964 : {
965 20 : perm.quick_push (std::make_pair (0, x));
966 20 : perm.quick_push (std::make_pair (1, x+1));
967 : }
968 :
969 20 : slp_tree vnode = vect_create_new_slp_node (2, SLP_TREE_CODE (even));
970 20 : SLP_TREE_CODE (vnode) = VEC_PERM_EXPR;
971 20 : SLP_TREE_LANE_PERMUTATION (vnode) = perm;
972 :
973 20 : SLP_TREE_CHILDREN (vnode).create (2);
974 20 : SLP_TREE_CHILDREN (vnode).quick_push (even);
975 20 : SLP_TREE_CHILDREN (vnode).quick_push (odd);
976 20 : SLP_TREE_REF_COUNT (even)++;
977 20 : SLP_TREE_REF_COUNT (odd)++;
978 20 : SLP_TREE_REF_COUNT (vnode) = 1;
979 :
980 20 : SLP_TREE_LANES (vnode) = SLP_TREE_LANES (rep);
981 40 : gcc_assert (perm.length () == SLP_TREE_LANES (vnode));
982 : /* Representation is set to that of the current node as the vectorizer
983 : can't deal with VEC_PERMs with no representation, as would be the
984 : case with invariants. */
985 20 : SLP_TREE_REPRESENTATIVE (vnode) = SLP_TREE_REPRESENTATIVE (rep);
986 20 : SLP_TREE_VECTYPE (vnode) = SLP_TREE_VECTYPE (rep);
987 20 : return vnode;
988 : }
989 :
990 : class complex_mul_pattern : public complex_pattern
991 : {
992 : protected:
993 20 : complex_mul_pattern (slp_tree *node, vec<slp_tree> *m_ops, internal_fn ifn)
994 40 : : complex_pattern (node, m_ops, ifn)
995 : {
996 20 : this->m_num_args = 2;
997 : }
998 :
999 : public:
1000 : void build (vec_info *) final override;
1001 : static internal_fn
1002 : matches (complex_operation_t op, slp_tree_to_load_perm_map_t *,
1003 : slp_compat_nodes_map_t *, slp_tree *, vec<slp_tree> *);
1004 :
1005 : static vect_pattern*
1006 : recognize (slp_tree_to_load_perm_map_t *, slp_compat_nodes_map_t *,
1007 : slp_tree *);
1008 :
1009 : static vect_pattern*
1010 20 : mkInstance (slp_tree *node, vec<slp_tree> *m_ops, internal_fn ifn)
1011 : {
1012 20 : return new complex_mul_pattern (node, m_ops, ifn);
1013 : }
1014 :
1015 : };
1016 :
1017 : /* Pattern matcher for trying to match complex multiply and complex multiply
1018 : and accumulate pattern in SLP tree. If the operation matches then IFN
1019 : is set to the operation it matched and the arguments to the two
1020 : replacement statements are put in m_ops.
1021 :
1022 : If no match is found then IFN is set to IFN_LAST and m_ops is unchanged.
1023 :
1024 : This function matches the patterns shaped as:
1025 :
1026 : double ax = (b[i+1] * a[i]);
1027 : double bx = (a[i+1] * b[i]);
1028 :
1029 : c[i] = c[i] - ax;
1030 : c[i+1] = c[i+1] + bx;
1031 :
1032 : If a match occurred then TRUE is returned, else FALSE. The initial match is
1033 : expected to be in OP1 and the initial match operands in args0. */
1034 :
1035 : internal_fn
1036 4722780 : complex_mul_pattern::matches (complex_operation_t op,
1037 : slp_tree_to_load_perm_map_t *perm_cache,
1038 : slp_compat_nodes_map_t *compat_cache,
1039 : slp_tree *node, vec<slp_tree> *ops)
1040 : {
1041 4722780 : internal_fn ifn = IFN_LAST;
1042 :
1043 4722780 : if (op != MINUS_PLUS)
1044 : return IFN_LAST;
1045 :
1046 2935 : auto childs = *ops;
1047 2935 : auto l0node = SLP_TREE_CHILDREN (childs[0]);
1048 :
1049 2935 : bool mul0 = vect_match_expression_p (l0node[0], MULT_EXPR);
1050 2935 : bool mul1 = vect_match_expression_p (l0node[1], MULT_EXPR);
1051 2935 : if (!mul0 && !mul1)
1052 : return IFN_LAST;
1053 :
1054 : /* Now operand2+4 may lead to another expression. */
1055 2019 : auto_vec<slp_tree> left_op, right_op;
1056 2019 : slp_tree add0 = NULL;
1057 :
