1 | /* Vectorizer |
2 | Copyright (C) 2003-2024 Free Software Foundation, Inc. |
3 | Contributed by Dorit Naishlos <dorit@il.ibm.com> |
4 | |
5 | This file is part of GCC. |
6 | |
7 | GCC is free software; you can redistribute it and/or modify it under |
8 | the terms of the GNU General Public License as published by the Free |
9 | Software Foundation; either version 3, or (at your option) any later |
10 | version. |
11 | |
12 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
13 | WARRANTY; without even the implied warranty of MERCHANTABILITY or |
14 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
15 | for more details. |
16 | |
17 | You should have received a copy of the GNU General Public License |
18 | along with GCC; see the file COPYING3. If not see |
19 | <http://www.gnu.org/licenses/>. */ |
20 | |
21 | #ifndef GCC_TREE_VECTORIZER_H |
22 | #define GCC_TREE_VECTORIZER_H |
23 | |
24 | typedef class _stmt_vec_info *stmt_vec_info; |
25 | typedef struct _slp_tree *slp_tree; |
26 | |
27 | #include "tree-data-ref.h" |
28 | #include "tree-hash-traits.h" |
29 | #include "target.h" |
30 | #include "internal-fn.h" |
31 | #include "tree-ssa-operands.h" |
32 | #include "gimple-match.h" |
33 | |
34 | /* Used for naming of new temporaries. */ |
35 | enum vect_var_kind { |
36 | vect_simple_var, |
37 | vect_pointer_var, |
38 | vect_scalar_var, |
39 | vect_mask_var |
40 | }; |
41 | |
42 | /* Defines type of operation. */ |
43 | enum operation_type { |
44 | unary_op = 1, |
45 | binary_op, |
46 | ternary_op |
47 | }; |
48 | |
49 | /* Define type of available alignment support. */ |
50 | enum dr_alignment_support { |
51 | dr_unaligned_unsupported, |
52 | dr_unaligned_supported, |
53 | dr_explicit_realign, |
54 | dr_explicit_realign_optimized, |
55 | dr_aligned |
56 | }; |
57 | |
58 | /* Define type of def-use cross-iteration cycle. */ |
59 | enum vect_def_type { |
60 | vect_uninitialized_def = 0, |
61 | vect_constant_def = 1, |
62 | vect_external_def, |
63 | vect_internal_def, |
64 | vect_induction_def, |
65 | vect_reduction_def, |
66 | vect_double_reduction_def, |
67 | vect_nested_cycle, |
68 | vect_first_order_recurrence, |
69 | vect_condition_def, |
70 | vect_unknown_def_type |
71 | }; |
72 | |
73 | /* Define operation type of linear/non-linear induction variable. */ |
74 | enum vect_induction_op_type { |
75 | vect_step_op_add = 0, |
76 | vect_step_op_neg, |
77 | vect_step_op_mul, |
78 | vect_step_op_shl, |
79 | vect_step_op_shr |
80 | }; |
81 | |
82 | /* Define type of reduction. */ |
83 | enum vect_reduction_type { |
84 | TREE_CODE_REDUCTION, |
85 | COND_REDUCTION, |
86 | INTEGER_INDUC_COND_REDUCTION, |
87 | CONST_COND_REDUCTION, |
88 | |
89 | /* Retain a scalar phi and use a FOLD_EXTRACT_LAST within the loop |
90 | to implement: |
91 | |
92 | for (int i = 0; i < VF; ++i) |
93 | res = cond[i] ? val[i] : res; */ |
94 | , |
95 | |
96 | /* Use a folding reduction within the loop to implement: |
97 | |
98 | for (int i = 0; i < VF; ++i) |
99 | res = res OP val[i]; |
100 | |
101 | (with no reassocation). */ |
102 | FOLD_LEFT_REDUCTION |
103 | }; |
104 | |
105 | #define VECTORIZABLE_CYCLE_DEF(D) (((D) == vect_reduction_def) \ |
106 | || ((D) == vect_double_reduction_def) \ |
107 | || ((D) == vect_nested_cycle)) |
108 | |
109 | /* Structure to encapsulate information about a group of like |
110 | instructions to be presented to the target cost model. */ |
111 | struct stmt_info_for_cost { |
112 | int count; |
113 | enum vect_cost_for_stmt kind; |
114 | enum vect_cost_model_location where; |
115 | stmt_vec_info stmt_info; |
116 | slp_tree node; |
117 | tree vectype; |
118 | int misalign; |
119 | }; |
120 | |
121 | typedef vec<stmt_info_for_cost> stmt_vector_for_cost; |
122 | |
123 | /* Maps base addresses to an innermost_loop_behavior and the stmt it was |
124 | derived from that gives the maximum known alignment for that base. */ |
125 | typedef hash_map<tree_operand_hash, |
126 | std::pair<stmt_vec_info, innermost_loop_behavior *> > |
127 | vec_base_alignments; |
128 | |
129 | /* Represents elements [START, START + LENGTH) of cyclical array OPS* |
130 | (i.e. OPS repeated to give at least START + LENGTH elements) */ |
131 | struct vect_scalar_ops_slice |
132 | { |
133 | tree op (unsigned int i) const; |
134 | bool all_same_p () const; |
135 | |
136 | vec<tree> *ops; |
137 | unsigned int start; |
138 | unsigned int length; |
139 | }; |
140 | |
141 | /* Return element I of the slice. */ |
142 | inline tree |
143 | vect_scalar_ops_slice::op (unsigned int i) const |
144 | { |
145 | return (*ops)[(i + start) % ops->length ()]; |
146 | } |
147 | |
148 | /* Hash traits for vect_scalar_ops_slice. */ |
149 | struct vect_scalar_ops_slice_hash : typed_noop_remove<vect_scalar_ops_slice> |
150 | { |
151 | typedef vect_scalar_ops_slice value_type; |
152 | typedef vect_scalar_ops_slice compare_type; |
153 | |
154 | static const bool empty_zero_p = true; |
155 | |
156 | static void mark_deleted (value_type &s) { s.length = ~0U; } |
157 | static void mark_empty (value_type &s) { s.length = 0; } |
158 | static bool is_deleted (const value_type &s) { return s.length == ~0U; } |
159 | static bool is_empty (const value_type &s) { return s.length == 0; } |
160 | static hashval_t hash (const value_type &); |
161 | static bool equal (const value_type &, const compare_type &); |
162 | }; |
163 | |
164 | /************************************************************************ |
165 | SLP |
166 | ************************************************************************/ |
167 | typedef vec<std::pair<unsigned, unsigned> > lane_permutation_t; |
168 | typedef auto_vec<std::pair<unsigned, unsigned>, 16> auto_lane_permutation_t; |
169 | typedef vec<unsigned> load_permutation_t; |
170 | typedef auto_vec<unsigned, 16> auto_load_permutation_t; |
171 | |
172 | /* A computation tree of an SLP instance. Each node corresponds to a group of |
173 | stmts to be packed in a SIMD stmt. */ |
174 | struct _slp_tree { |
175 | _slp_tree (); |
176 | ~_slp_tree (); |
177 | |
178 | void push_vec_def (gimple *def); |
179 | void push_vec_def (tree def) { vec_defs.quick_push (obj: def); } |
180 | |
181 | /* Nodes that contain def-stmts of this node statements operands. */ |
182 | vec<slp_tree> children; |
183 | |
184 | /* A group of scalar stmts to be vectorized together. */ |
185 | vec<stmt_vec_info> stmts; |
186 | /* A group of scalar operands to be vectorized together. */ |
187 | vec<tree> ops; |
188 | /* The representative that should be used for analysis and |
189 | code generation. */ |
190 | stmt_vec_info representative; |
191 | |
192 | /* Load permutation relative to the stores, NULL if there is no |
193 | permutation. */ |
194 | load_permutation_t load_permutation; |
195 | /* Lane permutation of the operands scalar lanes encoded as pairs |
196 | of { operand number, lane number }. The number of elements |
197 | denotes the number of output lanes. */ |
198 | lane_permutation_t lane_permutation; |
199 | |
200 | /* Selected SIMD clone's function info. First vector element |
201 | is SIMD clone's function decl, followed by a pair of trees (base + step) |
202 | for linear arguments (pair of NULLs for other arguments). */ |
203 | vec<tree> simd_clone_info; |
204 | |
205 | tree vectype; |
206 | /* Vectorized defs. */ |
207 | vec<tree> vec_defs; |
208 | /* Number of vector stmts that are created to replace the group of scalar |
209 | stmts. It is calculated during the transformation phase as the number of |
210 | scalar elements in one scalar iteration (GROUP_SIZE) multiplied by VF |
211 | divided by vector size. */ |
212 | unsigned int vec_stmts_size; |
213 | |
214 | /* Reference count in the SLP graph. */ |
215 | unsigned int refcnt; |
216 | /* The maximum number of vector elements for the subtree rooted |
217 | at this node. */ |
218 | poly_uint64 max_nunits; |
219 | /* The DEF type of this node. */ |
220 | enum vect_def_type def_type; |
221 | /* The number of scalar lanes produced by this node. */ |
222 | unsigned int lanes; |
223 | /* The operation of this node. */ |
224 | enum tree_code code; |
225 | |
226 | int vertex; |
227 | |
228 | /* If not NULL this is a cached failed SLP discovery attempt with |
229 | the lanes that failed during SLP discovery as 'false'. This is |
230 | a copy of the matches array. */ |
231 | bool *failed; |
232 | |
233 | /* Allocate from slp_tree_pool. */ |
234 | static void *operator new (size_t); |
235 | |
236 | /* Return memory to slp_tree_pool. */ |
237 | static void operator delete (void *, size_t); |
238 | |
239 | /* Linked list of nodes to release when we free the slp_tree_pool. */ |
240 | slp_tree next_node; |
241 | slp_tree prev_node; |
242 | }; |
243 | |
244 | /* The enum describes the type of operations that an SLP instance |
245 | can perform. */ |
246 | |
247 | enum slp_instance_kind { |
248 | slp_inst_kind_store, |
249 | slp_inst_kind_reduc_group, |
250 | slp_inst_kind_reduc_chain, |
251 | slp_inst_kind_bb_reduc, |
252 | slp_inst_kind_ctor |
253 | }; |
254 | |
255 | /* SLP instance is a sequence of stmts in a loop that can be packed into |
256 | SIMD stmts. */ |
257 | typedef class _slp_instance { |
258 | public: |
259 | /* The root of SLP tree. */ |
260 | slp_tree root; |
261 | |
262 | /* For vector constructors, the constructor stmt that the SLP tree is built |
263 | from, NULL otherwise. */ |
264 | vec<stmt_vec_info> root_stmts; |
265 | |
266 | /* For slp_inst_kind_bb_reduc the defs that were not vectorized, NULL |
267 | otherwise. */ |
268 | vec<tree> remain_defs; |
269 | |
270 | /* The unrolling factor required to vectorized this SLP instance. */ |
271 | poly_uint64 unrolling_factor; |
272 | |
273 | /* The group of nodes that contain loads of this SLP instance. */ |
274 | vec<slp_tree> loads; |
275 | |
276 | /* The SLP node containing the reduction PHIs. */ |
277 | slp_tree reduc_phis; |
278 | |
279 | /* Vector cost of this entry to the SLP graph. */ |
280 | stmt_vector_for_cost cost_vec; |
281 | |
282 | /* If this instance is the main entry of a subgraph the set of |
283 | entries into the same subgraph, including itself. */ |
284 | vec<_slp_instance *> subgraph_entries; |
285 | |
286 | /* The type of operation the SLP instance is performing. */ |
287 | slp_instance_kind kind; |
288 | |
289 | dump_user_location_t location () const; |
290 | } *slp_instance; |
291 | |
292 | |
293 | /* Access Functions. */ |
294 | #define SLP_INSTANCE_TREE(S) (S)->root |
295 | #define SLP_INSTANCE_UNROLLING_FACTOR(S) (S)->unrolling_factor |
296 | #define SLP_INSTANCE_LOADS(S) (S)->loads |
297 | #define SLP_INSTANCE_ROOT_STMTS(S) (S)->root_stmts |
298 | #define SLP_INSTANCE_REMAIN_DEFS(S) (S)->remain_defs |
299 | #define SLP_INSTANCE_KIND(S) (S)->kind |
300 | |
301 | #define SLP_TREE_CHILDREN(S) (S)->children |
302 | #define SLP_TREE_SCALAR_STMTS(S) (S)->stmts |
303 | #define SLP_TREE_SCALAR_OPS(S) (S)->ops |
304 | #define SLP_TREE_REF_COUNT(S) (S)->refcnt |
305 | #define SLP_TREE_VEC_DEFS(S) (S)->vec_defs |
306 | #define SLP_TREE_NUMBER_OF_VEC_STMTS(S) (S)->vec_stmts_size |
307 | #define SLP_TREE_LOAD_PERMUTATION(S) (S)->load_permutation |
308 | #define SLP_TREE_LANE_PERMUTATION(S) (S)->lane_permutation |
309 | #define SLP_TREE_SIMD_CLONE_INFO(S) (S)->simd_clone_info |
310 | #define SLP_TREE_DEF_TYPE(S) (S)->def_type |
311 | #define SLP_TREE_VECTYPE(S) (S)->vectype |
312 | #define SLP_TREE_REPRESENTATIVE(S) (S)->representative |
313 | #define SLP_TREE_LANES(S) (S)->lanes |
314 | #define SLP_TREE_CODE(S) (S)->code |
315 | |
316 | enum vect_partial_vector_style { |
317 | vect_partial_vectors_none, |
318 | vect_partial_vectors_while_ult, |
319 | vect_partial_vectors_avx512, |
320 | vect_partial_vectors_len |
321 | }; |
322 | |
323 | /* Key for map that records association between |
324 | scalar conditions and corresponding loop mask, and |
325 | is populated by vect_record_loop_mask. */ |
326 | |
327 | struct scalar_cond_masked_key |
328 | { |
329 | scalar_cond_masked_key (tree t, unsigned ncopies_) |
330 | : ncopies (ncopies_) |
331 | { |
332 | get_cond_ops_from_tree (t); |
333 | } |
334 | |
335 | void get_cond_ops_from_tree (tree); |
336 | |
337 | unsigned ncopies; |
338 | bool inverted_p; |
339 | tree_code code; |
340 | tree op0; |
341 | tree op1; |
342 | }; |
343 | |
344 | template<> |
345 | struct default_hash_traits<scalar_cond_masked_key> |
346 | { |
347 | typedef scalar_cond_masked_key compare_type; |
348 | typedef scalar_cond_masked_key value_type; |
349 | |
350 | static inline hashval_t |
351 | hash (value_type v) |
352 | { |
353 | inchash::hash h; |
354 | h.add_int (v: v.code); |
355 | inchash::add_expr (v.op0, h, 0); |
356 | inchash::add_expr (v.op1, h, 0); |
357 | h.add_int (v: v.ncopies); |
358 | h.add_flag (flag: v.inverted_p); |
359 | return h.end (); |
360 | } |
361 | |
362 | static inline bool |
363 | equal (value_type existing, value_type candidate) |
364 | { |
365 | return (existing.ncopies == candidate.ncopies |
366 | && existing.code == candidate.code |
367 | && existing.inverted_p == candidate.inverted_p |
368 | && operand_equal_p (existing.op0, candidate.op0, flags: 0) |
369 | && operand_equal_p (existing.op1, candidate.op1, flags: 0)); |
370 | } |
371 | |
372 | static const bool empty_zero_p = true; |
373 | |
374 | static inline void |
375 | mark_empty (value_type &v) |
376 | { |
377 | v.ncopies = 0; |
378 | v.inverted_p = false; |
379 | } |
380 | |
381 | static inline bool |
382 | is_empty (value_type v) |
383 | { |
384 | return v.ncopies == 0; |
385 | } |
386 | |
387 | static inline void mark_deleted (value_type &) {} |
388 | |
389 | static inline bool is_deleted (const value_type &) |
390 | { |
391 | return false; |
392 | } |
393 | |
394 | static inline void remove (value_type &) {} |
395 | }; |
396 | |
397 | typedef hash_set<scalar_cond_masked_key> scalar_cond_masked_set_type; |
398 | |
399 | /* Key and map that records association between vector conditions and |
400 | corresponding loop mask, and is populated by prepare_vec_mask. */ |
401 | |
402 | typedef pair_hash<tree_operand_hash, tree_operand_hash> tree_cond_mask_hash; |
403 | typedef hash_set<tree_cond_mask_hash> vec_cond_masked_set_type; |
404 | |
405 | /* Describes two objects whose addresses must be unequal for the vectorized |
406 | loop to be valid. */ |
407 | typedef std::pair<tree, tree> vec_object_pair; |
408 | |
409 | /* Records that vectorization is only possible if abs (EXPR) >= MIN_VALUE. |
410 | UNSIGNED_P is true if we can assume that abs (EXPR) == EXPR. */ |
