Branch data Line data Source code
1 : : /* Vectorizer
2 : : Copyright (C) 2003-2025 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 : : #include "dominance.h"
34 : :
35 : : /* Used for naming of new temporaries. */
36 : : enum vect_var_kind {
37 : : vect_simple_var,
38 : : vect_pointer_var,
39 : : vect_scalar_var,
40 : : vect_mask_var
41 : : };
42 : :
43 : : /* Defines type of operation. */
44 : : enum operation_type {
45 : : unary_op = 1,
46 : : binary_op,
47 : : ternary_op
48 : : };
49 : :
50 : : /* Define type of available alignment support. */
51 : : enum dr_alignment_support {
52 : : dr_unaligned_unsupported,
53 : : dr_unaligned_supported,
54 : : dr_explicit_realign,
55 : : dr_explicit_realign_optimized,
56 : : dr_aligned
57 : : };
58 : :
59 : : /* Define type of def-use cross-iteration cycle. */
60 : : enum vect_def_type {
61 : : vect_uninitialized_def = 0,
62 : : vect_constant_def = 1,
63 : : vect_external_def,
64 : : vect_internal_def,
65 : : vect_induction_def,
66 : : vect_reduction_def,
67 : : vect_double_reduction_def,
68 : : vect_nested_cycle,
69 : : vect_first_order_recurrence,
70 : : vect_condition_def,
71 : : vect_unknown_def_type
72 : : };
73 : :
74 : : /* Define operation type of linear/non-linear induction variable. */
75 : : enum vect_induction_op_type {
76 : : vect_step_op_add = 0,
77 : : vect_step_op_neg,
78 : : vect_step_op_mul,
79 : : vect_step_op_shl,
80 : : vect_step_op_shr
81 : : };
82 : :
83 : : /* Define type of reduction. */
84 : : enum vect_reduction_type {
85 : : TREE_CODE_REDUCTION,
86 : : COND_REDUCTION,
87 : : INTEGER_INDUC_COND_REDUCTION,
88 : : CONST_COND_REDUCTION,
89 : :
90 : : /* Retain a scalar phi and use a FOLD_EXTRACT_LAST within the loop
91 : : to implement:
92 : :
93 : : for (int i = 0; i < VF; ++i)
94 : : res = cond[i] ? val[i] : res; */
95 : : EXTRACT_LAST_REDUCTION,
96 : :
97 : : /* Use a folding reduction within the loop to implement:
98 : :
99 : : for (int i = 0; i < VF; ++i)
100 : : res = res OP val[i];
101 : :
102 : : (with no reassocation). */
103 : : FOLD_LEFT_REDUCTION
104 : : };
105 : :
106 : : #define VECTORIZABLE_CYCLE_DEF(D) (((D) == vect_reduction_def) \
107 : : || ((D) == vect_double_reduction_def) \
108 : : || ((D) == vect_nested_cycle))
109 : :
110 : : /* Structure to encapsulate information about a group of like
111 : : instructions to be presented to the target cost model. */
112 : : struct stmt_info_for_cost {
113 : : int count;
114 : : enum vect_cost_for_stmt kind;
115 : : enum vect_cost_model_location where;
116 : : stmt_vec_info stmt_info;
117 : : slp_tree node;
118 : : tree vectype;
119 : : int misalign;
120 : : };
121 : :
122 : : typedef vec<stmt_info_for_cost> stmt_vector_for_cost;
123 : :
124 : : /* Maps base addresses to an innermost_loop_behavior and the stmt it was
125 : : derived from that gives the maximum known alignment for that base. */
126 : : typedef hash_map<tree_operand_hash,
127 : : std::pair<stmt_vec_info, innermost_loop_behavior *> >
128 : : vec_base_alignments;
129 : :
130 : : /* Represents elements [START, START + LENGTH) of cyclical array OPS*
131 : : (i.e. OPS repeated to give at least START + LENGTH elements) */
132 : : struct vect_scalar_ops_slice
133 : : {
134 : : tree op (unsigned int i) const;
135 : : bool all_same_p () const;
136 : :
137 : : vec<tree> *ops;
138 : : unsigned int start;
139 : : unsigned int length;
140 : : };
141 : :
142 : : /* Return element I of the slice. */
143 : : inline tree
144 : 3217674 : vect_scalar_ops_slice::op (unsigned int i) const
145 : : {
146 : 6435348 : return (*ops)[(i + start) % ops->length ()];
147 : : }
148 : :
149 : : /* Hash traits for vect_scalar_ops_slice. */
150 : : struct vect_scalar_ops_slice_hash : typed_noop_remove<vect_scalar_ops_slice>
151 : : {
152 : : typedef vect_scalar_ops_slice value_type;
153 : : typedef vect_scalar_ops_slice compare_type;
154 : :