1058 : /* Check if we may be a multiply add. It's only valid to form FMAs
1059 : with -ffp-contract=fast. */
1060 2019 : if (!mul0
1061 1204 : && (flag_fp_contract_mode == FP_CONTRACT_FAST
1062 3 : || !FLOAT_TYPE_P (SLP_TREE_VECTYPE (*node)))
1063 3220 : && vect_match_expression_p (l0node[0], PLUS_EXPR))
1064 : {
1065 1144 : auto vals = SLP_TREE_CHILDREN (l0node[0]);
1066 : /* Check if it's a multiply, otherwise no idea what this is. */
1067 1144 : if (!(mul0 = vect_match_expression_p (vals[1], MULT_EXPR)))
1068 2019 : return IFN_LAST;
1069 :
1070 : /* Check if the ADD is linear, otherwise it's not valid complex FMA. */
1071 633 : if (linear_loads_p (perm_cache, vals[0]) != PERM_EVENODD)
1072 : return IFN_LAST;
1073 :
1074 18 : left_op.safe_splice (SLP_TREE_CHILDREN (vals[1]));
1075 18 : add0 = vals[0];
1076 : }
1077 : else
1078 875 : left_op.safe_splice (SLP_TREE_CHILDREN (l0node[0]));
1079 :
1080 893 : right_op.safe_splice (SLP_TREE_CHILDREN (l0node[1]));
1081 :
1082 893 : if (left_op.length () != 2
1083 789 : || right_op.length () != 2
1084 : || !mul0
1085 788 : || !mul1
1086 1619 : || linear_loads_p (perm_cache, left_op[1]) == PERM_ODDEVEN)
1087 113 : return IFN_LAST;
1088 :
1089 780 : enum _conj_status status;
1090 780 : auto_vec<slp_tree> res_ops;
1091 780 : if (!vect_validate_multiplication (perm_cache, compat_cache, left_op,
1092 : right_op, false, res_ops, &status))
1093 : {
1094 : /* Try swapping the order and re-trying since multiplication is
1095 : commutative. */
1096 700 : std::swap (left_op[0], left_op[1]);
1097 700 : std::swap (right_op[0], right_op[1]);
1098 700 : if (!vect_validate_multiplication (perm_cache, compat_cache, left_op,
1099 : right_op, false, res_ops, &status))
1100 : return IFN_LAST;
1101 : }
1102 :
1103 126 : if (status == CONJ_NONE)
1104 : {
1105 116 : if (add0)
1106 : ifn = IFN_COMPLEX_FMA;
1107 : else
1108 111 : ifn = IFN_COMPLEX_MUL;
1109 : }
1110 : else
1111 : {
1112 10 : if(add0)
1113 : ifn = IFN_COMPLEX_FMA_CONJ;
1114 : else
1115 5 : ifn = IFN_COMPLEX_MUL_CONJ;
1116 : }
1117 :
1118 126 : if (!vect_pattern_validate_optab (ifn, *node))
1119 : return IFN_LAST;
1120 :
1121 20 : ops->truncate (0);
1122 30 : ops->create (add0 ? 4 : 3);
1123 :
1124 20 : if (add0)
1125 10 : ops->quick_push (add0);
1126 :
1127 20 : complex_perm_kinds_t kind = linear_loads_p (perm_cache, res_ops[0]);
1128 20 : if (kind == PERM_EVENODD || kind == PERM_TOP)
1129 : {
1130 10 : ops->quick_push (res_ops[1]);
1131 10 : ops->quick_push (res_ops[3]);
1132 10 : ops->quick_push (res_ops[0]);
1133 : }
1134 10 : else if (kind == PERM_EVENEVEN && status != CONJ_SND)
1135 : {
1136 10 : ops->quick_push (res_ops[0]);
1137 10 : ops->quick_push (res_ops[2]);
1138 10 : ops->quick_push (res_ops[1]);
1139 : }
1140 : else
1141 : {
1142 0 : ops->quick_push (res_ops[0]);
1143 0 : ops->quick_push (res_ops[3]);
1144 0 : ops->quick_push (res_ops[1]);
1145 : }
1146 :
1147 : return ifn;
1148 2799 : }
1149 :
1150 : /* Attempt to recognize a complex mul pattern. */
1151 :
1152 : vect_pattern*
1153 0 : complex_mul_pattern::recognize (slp_tree_to_load_perm_map_t *perm_cache,
1154 : slp_compat_nodes_map_t *compat_cache,
1155 : slp_tree *node)
1156 : {
1157 0 : auto_vec<slp_tree> ops;
1158 0 : complex_operation_t op
1159 0 : = vect_detect_pair_op (*node, true, &ops);
1160 0 : internal_fn ifn
1161 0 : = complex_mul_pattern::matches (op, perm_cache, compat_cache, node, &ops);
1162 0 : if (ifn == IFN_LAST)
1163 : return NULL;