411 | class vec_lower_bound { |
412 | public: |
413 | vec_lower_bound () {} |
414 | vec_lower_bound (tree e, bool u, poly_uint64 m) |
415 | : expr (e), unsigned_p (u), min_value (m) {} |
416 | |
417 | tree expr; |
418 | bool unsigned_p; |
419 | poly_uint64 min_value; |
420 | }; |
421 | |
422 | /* Vectorizer state shared between different analyses like vector sizes |
423 | of the same CFG region. */ |
424 | class vec_info_shared { |
425 | public: |
426 | vec_info_shared(); |
427 | ~vec_info_shared(); |
428 | |
429 | void save_datarefs(); |
430 | void check_datarefs(); |
431 | |
432 | /* The number of scalar stmts. */ |
433 | unsigned n_stmts; |
434 | |
435 | /* All data references. Freed by free_data_refs, so not an auto_vec. */ |
436 | vec<data_reference_p> datarefs; |
437 | vec<data_reference> datarefs_copy; |
438 | |
439 | /* The loop nest in which the data dependences are computed. */ |
440 | auto_vec<loop_p> loop_nest; |
441 | |
442 | /* All data dependences. Freed by free_dependence_relations, so not |
443 | an auto_vec. */ |
444 | vec<ddr_p> ddrs; |
445 | }; |
446 | |
447 | /* Vectorizer state common between loop and basic-block vectorization. */ |
448 | class vec_info { |
449 | public: |
450 | typedef hash_set<int_hash<machine_mode, E_VOIDmode, E_BLKmode> > mode_set; |
451 | enum vec_kind { bb, loop }; |
452 | |
453 | vec_info (vec_kind, vec_info_shared *); |
454 | ~vec_info (); |
455 | |
456 | stmt_vec_info add_stmt (gimple *); |
457 | stmt_vec_info add_pattern_stmt (gimple *, stmt_vec_info); |
458 | stmt_vec_info lookup_stmt (gimple *); |
459 | stmt_vec_info lookup_def (tree); |
460 | stmt_vec_info lookup_single_use (tree); |
461 | class dr_vec_info *lookup_dr (data_reference *); |
462 | void move_dr (stmt_vec_info, stmt_vec_info); |
463 | void remove_stmt (stmt_vec_info); |
464 | void replace_stmt (gimple_stmt_iterator *, stmt_vec_info, gimple *); |
465 | void insert_on_entry (stmt_vec_info, gimple *); |
466 | void insert_seq_on_entry (stmt_vec_info, gimple_seq); |
467 | |
468 | /* The type of vectorization. */ |
469 | vec_kind kind; |
470 | |
471 | /* Shared vectorizer state. */ |
472 | vec_info_shared *shared; |
473 | |
474 | /* The mapping of GIMPLE UID to stmt_vec_info. */ |
475 | vec<stmt_vec_info> stmt_vec_infos; |
476 | /* Whether the above mapping is complete. */ |
477 | bool stmt_vec_info_ro; |
478 | |
479 | /* Whether we've done a transform we think OK to not update virtual |
480 | SSA form. */ |
481 | bool any_known_not_updated_vssa; |
482 | |
483 | /* The SLP graph. */ |
484 | auto_vec<slp_instance> slp_instances; |
485 | |
486 | /* Maps base addresses to an innermost_loop_behavior that gives the maximum |
487 | known alignment for that base. */ |
488 | vec_base_alignments base_alignments; |
489 | |
490 | /* All interleaving chains of stores, represented by the first |
491 | stmt in the chain. */ |
492 | auto_vec<stmt_vec_info> grouped_stores; |
493 | |
494 | /* The set of vector modes used in the vectorized region. */ |
495 | mode_set used_vector_modes; |
496 | |
497 | /* The argument we should pass to related_vector_mode when looking up |
498 | the vector mode for a scalar mode, or VOIDmode if we haven't yet |
499 | made any decisions about which vector modes to use. */ |
500 | machine_mode vector_mode; |
501 | |
502 | private: |
503 | stmt_vec_info new_stmt_vec_info (gimple *stmt); |
504 | void set_vinfo_for_stmt (gimple *, stmt_vec_info, bool = true); |
505 | void free_stmt_vec_infos (); |
506 | void free_stmt_vec_info (stmt_vec_info); |
507 | }; |
508 | |
509 | class _loop_vec_info; |
510 | class _bb_vec_info; |
511 | |
512 | template<> |
513 | template<> |
514 | inline bool |
515 | is_a_helper <_loop_vec_info *>::test (vec_info *i) |
516 | { |
517 | return i->kind == vec_info::loop; |
518 | } |
519 | |
520 | template<> |
521 | template<> |
522 | inline bool |
523 | is_a_helper <_bb_vec_info *>::test (vec_info *i) |
524 | { |
525 | return i->kind == vec_info::bb; |
526 | } |
527 | |
528 | /* In general, we can divide the vector statements in a vectorized loop |
529 | into related groups ("rgroups") and say that for each rgroup there is |
530 | some nS such that the rgroup operates on nS values from one scalar |
531 | iteration followed by nS values from the next. That is, if VF is the |
532 | vectorization factor of the loop, the rgroup operates on a sequence: |
533 | |
534 | (1,1) (1,2) ... (1,nS) (2,1) ... (2,nS) ... (VF,1) ... (VF,nS) |
535 | |
536 | where (i,j) represents a scalar value with index j in a scalar |
537 | iteration with index i. |
538 | |
539 | [ We use the term "rgroup" to emphasise that this grouping isn't |
540 | necessarily the same as the grouping of statements used elsewhere. |
541 | For example, if we implement a group of scalar loads using gather |
542 | loads, we'll use a separate gather load for each scalar load, and |
543 | thus each gather load will belong to its own rgroup. ] |
544 | |
545 | In general this sequence will occupy nV vectors concatenated |
546 | together. If these vectors have nL lanes each, the total number |
547 | of scalar values N is given by: |
548 | |
549 | N = nS * VF = nV * nL |
550 | |
551 | None of nS, VF, nV and nL are required to be a power of 2. nS and nV |
552 | are compile-time constants but VF and nL can be variable (if the target |
553 | supports variable-length vectors). |
554 | |
555 | In classical vectorization, each iteration of the vector loop would |
556 | handle exactly VF iterations of the original scalar loop. However, |
557 | in vector loops that are able to operate on partial vectors, a |
558 | particular iteration of the vector loop might handle fewer than VF |
559 | iterations of the scalar loop. The vector lanes that correspond to |
560 | iterations of the scalar loop are said to be "active" and the other |
561 | lanes are said to be "inactive". |
562 | |
563 | In such vector loops, many rgroups need to be controlled to ensure |
564 | that they have no effect for the inactive lanes. Conceptually, each |
565 | such rgroup needs a sequence of booleans in the same order as above, |
566 | but with each (i,j) replaced by a boolean that indicates whether |
567 | iteration i is active. This sequence occupies nV vector controls |
568 | that again have nL lanes each. Thus the control sequence as a whole |
569 | consists of VF independent booleans that are each repeated nS times. |
570 | |
571 | Taking mask-based approach as a partially-populated vectors example. |
572 | We make the simplifying assumption that if a sequence of nV masks is |
573 | suitable for one (nS,nL) pair, we can reuse it for (nS/2,nL/2) by |
574 | VIEW_CONVERTing it. This holds for all current targets that support |
575 | fully-masked loops. For example, suppose the scalar loop is: |
576 | |
577 | float *f; |
578 | double *d; |
579 | for (int i = 0; i < n; ++i) |
580 | { |
581 | f[i * 2 + 0] += 1.0f; |
582 | f[i * 2 + 1] += 2.0f; |
583 | d[i] += 3.0; |
584 | } |
585 | |
586 | and suppose that vectors have 256 bits. The vectorized f accesses |
587 | will belong to one rgroup and the vectorized d access to another: |
588 | |
589 | f rgroup: nS = 2, nV = 1, nL = 8 |
590 | d rgroup: nS = 1, nV = 1, nL = 4 |
591 | VF = 4 |
592 | |
593 | [ In this simple example the rgroups do correspond to the normal |
594 | SLP grouping scheme. ] |
595 | |
596 | If only the first three lanes are active, the masks we need are: |
597 | |
598 | f rgroup: 1 1 | 1 1 | 1 1 | 0 0 |
599 | d rgroup: 1 | 1 | 1 | 0 |
600 | |
601 | Here we can use a mask calculated for f's rgroup for d's, but not |
602 | vice versa. |
603 | |
604 | Thus for each value of nV, it is enough to provide nV masks, with the |
605 | mask being calculated based on the highest nL (or, equivalently, based |
606 | on the highest nS) required by any rgroup with that nV. We therefore |
607 | represent the entire collection of masks as a two-level table, with the |
608 | first level being indexed by nV - 1 (since nV == 0 doesn't exist) and |
609 | the second being indexed by the mask index 0 <= i < nV. */ |
610 | |
611 | /* The controls (like masks or lengths) needed by rgroups with nV vectors, |
612 | according to the description above. */ |
613 | struct rgroup_controls { |
614 | /* The largest nS for all rgroups that use these controls. |
615 | For vect_partial_vectors_avx512 this is the constant nscalars_per_iter |
616 | for all members of the group. */ |
617 | unsigned int max_nscalars_per_iter; |
618 | |
619 | /* For the largest nS recorded above, the loop controls divide each scalar |
620 | into FACTOR equal-sized pieces. This is useful if we need to split |
621 | element-based accesses into byte-based accesses. |
622 | For vect_partial_vectors_avx512 this records nV instead. */ |
623 | unsigned int factor; |
624 | |
625 | /* This is a vector type with MAX_NSCALARS_PER_ITER * VF / nV elements. |
626 | For mask-based controls, it is the type of the masks in CONTROLS. |
627 | For length-based controls, it can be any vector type that has the |
628 | specified number of elements; the type of the elements doesn't matter. */ |
629 | tree type; |
630 | |
631 | /* When there is no uniformly used LOOP_VINFO_RGROUP_COMPARE_TYPE this |
632 | is the rgroup specific type used. */ |
633 | tree compare_type; |
634 | |
635 | /* A vector of nV controls, in iteration order. */ |
636 | vec<tree> controls; |
637 | |
638 | /* In case of len_load and len_store with a bias there is only one |
639 | rgroup. This holds the adjusted loop length for the this rgroup. */ |
640 | tree bias_adjusted_ctrl; |
641 | }; |
642 | |
643 | struct vec_loop_masks |
644 | { |
645 | bool is_empty () const { return mask_set.is_empty (); } |
646 | |
647 | /* Set to record vectype, nvector pairs. */ |
648 | hash_set<pair_hash <nofree_ptr_hash <tree_node>, |
649 | int_hash<unsigned, 0>>> mask_set; |
650 | |
651 | /* rgroup_controls used for the partial vector scheme. */ |
652 | auto_vec<rgroup_controls> rgc_vec; |
653 | }; |
654 | |
655 | typedef auto_vec<rgroup_controls> vec_loop_lens; |
656 | |
657 | typedef auto_vec<std::pair<data_reference*, tree> > drs_init_vec; |
658 | |
659 | /* Information about a reduction accumulator from the main loop that could |
660 | conceivably be reused as the input to a reduction in an epilogue loop. */ |
661 | struct vect_reusable_accumulator { |
662 | /* The final value of the accumulator, which forms the input to the |
663 | reduction operation. */ |
664 | tree reduc_input; |
665 | |
666 | /* The stmt_vec_info that describes the reduction (i.e. the one for |
667 | which is_reduc_info is true). */ |
668 | stmt_vec_info reduc_info; |
669 | }; |
670 | |
671 | /*-----------------------------------------------------------------*/ |
672 | /* Info on vectorized loops. */ |
673 | /*-----------------------------------------------------------------*/ |
674 | typedef class _loop_vec_info : public vec_info { |
675 | public: |
676 | _loop_vec_info (class loop *, vec_info_shared *); |
677 | ~_loop_vec_info (); |
678 | |
679 | /* The loop to which this info struct refers to. */ |
680 | class loop *loop; |
681 | |
682 | /* The loop basic blocks. */ |
683 | basic_block *bbs; |
684 | |
685 | /* Number of latch executions. */ |
686 | tree num_itersm1; |
687 | /* Number of iterations. */ |
688 | tree num_iters; |
689 | /* Number of iterations of the original loop. */ |
690 | tree num_iters_unchanged; |
691 | /* Condition under which this loop is analyzed and versioned. */ |
692 | tree num_iters_assumptions; |
693 | |
694 | /* The cost of the vector code. */ |
695 | class vector_costs *vector_costs; |
696 | |
697 | /* The cost of the scalar code. */ |
698 | class vector_costs *scalar_costs; |
699 | |
700 | /* Threshold of number of iterations below which vectorization will not be |
701 | performed. It is calculated from MIN_PROFITABLE_ITERS and |
702 | param_min_vect_loop_bound. */ |
703 | unsigned int th; |
704 | |
705 | /* When applying loop versioning, the vector form should only be used |
706 | if the number of scalar iterations is >= this value, on top of all |
707 | the other requirements. Ignored when loop versioning is not being |
708 | used. */ |
709 | poly_uint64 versioning_threshold; |
710 | |
711 | /* Unrolling factor */ |
712 | poly_uint64 vectorization_factor; |
713 | |
714 | /* If this loop is an epilogue loop whose main loop can be skipped, |
715 | MAIN_LOOP_EDGE is the edge from the main loop to this loop's |
716 | preheader. SKIP_MAIN_LOOP_EDGE is then the edge that skips the |
717 | main loop and goes straight to this loop's preheader. |
718 | |
719 | Both fields are null otherwise. */ |
720 | edge main_loop_edge; |
721 | edge skip_main_loop_edge; |
722 | |
723 | /* If this loop is an epilogue loop that might be skipped after executing |
724 | the main loop, this edge is the one that skips the epilogue. */ |
725 | edge skip_this_loop_edge; |
726 | |
727 | /* The vectorized form of a standard reduction replaces the original |
728 | scalar code's final result (a loop-closed SSA PHI) with the result |
729 | of a vector-to-scalar reduction operation. After vectorization, |
730 | this variable maps these vector-to-scalar results to information |
731 | about the reductions that generated them. */ |
732 | hash_map<tree, vect_reusable_accumulator> reusable_accumulators; |
733 | |
734 | /* The number of times that the target suggested we unroll the vector loop |
735 | in order to promote more ILP. This value will be used to re-analyze the |
736 | loop for vectorization and if successful the value will be folded into |
737 | vectorization_factor (and therefore exactly divides |
738 | vectorization_factor). */ |
739 | unsigned int suggested_unroll_factor; |
740 | |
741 | /* Maximum runtime vectorization factor, or MAX_VECTORIZATION_FACTOR |
742 | if there is no particular limit. */ |
743 | unsigned HOST_WIDE_INT max_vectorization_factor; |
744 | |
745 | /* The masks that a fully-masked loop should use to avoid operating |
746 | on inactive scalars. */ |
747 | vec_loop_masks masks; |
748 | |
749 | /* The lengths that a loop with length should use to avoid operating |
750 | on inactive scalars. */ |
751 | vec_loop_lens lens; |
752 | |
753 | /* Set of scalar conditions that have loop mask applied. */ |
754 | scalar_cond_masked_set_type scalar_cond_masked_set; |
755 | |
756 | /* Set of vector conditions that have loop mask applied. */ |
757 | vec_cond_masked_set_type vec_cond_masked_set; |
758 | |
759 | /* If we are using a loop mask to align memory addresses, this variable |