155 : : static const bool empty_zero_p = true;
156 : :
157 : : static void mark_deleted (value_type &s) { s.length = ~0U; }
158 : 0 : static void mark_empty (value_type &s) { s.length = 0; }
159 : 725595 : static bool is_deleted (const value_type &s) { return s.length == ~0U; }
160 : 7705392 : static bool is_empty (const value_type &s) { return s.length == 0; }
161 : : static hashval_t hash (const value_type &);
162 : : static bool equal (const value_type &, const compare_type &);
163 : : };
164 : :
165 : : /* Describes how we're going to vectorize an individual load or store,
166 : : or a group of loads or stores. */
167 : : enum vect_memory_access_type {
168 : : /* An access to an invariant address. This is used only for loads. */
169 : : VMAT_INVARIANT,
170 : :
171 : : /* A simple contiguous access. */
172 : : VMAT_CONTIGUOUS,
173 : :
174 : : /* A contiguous access that goes down in memory rather than up,
175 : : with no additional permutation. This is used only for stores
176 : : of invariants. */
177 : : VMAT_CONTIGUOUS_DOWN,
178 : :
179 : : /* A simple contiguous access in which the elements need to be permuted
180 : : after loading or before storing. Only used for loop vectorization;
181 : : SLP uses separate permutes. */
182 : : VMAT_CONTIGUOUS_PERMUTE,
183 : :
184 : : /* A simple contiguous access in which the elements need to be reversed
185 : : after loading or before storing. */
186 : : VMAT_CONTIGUOUS_REVERSE,
187 : :
188 : : /* An access that uses IFN_LOAD_LANES or IFN_STORE_LANES. */
189 : : VMAT_LOAD_STORE_LANES,
190 : :
191 : : /* An access in which each scalar element is loaded or stored
192 : : individually. */
193 : : VMAT_ELEMENTWISE,
194 : :
195 : : /* A hybrid of VMAT_CONTIGUOUS and VMAT_ELEMENTWISE, used for grouped
196 : : SLP accesses. Each unrolled iteration uses a contiguous load
197 : : or store for the whole group, but the groups from separate iterations
198 : : are combined in the same way as for VMAT_ELEMENTWISE. */
199 : : VMAT_STRIDED_SLP,
200 : :
201 : : /* The access uses gather loads or scatter stores. */
202 : : VMAT_GATHER_SCATTER
203 : : };
204 : :
205 : : /************************************************************************
206 : : SLP
207 : : ************************************************************************/
208 : : typedef vec<std::pair<unsigned, unsigned> > lane_permutation_t;
209 : : typedef auto_vec<std::pair<unsigned, unsigned>, 16> auto_lane_permutation_t;
210 : : typedef vec<unsigned> load_permutation_t;
211 : : typedef auto_vec<unsigned, 16> auto_load_permutation_t;
212 : :
213 : : /* A computation tree of an SLP instance. Each node corresponds to a group of
214 : : stmts to be packed in a SIMD stmt. */
215 : : struct _slp_tree {
216 : : _slp_tree ();
217 : : ~_slp_tree ();
218 : :
219 : : void push_vec_def (gimple *def);
220 : 7693 : void push_vec_def (tree def) { vec_defs.quick_push (def); }
221 : :
222 : : /* Nodes that contain def-stmts of this node statements operands. */
223 : : vec<slp_tree> children;
224 : :
225 : : /* A group of scalar stmts to be vectorized together. */
226 : : vec<stmt_vec_info> stmts;
227 : : /* A group of scalar operands to be vectorized together. */
228 : : vec<tree> ops;
229 : : /* The representative that should be used for analysis and
230 : : code generation. */
231 : : stmt_vec_info representative;
232 : :
233 : : /* Load permutation relative to the stores, NULL if there is no
234 : : permutation. */
235 : : load_permutation_t load_permutation;
236 : : /* Lane permutation of the operands scalar lanes encoded as pairs
237 : : of { operand number, lane number }. The number of elements
238 : : denotes the number of output lanes. */
239 : : lane_permutation_t lane_permutation;
240 : :
241 : : /* Selected SIMD clone's function info. First vector element
242 : : is SIMD clone's function decl, followed by a pair of trees (base + step)