1164 :
1165 0 : return new complex_mul_pattern (node, &ops, ifn);
1166 0 : }
1167 :
1168 : /* Perform a replacement of the detected complex mul pattern with the new
1169 : instruction sequences. */
1170 :
1171 : void
1172 20 : complex_mul_pattern::build (vec_info *vinfo)
1173 : {
1174 20 : slp_tree node;
1175 20 : unsigned i;
1176 20 : switch (this->m_ifn)
1177 : {
1178 10 : case IFN_COMPLEX_MUL:
1179 10 : case IFN_COMPLEX_MUL_CONJ:
1180 10 : {
1181 10 : slp_tree newnode
1182 10 : = vect_build_combine_node (this->m_ops[0], this->m_ops[1],
1183 10 : *this->m_node);
1184 10 : SLP_TREE_REF_COUNT (this->m_ops[2])++;
1185 :
1186 30 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (*this->m_node), i, node)
1187 20 : vect_free_slp_tree (node);
1188 :
1189 : /* First re-arrange the children. */
1190 10 : SLP_TREE_CHILDREN (*this->m_node).reserve_exact (2);
1191 10 : SLP_TREE_CHILDREN (*this->m_node)[0] = this->m_ops[2];
1192 10 : SLP_TREE_CHILDREN (*this->m_node)[1] = newnode;
1193 10 : break;
1194 : }
1195 10 : case IFN_COMPLEX_FMA:
1196 10 : case IFN_COMPLEX_FMA_CONJ:
1197 10 : {
1198 10 : SLP_TREE_REF_COUNT (this->m_ops[0])++;
1199 10 : slp_tree newnode
1200 10 : = vect_build_combine_node (this->m_ops[1], this->m_ops[2],
1201 10 : *this->m_node);
1202 10 : SLP_TREE_REF_COUNT (this->m_ops[3])++;
1203 :
1204 30 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (*this->m_node), i, node)
1205 20 : vect_free_slp_tree (node);
1206 :
1207 : /* First re-arrange the children. */
1208 10 : SLP_TREE_CHILDREN (*this->m_node).safe_grow (3);
1209 10 : SLP_TREE_CHILDREN (*this->m_node)[0] = this->m_ops[3];
1210 10 : SLP_TREE_CHILDREN (*this->m_node)[1] = newnode;
1211 10 : SLP_TREE_CHILDREN (*this->m_node)[2] = this->m_ops[0];
1212 :
1213 : /* Tell the builder to expect an extra argument. */
1214 10 : this->m_num_args++;
1215 10 : break;
1216 : }
1217 0 : default:
1218 0 : gcc_unreachable ();
1219 : }
1220 :
1221 : /* And then rewrite the node itself. */
1222 20 : complex_pattern::build (vinfo);
1223 20 : }
1224 :
1225 : /*******************************************************************************
1226 : * complex_fms_pattern class
1227 : ******************************************************************************/
1228 :
1229 : class complex_fms_pattern : public complex_pattern
1230 : {
1231 : protected:
1232 0 : complex_fms_pattern (slp_tree *node, vec<slp_tree> *m_ops, internal_fn ifn)
1233 0 : : complex_pattern (node, m_ops, ifn)
1234 : {
1235 0 : this->m_num_args = 3;
1236 : }
1237 :
1238 : public:
1239 : void build (vec_info *) final override;
1240 : static internal_fn
1241 : matches (complex_operation_t op, slp_tree_to_load_perm_map_t *,
1242 : slp_compat_nodes_map_t *, slp_tree *, vec<slp_tree> *);
1243 :
1244 : static vect_pattern*
1245 : recognize (slp_tree_to_load_perm_map_t *, slp_compat_nodes_map_t *,
1246 : slp_tree *);
1247 :
1248 : static vect_pattern*
1249 0 : mkInstance (slp_tree *node, vec<slp_tree> *m_ops, internal_fn ifn)
1250 : {
1251 0 : return new complex_fms_pattern (node, m_ops, ifn);
1252 : }
1253 : };
1254 :
1255 :
1256 : /* Pattern matcher for trying to match complex multiply and subtract pattern
1257 : in SLP tree. If the operation matches then IFN is set to the operation
1258 : it matched and the arguments to the two replacement statements are put in
1259 : m_ops.
1260 :
1261 : If no match is found then IFN is set to IFN_LAST and m_ops is unchanged.