760 | contains the number of vector elements that we should skip in the |
761 | first iteration of the vector loop (i.e. the number of leading |
762 | elements that should be false in the first mask). */ |
763 | tree mask_skip_niters; |
764 | |
765 | /* The type that the loop control IV should be converted to before |
766 | testing which of the VF scalars are active and inactive. |
767 | Only meaningful if LOOP_VINFO_USING_PARTIAL_VECTORS_P. */ |
768 | tree rgroup_compare_type; |
769 | |
770 | /* For #pragma omp simd if (x) loops the x expression. If constant 0, |
771 | the loop should not be vectorized, if constant non-zero, simd_if_cond |
772 | shouldn't be set and loop vectorized normally, if SSA_NAME, the loop |
773 | should be versioned on that condition, using scalar loop if the condition |
774 | is false and vectorized loop otherwise. */ |
775 | tree simd_if_cond; |
776 | |
777 | /* The type that the vector loop control IV should have when |
778 | LOOP_VINFO_USING_PARTIAL_VECTORS_P is true. */ |
779 | tree rgroup_iv_type; |
780 | |
781 | /* The style used for implementing partial vectors when |
782 | LOOP_VINFO_USING_PARTIAL_VECTORS_P is true. */ |
783 | vect_partial_vector_style partial_vector_style; |
784 | |
785 | /* Unknown DRs according to which loop was peeled. */ |
786 | class dr_vec_info *unaligned_dr; |
787 | |
788 | /* peeling_for_alignment indicates whether peeling for alignment will take |
789 | place, and what the peeling factor should be: |
790 | peeling_for_alignment = X means: |
791 | If X=0: Peeling for alignment will not be applied. |
792 | If X>0: Peel first X iterations. |
793 | If X=-1: Generate a runtime test to calculate the number of iterations |
794 | to be peeled, using the dataref recorded in the field |
795 | unaligned_dr. */ |
796 | int peeling_for_alignment; |
797 | |
798 | /* The mask used to check the alignment of pointers or arrays. */ |
799 | int ptr_mask; |
800 | |
801 | /* Data Dependence Relations defining address ranges that are candidates |
802 | for a run-time aliasing check. */ |
803 | auto_vec<ddr_p> may_alias_ddrs; |
804 | |
805 | /* Data Dependence Relations defining address ranges together with segment |
806 | lengths from which the run-time aliasing check is built. */ |
807 | auto_vec<dr_with_seg_len_pair_t> comp_alias_ddrs; |
808 | |
809 | /* Check that the addresses of each pair of objects is unequal. */ |
810 | auto_vec<vec_object_pair> check_unequal_addrs; |
811 | |
812 | /* List of values that are required to be nonzero. This is used to check |
813 | whether things like "x[i * n] += 1;" are safe and eventually gets added |
814 | to the checks for lower bounds below. */ |
815 | auto_vec<tree> check_nonzero; |
816 | |
817 | /* List of values that need to be checked for a minimum value. */ |
818 | auto_vec<vec_lower_bound> lower_bounds; |
819 | |
820 | /* Statements in the loop that have data references that are candidates for a |
821 | runtime (loop versioning) misalignment check. */ |
822 | auto_vec<stmt_vec_info> may_misalign_stmts; |
823 | |
824 | /* Reduction cycles detected in the loop. Used in loop-aware SLP. */ |
825 | auto_vec<stmt_vec_info> reductions; |
826 | |
827 | /* All reduction chains in the loop, represented by the first |
828 | stmt in the chain. */ |
829 | auto_vec<stmt_vec_info> reduction_chains; |
830 | |
831 | /* Cost vector for a single scalar iteration. */ |
832 | auto_vec<stmt_info_for_cost> scalar_cost_vec; |
833 | |
834 | /* Map of IV base/step expressions to inserted name in the preheader. */ |
835 | hash_map<tree_operand_hash, tree> *ivexpr_map; |
836 | |
837 | /* Map of OpenMP "omp simd array" scan variables to corresponding |
838 | rhs of the store of the initializer. */ |
839 | hash_map<tree, tree> *scan_map; |
840 | |
841 | /* The unrolling factor needed to SLP the loop. In case of that pure SLP is |
842 | applied to the loop, i.e., no unrolling is needed, this is 1. */ |
843 | poly_uint64 slp_unrolling_factor; |
844 | |
845 | /* The factor used to over weight those statements in an inner loop |
846 | relative to the loop being vectorized. */ |
847 | unsigned int inner_loop_cost_factor; |
848 | |
849 | /* Is the loop vectorizable? */ |
850 | bool vectorizable; |
851 | |
852 | /* Records whether we still have the option of vectorizing this loop |
853 | using partially-populated vectors; in other words, whether it is |
854 | still possible for one iteration of the vector loop to handle |
855 | fewer than VF scalars. */ |
856 | bool can_use_partial_vectors_p; |
857 | |
858 | /* True if we've decided to use partially-populated vectors, so that |
859 | the vector loop can handle fewer than VF scalars. */ |
860 | bool using_partial_vectors_p; |
861 | |
862 | /* True if we've decided to use a decrementing loop control IV that counts |
863 | scalars. This can be done for any loop that: |
864 | |
865 | (a) uses length "controls"; and |
866 | (b) can iterate more than once. */ |
867 | bool using_decrementing_iv_p; |
868 | |
869 | /* True if we've decided to use output of select_vl to adjust IV of |
870 | both loop control and data reference pointer. This is only true |
871 | for single-rgroup control. */ |
872 | bool using_select_vl_p; |
873 | |
874 | /* True if we've decided to use partially-populated vectors for the |
875 | epilogue of loop. */ |
876 | bool epil_using_partial_vectors_p; |
877 | |
878 | /* The bias for len_load and len_store. For now, only 0 and -1 are |
879 | supported. -1 must be used when a backend does not support |
880 | len_load/len_store with a length of zero. */ |
881 | signed char partial_load_store_bias; |
882 | |
883 | /* When we have grouped data accesses with gaps, we may introduce invalid |
884 | memory accesses. We peel the last iteration of the loop to prevent |
885 | this. */ |
886 | bool peeling_for_gaps; |
887 | |
888 | /* When the number of iterations is not a multiple of the vector size |
889 | we need to peel off iterations at the end to form an epilogue loop. */ |
890 | bool peeling_for_niter; |
891 | |
892 | /* When the loop has early breaks that we can vectorize we need to peel |
893 | the loop for the break finding loop. */ |
894 | bool early_breaks; |
895 | |
896 | /* List of loop additional IV conditionals found in the loop. */ |
897 | auto_vec<gcond *> conds; |
898 | |
899 | /* Main loop IV cond. */ |
900 | gcond* loop_iv_cond; |
901 | |
902 | /* True if there are no loop carried data dependencies in the loop. |
903 | If loop->safelen <= 1, then this is always true, either the loop |
904 | didn't have any loop carried data dependencies, or the loop is being |
905 | vectorized guarded with some runtime alias checks, or couldn't |
906 | be vectorized at all, but then this field shouldn't be used. |
907 | For loop->safelen >= 2, the user has asserted that there are no |
908 | backward dependencies, but there still could be loop carried forward |
909 | dependencies in such loops. This flag will be false if normal |
910 | vectorizer data dependency analysis would fail or require versioning |
911 | for alias, but because of loop->safelen >= 2 it has been vectorized |
912 | even without versioning for alias. E.g. in: |
913 | #pragma omp simd |
914 | for (int i = 0; i < m; i++) |
915 | a[i] = a[i + k] * c; |
916 | (or #pragma simd or #pragma ivdep) we can vectorize this and it will |
917 | DTRT even for k > 0 && k < m, but without safelen we would not |
918 | vectorize this, so this field would be false. */ |
919 | bool no_data_dependencies; |
920 | |
921 | /* Mark loops having masked stores. */ |
922 | bool has_mask_store; |
923 | |
924 | /* Queued scaling factor for the scalar loop. */ |
925 | profile_probability scalar_loop_scaling; |
926 | |
927 | /* If if-conversion versioned this loop before conversion, this is the |
928 | loop version without if-conversion. */ |
929 | class loop *scalar_loop; |
930 | |
931 | /* For loops being epilogues of already vectorized loops |
932 | this points to the original vectorized loop. Otherwise NULL. */ |
933 | _loop_vec_info *orig_loop_info; |
934 | |
935 | /* Used to store loop_vec_infos of epilogues of this loop during |
936 | analysis. */ |
937 | vec<_loop_vec_info *> epilogue_vinfos; |
938 | |
939 | /* The controlling loop IV for the current loop when vectorizing. This IV |
940 | controls the natural exits of the loop. */ |
941 | edge vec_loop_iv_exit; |
942 | |
943 | /* The controlling loop IV for the epilogue loop when vectorizing. This IV |
944 | controls the natural exits of the loop. */ |
945 | edge vec_epilogue_loop_iv_exit; |
946 | |
947 | /* The controlling loop IV for the scalar loop being vectorized. This IV |
948 | controls the natural exits of the loop. */ |
949 | edge scalar_loop_iv_exit; |
950 | |
951 | /* Used to store the list of stores needing to be moved if doing early |
952 | break vectorization as they would violate the scalar loop semantics if |
953 | vectorized in their current location. These are stored in order that they |
954 | need to be moved. */ |
955 | auto_vec<gimple *> early_break_stores; |
956 | |
957 | /* The final basic block where to move statements to. In the case of |
958 | multiple exits this could be pretty far away. */ |
959 | basic_block early_break_dest_bb; |
960 | |
961 | /* Statements whose VUSES need updating if early break vectorization is to |
962 | happen. */ |
963 | auto_vec<gimple*> early_break_vuses; |
964 | } *loop_vec_info; |
965 | |
966 | /* Access Functions. */ |
967 | #define LOOP_VINFO_LOOP(L) (L)->loop |
968 | #define LOOP_VINFO_IV_EXIT(L) (L)->vec_loop_iv_exit |
969 | #define LOOP_VINFO_EPILOGUE_IV_EXIT(L) (L)->vec_epilogue_loop_iv_exit |
970 | #define LOOP_VINFO_SCALAR_IV_EXIT(L) (L)->scalar_loop_iv_exit |
971 | #define LOOP_VINFO_BBS(L) (L)->bbs |
972 | #define LOOP_VINFO_NITERSM1(L) (L)->num_itersm1 |
973 | #define LOOP_VINFO_NITERS(L) (L)->num_iters |
974 | /* Since LOOP_VINFO_NITERS and LOOP_VINFO_NITERSM1 can change after |
975 | prologue peeling retain total unchanged scalar loop iterations for |
976 | cost model. */ |
977 | #define LOOP_VINFO_NITERS_UNCHANGED(L) (L)->num_iters_unchanged |
978 | #define LOOP_VINFO_NITERS_ASSUMPTIONS(L) (L)->num_iters_assumptions |
979 | #define LOOP_VINFO_COST_MODEL_THRESHOLD(L) (L)->th |
980 | #define LOOP_VINFO_VERSIONING_THRESHOLD(L) (L)->versioning_threshold |
981 | #define LOOP_VINFO_VECTORIZABLE_P(L) (L)->vectorizable |
982 | #define LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P(L) (L)->can_use_partial_vectors_p |
983 | #define LOOP_VINFO_USING_PARTIAL_VECTORS_P(L) (L)->using_partial_vectors_p |
984 | #define LOOP_VINFO_USING_DECREMENTING_IV_P(L) (L)->using_decrementing_iv_p |
985 | #define LOOP_VINFO_USING_SELECT_VL_P(L) (L)->using_select_vl_p |
986 | #define LOOP_VINFO_EPIL_USING_PARTIAL_VECTORS_P(L) \ |
987 | (L)->epil_using_partial_vectors_p |
988 | #define LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS(L) (L)->partial_load_store_bias |
989 | #define LOOP_VINFO_VECT_FACTOR(L) (L)->vectorization_factor |
990 | #define LOOP_VINFO_MAX_VECT_FACTOR(L) (L)->max_vectorization_factor |
991 | #define LOOP_VINFO_MASKS(L) (L)->masks |
992 | #define LOOP_VINFO_LENS(L) (L)->lens |
993 | #define LOOP_VINFO_MASK_SKIP_NITERS(L) (L)->mask_skip_niters |
994 | #define LOOP_VINFO_RGROUP_COMPARE_TYPE(L) (L)->rgroup_compare_type |
995 | #define LOOP_VINFO_RGROUP_IV_TYPE(L) (L)->rgroup_iv_type |
996 | #define LOOP_VINFO_PARTIAL_VECTORS_STYLE(L) (L)->partial_vector_style |
997 | #define LOOP_VINFO_PTR_MASK(L) (L)->ptr_mask |
998 | #define LOOP_VINFO_N_STMTS(L) (L)->shared->n_stmts |
999 | #define LOOP_VINFO_LOOP_NEST(L) (L)->shared->loop_nest |
1000 | #define LOOP_VINFO_DATAREFS(L) (L)->shared->datarefs |
1001 | #define LOOP_VINFO_DDRS(L) (L)->shared->ddrs |
1002 | #define LOOP_VINFO_INT_NITERS(L) (TREE_INT_CST_LOW ((L)->num_iters)) |
1003 | #define LOOP_VINFO_PEELING_FOR_ALIGNMENT(L) (L)->peeling_for_alignment |
1004 | #define LOOP_VINFO_UNALIGNED_DR(L) (L)->unaligned_dr |
1005 | #define LOOP_VINFO_MAY_MISALIGN_STMTS(L) (L)->may_misalign_stmts |
1006 | #define LOOP_VINFO_MAY_ALIAS_DDRS(L) (L)->may_alias_ddrs |
1007 | #define LOOP_VINFO_COMP_ALIAS_DDRS(L) (L)->comp_alias_ddrs |
1008 | #define LOOP_VINFO_CHECK_UNEQUAL_ADDRS(L) (L)->check_unequal_addrs |
1009 | #define LOOP_VINFO_CHECK_NONZERO(L) (L)->check_nonzero |
1010 | #define LOOP_VINFO_LOWER_BOUNDS(L) (L)->lower_bounds |
1011 | #define LOOP_VINFO_GROUPED_STORES(L) (L)->grouped_stores |
1012 | #define LOOP_VINFO_SLP_INSTANCES(L) (L)->slp_instances |
1013 | #define LOOP_VINFO_SLP_UNROLLING_FACTOR(L) (L)->slp_unrolling_factor |
1014 | #define LOOP_VINFO_REDUCTIONS(L) (L)->reductions |
1015 | #define LOOP_VINFO_REDUCTION_CHAINS(L) (L)->reduction_chains |
1016 | #define LOOP_VINFO_PEELING_FOR_GAPS(L) (L)->peeling_for_gaps |
1017 | #define LOOP_VINFO_PEELING_FOR_NITER(L) (L)->peeling_for_niter |
1018 | #define LOOP_VINFO_EARLY_BREAKS(L) (L)->early_breaks |
1019 | #define LOOP_VINFO_EARLY_BRK_STORES(L) (L)->early_break_stores |
1020 | #define LOOP_VINFO_EARLY_BREAKS_VECT_PEELED(L) \ |
1021 | (single_pred ((L)->loop->latch) != (L)->vec_loop_iv_exit->src) |
1022 | #define LOOP_VINFO_EARLY_BRK_DEST_BB(L) (L)->early_break_dest_bb |
1023 | #define LOOP_VINFO_EARLY_BRK_VUSES(L) (L)->early_break_vuses |
1024 | #define LOOP_VINFO_LOOP_CONDS(L) (L)->conds |
1025 | #define LOOP_VINFO_LOOP_IV_COND(L) (L)->loop_iv_cond |
1026 | #define LOOP_VINFO_NO_DATA_DEPENDENCIES(L) (L)->no_data_dependencies |
1027 | #define LOOP_VINFO_SCALAR_LOOP(L) (L)->scalar_loop |
1028 | #define LOOP_VINFO_SCALAR_LOOP_SCALING(L) (L)->scalar_loop_scaling |
1029 | #define LOOP_VINFO_HAS_MASK_STORE(L) (L)->has_mask_store |
1030 | #define LOOP_VINFO_SCALAR_ITERATION_COST(L) (L)->scalar_cost_vec |
1031 | #define LOOP_VINFO_ORIG_LOOP_INFO(L) (L)->orig_loop_info |
1032 | #define LOOP_VINFO_SIMD_IF_COND(L) (L)->simd_if_cond |
1033 | #define LOOP_VINFO_INNER_LOOP_COST_FACTOR(L) (L)->inner_loop_cost_factor |
1034 | |
1035 | #define LOOP_VINFO_FULLY_MASKED_P(L) \ |
1036 | (LOOP_VINFO_USING_PARTIAL_VECTORS_P (L) \ |
1037 | && !LOOP_VINFO_MASKS (L).is_empty ()) |
1038 | |
1039 | #define LOOP_VINFO_FULLY_WITH_LENGTH_P(L) \ |
1040 | (LOOP_VINFO_USING_PARTIAL_VECTORS_P (L) \ |
1041 | && !LOOP_VINFO_LENS (L).is_empty ()) |
1042 | |
1043 | #define LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT(L) \ |
1044 | ((L)->may_misalign_stmts.length () > 0) |
1045 | #define LOOP_REQUIRES_VERSIONING_FOR_ALIAS(L) \ |
1046 | ((L)->comp_alias_ddrs.length () > 0 \ |
1047 | || (L)->check_unequal_addrs.length () > 0 \ |