243 : : for linear arguments (pair of NULLs for other arguments). */
244 : : vec<tree> simd_clone_info;
245 : :
246 : : tree vectype;
247 : : /* Vectorized defs. */
248 : : vec<tree> vec_defs;
249 : : /* Number of vector stmts that are created to replace the group of scalar
250 : : stmts. It is calculated during the transformation phase as the number of
251 : : scalar elements in one scalar iteration (GROUP_SIZE) multiplied by VF
252 : : divided by vector size. */
253 : : unsigned int vec_stmts_size;
254 : :
255 : : /* Reference count in the SLP graph. */
256 : : unsigned int refcnt;
257 : : /* The maximum number of vector elements for the subtree rooted
258 : : at this node. */
259 : : poly_uint64 max_nunits;
260 : : /* The DEF type of this node. */
261 : : enum vect_def_type def_type;
262 : : /* The number of scalar lanes produced by this node. */
263 : : unsigned int lanes;
264 : : /* The operation of this node. */
265 : : enum tree_code code;
266 : : /* Whether uses of this load or feeders of this store are suitable
267 : : for load/store-lanes. */
268 : : bool ldst_lanes;
269 : : /* For BB vect, flag to indicate this load node should be vectorized
270 : : as to avoid STLF fails because of related stores. */
271 : : bool avoid_stlf_fail;
272 : :
273 : : int vertex;
274 : :
275 : : /* Classifies how the load or store is going to be implemented
276 : : for loop vectorization. */
277 : : vect_memory_access_type memory_access_type;
278 : :
279 : : /* If not NULL this is a cached failed SLP discovery attempt with
280 : : the lanes that failed during SLP discovery as 'false'. This is
281 : : a copy of the matches array. */
282 : : bool *failed;
283 : :
284 : : /* Allocate from slp_tree_pool. */
285 : : static void *operator new (size_t);
286 : :
287 : : /* Return memory to slp_tree_pool. */
288 : : static void operator delete (void *, size_t);
289 : :
290 : : /* Linked list of nodes to release when we free the slp_tree_pool. */
291 : : slp_tree next_node;
292 : : slp_tree prev_node;
293 : : };
294 : :
295 : : /* The enum describes the type of operations that an SLP instance
296 : : can perform. */
297 : :
298 : : enum slp_instance_kind {
299 : : slp_inst_kind_store,
300 : : slp_inst_kind_reduc_group,
301 : : slp_inst_kind_reduc_chain,
302 : : slp_inst_kind_bb_reduc,
303 : : slp_inst_kind_ctor,
304 : : slp_inst_kind_gcond
305 : : };
306 : :
307 : : /* SLP instance is a sequence of stmts in a loop that can be packed into
308 : : SIMD stmts. */
309 : : typedef class _slp_instance {
310 : : public:
311 : : /* The root of SLP tree. */
312 : : slp_tree root;
313 : :
314 : : /* For vector constructors, the constructor stmt that the SLP tree is built
315 : : from, NULL otherwise. */
316 : : vec<stmt_vec_info> root_stmts;
317 : :
318 : : /* For slp_inst_kind_bb_reduc the defs that were not vectorized, NULL
319 : : otherwise. */
320 : : vec<tree> remain_defs;
321 : :
322 : : /* The group of nodes that contain loads of this SLP instance. */
323 : : vec<slp_tree> loads;
324 : :
325 : : /* The SLP node containing the reduction PHIs. */
326 : : slp_tree reduc_phis;
327 : :
328 : : /* Vector cost of this entry to the SLP graph. */
329 : : stmt_vector_for_cost cost_vec;
330 : :
331 : : /* If this instance is the main entry of a subgraph the set of
332 : : entries into the same subgraph, including itself. */
333 : : vec<_slp_instance *> subgraph_entries;
334 : :
335 : : /* The type of operation the SLP instance is performing. */
336 : : slp_instance_kind kind;
337 : :
338 : : dump_user_location_t location () const;
339 : : } *slp_instance;
340 : :
341 : :
342 : : /* Access Functions. */
343 : : #define SLP_INSTANCE_TREE(S) (S)->root
344 : : #define SLP_INSTANCE_LOADS(S) (S)->loads
345 : : #define SLP_INSTANCE_ROOT_STMTS(S) (S)->root_stmts
346 : : #define SLP_INSTANCE_REMAIN_DEFS(S) (S)->remain_defs
347 : : #define SLP_INSTANCE_KIND(S) (S)->kind