1262 :
1263 : This function matches the patterns shaped as:
1264 :
1265 : double ax = (b[i+1] * a[i]) + (b[i] * a[i]);
1266 : double bx = (a[i+1] * b[i]) - (a[i+1] * b[i+1]);
1267 :
1268 : c[i] = c[i] - ax;
1269 : c[i+1] = c[i+1] + bx;
1270 :
1271 : If a match occurred then TRUE is returned, else FALSE. The initial match is
1272 : expected to be in OP1 and the initial match operands in args0. */
1273 :
1274 : internal_fn
1275 4722780 : complex_fms_pattern::matches (complex_operation_t op,
1276 : slp_tree_to_load_perm_map_t *perm_cache,
1277 : slp_compat_nodes_map_t *compat_cache,
1278 : slp_tree * ref_node, vec<slp_tree> *ops)
1279 : {
1280 4722780 : internal_fn ifn = IFN_LAST;
1281 :
1282 : /* We need to ignore the two_operands nodes that may also match,
1283 : for that we can check if they have any scalar statements and also
1284 : check that it's not a permute node as we're looking for a normal
1285 : MINUS_EXPR operation. */
1286 4722780 : if (op != CMPLX_NONE)
1287 : return IFN_LAST;
1288 :
1289 4716711 : slp_tree root = *ref_node;
1290 4716711 : if (!vect_match_expression_p (root, MINUS_EXPR))
1291 : return IFN_LAST;
1292 :
1293 : /* TODO: Support invariants here, with the new layout CADD now
1294 : can match before we get a chance to try CFMS. */
1295 59246 : auto nodes = SLP_TREE_CHILDREN (root);
1296 59246 : if (!vect_match_expression_p (nodes[1], MULT_EXPR)
1297 69937 : || vect_detect_pair_op (nodes[0]) != PLUS_MINUS)
1298 59227 : return IFN_LAST;
1299 :
1300 19 : auto childs = SLP_TREE_CHILDREN (nodes[0]);
1301 19 : auto l0node = SLP_TREE_CHILDREN (childs[0]);
1302 :
1303 : /* Now operand2+4 may lead to another expression. */
1304 19 : auto_vec<slp_tree> left_op, right_op;
1305 19 : left_op.safe_splice (SLP_TREE_CHILDREN (l0node[1]));
1306 19 : right_op.safe_splice (SLP_TREE_CHILDREN (nodes[1]));
1307 :
1308 : /* If these nodes don't have any children then they're
1309 : not ones we're interested in. */
1310 19 : if (left_op.length () != 2
1311 13 : || right_op.length () != 2
1312 26 : || !vect_match_expression_p (l0node[1], MULT_EXPR))
1313 6 : return IFN_LAST;
1314 :
1315 13 : enum _conj_status status;
1316 13 : auto_vec<slp_tree> res_ops;
1317 13 : if (!vect_validate_multiplication (perm_cache, compat_cache, right_op,
1318 : left_op, true, res_ops, &status))
1319 : {
1320 : /* Try swapping the order and re-trying since multiplication is
1321 : commutative. */
1322 13 : std::swap (left_op[0], left_op[1]);
1323 13 : std::swap (right_op[0], right_op[1]);
1324 13 : auto_vec<slp_tree> res_ops;
1325 13 : if (!vect_validate_multiplication (perm_cache, compat_cache, right_op,
1326 : left_op, true, res_ops, &status))
1327 13 : return IFN_LAST;
1328 13 : }
1329 :
1330 0 : if (status == CONJ_NONE)
1331 : ifn = IFN_COMPLEX_FMS;
1332 : else
1333 0 : ifn = IFN_COMPLEX_FMS_CONJ;
1334 :
1335 0 : if (!vect_pattern_validate_optab (ifn, *ref_node))
1336 : return IFN_LAST;
1337 :
1338 0 : ops->truncate (0);
1339 0 : ops->create (4);
1340 :
1341 0 : complex_perm_kinds_t kind = linear_loads_p (perm_cache, res_ops[2]);
1342 0 : if (kind == PERM_EVENODD)
1343 : {
1344 0 : ops->quick_push (l0node[0]);
1345 0 : ops->quick_push (res_ops[2]);
1346 0 : ops->quick_push (res_ops[3]);
1347 0 : ops->quick_push (res_ops[1]);
1348 : }
1349 : else
1350 : {
1351 0 : ops->quick_push (l0node[0]);
1352 0 : ops->quick_push (res_ops[3]);
1353 0 : ops->quick_push (res_ops[2]);
1354 0 : ops->quick_push (res_ops[0]);
1355 : }
1356 :
1357 : return ifn;
1358 32 : }
1359 :
1360 : /* Attempt to recognize a complex mul pattern. */
1361 :
1362 : vect_pattern*