1048 | || (L)->lower_bounds.length () > 0) |
1049 | #define LOOP_REQUIRES_VERSIONING_FOR_NITERS(L) \ |
1050 | (LOOP_VINFO_NITERS_ASSUMPTIONS (L)) |
1051 | #define LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND(L) \ |
1052 | (LOOP_VINFO_SIMD_IF_COND (L)) |
1053 | #define LOOP_REQUIRES_VERSIONING(L) \ |
1054 | (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (L) \ |
1055 | || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (L) \ |
1056 | || LOOP_REQUIRES_VERSIONING_FOR_NITERS (L) \ |
1057 | || LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND (L)) |
1058 | |
1059 | #define LOOP_VINFO_NITERS_KNOWN_P(L) \ |
1060 | (tree_fits_shwi_p ((L)->num_iters) && tree_to_shwi ((L)->num_iters) > 0) |
1061 | |
1062 | #define LOOP_VINFO_EPILOGUE_P(L) \ |
1063 | (LOOP_VINFO_ORIG_LOOP_INFO (L) != NULL) |
1064 | |
1065 | #define LOOP_VINFO_ORIG_MAX_VECT_FACTOR(L) \ |
1066 | (LOOP_VINFO_MAX_VECT_FACTOR (LOOP_VINFO_ORIG_LOOP_INFO (L))) |
1067 | |
1068 | /* Wrapper for loop_vec_info, for tracking success/failure, where a non-NULL |
1069 | value signifies success, and a NULL value signifies failure, supporting |
1070 | propagating an opt_problem * describing the failure back up the call |
1071 | stack. */ |
1072 | typedef opt_pointer_wrapper <loop_vec_info> opt_loop_vec_info; |
1073 | |
1074 | inline loop_vec_info |
1075 | loop_vec_info_for_loop (class loop *loop) |
1076 | { |
1077 | return (loop_vec_info) loop->aux; |
1078 | } |
1079 | |
1080 | struct slp_root |
1081 | { |
1082 | slp_root (slp_instance_kind kind_, vec<stmt_vec_info> stmts_, |
1083 | vec<stmt_vec_info> roots_, vec<tree> remain_ = vNULL) |
1084 | : kind(kind_), stmts(stmts_), roots(roots_), remain(remain_) {} |
1085 | slp_instance_kind kind; |
1086 | vec<stmt_vec_info> stmts; |
1087 | vec<stmt_vec_info> roots; |
1088 | vec<tree> remain; |
1089 | }; |
1090 | |
1091 | typedef class _bb_vec_info : public vec_info |
1092 | { |
1093 | public: |
1094 | _bb_vec_info (vec<basic_block> bbs, vec_info_shared *); |
1095 | ~_bb_vec_info (); |
1096 | |
1097 | /* The region we are operating on. bbs[0] is the entry, excluding |
1098 | its PHI nodes. In the future we might want to track an explicit |
1099 | entry edge to cover bbs[0] PHI nodes and have a region entry |
1100 | insert location. */ |
1101 | vec<basic_block> bbs; |
1102 | |
1103 | vec<slp_root> roots; |
1104 | } *bb_vec_info; |
1105 | |
1106 | #define BB_VINFO_BB(B) (B)->bb |
1107 | #define BB_VINFO_GROUPED_STORES(B) (B)->grouped_stores |
1108 | #define BB_VINFO_SLP_INSTANCES(B) (B)->slp_instances |
1109 | #define BB_VINFO_DATAREFS(B) (B)->shared->datarefs |
1110 | #define BB_VINFO_DDRS(B) (B)->shared->ddrs |
1111 | |
1112 | /*-----------------------------------------------------------------*/ |
1113 | /* Info on vectorized defs. */ |
1114 | /*-----------------------------------------------------------------*/ |
1115 | enum stmt_vec_info_type { |
1116 | undef_vec_info_type = 0, |
1117 | load_vec_info_type, |
1118 | store_vec_info_type, |
1119 | shift_vec_info_type, |
1120 | op_vec_info_type, |
1121 | call_vec_info_type, |
1122 | call_simd_clone_vec_info_type, |
1123 | assignment_vec_info_type, |
1124 | condition_vec_info_type, |
1125 | comparison_vec_info_type, |
1126 | reduc_vec_info_type, |
1127 | induc_vec_info_type, |
1128 | type_promotion_vec_info_type, |
1129 | type_demotion_vec_info_type, |
1130 | type_conversion_vec_info_type, |
1131 | cycle_phi_info_type, |
1132 | lc_phi_info_type, |
1133 | phi_info_type, |
1134 | recurr_info_type, |
1135 | loop_exit_ctrl_vec_info_type |
1136 | }; |
1137 | |
1138 | /* Indicates whether/how a variable is used in the scope of loop/basic |
1139 | block. */ |
1140 | enum vect_relevant { |
1141 | vect_unused_in_scope = 0, |
1142 | |
1143 | /* The def is only used outside the loop. */ |
1144 | vect_used_only_live, |
1145 | /* The def is in the inner loop, and the use is in the outer loop, and the |
1146 | use is a reduction stmt. */ |
1147 | vect_used_in_outer_by_reduction, |
1148 | /* The def is in the inner loop, and the use is in the outer loop (and is |
1149 | not part of reduction). */ |
1150 | vect_used_in_outer, |
1151 | |
1152 | /* defs that feed computations that end up (only) in a reduction. These |
1153 | defs may be used by non-reduction stmts, but eventually, any |
1154 | computations/values that are affected by these defs are used to compute |
1155 | a reduction (i.e. don't get stored to memory, for example). We use this |
1156 | to identify computations that we can change the order in which they are |
1157 | computed. */ |
1158 | vect_used_by_reduction, |
1159 | |
1160 | vect_used_in_scope |
1161 | }; |
1162 | |
1163 | /* The type of vectorization that can be applied to the stmt: regular loop-based |
1164 | vectorization; pure SLP - the stmt is a part of SLP instances and does not |
1165 | have uses outside SLP instances; or hybrid SLP and loop-based - the stmt is |
1166 | a part of SLP instance and also must be loop-based vectorized, since it has |
1167 | uses outside SLP sequences. |
1168 | |
1169 | In the loop context the meanings of pure and hybrid SLP are slightly |
1170 | different. By saying that pure SLP is applied to the loop, we mean that we |
1171 | exploit only intra-iteration parallelism in the loop; i.e., the loop can be |
1172 | vectorized without doing any conceptual unrolling, cause we don't pack |
1173 | together stmts from different iterations, only within a single iteration. |
1174 | Loop hybrid SLP means that we exploit both intra-iteration and |
1175 | inter-iteration parallelism (e.g., number of elements in the vector is 4 |
1176 | and the slp-group-size is 2, in which case we don't have enough parallelism |
1177 | within an iteration, so we obtain the rest of the parallelism from subsequent |
1178 | iterations by unrolling the loop by 2). */ |
1179 | enum slp_vect_type { |
1180 | loop_vect = 0, |
1181 | pure_slp, |
1182 | hybrid |
1183 | }; |
1184 | |
1185 | /* Says whether a statement is a load, a store of a vectorized statement |
1186 | result, or a store of an invariant value. */ |
1187 | enum vec_load_store_type { |
1188 | VLS_LOAD, |
1189 | VLS_STORE, |
1190 | VLS_STORE_INVARIANT |
1191 | }; |
1192 | |
1193 | /* Describes how we're going to vectorize an individual load or store, |
1194 | or a group of loads or stores. */ |
1195 | enum vect_memory_access_type { |
1196 | /* An access to an invariant address. This is used only for loads. */ |
1197 | VMAT_INVARIANT, |
1198 | |
1199 | /* A simple contiguous access. */ |
1200 | VMAT_CONTIGUOUS, |
1201 | |
1202 | /* A contiguous access that goes down in memory rather than up, |
1203 | with no additional permutation. This is used only for stores |
1204 | of invariants. */ |
1205 | VMAT_CONTIGUOUS_DOWN, |
1206 | |
1207 | /* A simple contiguous access in which the elements need to be permuted |
1208 | after loading or before storing. Only used for loop vectorization; |
1209 | SLP uses separate permutes. */ |
1210 | VMAT_CONTIGUOUS_PERMUTE, |
1211 | |
1212 | /* A simple contiguous access in which the elements need to be reversed |
1213 | after loading or before storing. */ |
1214 | VMAT_CONTIGUOUS_REVERSE, |
1215 | |
1216 | /* An access that uses IFN_LOAD_LANES or IFN_STORE_LANES. */ |
1217 | VMAT_LOAD_STORE_LANES, |
1218 | |
1219 | /* An access in which each scalar element is loaded or stored |
1220 | individually. */ |
1221 | VMAT_ELEMENTWISE, |
1222 | |
1223 | /* A hybrid of VMAT_CONTIGUOUS and VMAT_ELEMENTWISE, used for grouped |
1224 | SLP accesses. Each unrolled iteration uses a contiguous load |
1225 | or store for the whole group, but the groups from separate iterations |
1226 | are combined in the same way as for VMAT_ELEMENTWISE. */ |
1227 | VMAT_STRIDED_SLP, |
1228 | |
1229 | /* The access uses gather loads or scatter stores. */ |
1230 | VMAT_GATHER_SCATTER |
1231 | }; |
1232 | |
1233 | class dr_vec_info { |
1234 | public: |
1235 | /* The data reference itself. */ |
1236 | data_reference *dr; |
1237 | /* The statement that contains the data reference. */ |
1238 | stmt_vec_info stmt; |
1239 | /* The analysis group this DR belongs to when doing BB vectorization. |
1240 | DRs of the same group belong to the same conditional execution context. */ |
1241 | unsigned group; |
1242 | /* The misalignment in bytes of the reference, or -1 if not known. */ |
1243 | int misalignment; |
1244 | /* The byte alignment that we'd ideally like the reference to have, |
1245 | and the value that misalignment is measured against. */ |
1246 | poly_uint64 target_alignment; |
1247 | /* If true the alignment of base_decl needs to be increased. */ |
1248 | bool base_misaligned; |
1249 | tree base_decl; |
1250 | |
1251 | /* Stores current vectorized loop's offset. To be added to the DR's |
1252 | offset to calculate current offset of data reference. */ |
1253 | tree offset; |
1254 | }; |
1255 | |
1256 | typedef struct data_reference *dr_p; |
1257 | |
1258 | class _stmt_vec_info { |
1259 | public: |
1260 | |
1261 | enum stmt_vec_info_type type; |
1262 | |
1263 | /* Indicates whether this stmts is part of a computation whose result is |
1264 | used outside the loop. */ |
1265 | bool live; |
1266 | |
1267 | /* Stmt is part of some pattern (computation idiom) */ |
1268 | bool in_pattern_p; |
1269 | |
1270 | /* True if the statement was created during pattern recognition as |
1271 | part of the replacement for RELATED_STMT. This implies that the |
1272 | statement isn't part of any basic block, although for convenience |
1273 | its gimple_bb is the same as for RELATED_STMT. */ |
1274 | bool pattern_stmt_p; |
1275 | |
1276 | /* Is this statement vectorizable or should it be skipped in (partial) |
1277 | vectorization. */ |
1278 | bool vectorizable; |
1279 | |
1280 | /* The stmt to which this info struct refers to. */ |
1281 | gimple *stmt; |
1282 | |
1283 | /* The vector type to be used for the LHS of this statement. */ |
1284 | tree vectype; |
1285 | |
1286 | /* The vectorized stmts. */ |
1287 | vec<gimple *> vec_stmts; |
1288 | |
1289 | /* The following is relevant only for stmts that contain a non-scalar |
1290 | data-ref (array/pointer/struct access). A GIMPLE stmt is expected to have |
1291 | at most one such data-ref. */ |
1292 | |
1293 | dr_vec_info dr_aux; |
1294 | |
1295 | /* Information about the data-ref relative to this loop |
1296 | nest (the loop that is being considered for vectorization). */ |
1297 | innermost_loop_behavior dr_wrt_vec_loop; |
1298 | |
1299 | /* For loop PHI nodes, the base and evolution part of it. This makes sure |
1300 | this information is still available in vect_update_ivs_after_vectorizer |
1301 | where we may not be able to re-analyze the PHI nodes evolution as |
1302 | peeling for the prologue loop can make it unanalyzable. The evolution |
1303 | part is still correct after peeling, but the base may have changed from |
1304 | the version here. */ |
1305 | tree loop_phi_evolution_base_unchanged; |
1306 | tree loop_phi_evolution_part; |
1307 | enum vect_induction_op_type loop_phi_evolution_type; |
1308 | |
1309 | /* Used for various bookkeeping purposes, generally holding a pointer to |
1310 | some other stmt S that is in some way "related" to this stmt. |
1311 | Current use of this field is: |
1312 | If this stmt is part of a pattern (i.e. the field 'in_pattern_p' is |
1313 | true): S is the "pattern stmt" that represents (and replaces) the |
1314 | sequence of stmts that constitutes the pattern. Similarly, the |
1315 | related_stmt of the "pattern stmt" points back to this stmt (which is |
1316 | the last stmt in the original sequence of stmts that constitutes the |
1317 | pattern). */ |
1318 | stmt_vec_info related_stmt; |
1319 | |
1320 | /* Used to keep a sequence of def stmts of a pattern stmt if such exists. |
1321 | The sequence is attached to the original statement rather than the |
1322 | pattern statement. */ |
1323 | gimple_seq pattern_def_seq; |
1324 | |
1325 | /* Selected SIMD clone's function info. First vector element |
1326 | is SIMD clone's function decl, followed by a pair of trees (base + step) |
1327 | for linear arguments (pair of NULLs for other arguments). */ |
1328 | vec<tree> simd_clone_info; |
1329 | |
1330 | /* Classify the def of this stmt. */ |
1331 | enum vect_def_type def_type; |
1332 | |
1333 | /* Whether the stmt is SLPed, loop-based vectorized, or both. */ |
1334 | enum slp_vect_type slp_type; |
1335 | |
1336 | /* Interleaving and reduction chains info. */ |
1337 | /* First element in the group. */ |
1338 | stmt_vec_info first_element; |
1339 | /* Pointer to the next element in the group. */ |
1340 | stmt_vec_info next_element; |
1341 | /* The size of the group. */ |
1342 | unsigned int size; |
1343 | /* For stores, number of stores from this group seen. We vectorize the last |
1344 | one. */ |
1345 | unsigned int store_count; |
1346 | /* For loads only, the gap from the previous load. For consecutive loads, GAP |
1347 | is 1. */ |
1348 | unsigned int gap; |
1349 | |
1350 | /* The minimum negative dependence distance this stmt participates in |
1351 | or zero if none. */ |
1352 | unsigned int min_neg_dist; |
1353 | |
1354 | /* Not all stmts in the loop need to be vectorized. e.g, the increment |
1355 | of the loop induction variable and computation of array indexes. relevant |
1356 | indicates whether the stmt needs to be vectorized. */ |
1357 | enum vect_relevant relevant; |
1358 | |
1359 | /* For loads if this is a gather, for stores if this is a scatter. */ |
1360 | bool gather_scatter_p; |
1361 | |
1362 | /* True if this is an access with loop-invariant stride. */ |
1363 | bool strided_p; |
1364 | |
1365 | /* For both loads and stores. */ |
1366 | unsigned simd_lane_access_p : 3; |
1367 | |
1368 | /* Classifies how the load or store is going to be implemented |
1369 | for loop vectorization. */ |
1370 | vect_memory_access_type memory_access_type; |
1371 | |
1372 | /* For INTEGER_INDUC_COND_REDUCTION, the initial value to be used. */ |
1373 | tree induc_cond_initial_val; |
1374 | |
1375 | /* If not NULL the value to be added to compute final reduction value. */ |
1376 | tree reduc_epilogue_adjustment; |
1377 | |
1378 | /* On a reduction PHI the reduction type as detected by |
1379 | vect_is_simple_reduction and vectorizable_reduction. */ |
1380 | enum vect_reduction_type reduc_type; |
1381 | |
1382 | /* The original reduction code, to be used in the epilogue. */ |
1383 | code_helper reduc_code; |
1384 | /* An internal function we should use in the epilogue. */ |