348 : :
349 : : #define SLP_TREE_CHILDREN(S) (S)->children
350 : : #define SLP_TREE_SCALAR_STMTS(S) (S)->stmts
351 : : #define SLP_TREE_SCALAR_OPS(S) (S)->ops
352 : : #define SLP_TREE_REF_COUNT(S) (S)->refcnt
353 : : #define SLP_TREE_VEC_DEFS(S) (S)->vec_defs
354 : : #define SLP_TREE_NUMBER_OF_VEC_STMTS(S) (S)->vec_stmts_size
355 : : #define SLP_TREE_LOAD_PERMUTATION(S) (S)->load_permutation
356 : : #define SLP_TREE_LANE_PERMUTATION(S) (S)->lane_permutation
357 : : #define SLP_TREE_SIMD_CLONE_INFO(S) (S)->simd_clone_info
358 : : #define SLP_TREE_DEF_TYPE(S) (S)->def_type
359 : : #define SLP_TREE_VECTYPE(S) (S)->vectype
360 : : #define SLP_TREE_REPRESENTATIVE(S) (S)->representative
361 : : #define SLP_TREE_LANES(S) (S)->lanes
362 : : #define SLP_TREE_CODE(S) (S)->code
363 : : #define SLP_TREE_MEMORY_ACCESS_TYPE(S) (S)->memory_access_type
364 : :
365 : : enum vect_partial_vector_style {
366 : : vect_partial_vectors_none,
367 : : vect_partial_vectors_while_ult,
368 : : vect_partial_vectors_avx512,
369 : : vect_partial_vectors_len
370 : : };
371 : :
372 : : /* Key for map that records association between
373 : : scalar conditions and corresponding loop mask, and
374 : : is populated by vect_record_loop_mask. */
375 : :
376 : : struct scalar_cond_masked_key
377 : : {
378 : 51091 : scalar_cond_masked_key (tree t, unsigned ncopies_)
379 : 51091 : : ncopies (ncopies_)
380 : : {
381 : 51091 : get_cond_ops_from_tree (t);
382 : : }
383 : :
384 : : void get_cond_ops_from_tree (tree);
385 : :
386 : : unsigned ncopies;
387 : : bool inverted_p;
388 : : tree_code code;
389 : : tree op0;
390 : : tree op1;
391 : : };
392 : :
393 : : template<>
394 : : struct default_hash_traits<scalar_cond_masked_key>
395 : : {
396 : : typedef scalar_cond_masked_key compare_type;
397 : : typedef scalar_cond_masked_key value_type;
398 : :
399 : : static inline hashval_t
400 : 58944 : hash (value_type v)
401 : : {
402 : 58944 : inchash::hash h;
403 : 58944 : h.add_int (v.code);
404 : 58944 : inchash::add_expr (v.op0, h, 0);
405 : 58944 : inchash::add_expr (v.op1, h, 0);
406 : 58944 : h.add_int (v.ncopies);
407 : 58944 : h.add_flag (v.inverted_p);
408 : 58944 : return h.end ();
409 : : }
410 : :
411 : : static inline bool
412 : 8729 : equal (value_type existing, value_type candidate)
413 : : {
414 : 8729 : return (existing.ncopies == candidate.ncopies
415 : 8725 : && existing.code == candidate.code
416 : 5065 : && existing.inverted_p == candidate.inverted_p
417 : 3517 : && operand_equal_p (existing.op0, candidate.op0, 0)
418 : 10669 : && operand_equal_p (existing.op1, candidate.op1, 0));
419 : : }
420 : :
421 : : static const bool empty_zero_p = true;
422 : :
423 : : static inline void
424 : 0 : mark_empty (value_type &v)
425 : : {
426 : 0 : v.ncopies = 0;
427 : 0 : v.inverted_p = false;
428 : : }
429 : :
430 : : static inline bool
431 : 6841316 : is_empty (value_type v)
432 : : {
433 : 6790241 : return v.ncopies == 0;
434 : : }
435 : :
436 : : static inline void mark_deleted (value_type &) {}
437 : :
438 : : static inline bool is_deleted (const value_type &)
439 : : {
440 : : return false;
441 : : }
442 : :
443 : 44301 : static inline void remove (value_type &) {}
444 : : };
445 : :
446 : : typedef hash_set<scalar_cond_masked_key> scalar_cond_masked_set_type;
447 : :
448 : : /* Key and map that records association between vector conditions and
449 : : corresponding loop mask, and is populated by prepare_vec_mask. */
450 : :
451 : : typedef pair_hash<tree_operand_hash, tree_operand_hash> tree_cond_mask_hash;
452 : : typedef hash_set<tree_cond_mask_hash> vec_cond_masked_set_type;
453 : :
454 : : /* Describes two objects whose addresses must be unequal for the vectorized
455 : : loop to be valid. */
456 : : typedef std::pair<tree, tree> vec_object_pair;
457 : :
458 : : /* Records that vectorization is only possible if abs (EXPR) >= MIN_VALUE.