1363 0 : complex_fms_pattern::recognize (slp_tree_to_load_perm_map_t *perm_cache,
1364 : slp_compat_nodes_map_t *compat_cache,
1365 : slp_tree *node)
1366 : {
1367 0 : auto_vec<slp_tree> ops;
1368 0 : complex_operation_t op
1369 0 : = vect_detect_pair_op (*node, true, &ops);
1370 0 : internal_fn ifn
1371 0 : = complex_fms_pattern::matches (op, perm_cache, compat_cache, node, &ops);
1372 0 : if (ifn == IFN_LAST)
1373 : return NULL;
1374 :
1375 0 : return new complex_fms_pattern (node, &ops, ifn);
1376 0 : }
1377 :
1378 : /* Perform a replacement of the detected complex mul pattern with the new
1379 : instruction sequences. */
1380 :
1381 : void
1382 0 : complex_fms_pattern::build (vec_info *vinfo)
1383 : {
1384 0 : slp_tree node;
1385 0 : unsigned i;
1386 0 : slp_tree newnode =
1387 0 : vect_build_combine_node (this->m_ops[2], this->m_ops[3], *this->m_node);
1388 0 : SLP_TREE_REF_COUNT (this->m_ops[0])++;
1389 0 : SLP_TREE_REF_COUNT (this->m_ops[1])++;
1390 :
1391 0 : FOR_EACH_VEC_ELT (SLP_TREE_CHILDREN (*this->m_node), i, node)
1392 0 : vect_free_slp_tree (node);
1393 :
1394 0 : SLP_TREE_CHILDREN (*this->m_node).release ();
1395 0 : SLP_TREE_CHILDREN (*this->m_node).create (3);
1396 :
1397 : /* First re-arrange the children. */
1398 0 : SLP_TREE_CHILDREN (*this->m_node).quick_push (this->m_ops[1]);
1399 0 : SLP_TREE_CHILDREN (*this->m_node).quick_push (newnode);
1400 0 : SLP_TREE_CHILDREN (*this->m_node).quick_push (this->m_ops[0]);
1401 :
1402 : /* And then rewrite the node itself. */
1403 0 : complex_pattern::build (vinfo);
1404 0 : }
1405 :
1406 : /*******************************************************************************
1407 : * complex_operations_pattern class
1408 : ******************************************************************************/
1409 :
1410 : /* This function combines all the existing pattern matchers above into one class
1411 : that shares the functionality between them. The initial match is shared
1412 : between all complex operations. */
1413 :
1414 : class complex_operations_pattern : public complex_pattern
1415 : {
1416 : protected:
1417 : complex_operations_pattern (slp_tree *node, vec<slp_tree> *m_ops,
1418 : internal_fn ifn)
1419 : : complex_pattern (node, m_ops, ifn)
1420 : {
1421 : this->m_num_args = 0;
1422 : }
1423 :
1424 : public:
1425 : void build (vec_info *) final override;
1426 : static internal_fn
1427 : matches (complex_operation_t op, slp_tree_to_load_perm_map_t *,
1428 : slp_compat_nodes_map_t *, slp_tree *, vec<slp_tree> *);
1429 :
1430 : static vect_pattern*
1431 : recognize (slp_tree_to_load_perm_map_t *, slp_compat_nodes_map_t *,
1432 : slp_tree *);
1433 : };
1434 :
1435 : /* Dummy matches implementation for proxy object. */
1436 :
1437 : internal_fn
1438 0 : complex_operations_pattern::
1439 : matches (complex_operation_t /* op */,
1440 : slp_tree_to_load_perm_map_t * /* perm_cache */,
1441 : slp_compat_nodes_map_t * /* compat_cache */,
1442 : slp_tree * /* ref_node */, vec<slp_tree> * /* ops */)
1443 : {
1444 0 : return IFN_LAST;
1445 : }
1446 :
1447 : /* Attempt to recognize a complex mul pattern. */
1448 :
1449 : vect_pattern*
1450 4722780 : complex_operations_pattern::recognize (slp_tree_to_load_perm_map_t *perm_cache,
1451 : slp_compat_nodes_map_t *ccache,
1452 : slp_tree *node)
1453 : {
1454 4722780 : auto_vec<slp_tree> ops;
1455 4722780 : complex_operation_t op
1456 4722780 : = vect_detect_pair_op (*node, true, &ops);
1457 4722780 : internal_fn ifn = IFN_LAST;
1458 :
1459 4722780 : ifn = complex_fms_pattern::matches (op, perm_cache, ccache, node, &ops);
1460 4722780 : if (ifn != IFN_LAST)