1385 | internal_fn reduc_fn; |
1386 | |
1387 | /* On a stmt participating in the reduction the index of the operand |
1388 | on the reduction SSA cycle. */ |
1389 | int reduc_idx; |
1390 | |
1391 | /* On a reduction PHI the def returned by vect_force_simple_reduction. |
1392 | On the def returned by vect_force_simple_reduction the |
1393 | corresponding PHI. */ |
1394 | stmt_vec_info reduc_def; |
1395 | |
1396 | /* The vector input type relevant for reduction vectorization. */ |
1397 | tree reduc_vectype_in; |
1398 | |
1399 | /* The vector type for performing the actual reduction. */ |
1400 | tree reduc_vectype; |
1401 | |
1402 | /* If IS_REDUC_INFO is true and if the vector code is performing |
1403 | N scalar reductions in parallel, this variable gives the initial |
1404 | scalar values of those N reductions. */ |
1405 | vec<tree> reduc_initial_values; |
1406 | |
1407 | /* If IS_REDUC_INFO is true and if the vector code is performing |
1408 | N scalar reductions in parallel, this variable gives the vectorized code's |
1409 | final (scalar) result for each of those N reductions. In other words, |
1410 | REDUC_SCALAR_RESULTS[I] replaces the original scalar code's loop-closed |
1411 | SSA PHI for reduction number I. */ |
1412 | vec<tree> reduc_scalar_results; |
1413 | |
1414 | /* Only meaningful if IS_REDUC_INFO. If non-null, the reduction is |
1415 | being performed by an epilogue loop and we have decided to reuse |
1416 | this accumulator from the main loop. */ |
1417 | vect_reusable_accumulator *reused_accumulator; |
1418 | |
1419 | /* Whether we force a single cycle PHI during reduction vectorization. */ |
1420 | bool force_single_cycle; |
1421 | |
1422 | /* Whether on this stmt reduction meta is recorded. */ |
1423 | bool is_reduc_info; |
1424 | |
1425 | /* If nonzero, the lhs of the statement could be truncated to this |
1426 | many bits without affecting any users of the result. */ |
1427 | unsigned int min_output_precision; |
1428 | |
1429 | /* If nonzero, all non-boolean input operands have the same precision, |
1430 | and they could each be truncated to this many bits without changing |
1431 | the result. */ |
1432 | unsigned int min_input_precision; |
1433 | |
1434 | /* If OPERATION_BITS is nonzero, the statement could be performed on |
1435 | an integer with the sign and number of bits given by OPERATION_SIGN |
1436 | and OPERATION_BITS without changing the result. */ |
1437 | unsigned int operation_precision; |
1438 | signop operation_sign; |
1439 | |
1440 | /* If the statement produces a boolean result, this value describes |
1441 | how we should choose the associated vector type. The possible |
1442 | values are: |
1443 | |
1444 | - an integer precision N if we should use the vector mask type |
1445 | associated with N-bit integers. This is only used if all relevant |
1446 | input booleans also want the vector mask type for N-bit integers, |
1447 | or if we can convert them into that form by pattern-matching. |
1448 | |
1449 | - ~0U if we considered choosing a vector mask type but decided |
1450 | to treat the boolean as a normal integer type instead. |
1451 | |
1452 | - 0 otherwise. This means either that the operation isn't one that |
1453 | could have a vector mask type (and so should have a normal vector |
1454 | type instead) or that we simply haven't made a choice either way. */ |
1455 | unsigned int mask_precision; |
1456 | |
1457 | /* True if this is only suitable for SLP vectorization. */ |
1458 | bool slp_vect_only_p; |
1459 | |
1460 | /* True if this is a pattern that can only be handled by SLP |
1461 | vectorization. */ |
1462 | bool slp_vect_pattern_only_p; |
1463 | }; |
1464 | |
1465 | /* Information about a gather/scatter call. */ |
1466 | struct gather_scatter_info { |
1467 | /* The internal function to use for the gather/scatter operation, |
1468 | or IFN_LAST if a built-in function should be used instead. */ |
1469 | internal_fn ifn; |
1470 | |
1471 | /* The FUNCTION_DECL for the built-in gather/scatter function, |
1472 | or null if an internal function should be used instead. */ |
1473 | tree decl; |
1474 | |
1475 | /* The loop-invariant base value. */ |
1476 | tree base; |
1477 | |
1478 | /* The original scalar offset, which is a non-loop-invariant SSA_NAME. */ |
1479 | tree offset; |
1480 | |
1481 | /* Each offset element should be multiplied by this amount before |
1482 | being added to the base. */ |
1483 | int scale; |
1484 | |
1485 | /* The definition type for the vectorized offset. */ |
1486 | enum vect_def_type offset_dt; |
1487 | |
1488 | /* The type of the vectorized offset. */ |
1489 | tree offset_vectype; |
1490 | |
1491 | /* The type of the scalar elements after loading or before storing. */ |
1492 | tree element_type; |
1493 | |
1494 | /* The type of the scalar elements being loaded or stored. */ |
1495 | tree memory_type; |
1496 | }; |
1497 | |
1498 | /* Access Functions. */ |
1499 | #define STMT_VINFO_TYPE(S) (S)->type |
1500 | #define STMT_VINFO_STMT(S) (S)->stmt |
1501 | #define STMT_VINFO_RELEVANT(S) (S)->relevant |
1502 | #define STMT_VINFO_LIVE_P(S) (S)->live |
1503 | #define STMT_VINFO_VECTYPE(S) (S)->vectype |
1504 | #define STMT_VINFO_VEC_STMTS(S) (S)->vec_stmts |
1505 | #define STMT_VINFO_VECTORIZABLE(S) (S)->vectorizable |
1506 | #define STMT_VINFO_DATA_REF(S) ((S)->dr_aux.dr + 0) |
1507 | #define STMT_VINFO_GATHER_SCATTER_P(S) (S)->gather_scatter_p |
1508 | #define STMT_VINFO_STRIDED_P(S) (S)->strided_p |
1509 | #define STMT_VINFO_MEMORY_ACCESS_TYPE(S) (S)->memory_access_type |
1510 | #define STMT_VINFO_SIMD_LANE_ACCESS_P(S) (S)->simd_lane_access_p |
1511 | #define STMT_VINFO_VEC_INDUC_COND_INITIAL_VAL(S) (S)->induc_cond_initial_val |
1512 | #define STMT_VINFO_REDUC_EPILOGUE_ADJUSTMENT(S) (S)->reduc_epilogue_adjustment |
1513 | #define STMT_VINFO_REDUC_IDX(S) (S)->reduc_idx |
1514 | #define STMT_VINFO_FORCE_SINGLE_CYCLE(S) (S)->force_single_cycle |
1515 | |
1516 | #define STMT_VINFO_DR_WRT_VEC_LOOP(S) (S)->dr_wrt_vec_loop |
1517 | #define STMT_VINFO_DR_BASE_ADDRESS(S) (S)->dr_wrt_vec_loop.base_address |
1518 | #define STMT_VINFO_DR_INIT(S) (S)->dr_wrt_vec_loop.init |
1519 | #define STMT_VINFO_DR_OFFSET(S) (S)->dr_wrt_vec_loop.offset |
1520 | #define STMT_VINFO_DR_STEP(S) (S)->dr_wrt_vec_loop.step |
1521 | #define STMT_VINFO_DR_BASE_ALIGNMENT(S) (S)->dr_wrt_vec_loop.base_alignment |
1522 | #define STMT_VINFO_DR_BASE_MISALIGNMENT(S) \ |
1523 | (S)->dr_wrt_vec_loop.base_misalignment |
1524 | #define STMT_VINFO_DR_OFFSET_ALIGNMENT(S) \ |
1525 | (S)->dr_wrt_vec_loop.offset_alignment |
1526 | #define STMT_VINFO_DR_STEP_ALIGNMENT(S) \ |
1527 | (S)->dr_wrt_vec_loop.step_alignment |
1528 | |
1529 | #define STMT_VINFO_DR_INFO(S) \ |
1530 | (gcc_checking_assert ((S)->dr_aux.stmt == (S)), &(S)->dr_aux) |
1531 | |
1532 | #define STMT_VINFO_IN_PATTERN_P(S) (S)->in_pattern_p |
1533 | #define STMT_VINFO_RELATED_STMT(S) (S)->related_stmt |
1534 | #define STMT_VINFO_PATTERN_DEF_SEQ(S) (S)->pattern_def_seq |
1535 | #define STMT_VINFO_SIMD_CLONE_INFO(S) (S)->simd_clone_info |
1536 | #define STMT_VINFO_DEF_TYPE(S) (S)->def_type |
1537 | #define STMT_VINFO_GROUPED_ACCESS(S) \ |
1538 | ((S)->dr_aux.dr && DR_GROUP_FIRST_ELEMENT(S)) |
1539 | #define STMT_VINFO_LOOP_PHI_EVOLUTION_BASE_UNCHANGED(S) (S)->loop_phi_evolution_base_unchanged |
1540 | #define STMT_VINFO_LOOP_PHI_EVOLUTION_PART(S) (S)->loop_phi_evolution_part |
1541 | #define STMT_VINFO_LOOP_PHI_EVOLUTION_TYPE(S) (S)->loop_phi_evolution_type |
1542 | #define STMT_VINFO_MIN_NEG_DIST(S) (S)->min_neg_dist |
1543 | #define STMT_VINFO_REDUC_TYPE(S) (S)->reduc_type |
1544 | #define STMT_VINFO_REDUC_CODE(S) (S)->reduc_code |
1545 | #define STMT_VINFO_REDUC_FN(S) (S)->reduc_fn |
1546 | #define STMT_VINFO_REDUC_DEF(S) (S)->reduc_def |
1547 | #define STMT_VINFO_REDUC_VECTYPE(S) (S)->reduc_vectype |
1548 | #define STMT_VINFO_REDUC_VECTYPE_IN(S) (S)->reduc_vectype_in |
1549 | #define STMT_VINFO_SLP_VECT_ONLY(S) (S)->slp_vect_only_p |
1550 | #define STMT_VINFO_SLP_VECT_ONLY_PATTERN(S) (S)->slp_vect_pattern_only_p |
1551 | |
1552 | #define DR_GROUP_FIRST_ELEMENT(S) \ |
1553 | (gcc_checking_assert ((S)->dr_aux.dr), (S)->first_element) |
1554 | #define DR_GROUP_NEXT_ELEMENT(S) \ |
1555 | (gcc_checking_assert ((S)->dr_aux.dr), (S)->next_element) |
1556 | #define DR_GROUP_SIZE(S) \ |
1557 | (gcc_checking_assert ((S)->dr_aux.dr), (S)->size) |
1558 | #define DR_GROUP_STORE_COUNT(S) \ |
1559 | (gcc_checking_assert ((S)->dr_aux.dr), (S)->store_count) |
1560 | #define DR_GROUP_GAP(S) \ |
1561 | (gcc_checking_assert ((S)->dr_aux.dr), (S)->gap) |
1562 | |
1563 | #define REDUC_GROUP_FIRST_ELEMENT(S) \ |
1564 | (gcc_checking_assert (!(S)->dr_aux.dr), (S)->first_element) |
1565 | #define REDUC_GROUP_NEXT_ELEMENT(S) \ |
1566 | (gcc_checking_assert (!(S)->dr_aux.dr), (S)->next_element) |
1567 | #define REDUC_GROUP_SIZE(S) \ |
1568 | (gcc_checking_assert (!(S)->dr_aux.dr), (S)->size) |
1569 | |
1570 | #define STMT_VINFO_RELEVANT_P(S) ((S)->relevant != vect_unused_in_scope) |
1571 | |
1572 | #define HYBRID_SLP_STMT(S) ((S)->slp_type == hybrid) |
1573 | #define PURE_SLP_STMT(S) ((S)->slp_type == pure_slp) |
1574 | #define STMT_SLP_TYPE(S) (S)->slp_type |
1575 | |
1576 | /* Contains the scalar or vector costs for a vec_info. */ |
1577 | class vector_costs |
1578 | { |
1579 | public: |
1580 | vector_costs (vec_info *, bool); |
1581 | virtual ~vector_costs () {} |
1582 | |
1583 | /* Update the costs in response to adding COUNT copies of a statement. |
1584 | |
1585 | - WHERE specifies whether the cost occurs in the loop prologue, |
1586 | the loop body, or the loop epilogue. |
1587 | - KIND is the kind of statement, which is always meaningful. |
1588 | - STMT_INFO or NODE, if nonnull, describe the statement that will be |
1589 | vectorized. |
1590 | - VECTYPE, if nonnull, is the vector type that the vectorized |
1591 | statement will operate on. Note that this should be used in |
1592 | preference to STMT_VINFO_VECTYPE (STMT_INFO) since the latter |
1593 | is not correct for SLP. |
1594 | - for unaligned_load and unaligned_store statements, MISALIGN is |
1595 | the byte misalignment of the load or store relative to the target's |
1596 | preferred alignment for VECTYPE, or DR_MISALIGNMENT_UNKNOWN |
1597 | if the misalignment is not known. |
1598 | |
1599 | Return the calculated cost as well as recording it. The return |
1600 | value is used for dumping purposes. */ |
1601 | virtual unsigned int add_stmt_cost (int count, vect_cost_for_stmt kind, |
1602 | stmt_vec_info stmt_info, |
1603 | slp_tree node, |
1604 | tree vectype, int misalign, |
1605 | vect_cost_model_location where); |
1606 | |
1607 | /* Finish calculating the cost of the code. The results can be |
1608 | read back using the functions below. |
1609 | |
1610 | If the costs describe vector code, SCALAR_COSTS gives the costs |
1611 | of the corresponding scalar code, otherwise it is null. */ |
1612 | virtual void finish_cost (const vector_costs *scalar_costs); |
1613 | |
1614 | /* The costs in THIS and OTHER both describe ways of vectorizing |
1615 | a main loop. Return true if the costs described by THIS are |
1616 | cheaper than the costs described by OTHER. Return false if any |
1617 | of the following are true: |
1618 | |
1619 | - THIS and OTHER are of equal cost |
1620 | - OTHER is better than THIS |
1621 | - we can't be sure about the relative costs of THIS and OTHER. */ |
1622 | virtual bool better_main_loop_than_p (const vector_costs *other) const; |
1623 | |
1624 | /* Likewise, but the costs in THIS and OTHER both describe ways of |
1625 | vectorizing an epilogue loop of MAIN_LOOP. */ |
1626 | virtual bool better_epilogue_loop_than_p (const vector_costs *other, |
1627 | loop_vec_info main_loop) const; |
1628 | |
1629 | unsigned int prologue_cost () const; |
1630 | unsigned int body_cost () const; |
1631 | unsigned int epilogue_cost () const; |
1632 | unsigned int outside_cost () const; |
1633 | unsigned int total_cost () const; |
1634 | unsigned int suggested_unroll_factor () const; |
1635 | |
1636 | protected: |
1637 | unsigned int record_stmt_cost (stmt_vec_info, vect_cost_model_location, |
1638 | unsigned int); |
1639 | unsigned int adjust_cost_for_freq (stmt_vec_info, vect_cost_model_location, |
1640 | unsigned int); |
1641 | int compare_inside_loop_cost (const vector_costs *) const; |
1642 | int compare_outside_loop_cost (const vector_costs *) const; |
1643 | |
1644 | /* The region of code that we're considering vectorizing. */ |
1645 | vec_info *m_vinfo; |
1646 | |
1647 | /* True if we're costing the scalar code, false if we're costing |
1648 | the vector code. */ |
1649 | bool m_costing_for_scalar; |
1650 | |
1651 | /* The costs of the three regions, indexed by vect_cost_model_location. */ |
1652 | unsigned int m_costs[3]; |
1653 | |
1654 | /* The suggested unrolling factor determined at finish_cost. */ |
1655 | unsigned int m_suggested_unroll_factor; |
1656 | |
1657 | /* True if finish_cost has been called. */ |
1658 | bool m_finished; |
1659 | }; |
1660 | |
1661 | /* Create costs for VINFO. COSTING_FOR_SCALAR is true if the costs |
1662 | are for scalar code, false if they are for vector code. */ |
1663 | |
1664 | inline |
1665 | vector_costs::vector_costs (vec_info *vinfo, bool costing_for_scalar) |
1666 | : m_vinfo (vinfo), |
1667 | m_costing_for_scalar (costing_for_scalar), |
1668 | m_costs (), |
1669 | m_suggested_unroll_factor(1), |
1670 | m_finished (false) |
1671 | { |
1672 | } |
1673 | |
1674 | /* Return the cost of the prologue code (in abstract units). */ |
1675 | |
1676 | inline unsigned int |
1677 | vector_costs::prologue_cost () const |
1678 | { |
1679 | gcc_checking_assert (m_finished); |
1680 | return m_costs[vect_prologue]; |
1681 | } |
1682 | |
1683 | /* Return the cost of the body code (in abstract units). */ |
1684 | |
1685 | inline unsigned int |
1686 | vector_costs::body_cost () const |
1687 | { |
1688 | gcc_checking_assert (m_finished); |
1689 | return m_costs[vect_body]; |
1690 | } |
1691 | |
1692 | /* Return the cost of the epilogue code (in abstract units). */ |
1693 | |
1694 | inline unsigned int |
1695 | vector_costs::epilogue_cost () const |
1696 | { |
1697 | gcc_checking_assert (m_finished); |
1698 | return m_costs[vect_epilogue]; |
1699 | } |
1700 | |
1701 | /* Return the cost of the prologue and epilogue code (in abstract units). */ |
1702 | |
1703 | inline unsigned int |
1704 | vector_costs::outside_cost () const |
1705 | { |
1706 | return prologue_cost () + epilogue_cost (); |