459 : : UNSIGNED_P is true if we can assume that abs (EXPR) == EXPR. */
460 : : class vec_lower_bound {
461 : : public:
462 : : vec_lower_bound () {}
463 : 1444 : vec_lower_bound (tree e, bool u, poly_uint64 m)
464 : 1444 : : expr (e), unsigned_p (u), min_value (m) {}
465 : :
466 : : tree expr;
467 : : bool unsigned_p;
468 : : poly_uint64 min_value;
469 : : };
470 : :
471 : : /* Vectorizer state shared between different analyses like vector sizes
472 : : of the same CFG region. */
473 : : class vec_info_shared {
474 : : public:
475 : : vec_info_shared();
476 : : ~vec_info_shared();
477 : :
478 : : void save_datarefs();
479 : : void check_datarefs();
480 : :
481 : : /* All data references. Freed by free_data_refs, so not an auto_vec. */
482 : : vec<data_reference_p> datarefs;
483 : : vec<data_reference> datarefs_copy;
484 : :
485 : : /* The loop nest in which the data dependences are computed. */
486 : : auto_vec<loop_p> loop_nest;
487 : :
488 : : /* All data dependences. Freed by free_dependence_relations, so not
489 : : an auto_vec. */
490 : : vec<ddr_p> ddrs;
491 : : };
492 : :
493 : : /* Vectorizer state common between loop and basic-block vectorization. */
494 : : class vec_info {
495 : : public:
496 : : typedef hash_set<int_hash<machine_mode, E_VOIDmode, E_BLKmode> > mode_set;
497 : : enum vec_kind { bb, loop };
498 : :
499 : : vec_info (vec_kind, vec_info_shared *);
500 : : ~vec_info ();
501 : :
502 : : stmt_vec_info add_stmt (gimple *);
503 : : stmt_vec_info add_pattern_stmt (gimple *, stmt_vec_info);
504 : : stmt_vec_info resync_stmt_addr (gimple *);
505 : : stmt_vec_info lookup_stmt (gimple *);
506 : : stmt_vec_info lookup_def (tree);
507 : : stmt_vec_info lookup_single_use (tree);
508 : : class dr_vec_info *lookup_dr (data_reference *);
509 : : void move_dr (stmt_vec_info, stmt_vec_info);
510 : : void remove_stmt (stmt_vec_info);
511 : : void replace_stmt (gimple_stmt_iterator *, stmt_vec_info, gimple *);
512 : : void insert_on_entry (stmt_vec_info, gimple *);
513 : : void insert_seq_on_entry (stmt_vec_info, gimple_seq);
514 : :
515 : : /* The type of vectorization. */
516 : : vec_kind kind;
517 : :
518 : : /* Shared vectorizer state. */
519 : : vec_info_shared *shared;
520 : :
521 : : /* The mapping of GIMPLE UID to stmt_vec_info. */
522 : : vec<stmt_vec_info> stmt_vec_infos;
523 : : /* Whether the above mapping is complete. */
524 : : bool stmt_vec_info_ro;
525 : :
526 : : /* Whether we've done a transform we think OK to not update virtual
527 : : SSA form. */
528 : : bool any_known_not_updated_vssa;
529 : :
530 : : /* The SLP graph. */
531 : : auto_vec<slp_instance> slp_instances;
532 : :
533 : : /* Maps base addresses to an innermost_loop_behavior that gives the maximum
534 : : known alignment for that base. */
535 : : vec_base_alignments base_alignments;
536 : :
537 : : /* All interleaving chains of stores, represented by the first
538 : : stmt in the chain. */
539 : : auto_vec<stmt_vec_info> grouped_stores;
540 : :
541 : : /* The set of vector modes used in the vectorized region. */
542 : : mode_set used_vector_modes;
543 : :
544 : : /* The argument we should pass to related_vector_mode when looking up
545 : : the vector mode for a scalar mode, or VOIDmode if we haven't yet
546 : : made any decisions about which vector modes to use. */
547 : : machine_mode vector_mode;
548 : :
549 : : /* The basic blocks in the vectorization region. For _loop_vec_info,
550 : : the memory is internally managed, while for _bb_vec_info, it points
551 : : to element space of an external auto_vec<>. This inconsistency is
552 : : not a good class design pattern. TODO: improve it with an unified
553 : : auto_vec<> whose lifetime is confined to vec_info object. */
554 : : basic_block *bbs;
555 : :
556 : : /* The count of the basic blocks in the vectorization region. */
557 : : unsigned int nbbs;
558 : :
559 : : /* Used to keep a sequence of def stmts of a pattern stmt that are loop
560 : : invariant if they exists.
561 : : The sequence is emitted in the loop preheader should the loop be vectorized
562 : : and are reset when undoing patterns. */
563 : : gimple_seq inv_pattern_def_seq;
564 : :
565 : : private:
566 : : stmt_vec_info new_stmt_vec_info (gimple *stmt);
567 : : void set_vinfo_for_stmt (gimple *, stmt_vec_info, bool = true);
568 : : void free_stmt_vec_infos ();
569 : : void free_stmt_vec_info (stmt_vec_info);
570 : : };
571 : :
572 : : class _loop_vec_info;
573 : : class _bb_vec_info;
574 : :
575 : : template<>
576 : : template<>
577 : : inline bool
578 : 362835968 : is_a_helper <_loop_vec_info *>::test (vec_info *i)
579 : : {
580 : 362200535 : return i->kind == vec_info::loop;
581 : : }
582 : :
583 : : template<>
584 : : template<>
585 : : inline bool
586 : 66362090 : is_a_helper <_bb_vec_info *>::test (vec_info *i)
587 : : {
588 : 66362090 : return i->kind == vec_info::bb;
589 : : }
590 : :
591 : : /* In general, we can divide the vector statements in a vectorized loop
592 : : into related groups ("rgroups") and say that for each rgroup there is
593 : : some nS such that the rgroup operates on nS values from one scalar
594 : : iteration followed by nS values from the next. That is, if VF is the
595 : : vectorization factor of the loop, the rgroup operates on a sequence:
596 : :
597 : : (1,1) (1,2) ... (1,nS) (2,1) ... (2,nS) ... (VF,1) ... (VF,nS)
598 : :
599 : : where (i,j) represents a scalar value with index j in a scalar
600 : : iteration with index i.