1461 0 : return complex_fms_pattern::mkInstance (node, &ops, ifn);
1462 :
1463 4722780 : ifn = complex_mul_pattern::matches (op, perm_cache, ccache, node, &ops);
1464 4722780 : if (ifn != IFN_LAST)
1465 20 : return complex_mul_pattern::mkInstance (node, &ops, ifn);
1466 :
1467 4722760 : ifn = complex_add_pattern::matches (op, perm_cache, ccache, node, &ops);
1468 4722760 : if (ifn != IFN_LAST)
1469 0 : return complex_add_pattern::mkInstance (node, &ops, ifn);
1470 :
1471 : return NULL;
1472 4722780 : }
1473 :
1474 : /* Dummy implementation of build. */
1475 :
1476 : void
1477 0 : complex_operations_pattern::build (vec_info * /* vinfo */)
1478 : {
1479 0 : gcc_unreachable ();
1480 : }
1481 :
1482 :
1483 : /* The addsub_pattern. */
1484 :
1485 : class addsub_pattern : public vect_pattern
1486 : {
1487 : public:
1488 1076 : addsub_pattern (slp_tree *node, internal_fn ifn)
1489 1076 : : vect_pattern (node, NULL, ifn) {};
1490 :
1491 : void build (vec_info *) final override;
1492 :
1493 : static vect_pattern*
1494 : recognize (slp_tree_to_load_perm_map_t *, slp_compat_nodes_map_t *,
1495 : slp_tree *);
1496 : };
1497 :
1498 : vect_pattern *
1499 4722780 : addsub_pattern::recognize (slp_tree_to_load_perm_map_t *,
1500 : slp_compat_nodes_map_t *, slp_tree *node_)
1501 : {
1502 4722780 : slp_tree node = *node_;
1503 4722780 : if (!SLP_TREE_PERMUTE_P (node)
1504 20476 : || SLP_TREE_CHILDREN (node).length () != 2
1505 4740961 : || SLP_TREE_LANE_PERMUTATION (node).length () % 2)
1506 : return NULL;
1507 :
1508 : /* Match a blend of a plus and a minus op with the same number of plus and
1509 : minus lanes on the same operands. */
1510 13429 : unsigned l0 = SLP_TREE_LANE_PERMUTATION (node)[0].first;
1511 13429 : unsigned l1 = SLP_TREE_LANE_PERMUTATION (node)[1].first;
1512 13429 : if (l0 == l1)
1513 : return NULL;
1514 11305 : bool fma_p = false;
1515 11305 : bool l0add_p = vect_match_expression_p (SLP_TREE_CHILDREN (node)[l0],
1516 11305 : PLUS_EXPR);
1517 11305 : if (!l0add_p
1518 11305 : && !vect_match_expression_p (SLP_TREE_CHILDREN (node)[l0], MINUS_EXPR))
1519 : {
1520 4309 : l0add_p = vect_match_expression_p (SLP_TREE_CHILDREN (node)[l0], CFN_FMA);
1521 4309 : if (!l0add_p
1522 4309 : && !vect_match_expression_p (SLP_TREE_CHILDREN (node)[l0], CFN_FMS))
1523 4307 : return NULL;
1524 : fma_p = true;
1525 : }
1526 6998 : bool l1add_p = vect_match_expression_p (SLP_TREE_CHILDREN (node)[l1],
1527 6998 : PLUS_EXPR);
1528 6998 : if (l1add_p && fma_p)
1529 : return NULL;
1530 6998 : if (!l1add_p
1531 6998 : && !vect_match_expression_p (SLP_TREE_CHILDREN (node)[l1], MINUS_EXPR))
1532 : {
1533 718 : if (!fma_p)
1534 : return NULL;
1535 2 : l1add_p = vect_match_expression_p (SLP_TREE_CHILDREN (node)[l1], CFN_FMA);
1536 2 : if (!l1add_p
1537 2 : && !vect_match_expression_p (SLP_TREE_CHILDREN (node)[l1], CFN_FMS))
1538 0 : return NULL;
1539 : }
1540 6280 : else if (!l1add_p && fma_p)
1541 : return NULL;
1542 :
1543 6282 : slp_tree l0node = SLP_TREE_CHILDREN (node)[l0];
1544 6282 : slp_tree l1node = SLP_TREE_CHILDREN (node)[l1];
1545 6282 : if (!((SLP_TREE_CHILDREN (l0node)[0] == SLP_TREE_CHILDREN (l1node)[0]
1546 5949 : && SLP_TREE_CHILDREN (l0node)[1] == SLP_TREE_CHILDREN (l1node)[1])
1547 347 : || (SLP_TREE_CHILDREN (l0node)[0] == SLP_TREE_CHILDREN (l1node)[1]
1548 0 : && SLP_TREE_CHILDREN (l0node)[1] == SLP_TREE_CHILDREN (l1node)[0])))
1549 : return NULL;
1550 :
1551 20985 : for (unsigned i = 0; i < SLP_TREE_LANE_PERMUTATION (node).length (); ++i)
1552 : {
1553 15196 : std::pair<unsigned, unsigned> perm = SLP_TREE_LANE_PERMUTATION (node)[i];