1707 | } |
1708 | |
1709 | /* Return the cost of the prologue, body and epilogue code |
1710 | (in abstract units). */ |
1711 | |
1712 | inline unsigned int |
1713 | vector_costs::total_cost () const |
1714 | { |
1715 | return body_cost () + outside_cost (); |
1716 | } |
1717 | |
1718 | /* Return the suggested unroll factor. */ |
1719 | |
1720 | inline unsigned int |
1721 | vector_costs::suggested_unroll_factor () const |
1722 | { |
1723 | gcc_checking_assert (m_finished); |
1724 | return m_suggested_unroll_factor; |
1725 | } |
1726 | |
1727 | #define VECT_MAX_COST 1000 |
1728 | |
1729 | /* The maximum number of intermediate steps required in multi-step type |
1730 | conversion. */ |
1731 | #define MAX_INTERM_CVT_STEPS 3 |
1732 | |
1733 | #define MAX_VECTORIZATION_FACTOR INT_MAX |
1734 | |
1735 | /* Nonzero if TYPE represents a (scalar) boolean type or type |
1736 | in the middle-end compatible with it (unsigned precision 1 integral |
1737 | types). Used to determine which types should be vectorized as |
1738 | VECTOR_BOOLEAN_TYPE_P. */ |
1739 | |
1740 | #define VECT_SCALAR_BOOLEAN_TYPE_P(TYPE) \ |
1741 | (TREE_CODE (TYPE) == BOOLEAN_TYPE \ |
1742 | || ((TREE_CODE (TYPE) == INTEGER_TYPE \ |
1743 | || TREE_CODE (TYPE) == ENUMERAL_TYPE) \ |
1744 | && TYPE_PRECISION (TYPE) == 1 \ |
1745 | && TYPE_UNSIGNED (TYPE))) |
1746 | |
1747 | inline bool |
1748 | nested_in_vect_loop_p (class loop *loop, stmt_vec_info stmt_info) |
1749 | { |
1750 | return (loop->inner |
1751 | && (loop->inner == (gimple_bb (g: stmt_info->stmt))->loop_father)); |
1752 | } |
1753 | |
1754 | /* PHI is either a scalar reduction phi or a scalar induction phi. |
1755 | Return the initial value of the variable on entry to the containing |
1756 | loop. */ |
1757 | |
1758 | inline tree |
1759 | vect_phi_initial_value (gphi *phi) |
1760 | { |
1761 | basic_block bb = gimple_bb (g: phi); |
1762 | edge pe = loop_preheader_edge (bb->loop_father); |
1763 | gcc_assert (pe->dest == bb); |
1764 | return PHI_ARG_DEF_FROM_EDGE (phi, pe); |
1765 | } |
1766 | |
1767 | /* Return true if STMT_INFO should produce a vector mask type rather than |
1768 | a normal nonmask type. */ |
1769 | |
1770 | inline bool |
1771 | vect_use_mask_type_p (stmt_vec_info stmt_info) |
1772 | { |
1773 | return stmt_info->mask_precision && stmt_info->mask_precision != ~0U; |
1774 | } |
1775 | |
1776 | /* Return TRUE if a statement represented by STMT_INFO is a part of a |
1777 | pattern. */ |
1778 | |
1779 | inline bool |
1780 | is_pattern_stmt_p (stmt_vec_info stmt_info) |
1781 | { |
1782 | return stmt_info->pattern_stmt_p; |
1783 | } |
1784 | |
1785 | /* If STMT_INFO is a pattern statement, return the statement that it |
1786 | replaces, otherwise return STMT_INFO itself. */ |
1787 | |
1788 | inline stmt_vec_info |
1789 | vect_orig_stmt (stmt_vec_info stmt_info) |
1790 | { |
1791 | if (is_pattern_stmt_p (stmt_info)) |
1792 | return STMT_VINFO_RELATED_STMT (stmt_info); |
1793 | return stmt_info; |
1794 | } |
1795 | |
1796 | /* Return the later statement between STMT1_INFO and STMT2_INFO. */ |
1797 | |
1798 | inline stmt_vec_info |
1799 | get_later_stmt (stmt_vec_info stmt1_info, stmt_vec_info stmt2_info) |
1800 | { |
1801 | if (gimple_uid (g: vect_orig_stmt (stmt_info: stmt1_info)->stmt) |
1802 | > gimple_uid (g: vect_orig_stmt (stmt_info: stmt2_info)->stmt)) |
1803 | return stmt1_info; |
1804 | else |
1805 | return stmt2_info; |
1806 | } |
1807 | |
1808 | /* If STMT_INFO has been replaced by a pattern statement, return the |
1809 | replacement statement, otherwise return STMT_INFO itself. */ |
1810 | |
1811 | inline stmt_vec_info |
1812 | vect_stmt_to_vectorize (stmt_vec_info stmt_info) |
1813 | { |
1814 | if (STMT_VINFO_IN_PATTERN_P (stmt_info)) |
1815 | return STMT_VINFO_RELATED_STMT (stmt_info); |
1816 | return stmt_info; |
1817 | } |
1818 | |
1819 | /* Return true if BB is a loop header. */ |
1820 | |
1821 | inline bool |
1822 | (basic_block bb) |
1823 | { |
1824 | if (bb == (bb->loop_father)->header) |
1825 | return true; |
1826 | |
1827 | return false; |
1828 | } |
1829 | |
1830 | /* Return pow2 (X). */ |
1831 | |
1832 | inline int |
1833 | vect_pow2 (int x) |
1834 | { |
1835 | int i, res = 1; |
1836 | |
1837 | for (i = 0; i < x; i++) |
1838 | res *= 2; |
1839 | |
1840 | return res; |
1841 | } |
1842 | |
1843 | /* Alias targetm.vectorize.builtin_vectorization_cost. */ |
1844 | |
1845 | inline int |
1846 | builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost, |
1847 | tree vectype, int misalign) |
1848 | { |
1849 | return targetm.vectorize.builtin_vectorization_cost (type_of_cost, |
1850 | vectype, misalign); |
1851 | } |
1852 | |
1853 | /* Get cost by calling cost target builtin. */ |
1854 | |
1855 | inline |
1856 | int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost) |
1857 | { |
1858 | return builtin_vectorization_cost (type_of_cost, NULL, misalign: 0); |
1859 | } |
1860 | |
1861 | /* Alias targetm.vectorize.init_cost. */ |
1862 | |
1863 | inline vector_costs * |
1864 | init_cost (vec_info *vinfo, bool costing_for_scalar) |
1865 | { |
1866 | return targetm.vectorize.create_costs (vinfo, costing_for_scalar); |
1867 | } |
1868 | |
1869 | extern void dump_stmt_cost (FILE *, int, enum vect_cost_for_stmt, |
1870 | stmt_vec_info, slp_tree, tree, int, unsigned, |
1871 | enum vect_cost_model_location); |
1872 | |
1873 | /* Alias targetm.vectorize.add_stmt_cost. */ |
1874 | |
1875 | inline unsigned |
1876 | add_stmt_cost (vector_costs *costs, int count, |
1877 | enum vect_cost_for_stmt kind, |
1878 | stmt_vec_info stmt_info, slp_tree node, |
1879 | tree vectype, int misalign, |
1880 | enum vect_cost_model_location where) |
1881 | { |
1882 | unsigned cost = costs->add_stmt_cost (count, kind, stmt_info, node, vectype, |
1883 | misalign, where); |
1884 | if (dump_file && (dump_flags & TDF_DETAILS)) |
1885 | dump_stmt_cost (dump_file, count, kind, stmt_info, node, vectype, misalign, |
1886 | cost, where); |
1887 | return cost; |
1888 | } |
1889 | |
1890 | inline unsigned |
1891 | add_stmt_cost (vector_costs *costs, int count, enum vect_cost_for_stmt kind, |
1892 | enum vect_cost_model_location where) |
1893 | { |
1894 | gcc_assert (kind == cond_branch_taken || kind == cond_branch_not_taken |
1895 | || kind == scalar_stmt); |
1896 | return add_stmt_cost (costs, count, kind, NULL, NULL, NULL_TREE, misalign: 0, where); |
1897 | } |
1898 | |
1899 | /* Alias targetm.vectorize.add_stmt_cost. */ |
1900 | |
1901 | inline unsigned |
1902 | add_stmt_cost (vector_costs *costs, stmt_info_for_cost *i) |
1903 | { |
1904 | return add_stmt_cost (costs, count: i->count, kind: i->kind, stmt_info: i->stmt_info, node: i->node, |
1905 | vectype: i->vectype, misalign: i->misalign, where: i->where); |
1906 | } |
1907 | |
1908 | /* Alias targetm.vectorize.finish_cost. */ |
1909 | |
1910 | inline void |
1911 | finish_cost (vector_costs *costs, const vector_costs *scalar_costs, |
1912 | unsigned *prologue_cost, unsigned *body_cost, |
1913 | unsigned *epilogue_cost, unsigned *suggested_unroll_factor = NULL) |
1914 | { |
1915 | costs->finish_cost (scalar_costs); |
1916 | *prologue_cost = costs->prologue_cost (); |
1917 | *body_cost = costs->body_cost (); |
1918 | *epilogue_cost = costs->epilogue_cost (); |
1919 | if (suggested_unroll_factor) |
1920 | *suggested_unroll_factor = costs->suggested_unroll_factor (); |
1921 | } |
1922 | |
1923 | inline void |
1924 | add_stmt_costs (vector_costs *costs, stmt_vector_for_cost *cost_vec) |
1925 | { |
1926 | stmt_info_for_cost *cost; |
1927 | unsigned i; |
1928 | FOR_EACH_VEC_ELT (*cost_vec, i, cost) |
1929 | add_stmt_cost (costs, count: cost->count, kind: cost->kind, stmt_info: cost->stmt_info, |
1930 | node: cost->node, vectype: cost->vectype, misalign: cost->misalign, where: cost->where); |
1931 | } |
1932 | |
1933 | /*-----------------------------------------------------------------*/ |
1934 | /* Info on data references alignment. */ |
1935 | /*-----------------------------------------------------------------*/ |
1936 | #define DR_MISALIGNMENT_UNKNOWN (-1) |
1937 | #define DR_MISALIGNMENT_UNINITIALIZED (-2) |
1938 | |
1939 | inline void |
1940 | set_dr_misalignment (dr_vec_info *dr_info, int val) |
1941 | { |
1942 | dr_info->misalignment = val; |
1943 | } |
1944 | |
1945 | extern int dr_misalignment (dr_vec_info *dr_info, tree vectype, |
1946 | poly_int64 offset = 0); |
1947 | |
1948 | #define SET_DR_MISALIGNMENT(DR, VAL) set_dr_misalignment (DR, VAL) |
1949 | |
1950 | /* Only defined once DR_MISALIGNMENT is defined. */ |
1951 | inline const poly_uint64 |
1952 | dr_target_alignment (dr_vec_info *dr_info) |
1953 | { |
1954 | if (STMT_VINFO_GROUPED_ACCESS (dr_info->stmt)) |
1955 | dr_info = STMT_VINFO_DR_INFO (DR_GROUP_FIRST_ELEMENT (dr_info->stmt)); |
1956 | return dr_info->target_alignment; |
1957 | } |
1958 | #define DR_TARGET_ALIGNMENT(DR) dr_target_alignment (DR) |
1959 | |
1960 | inline void |
1961 | set_dr_target_alignment (dr_vec_info *dr_info, poly_uint64 val) |
1962 | { |
1963 | dr_info->target_alignment = val; |
1964 | } |
1965 | #define SET_DR_TARGET_ALIGNMENT(DR, VAL) set_dr_target_alignment (DR, VAL) |
1966 | |
1967 | /* Return true if data access DR_INFO is aligned to the targets |
1968 | preferred alignment for VECTYPE (which may be less than a full vector). */ |
1969 | |
1970 | inline bool |
1971 | aligned_access_p (dr_vec_info *dr_info, tree vectype) |
1972 | { |
1973 | return (dr_misalignment (dr_info, vectype) == 0); |
1974 | } |
1975 | |
1976 | /* Return TRUE if the (mis-)alignment of the data access is known with |
1977 | respect to the targets preferred alignment for VECTYPE, and FALSE |
1978 | otherwise. */ |
1979 | |
1980 | inline bool |
1981 | known_alignment_for_access_p (dr_vec_info *dr_info, tree vectype) |
1982 | { |
1983 | return (dr_misalignment (dr_info, vectype) != DR_MISALIGNMENT_UNKNOWN); |
1984 | } |
1985 | |
1986 | /* Return the minimum alignment in bytes that the vectorized version |
1987 | of DR_INFO is guaranteed to have. */ |
1988 | |
1989 | inline unsigned int |
1990 | vect_known_alignment_in_bytes (dr_vec_info *dr_info, tree vectype) |
1991 | { |
1992 | int misalignment = dr_misalignment (dr_info, vectype); |
1993 | if (misalignment == DR_MISALIGNMENT_UNKNOWN) |
1994 | return TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr_info->dr))); |
1995 | else if (misalignment == 0) |
1996 | return known_alignment (DR_TARGET_ALIGNMENT (dr_info)); |
1997 | return misalignment & -misalignment; |
1998 | } |
1999 | |
2000 | /* Return the behavior of DR_INFO with respect to the vectorization context |
2001 | (which for outer loop vectorization might not be the behavior recorded |
2002 | in DR_INFO itself). */ |
2003 | |
2004 | inline innermost_loop_behavior * |
2005 | vect_dr_behavior (vec_info *vinfo, dr_vec_info *dr_info) |
2006 | { |
2007 | stmt_vec_info stmt_info = dr_info->stmt; |
2008 | loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (p: vinfo); |
2009 | if (loop_vinfo == NULL |
2010 | || !nested_in_vect_loop_p (LOOP_VINFO_LOOP (loop_vinfo), stmt_info)) |
2011 | return &DR_INNERMOST (dr_info->dr); |
2012 | else |
2013 | return &STMT_VINFO_DR_WRT_VEC_LOOP (stmt_info); |
2014 | } |
2015 | |
2016 | /* Return the offset calculated by adding the offset of this DR_INFO to the |
2017 | corresponding data_reference's offset. If CHECK_OUTER then use |
2018 | vect_dr_behavior to select the appropriate data_reference to use. */ |
2019 | |
2020 | inline tree |
2021 | get_dr_vinfo_offset (vec_info *vinfo, |
2022 | dr_vec_info *dr_info, bool check_outer = false) |
2023 | { |
2024 | innermost_loop_behavior *base; |
2025 | if (check_outer) |
2026 | base = vect_dr_behavior (vinfo, dr_info); |
2027 | else |
2028 | base = &dr_info->dr->innermost; |
2029 | |
2030 | tree offset = base->offset; |
2031 | |
2032 | if (!dr_info->offset) |
2033 | return offset; |
2034 | |
2035 | offset = fold_convert (sizetype, offset); |
2036 | return fold_build2 (PLUS_EXPR, TREE_TYPE (dr_info->offset), offset, |
2037 | dr_info->offset); |
2038 | } |
2039 | |
2040 | |
2041 | /* Return the vect cost model for LOOP. */ |
2042 | inline enum vect_cost_model |
2043 | loop_cost_model (loop_p loop) |
2044 | { |
2045 | if (loop != NULL |
2046 | && loop->force_vectorize |
2047 | && flag_simd_cost_model != VECT_COST_MODEL_DEFAULT) |
2048 | return flag_simd_cost_model; |
2049 | return flag_vect_cost_model; |
2050 | } |
2051 | |
2052 | /* Return true if the vect cost model is unlimited. */ |
2053 | inline bool |
2054 | unlimited_cost_model (loop_p loop) |
2055 | { |
2056 | return loop_cost_model (loop) == VECT_COST_MODEL_UNLIMITED; |
2057 | } |
2058 | |
2059 | /* Return true if the loop described by LOOP_VINFO is fully-masked and |
2060 | if the first iteration should use a partial mask in order to achieve |
2061 | alignment. */ |
2062 | |
2063 | inline bool |
2064 | vect_use_loop_mask_for_alignment_p (loop_vec_info loop_vinfo) |
2065 | { |
2066 | return (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) |
2067 | && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo)); |
2068 | } |
2069 | |
2070 | /* Return the number of vectors of type VECTYPE that are needed to get |
2071 | NUNITS elements. NUNITS should be based on the vectorization factor, |
2072 | so it is always a known multiple of the number of elements in VECTYPE. */ |
2073 | |
2074 | inline unsigned int |
2075 | vect_get_num_vectors (poly_uint64 nunits, tree vectype) |
2076 | { |
2077 | return exact_div (a: nunits, b: TYPE_VECTOR_SUBPARTS (node: vectype)).to_constant (); |
2078 | } |
2079 | |
2080 | /* Return the number of copies needed for loop vectorization when |
2081 | a statement operates on vectors of type VECTYPE. This is the |
2082 | vectorization factor divided by the number of elements in |
2083 | VECTYPE and is always known at compile time. */ |
2084 | |
2085 | inline unsigned int |
2086 | vect_get_num_copies (loop_vec_info loop_vinfo, tree vectype) |
2087 | { |
2088 | return vect_get_num_vectors (LOOP_VINFO_VECT_FACTOR (loop_vinfo), vectype); |
2089 | } |
2090 | |
2091 | /* Update maximum unit count *MAX_NUNITS so that it accounts for |
2092 | NUNITS. *MAX_NUNITS can be 1 if we haven't yet recorded anything. */ |
2093 | |
2094 | inline void |
2095 | vect_update_max_nunits (poly_uint64 *max_nunits, poly_uint64 nunits) |
2096 | { |
2097 | /* All unit counts have the form vec_info::vector_size * X for some |
2098 | rational X, so two unit sizes must have a common multiple. |
2099 | Everything is a multiple of the initial value of 1. */ |