601 : :
602 : : [ We use the term "rgroup" to emphasise that this grouping isn't
603 : : necessarily the same as the grouping of statements used elsewhere.
604 : : For example, if we implement a group of scalar loads using gather
605 : : loads, we'll use a separate gather load for each scalar load, and
606 : : thus each gather load will belong to its own rgroup. ]
607 : :
608 : : In general this sequence will occupy nV vectors concatenated
609 : : together. If these vectors have nL lanes each, the total number
610 : : of scalar values N is given by:
611 : :
612 : : N = nS * VF = nV * nL
613 : :
614 : : None of nS, VF, nV and nL are required to be a power of 2. nS and nV
615 : : are compile-time constants but VF and nL can be variable (if the target
616 : : supports variable-length vectors).
617 : :
618 : : In classical vectorization, each iteration of the vector loop would
619 : : handle exactly VF iterations of the original scalar loop. However,
620 : : in vector loops that are able to operate on partial vectors, a
621 : : particular iteration of the vector loop might handle fewer than VF
622 : : iterations of the scalar loop. The vector lanes that correspond to
623 : : iterations of the scalar loop are said to be "active" and the other
624 : : lanes are said to be "inactive".
625 : :
626 : : In such vector loops, many rgroups need to be controlled to ensure
627 : : that they have no effect for the inactive lanes. Conceptually, each
628 : : such rgroup needs a sequence of booleans in the same order as above,
629 : : but with each (i,j) replaced by a boolean that indicates whether
630 : : iteration i is active. This sequence occupies nV vector controls
631 : : that again have nL lanes each. Thus the control sequence as a whole
632 : : consists of VF independent booleans that are each repeated nS times.
633 : :
634 : : Taking mask-based approach as a partially-populated vectors example.
635 : : We make the simplifying assumption that if a sequence of nV masks is
636 : : suitable for one (nS,nL) pair, we can reuse it for (nS/2,nL/2) by
637 : : VIEW_CONVERTing it. This holds for all current targets that support
638 : : fully-masked loops. For example, suppose the scalar loop is:
639 : :
640 : : float *f;
641 : : double *d;
642 : : for (int i = 0; i < n; ++i)
643 : : {
644 : : f[i * 2 + 0] += 1.0f;
645 : : f[i * 2 + 1] += 2.0f;
646 : : d[i] += 3.0;
647 : : }
648 : :
649 : : and suppose that vectors have 256 bits. The vectorized f accesses
650 : : will belong to one rgroup and the vectorized d access to another:
651 : :
652 : : f rgroup: nS = 2, nV = 1, nL = 8
653 : : d rgroup: nS = 1, nV = 1, nL = 4
654 : : VF = 4
655 : :
656 : : [ In this simple example the rgroups do correspond to the normal
657 : : SLP grouping scheme. ]
658 : :
659 : : If only the first three lanes are active, the masks we need are:
660 : :
661 : : f rgroup: 1 1 | 1 1 | 1 1 | 0 0
662 : : d rgroup: 1 | 1 | 1 | 0
663 : :
664 : : Here we can use a mask calculated for f's rgroup for d's, but not
665 : : vice versa.
666 : :
667 : : Thus for each value of nV, it is enough to provide nV masks, with the
668 : : mask being calculated based on the highest nL (or, equivalently, based
669 : : on the highest nS) required by any rgroup with that nV. We therefore
670 : : represent the entire collection of masks as a two-level table, with the
671 : : first level being indexed by nV - 1 (since nV == 0 doesn't exist) and
672 : : the second being indexed by the mask index 0 <= i < nV. */
673 : :
674 : : /* The controls (like masks or lengths) needed by rgroups with nV vectors,
675 : : according to the description above. */
676 : : struct rgroup_controls {
677 : : /* The largest nS for all rgroups that use these controls.
678 : : For vect_partial_vectors_avx512 this is the constant nscalars_per_iter
679 : : for all members of the group. */
680 : : unsigned int max_nscalars_per_iter;
681 : :
682 : : /* For the largest nS recorded above, the loop controls divide each scalar
683 : : into FACTOR equal-sized pieces. This is useful if we need to split
684 : : element-based accesses into byte-based accesses.