1554 : /* It has to be alternating -, +, -,
1555 : While we could permute the .ADDSUB inputs and the .ADDSUB output
1556 : that's only profitable over the add + sub + blend if at least
1557 : one of the permute is optimized which we can't determine here. */
1558 22842 : if (perm.first != ((i & 1) ? l1 : l0)
1559 15096 : || perm.second != i)
1560 4721704 : return NULL;
1561 : }
1562 :
1563 : /* Now we have either { -, +, -, + ... } (!l0add_p) or { +, -, +, - ... }
1564 : (l0add_p), see whether we have FMA variants. We can only form FMAs
1565 : if allowed via -ffp-contract=fast or if they were FMA before. */
1566 5789 : if (!fma_p
1567 5787 : && flag_fp_contract_mode != FP_CONTRACT_FAST
1568 5820 : && FLOAT_TYPE_P (SLP_TREE_VECTYPE (l0node)))
1569 : ;
1570 5758 : else if (!l0add_p
1571 5758 : && (fma_p
1572 2885 : || vect_match_expression_p (SLP_TREE_CHILDREN (l0node)[0],
1573 2885 : MULT_EXPR)))
1574 : {
1575 : /* (c * d) -+ a */
1576 780 : if (vect_pattern_validate_optab (IFN_VEC_FMADDSUB, node))
1577 23 : return new addsub_pattern (node_, IFN_VEC_FMADDSUB);
1578 : }
1579 4978 : else if (l0add_p
1580 4978 : && (fma_p
1581 4976 : || vect_match_expression_p (SLP_TREE_CHILDREN (l1node)[0],
1582 2871 : MULT_EXPR)))
1583 : {
1584 : /* (c * d) +- a */
1585 538 : if (vect_pattern_validate_optab (IFN_VEC_FMSUBADD, node))
1586 18 : return new addsub_pattern (node_, IFN_VEC_FMSUBADD);
1587 : }
1588 :
1589 5748 : if (!fma_p && !l0add_p && vect_pattern_validate_optab (IFN_VEC_ADDSUB, node))
1590 1035 : return new addsub_pattern (node_, IFN_VEC_ADDSUB);
1591 :
1592 : return NULL;
1593 : }
1594 :
1595 : void
1596 1076 : addsub_pattern::build (vec_info *vinfo)
1597 : {
1598 1076 : slp_tree node = *m_node;
1599 :
1600 1076 : unsigned l0 = SLP_TREE_LANE_PERMUTATION (node)[0].first;
1601 1076 : unsigned l1 = SLP_TREE_LANE_PERMUTATION (node)[1].first;
1602 :
1603 1076 : switch (m_ifn)
1604 : {
1605 1035 : case IFN_VEC_ADDSUB:
1606 1035 : {
1607 1035 : slp_tree sub = SLP_TREE_CHILDREN (node)[l0];
1608 1035 : slp_tree add = SLP_TREE_CHILDREN (node)[l1];
1609 :
1610 : /* Modify the blend node in-place. */
1611 1035 : SLP_TREE_CHILDREN (node)[0] = SLP_TREE_CHILDREN (sub)[0];
1612 1035 : SLP_TREE_CHILDREN (node)[1] = SLP_TREE_CHILDREN (sub)[1];
1613 1035 : SLP_TREE_REF_COUNT (SLP_TREE_CHILDREN (node)[0])++;
1614 1035 : SLP_TREE_REF_COUNT (SLP_TREE_CHILDREN (node)[1])++;
1615 :
1616 : /* Build IFN_VEC_ADDSUB from the sub representative operands. */
1617 1035 : stmt_vec_info rep = SLP_TREE_REPRESENTATIVE (sub);
1618 1035 : gcall *call = gimple_build_call_internal (IFN_VEC_ADDSUB, 2,
1619 : gimple_assign_rhs1 (rep->stmt),
1620 1035 : gimple_assign_rhs2 (rep->stmt));
1621 1035 : gimple_call_set_lhs (call, make_ssa_name
1622 1035 : (TREE_TYPE (gimple_assign_lhs (rep->stmt))));
1623 1035 : gimple_call_set_nothrow (call, true);
1624 1035 : gimple_set_bb (call, gimple_bb (rep->stmt));
1625 1035 : stmt_vec_info new_rep = vinfo->add_pattern_stmt (call, rep);
1626 1035 : SLP_TREE_REPRESENTATIVE (node) = new_rep;
1627 1035 : STMT_VINFO_RELEVANT (new_rep) = vect_used_in_scope;
1628 1035 : STMT_SLP_TYPE (new_rep) = pure_slp;
1629 1035 : STMT_VINFO_VECTYPE (new_rep) = SLP_TREE_VECTYPE (node);
1630 1035 : STMT_VINFO_SLP_VECT_ONLY_PATTERN (new_rep) = true;
1631 1035 : STMT_VINFO_REDUC_DEF (new_rep) = STMT_VINFO_REDUC_DEF (vect_orig_stmt (rep));
1632 1035 : SLP_TREE_CODE (node) = ERROR_MARK;
1633 1035 : SLP_TREE_LANE_PERMUTATION (node).release ();
1634 :
1635 1035 : vect_free_slp_tree (sub);
1636 1035 : vect_free_slp_tree (add);
1637 1035 : break;
1638 : }
1639 41 : case IFN_VEC_FMADDSUB:
1640 41 : case IFN_VEC_FMSUBADD:
1641 41 : {
1642 41 : slp_tree sub, add;
1643 41 : if (m_ifn == IFN_VEC_FMADDSUB)
1644 : {
1645 23 : sub = SLP_TREE_CHILDREN (node)[l0];
1646 23 : add = SLP_TREE_CHILDREN (node)[l1];
1647 : }
1648 : else /* m_ifn == IFN_VEC_FMSUBADD */
1649 : {
1650 18 : sub = SLP_TREE_CHILDREN (node)[l1];
1651 18 : add = SLP_TREE_CHILDREN (node)[l0];
1652 : }
1653 : /* Modify the blend node in-place. */
1654 41 : SLP_TREE_CHILDREN (node).safe_grow (3, true);
1655 41 : gcall *call;
1656 41 : stmt_vec_info srep = SLP_TREE_REPRESENTATIVE (sub);
1657 41 : if (vect_match_expression_p (add, CFN_FMA))
1658 : {
1659 2 : SLP_TREE_CHILDREN (node)[0] = SLP_TREE_CHILDREN (add)[0];
1660 2 : SLP_TREE_CHILDREN (node)[1] = SLP_TREE_CHILDREN (add)[1];
1661 2 : SLP_TREE_CHILDREN (node)[2] = SLP_TREE_CHILDREN (add)[2];
1662 : /* Build IFN_VEC_FMADDSUB from the fms representative
1663 : operands. */
1664 2 : call = gimple_build_call_internal (m_ifn, 3,
1665 : gimple_call_arg (srep->stmt, 0),
1666 : gimple_call_arg (srep->stmt, 1),
1667 2 : gimple_call_arg (srep->stmt, 2));
1668 : }
1669 : else
1670 : {
1671 39 : slp_tree mul = SLP_TREE_CHILDREN (sub)[0];
1672 39 : SLP_TREE_CHILDREN (node)[0] = SLP_TREE_CHILDREN (mul)[0];
1673 39 : SLP_TREE_CHILDREN (node)[1] = SLP_TREE_CHILDREN (mul)[1];
1674 39 : SLP_TREE_CHILDREN (node)[2] = SLP_TREE_CHILDREN (sub)[1];
1675 : /* Build IFN_VEC_FMADDSUB from the mul/sub representative
1676 : operands. */
1677 39 : stmt_vec_info mrep = SLP_TREE_REPRESENTATIVE (mul);
1678 39 : call = gimple_build_call_internal (m_ifn, 3,
1679 : gimple_assign_rhs1 (mrep->stmt),
1680 39 : gimple_assign_rhs2 (mrep->stmt),
1681 39 : gimple_assign_rhs2 (srep->stmt));
1682 : }
1683 41 : SLP_TREE_REF_COUNT (SLP_TREE_CHILDREN (node)[0])++;
1684 41 : SLP_TREE_REF_COUNT (SLP_TREE_CHILDREN (node)[1])++;
1685 41 : SLP_TREE_REF_COUNT (SLP_TREE_CHILDREN (node)[2])++;
1686 :
1687 41 : gimple_call_set_lhs (call, make_ssa_name
1688 41 : (TREE_TYPE (gimple_get_lhs (srep->stmt))));
1689 41 : gimple_call_set_nothrow (call, true);
1690 41 : gimple_set_bb (call, gimple_bb (srep->stmt));
1691 41 : stmt_vec_info new_rep = vinfo->add_pattern_stmt (call, srep);
1692 41 : SLP_TREE_REPRESENTATIVE (node) = new_rep;
1693 41 : STMT_VINFO_RELEVANT (new_rep) = vect_used_in_scope;
1694 41 : STMT_SLP_TYPE (new_rep) = pure_slp;
1695 41 : STMT_VINFO_VECTYPE (new_rep) = SLP_TREE_VECTYPE (node);
1696 41 : STMT_VINFO_SLP_VECT_ONLY_PATTERN (new_rep) = true;
1697 41 : STMT_VINFO_REDUC_DEF (new_rep) = STMT_VINFO_REDUC_DEF (vect_orig_stmt (srep));
1698 41 : SLP_TREE_CODE (node) = ERROR_MARK;
1699 41 : SLP_TREE_LANE_PERMUTATION (node).release ();
1700 :
1701 41 : vect_free_slp_tree (sub);
1702 41 : vect_free_slp_tree (add);
1703 41 : break;
1704 : }
1705 1076 : default:;
1706 : }
1707 1076 : }
1708 :
1709 : /*******************************************************************************
1710 : * Pattern matching definitions
1711 : ******************************************************************************/
1712 :
1713 : #define SLP_PATTERN(x) &x::recognize
1714 : vect_pattern_decl_t slp_patterns[]
1715 : {
1716 : /* For least amount of back-tracking and more efficient matching
1717 : order patterns from the largest to the smallest. Especially if they
1718 : overlap in what they can detect. */
1719 :
1720 : SLP_PATTERN (complex_operations_pattern),
1721 : SLP_PATTERN (addsub_pattern)
1722 : };
1723 : #undef SLP_PATTERN
1724 :
1725 : /* Set the number of SLP pattern matchers available. */
1726 : size_t num__slp_patterns = ARRAY_SIZE (slp_patterns);
|