2100 | *max_nunits = force_common_multiple (a: *max_nunits, b: nunits); |
2101 | } |
2102 | |
2103 | /* Update maximum unit count *MAX_NUNITS so that it accounts for |
2104 | the number of units in vector type VECTYPE. *MAX_NUNITS can be 1 |
2105 | if we haven't yet recorded any vector types. */ |
2106 | |
2107 | inline void |
2108 | vect_update_max_nunits (poly_uint64 *max_nunits, tree vectype) |
2109 | { |
2110 | vect_update_max_nunits (max_nunits, nunits: TYPE_VECTOR_SUBPARTS (node: vectype)); |
2111 | } |
2112 | |
2113 | /* Return the vectorization factor that should be used for costing |
2114 | purposes while vectorizing the loop described by LOOP_VINFO. |
2115 | Pick a reasonable estimate if the vectorization factor isn't |
2116 | known at compile time. */ |
2117 | |
2118 | inline unsigned int |
2119 | vect_vf_for_cost (loop_vec_info loop_vinfo) |
2120 | { |
2121 | return estimated_poly_value (LOOP_VINFO_VECT_FACTOR (loop_vinfo)); |
2122 | } |
2123 | |
2124 | /* Estimate the number of elements in VEC_TYPE for costing purposes. |
2125 | Pick a reasonable estimate if the exact number isn't known at |
2126 | compile time. */ |
2127 | |
2128 | inline unsigned int |
2129 | vect_nunits_for_cost (tree vec_type) |
2130 | { |
2131 | return estimated_poly_value (x: TYPE_VECTOR_SUBPARTS (node: vec_type)); |
2132 | } |
2133 | |
2134 | /* Return the maximum possible vectorization factor for LOOP_VINFO. */ |
2135 | |
2136 | inline unsigned HOST_WIDE_INT |
2137 | vect_max_vf (loop_vec_info loop_vinfo) |
2138 | { |
2139 | unsigned HOST_WIDE_INT vf; |
2140 | if (LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (const_value: &vf)) |
2141 | return vf; |
2142 | return MAX_VECTORIZATION_FACTOR; |
2143 | } |
2144 | |
2145 | /* Return the size of the value accessed by unvectorized data reference |
2146 | DR_INFO. This is only valid once STMT_VINFO_VECTYPE has been calculated |
2147 | for the associated gimple statement, since that guarantees that DR_INFO |
2148 | accesses either a scalar or a scalar equivalent. ("Scalar equivalent" |
2149 | here includes things like V1SI, which can be vectorized in the same way |
2150 | as a plain SI.) */ |
2151 | |
2152 | inline unsigned int |
2153 | vect_get_scalar_dr_size (dr_vec_info *dr_info) |
2154 | { |
2155 | return tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_info->dr)))); |
2156 | } |
2157 | |
2158 | /* Return true if LOOP_VINFO requires a runtime check for whether the |
2159 | vector loop is profitable. */ |
2160 | |
2161 | inline bool |
2162 | vect_apply_runtime_profitability_check_p (loop_vec_info loop_vinfo) |
2163 | { |
2164 | unsigned int th = LOOP_VINFO_COST_MODEL_THRESHOLD (loop_vinfo); |
2165 | return (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) |
2166 | && th >= vect_vf_for_cost (loop_vinfo)); |
2167 | } |
2168 | |
2169 | /* Source location + hotness information. */ |
2170 | extern dump_user_location_t vect_location; |
2171 | |
2172 | /* A macro for calling: |
2173 | dump_begin_scope (MSG, vect_location); |
2174 | via an RAII object, thus printing "=== MSG ===\n" to the dumpfile etc, |
2175 | and then calling |
2176 | dump_end_scope (); |
2177 | once the object goes out of scope, thus capturing the nesting of |
2178 | the scopes. |
2179 | |
2180 | These scopes affect dump messages within them: dump messages at the |
2181 | top level implicitly default to MSG_PRIORITY_USER_FACING, whereas those |
2182 | in a nested scope implicitly default to MSG_PRIORITY_INTERNALS. */ |
2183 | |
2184 | #define DUMP_VECT_SCOPE(MSG) \ |
2185 | AUTO_DUMP_SCOPE (MSG, vect_location) |
2186 | |
2187 | /* A sentinel class for ensuring that the "vect_location" global gets |
2188 | reset at the end of a scope. |
2189 | |
2190 | The "vect_location" global is used during dumping and contains a |
2191 | location_t, which could contain references to a tree block via the |
2192 | ad-hoc data. This data is used for tracking inlining information, |
2193 | but it's not a GC root; it's simply assumed that such locations never |
2194 | get accessed if the blocks are optimized away. |
2195 | |
2196 | Hence we need to ensure that such locations are purged at the end |
2197 | of any operations using them (e.g. via this class). */ |
2198 | |
2199 | class auto_purge_vect_location |
2200 | { |
2201 | public: |
2202 | ~auto_purge_vect_location (); |
2203 | }; |
2204 | |
2205 | /*-----------------------------------------------------------------*/ |
2206 | /* Function prototypes. */ |
2207 | /*-----------------------------------------------------------------*/ |
2208 | |
2209 | /* Simple loop peeling and versioning utilities for vectorizer's purposes - |
2210 | in tree-vect-loop-manip.cc. */ |
2211 | extern void vect_set_loop_condition (class loop *, edge, loop_vec_info, |
2212 | tree, tree, tree, bool); |
2213 | extern bool slpeel_can_duplicate_loop_p (const class loop *, const_edge, |
2214 | const_edge); |
2215 | class loop *slpeel_tree_duplicate_loop_to_edge_cfg (class loop *, edge, |
2216 | class loop *, edge, |
2217 | edge, edge *, bool = true, |
2218 | vec<basic_block> * = NULL); |
2219 | class loop *vect_loop_versioning (loop_vec_info, gimple *); |
2220 | extern class loop *vect_do_peeling (loop_vec_info, tree, tree, |
2221 | tree *, tree *, tree *, int, bool, bool, |
2222 | tree *); |
2223 | extern tree vect_get_main_loop_result (loop_vec_info, tree, tree); |
2224 | extern void vect_prepare_for_masked_peels (loop_vec_info); |
2225 | extern dump_user_location_t find_loop_location (class loop *); |
2226 | extern bool vect_can_advance_ivs_p (loop_vec_info); |
2227 | extern void vect_update_inits_of_drs (loop_vec_info, tree, tree_code); |
2228 | extern edge vec_init_loop_exit_info (class loop *); |
2229 | extern void vect_iv_increment_position (edge, gimple_stmt_iterator *, bool *); |
2230 | |
2231 | /* In tree-vect-stmts.cc. */ |
2232 | extern tree get_related_vectype_for_scalar_type (machine_mode, tree, |
2233 | poly_uint64 = 0); |
2234 | extern tree get_vectype_for_scalar_type (vec_info *, tree, unsigned int = 0); |
2235 | extern tree get_vectype_for_scalar_type (vec_info *, tree, slp_tree); |
2236 | extern tree get_mask_type_for_scalar_type (vec_info *, tree, unsigned int = 0); |
2237 | extern tree get_mask_type_for_scalar_type (vec_info *, tree, slp_tree); |
2238 | extern tree get_same_sized_vectype (tree, tree); |
2239 | extern bool vect_chooses_same_modes_p (vec_info *, machine_mode); |
2240 | extern bool vect_get_loop_mask_type (loop_vec_info); |
2241 | extern bool vect_is_simple_use (tree, vec_info *, enum vect_def_type *, |
2242 | stmt_vec_info * = NULL, gimple ** = NULL); |
2243 | extern bool vect_is_simple_use (tree, vec_info *, enum vect_def_type *, |
2244 | tree *, stmt_vec_info * = NULL, |
2245 | gimple ** = NULL); |
2246 | extern bool vect_is_simple_use (vec_info *, stmt_vec_info, slp_tree, |
2247 | unsigned, tree *, slp_tree *, |
2248 | enum vect_def_type *, |
2249 | tree *, stmt_vec_info * = NULL); |
2250 | extern bool vect_maybe_update_slp_op_vectype (slp_tree, tree); |
2251 | extern tree perm_mask_for_reverse (tree); |
2252 | extern bool supportable_widening_operation (vec_info*, code_helper, |
2253 | stmt_vec_info, tree, tree, |
2254 | code_helper*, code_helper*, |
2255 | int*, vec<tree> *); |
2256 | extern bool supportable_narrowing_operation (code_helper, tree, tree, |
2257 | code_helper *, int *, |
2258 | vec<tree> *); |
2259 | |
2260 | extern unsigned record_stmt_cost (stmt_vector_for_cost *, int, |
2261 | enum vect_cost_for_stmt, stmt_vec_info, |
2262 | tree, int, enum vect_cost_model_location); |
2263 | extern unsigned record_stmt_cost (stmt_vector_for_cost *, int, |
2264 | enum vect_cost_for_stmt, slp_tree, |
2265 | tree, int, enum vect_cost_model_location); |
2266 | extern unsigned record_stmt_cost (stmt_vector_for_cost *, int, |
2267 | enum vect_cost_for_stmt, |
2268 | enum vect_cost_model_location); |
2269 | |
2270 | /* Overload of record_stmt_cost with VECTYPE derived from STMT_INFO. */ |
2271 | |
2272 | inline unsigned |
2273 | record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, |
2274 | enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, |
2275 | int misalign, enum vect_cost_model_location where) |
2276 | { |
2277 | return record_stmt_cost (body_cost_vec, count, kind, stmt_info, |
2278 | STMT_VINFO_VECTYPE (stmt_info), misalign, where); |
2279 | } |
2280 | |
2281 | extern void vect_finish_replace_stmt (vec_info *, stmt_vec_info, gimple *); |
2282 | extern void vect_finish_stmt_generation (vec_info *, stmt_vec_info, gimple *, |
2283 | gimple_stmt_iterator *); |
2284 | extern opt_result vect_mark_stmts_to_be_vectorized (loop_vec_info, bool *); |
2285 | extern tree vect_get_store_rhs (stmt_vec_info); |
2286 | void vect_get_vec_defs_for_operand (vec_info *vinfo, stmt_vec_info, unsigned, |
2287 | tree op, vec<tree> *, tree = NULL); |
2288 | void vect_get_vec_defs (vec_info *, stmt_vec_info, slp_tree, unsigned, |
2289 | tree, vec<tree> *, |
2290 | tree = NULL, vec<tree> * = NULL, |
2291 | tree = NULL, vec<tree> * = NULL, |
2292 | tree = NULL, vec<tree> * = NULL); |
2293 | void vect_get_vec_defs (vec_info *, stmt_vec_info, slp_tree, unsigned, |
2294 | tree, tree, vec<tree> *, |
2295 | tree = NULL, tree = NULL, vec<tree> * = NULL, |
2296 | tree = NULL, tree = NULL, vec<tree> * = NULL, |
2297 | tree = NULL, tree = NULL, vec<tree> * = NULL); |
2298 | extern tree vect_init_vector (vec_info *, stmt_vec_info, tree, tree, |
2299 | gimple_stmt_iterator *); |
2300 | extern tree vect_get_slp_vect_def (slp_tree, unsigned); |
2301 | extern bool vect_transform_stmt (vec_info *, stmt_vec_info, |
2302 | gimple_stmt_iterator *, |
2303 | slp_tree, slp_instance); |
2304 | extern void vect_remove_stores (vec_info *, stmt_vec_info); |
2305 | extern bool vect_nop_conversion_p (stmt_vec_info); |
2306 | extern opt_result vect_analyze_stmt (vec_info *, stmt_vec_info, bool *, |
2307 | slp_tree, |
2308 | slp_instance, stmt_vector_for_cost *); |
2309 | extern void vect_get_load_cost (vec_info *, stmt_vec_info, int, |
2310 | dr_alignment_support, int, bool, |
2311 | unsigned int *, unsigned int *, |
2312 | stmt_vector_for_cost *, |
2313 | stmt_vector_for_cost *, bool); |
2314 | extern void vect_get_store_cost (vec_info *, stmt_vec_info, int, |
2315 | dr_alignment_support, int, |
2316 | unsigned int *, stmt_vector_for_cost *); |
2317 | extern bool vect_supportable_shift (vec_info *, enum tree_code, tree); |
2318 | extern tree vect_gen_perm_mask_any (tree, const vec_perm_indices &); |
2319 | extern tree vect_gen_perm_mask_checked (tree, const vec_perm_indices &); |
2320 | extern void optimize_mask_stores (class loop*); |
2321 | extern tree vect_gen_while (gimple_seq *, tree, tree, tree, |
2322 | const char * = nullptr); |
2323 | extern tree vect_gen_while_not (gimple_seq *, tree, tree, tree); |
2324 | extern opt_result vect_get_vector_types_for_stmt (vec_info *, |
2325 | stmt_vec_info, tree *, |
2326 | tree *, unsigned int = 0); |
2327 | extern opt_tree vect_get_mask_type_for_stmt (stmt_vec_info, unsigned int = 0); |
2328 | |
2329 | /* In tree-if-conv.cc. */ |
2330 | extern bool ref_within_array_bound (gimple *, tree); |
2331 | |
2332 | /* In tree-vect-data-refs.cc. */ |
2333 | extern bool vect_can_force_dr_alignment_p (const_tree, poly_uint64); |
2334 | extern enum dr_alignment_support vect_supportable_dr_alignment |
2335 | (vec_info *, dr_vec_info *, tree, int); |
2336 | extern tree vect_get_smallest_scalar_type (stmt_vec_info, tree); |
2337 | extern opt_result vect_analyze_data_ref_dependences (loop_vec_info, unsigned int *); |
2338 | extern bool vect_slp_analyze_instance_dependence (vec_info *, slp_instance); |
2339 | extern opt_result vect_enhance_data_refs_alignment (loop_vec_info); |
2340 | extern opt_result vect_analyze_data_refs_alignment (loop_vec_info); |
2341 | extern bool vect_slp_analyze_instance_alignment (vec_info *, slp_instance); |
2342 | extern opt_result vect_analyze_data_ref_accesses (vec_info *, vec<int> *); |
2343 | extern opt_result vect_prune_runtime_alias_test_list (loop_vec_info); |
2344 | extern bool vect_gather_scatter_fn_p (vec_info *, bool, bool, tree, tree, |
2345 | tree, int, internal_fn *, tree *); |
2346 | extern bool vect_check_gather_scatter (stmt_vec_info, loop_vec_info, |
2347 | gather_scatter_info *); |
2348 | extern opt_result vect_find_stmt_data_reference (loop_p, gimple *, |
2349 | vec<data_reference_p> *, |
2350 | vec<int> *, int); |
2351 | extern opt_result vect_analyze_data_refs (vec_info *, poly_uint64 *, bool *); |
2352 | extern void vect_record_base_alignments (vec_info *); |
2353 | extern tree vect_create_data_ref_ptr (vec_info *, |
2354 | stmt_vec_info, tree, class loop *, tree, |
2355 | tree *, gimple_stmt_iterator *, |
2356 | gimple **, bool, |
2357 | tree = NULL_TREE); |
2358 | extern tree bump_vector_ptr (vec_info *, tree, gimple *, gimple_stmt_iterator *, |
2359 | stmt_vec_info, tree); |
2360 | extern void vect_copy_ref_info (tree, tree); |
2361 | extern tree vect_create_destination_var (tree, tree); |
2362 | extern bool vect_grouped_store_supported (tree, unsigned HOST_WIDE_INT); |
2363 | extern internal_fn vect_store_lanes_supported (tree, unsigned HOST_WIDE_INT, bool); |
2364 | extern bool vect_grouped_load_supported (tree, bool, unsigned HOST_WIDE_INT); |
2365 | extern internal_fn vect_load_lanes_supported (tree, unsigned HOST_WIDE_INT, bool); |
2366 | extern void vect_permute_store_chain (vec_info *, vec<tree> &, |
2367 | unsigned int, stmt_vec_info, |
2368 | gimple_stmt_iterator *, vec<tree> *); |
2369 | extern tree vect_setup_realignment (vec_info *, |
2370 | stmt_vec_info, gimple_stmt_iterator *, |
2371 | tree *, enum dr_alignment_support, tree, |
2372 | class loop **); |
2373 | extern void vect_transform_grouped_load (vec_info *, stmt_vec_info, vec<tree>, |
2374 | int, gimple_stmt_iterator *); |
2375 | extern void vect_record_grouped_load_vectors (vec_info *, |
2376 | stmt_vec_info, vec<tree>); |
2377 | extern tree vect_get_new_vect_var (tree, enum vect_var_kind, const char *); |
2378 | extern tree vect_get_new_ssa_name (tree, enum vect_var_kind, |
2379 | const char * = NULL); |
2380 | extern tree vect_create_addr_base_for_vector_ref (vec_info *, |
2381 | stmt_vec_info, gimple_seq *, |
2382 | tree); |
2383 | |
2384 | /* In tree-vect-loop.cc. */ |
2385 | extern tree neutral_op_for_reduction (tree, code_helper, tree, bool = true); |
2386 | extern widest_int vect_iv_limit_for_partial_vectors (loop_vec_info loop_vinfo); |