685 : : For vect_partial_vectors_avx512 this records nV instead. */
686 : : unsigned int factor;
687 : :
688 : : /* This is a vector type with MAX_NSCALARS_PER_ITER * VF / nV elements.
689 : : For mask-based controls, it is the type of the masks in CONTROLS.
690 : : For length-based controls, it can be any vector type that has the
691 : : specified number of elements; the type of the elements doesn't matter. */
692 : : tree type;
693 : :
694 : : /* When there is no uniformly used LOOP_VINFO_RGROUP_COMPARE_TYPE this
695 : : is the rgroup specific type used. */
696 : : tree compare_type;
697 : :
698 : : /* A vector of nV controls, in iteration order. */
699 : : vec<tree> controls;
700 : :
701 : : /* In case of len_load and len_store with a bias there is only one
702 : : rgroup. This holds the adjusted loop length for the this rgroup. */
703 : : tree bias_adjusted_ctrl;
704 : : };
705 : :
706 : 419214 : struct vec_loop_masks
707 : : {
708 : 360294 : bool is_empty () const { return mask_set.is_empty (); }
709 : :
710 : : /* Set to record vectype, nvector pairs. */
711 : : hash_set<pair_hash <nofree_ptr_hash <tree_node>,
712 : : int_hash<unsigned, 0>>> mask_set;
713 : :
714 : : /* rgroup_controls used for the partial vector scheme. */
715 : : auto_vec<rgroup_controls> rgc_vec;
716 : : };
717 : :
718 : : typedef auto_vec<rgroup_controls> vec_loop_lens;
719 : :
720 : : typedef auto_vec<std::pair<data_reference*, tree> > drs_init_vec;
721 : :
722 : : /* Information about a reduction accumulator from the main loop that could
723 : : conceivably be reused as the input to a reduction in an epilogue loop. */
724 : : struct vect_reusable_accumulator {
725 : : /* The final value of the accumulator, which forms the input to the
726 : : reduction operation. */
727 : : tree reduc_input;
728 : :
729 : : /* The stmt_vec_info that describes the reduction (i.e. the one for
730 : : which is_reduc_info is true). */
731 : : stmt_vec_info reduc_info;
732 : : };
733 : :
734 : : /*-----------------------------------------------------------------*/
735 : : /* Info on vectorized loops. */
736 : : /*-----------------------------------------------------------------*/
737 : : typedef class _loop_vec_info : public vec_info {
738 : : public:
739 : : _loop_vec_info (class loop *, vec_info_shared *);
740 : : ~_loop_vec_info ();
741 : :
742 : : /* The loop to which this info struct refers to. */
743 : : class loop *loop;
744 : :
745 : : /* Number of latch executions. */
746 : : tree num_itersm1;
747 : : /* Number of iterations. */
748 : : tree num_iters;
749 : : /* Number of iterations of the original loop. */
750 : : tree num_iters_unchanged;
751 : : /* Condition under which this loop is analyzed and versioned. */
752 : : tree num_iters_assumptions;
753 : :
754 : : /* The cost of the vector code. */
755 : : class vector_costs *vector_costs;
756 : :
757 : : /* The cost of the scalar code. */
758 : : class vector_costs *scalar_costs;
759 : :
760 : : /* Threshold of number of iterations below which vectorization will not be
761 : : performed. It is calculated from MIN_PROFITABLE_ITERS and
762 : : param_min_vect_loop_bound. */
763 : : unsigned int th;
764 : :
765 : : /* When applying loop versioning, the vector form should only be used
766 : : if the number of scalar iterations is >= this value, on top of all
767 : : the other requirements. Ignored when loop versioning is not being
768 : : used. */
769 : : poly_uint64 versioning_threshold;
770 : :
771 : : /* Unrolling factor */
772 : : poly_uint64 vectorization_factor;
773 : :
774 : : /* If this loop is an epilogue loop whose main loop can be skipped,
775 : : MAIN_LOOP_EDGE is the edge from the main loop to this loop's
776 : : preheader. SKIP_MAIN_LOOP_EDGE is then the edge that skips the
777 : : main loop and goes straight to this loop's preheader.