2387 | bool vect_rgroup_iv_might_wrap_p (loop_vec_info, rgroup_controls *); |
2388 | /* Used in tree-vect-loop-manip.cc */ |
2389 | extern opt_result vect_determine_partial_vectors_and_peeling (loop_vec_info); |
2390 | /* Used in gimple-loop-interchange.c and tree-parloops.cc. */ |
2391 | extern bool check_reduction_path (dump_user_location_t, loop_p, gphi *, tree, |
2392 | enum tree_code); |
2393 | extern bool needs_fold_left_reduction_p (tree, code_helper); |
2394 | /* Drive for loop analysis stage. */ |
2395 | extern opt_loop_vec_info vect_analyze_loop (class loop *, vec_info_shared *); |
2396 | extern tree vect_build_loop_niters (loop_vec_info, bool * = NULL); |
2397 | extern void vect_gen_vector_loop_niters (loop_vec_info, tree, tree *, |
2398 | tree *, bool); |
2399 | extern tree vect_halve_mask_nunits (tree, machine_mode); |
2400 | extern tree vect_double_mask_nunits (tree, machine_mode); |
2401 | extern void vect_record_loop_mask (loop_vec_info, vec_loop_masks *, |
2402 | unsigned int, tree, tree); |
2403 | extern tree vect_get_loop_mask (loop_vec_info, gimple_stmt_iterator *, |
2404 | vec_loop_masks *, |
2405 | unsigned int, tree, unsigned int); |
2406 | extern void vect_record_loop_len (loop_vec_info, vec_loop_lens *, unsigned int, |
2407 | tree, unsigned int); |
2408 | extern tree vect_get_loop_len (loop_vec_info, gimple_stmt_iterator *, |
2409 | vec_loop_lens *, unsigned int, tree, |
2410 | unsigned int, unsigned int); |
2411 | extern gimple_seq vect_gen_len (tree, tree, tree, tree); |
2412 | extern stmt_vec_info info_for_reduction (vec_info *, stmt_vec_info); |
2413 | extern bool reduction_fn_for_scalar_code (code_helper, internal_fn *); |
2414 | |
2415 | /* Drive for loop transformation stage. */ |
2416 | extern class loop *vect_transform_loop (loop_vec_info, gimple *); |
2417 | struct vect_loop_form_info |
2418 | { |
2419 | tree number_of_iterations; |
2420 | tree number_of_iterationsm1; |
2421 | tree assumptions; |
2422 | auto_vec<gcond *> conds; |
2423 | gcond *inner_loop_cond; |
2424 | edge loop_exit; |
2425 | }; |
2426 | extern opt_result vect_analyze_loop_form (class loop *, vect_loop_form_info *); |
2427 | extern loop_vec_info vect_create_loop_vinfo (class loop *, vec_info_shared *, |
2428 | const vect_loop_form_info *, |
2429 | loop_vec_info = nullptr); |
2430 | extern bool vectorizable_live_operation (vec_info *, stmt_vec_info, |
2431 | slp_tree, slp_instance, int, |
2432 | bool, stmt_vector_for_cost *); |
2433 | extern bool vectorizable_reduction (loop_vec_info, stmt_vec_info, |
2434 | slp_tree, slp_instance, |
2435 | stmt_vector_for_cost *); |
2436 | extern bool vectorizable_induction (loop_vec_info, stmt_vec_info, |
2437 | gimple **, slp_tree, |
2438 | stmt_vector_for_cost *); |
2439 | extern bool vect_transform_reduction (loop_vec_info, stmt_vec_info, |
2440 | gimple_stmt_iterator *, |
2441 | gimple **, slp_tree); |
2442 | extern bool vect_transform_cycle_phi (loop_vec_info, stmt_vec_info, |
2443 | gimple **, |
2444 | slp_tree, slp_instance); |
2445 | extern bool vectorizable_lc_phi (loop_vec_info, stmt_vec_info, |
2446 | gimple **, slp_tree); |
2447 | extern bool vectorizable_phi (vec_info *, stmt_vec_info, gimple **, slp_tree, |
2448 | stmt_vector_for_cost *); |
2449 | extern bool vectorizable_recurr (loop_vec_info, stmt_vec_info, |
2450 | gimple **, slp_tree, stmt_vector_for_cost *); |
2451 | extern bool vect_emulated_vector_p (tree); |
2452 | extern bool vect_can_vectorize_without_simd_p (tree_code); |
2453 | extern bool vect_can_vectorize_without_simd_p (code_helper); |
2454 | extern int vect_get_known_peeling_cost (loop_vec_info, int, int *, |
2455 | stmt_vector_for_cost *, |
2456 | stmt_vector_for_cost *, |
2457 | stmt_vector_for_cost *); |
2458 | extern tree cse_and_gimplify_to_preheader (loop_vec_info, tree); |
2459 | |
2460 | /* Nonlinear induction. */ |
2461 | extern tree vect_peel_nonlinear_iv_init (gimple_seq*, tree, tree, |
2462 | tree, enum vect_induction_op_type); |
2463 | |
2464 | /* In tree-vect-slp.cc. */ |
2465 | extern void vect_slp_init (void); |
2466 | extern void vect_slp_fini (void); |
2467 | extern void vect_free_slp_instance (slp_instance); |
2468 | extern bool vect_transform_slp_perm_load (vec_info *, slp_tree, const vec<tree> &, |
2469 | gimple_stmt_iterator *, poly_uint64, |
2470 | bool, unsigned *, |
2471 | unsigned * = nullptr, bool = false); |
2472 | extern bool vect_slp_analyze_operations (vec_info *); |
2473 | extern void vect_schedule_slp (vec_info *, const vec<slp_instance> &); |
2474 | extern opt_result vect_analyze_slp (vec_info *, unsigned); |
2475 | extern bool vect_make_slp_decision (loop_vec_info); |
2476 | extern void vect_detect_hybrid_slp (loop_vec_info); |
2477 | extern void vect_optimize_slp (vec_info *); |
2478 | extern void vect_gather_slp_loads (vec_info *); |
2479 | extern void vect_get_slp_defs (slp_tree, vec<tree> *); |
2480 | extern void vect_get_slp_defs (vec_info *, slp_tree, vec<vec<tree> > *, |
2481 | unsigned n = -1U); |
2482 | extern bool vect_slp_if_converted_bb (basic_block bb, loop_p orig_loop); |
2483 | extern bool vect_slp_function (function *); |
2484 | extern stmt_vec_info vect_find_last_scalar_stmt_in_slp (slp_tree); |
2485 | extern stmt_vec_info vect_find_first_scalar_stmt_in_slp (slp_tree); |
2486 | extern bool is_simple_and_all_uses_invariant (stmt_vec_info, loop_vec_info); |
2487 | extern bool can_duplicate_and_interleave_p (vec_info *, unsigned int, tree, |
2488 | unsigned int * = NULL, |
2489 | tree * = NULL, tree * = NULL); |
2490 | extern void duplicate_and_interleave (vec_info *, gimple_seq *, tree, |
2491 | const vec<tree> &, unsigned int, vec<tree> &); |
2492 | extern int vect_get_place_in_interleaving_chain (stmt_vec_info, stmt_vec_info); |
2493 | extern slp_tree vect_create_new_slp_node (unsigned, tree_code); |
2494 | extern void vect_free_slp_tree (slp_tree); |
2495 | extern bool compatible_calls_p (gcall *, gcall *); |
2496 | extern int vect_slp_child_index_for_operand (const gimple *, int op, bool); |
2497 | |
2498 | /* In tree-vect-patterns.cc. */ |
2499 | extern void |
2500 | vect_mark_pattern_stmts (vec_info *, stmt_vec_info, gimple *, tree); |
2501 | extern bool vect_get_range_info (tree, wide_int*, wide_int*); |
2502 | |
2503 | /* Pattern recognition functions. |
2504 | Additional pattern recognition functions can (and will) be added |
2505 | in the future. */ |
2506 | void vect_pattern_recog (vec_info *); |
2507 | |
2508 | /* In tree-vectorizer.cc. */ |
2509 | unsigned vectorize_loops (void); |
2510 | void vect_free_loop_info_assumptions (class loop *); |
2511 | gimple *vect_loop_vectorized_call (class loop *, gcond **cond = NULL); |
2512 | bool vect_stmt_dominates_stmt_p (gimple *, gimple *); |
2513 | |
2514 | /* SLP Pattern matcher types, tree-vect-slp-patterns.cc. */ |
2515 | |
2516 | /* Forward declaration of possible two operands operation that can be matched |
2517 | by the complex numbers pattern matchers. */ |
2518 | enum _complex_operation : unsigned; |
2519 | |
2520 | /* All possible load permute values that could result from the partial data-flow |
2521 | analysis. */ |
2522 | typedef enum _complex_perm_kinds { |
2523 | PERM_UNKNOWN, |
2524 | PERM_EVENODD, |
2525 | PERM_ODDEVEN, |
2526 | PERM_ODDODD, |
2527 | PERM_EVENEVEN, |
2528 | /* Can be combined with any other PERM values. */ |
2529 | PERM_TOP |
2530 | } complex_perm_kinds_t; |
2531 | |
2532 | /* Cache from nodes to the load permutation they represent. */ |
2533 | typedef hash_map <slp_tree, complex_perm_kinds_t> |
2534 | slp_tree_to_load_perm_map_t; |
2535 | |
2536 | /* Cache from nodes pair to being compatible or not. */ |
2537 | typedef pair_hash <nofree_ptr_hash <_slp_tree>, |
2538 | nofree_ptr_hash <_slp_tree>> slp_node_hash; |
2539 | typedef hash_map <slp_node_hash, bool> slp_compat_nodes_map_t; |
2540 | |
2541 | |
2542 | /* Vector pattern matcher base class. All SLP pattern matchers must inherit |
2543 | from this type. */ |
2544 | |
2545 | class vect_pattern |
2546 | { |
2547 | protected: |
2548 | /* The number of arguments that the IFN requires. */ |
2549 | unsigned m_num_args; |
2550 | |
2551 | /* The internal function that will be used when a pattern is created. */ |
2552 | internal_fn m_ifn; |
2553 | |
2554 | /* The current node being inspected. */ |
2555 | slp_tree *m_node; |
2556 | |
2557 | /* The list of operands to be the children for the node produced when the |
2558 | internal function is created. */ |
2559 | vec<slp_tree> m_ops; |
2560 | |
2561 | /* Default constructor where NODE is the root of the tree to inspect. */ |
2562 | vect_pattern (slp_tree *node, vec<slp_tree> *m_ops, internal_fn ifn) |
2563 | { |
2564 | this->m_ifn = ifn; |
2565 | this->m_node = node; |
2566 | this->m_ops.create (nelems: 0); |
2567 | if (m_ops) |
2568 | this->m_ops.safe_splice (src: *m_ops); |
2569 | } |
2570 | |
2571 | public: |
2572 | |
2573 | /* Create a new instance of the pattern matcher class of the given type. */ |
2574 | static vect_pattern* recognize (slp_tree_to_load_perm_map_t *, |
2575 | slp_compat_nodes_map_t *, slp_tree *); |
2576 | |
2577 | /* Build the pattern from the data collected so far. */ |
2578 | virtual void build (vec_info *) = 0; |
2579 | |
2580 | /* Default destructor. */ |
2581 | virtual ~vect_pattern () |
2582 | { |
2583 | this->m_ops.release (); |
2584 | } |
2585 | }; |
2586 | |
2587 | /* Function pointer to create a new pattern matcher from a generic type. */ |
2588 | typedef vect_pattern* (*vect_pattern_decl_t) (slp_tree_to_load_perm_map_t *, |
2589 | slp_compat_nodes_map_t *, |
2590 | slp_tree *); |
2591 | |
2592 | /* List of supported pattern matchers. */ |
2593 | extern vect_pattern_decl_t slp_patterns[]; |
2594 | |
2595 | /* Number of supported pattern matchers. */ |
2596 | extern size_t num__slp_patterns; |
2597 | |
2598 | /* ---------------------------------------------------------------------- |
2599 | Target support routines |
2600 | ----------------------------------------------------------------------- |
2601 | The following routines are provided to simplify costing decisions in |
2602 | target code. Please add more as needed. */ |
2603 | |
2604 | /* Return true if an operaton of kind KIND for STMT_INFO represents |
2605 | the extraction of an element from a vector in preparation for |
2606 | storing the element to memory. */ |
2607 | inline bool |
2608 | (vect_cost_for_stmt kind, stmt_vec_info stmt_info) |
2609 | { |
2610 | return (kind == vec_to_scalar |
2611 | && STMT_VINFO_DATA_REF (stmt_info) |
2612 | && DR_IS_WRITE (STMT_VINFO_DATA_REF (stmt_info))); |
2613 | } |
2614 | |
2615 | /* Return true if STMT_INFO represents part of a reduction. */ |
2616 | inline bool |
2617 | vect_is_reduction (stmt_vec_info stmt_info) |
2618 | { |
2619 | return STMT_VINFO_REDUC_IDX (stmt_info) >= 0; |
2620 | } |
2621 | |
2622 | /* If STMT_INFO describes a reduction, return the vect_reduction_type |
2623 | of the reduction it describes, otherwise return -1. */ |
2624 | inline int |
2625 | vect_reduc_type (vec_info *vinfo, stmt_vec_info stmt_info) |
2626 | { |
2627 | if (loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (p: vinfo)) |
2628 | if (STMT_VINFO_REDUC_DEF (stmt_info)) |
2629 | { |
2630 | stmt_vec_info reduc_info = info_for_reduction (loop_vinfo, stmt_info); |
2631 | return int (STMT_VINFO_REDUC_TYPE (reduc_info)); |
2632 | } |
2633 | return -1; |
2634 | } |
2635 | |
2636 | /* If STMT_INFO is a COND_EXPR that includes an embedded comparison, return the |
2637 | scalar type of the values being compared. Return null otherwise. */ |
2638 | inline tree |
2639 | vect_embedded_comparison_type (stmt_vec_info stmt_info) |
2640 | { |
2641 | if (auto *assign = dyn_cast<gassign *> (p: stmt_info->stmt)) |
2642 | if (gimple_assign_rhs_code (gs: assign) == COND_EXPR) |
2643 | { |
2644 | tree cond = gimple_assign_rhs1 (gs: assign); |
2645 | if (COMPARISON_CLASS_P (cond)) |
2646 | return TREE_TYPE (TREE_OPERAND (cond, 0)); |
2647 | } |
2648 | return NULL_TREE; |
2649 | } |
2650 | |
2651 | /* If STMT_INFO is a comparison or contains an embedded comparison, return the |
2652 | scalar type of the values being compared. Return null otherwise. */ |
2653 | inline tree |
2654 | vect_comparison_type (stmt_vec_info stmt_info) |
2655 | { |
2656 | if (auto *assign = dyn_cast<gassign *> (p: stmt_info->stmt)) |
2657 | if (TREE_CODE_CLASS (gimple_assign_rhs_code (assign)) == tcc_comparison) |
2658 | return TREE_TYPE (gimple_assign_rhs1 (assign)); |
2659 | return vect_embedded_comparison_type (stmt_info); |
2660 | } |
2661 | |
2662 | /* Return true if STMT_INFO extends the result of a load. */ |
2663 | inline bool |
2664 | vect_is_extending_load (class vec_info *vinfo, stmt_vec_info stmt_info) |
2665 | { |
2666 | /* Although this is quite large for an inline function, this part |
2667 | at least should be inline. */ |
2668 | gassign *assign = dyn_cast <gassign *> (p: stmt_info->stmt); |
2669 | if (!assign || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (assign))) |
2670 | return false; |
2671 | |
2672 | tree rhs = gimple_assign_rhs1 (gs: stmt_info->stmt); |
2673 | tree lhs_type = TREE_TYPE (gimple_assign_lhs (assign)); |
2674 | tree rhs_type = TREE_TYPE (rhs); |
2675 | if (!INTEGRAL_TYPE_P (lhs_type) |
2676 | || !INTEGRAL_TYPE_P (rhs_type) |
2677 | || TYPE_PRECISION (lhs_type) <= TYPE_PRECISION (rhs_type)) |
2678 | return false; |
2679 | |
2680 | stmt_vec_info def_stmt_info = vinfo->lookup_def (rhs); |
2681 | return (def_stmt_info |
2682 | && STMT_VINFO_DATA_REF (def_stmt_info) |
2683 | && DR_IS_READ (STMT_VINFO_DATA_REF (def_stmt_info))); |
2684 | } |
2685 | |
2686 | /* Return true if STMT_INFO is an integer truncation. */ |
2687 | inline bool |
2688 | vect_is_integer_truncation (stmt_vec_info stmt_info) |
2689 | { |
2690 | gassign *assign = dyn_cast <gassign *> (p: stmt_info->stmt); |
2691 | if (!assign || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (assign))) |
2692 | return false; |
2693 | |
2694 | tree lhs_type = TREE_TYPE (gimple_assign_lhs (assign)); |
2695 | tree rhs_type = TREE_TYPE (gimple_assign_rhs1 (assign)); |
2696 | return (INTEGRAL_TYPE_P (lhs_type) |
2697 | && INTEGRAL_TYPE_P (rhs_type) |
2698 | && TYPE_PRECISION (lhs_type) < TYPE_PRECISION (rhs_type)); |
2699 | } |
2700 | |
2701 | /* Build a GIMPLE_ASSIGN or GIMPLE_CALL with the tree_code, |
2702 | or internal_fn contained in ch, respectively. */ |
2703 | gimple * vect_gimple_build (tree, code_helper, tree, tree = NULL_TREE); |
2704 | #endif /* GCC_TREE_VECTORIZER_H */ |
2705 | |