778 : :
779 : : Both fields are null otherwise. */
780 : : edge main_loop_edge;
781 : : edge skip_main_loop_edge;
782 : :
783 : : /* If this loop is an epilogue loop that might be skipped after executing
784 : : the main loop, this edge is the one that skips the epilogue. */
785 : : edge skip_this_loop_edge;
786 : :
787 : : /* The vectorized form of a standard reduction replaces the original
788 : : scalar code's final result (a loop-closed SSA PHI) with the result
789 : : of a vector-to-scalar reduction operation. After vectorization,
790 : : this variable maps these vector-to-scalar results to information
791 : : about the reductions that generated them. */
792 : : hash_map<tree, vect_reusable_accumulator> reusable_accumulators;
793 : :
794 : : /* The number of times that the target suggested we unroll the vector loop
795 : : in order to promote more ILP. This value will be used to re-analyze the
796 : : loop for vectorization and if successful the value will be folded into
797 : : vectorization_factor (and therefore exactly divides
798 : : vectorization_factor). */
799 : : unsigned int suggested_unroll_factor;
800 : :
801 : : /* Maximum runtime vectorization factor, or MAX_VECTORIZATION_FACTOR
802 : : if there is no particular limit. */
803 : : unsigned HOST_WIDE_INT max_vectorization_factor;
804 : :
805 : : /* The masks that a fully-masked loop should use to avoid operating
806 : : on inactive scalars. */
807 : : vec_loop_masks masks;
808 : :
809 : : /* The lengths that a loop with length should use to avoid operating
810 : : on inactive scalars. */
811 : : vec_loop_lens lens;
812 : :
813 : : /* Set of scalar conditions that have loop mask applied. */
814 : : scalar_cond_masked_set_type scalar_cond_masked_set;
815 : :
816 : : /* Set of vector conditions that have loop mask applied. */
817 : : vec_cond_masked_set_type vec_cond_masked_set;
818 : :
819 : : /* If we are using a loop mask to align memory addresses, this variable
820 : : contains the number of vector elements that we should skip in the
821 : : first iteration of the vector loop (i.e. the number of leading
822 : : elements that should be false in the first mask). */
823 : : tree mask_skip_niters;
824 : :
825 : : /* If we are using a loop mask to align memory addresses and we're in an
826 : : early break loop then this variable contains the number of elements that
827 : : were skipped during the initial iteration of the loop. */
828 : : tree mask_skip_niters_pfa_offset;
829 : :
830 : : /* The type that the loop control IV should be converted to before
831 : : testing which of the VF scalars are active and inactive.
832 : : Only meaningful if LOOP_VINFO_USING_PARTIAL_VECTORS_P. */
833 : : tree rgroup_compare_type;
834 : :
835 : : /* For #pragma omp simd if (x) loops the x expression. If constant 0,
836 : : the loop should not be vectorized, if constant non-zero, simd_if_cond
837 : : shouldn't be set and loop vectorized normally, if SSA_NAME, the loop
838 : : should be versioned on that condition, using scalar loop if the condition
839 : : is false and vectorized loop otherwise. */
840 : : tree simd_if_cond;
841 : :
842 : : /* The type that the vector loop control IV should have when
843 : : LOOP_VINFO_USING_PARTIAL_VECTORS_P is true. */
844 : : tree rgroup_iv_type;
845 : :
846 : : /* The style used for implementing partial vectors when
847 : : LOOP_VINFO_USING_PARTIAL_VECTORS_P is true. */
848 : : vect_partial_vector_style partial_vector_style;
849 : :
850 : : /* Unknown DRs according to which loop was peeled. */
851 : : class dr_vec_info *unaligned_dr;
852 : :
853 : : /* peeling_for_alignment indicates whether peeling for alignment will take
854 : : place, and what the peeling factor should be:
855 : : peeling_for_alignment = X means:
856 : : If X=0: Peeling for alignment will not be applied.
857 : : If X>0: Peel first X iterations.
858 : : If X=-1: Generate a runtime test to calculate the number of iterations
859 : : to be peeled, using the dataref recorded in the field
860 : : unaligned_dr. */
861 : : int peeling_for_alignment;
862 : :
863 : : /* The mask used to check the alignment of pointers or arrays. */
864 : : int ptr_mask;
865 : :
866 : : /* Indicates whether the loop has any non-linear IV. */
867 : : bool nonlinear_iv;
868 : :
869 : : /* Data Dependence Relations defining address ranges that are candidates
870 : : for a run-time aliasing check. */
871 : : auto_vec<ddr_p> may_alias_ddrs;
872 : :
873 : : /* Data Dependence Relations defining address ranges together with segment
874 : : lengths from which the run-time aliasing check is built. */
875 : : auto_vec<dr_with_seg_len_pair_t> comp_alias_ddrs;
876 : :
877 : : /* Check that the addresses of each pair of objects is unequal. */
878 : : auto_vec<vec_object_pair> check_unequal_addrs;
879 : :
880 : : /* List of values that are required to be nonzero. This is used to check
881 : : whether things like "x[i * n] += 1;" are safe and eventually gets added
882 : : to the checks for lower bounds below. */
883 : : auto_vec<tree> check_nonzero;
884 : :
885 : : /* List of values that need to be checked for a minimum value. */
886 : : auto_vec<vec_lower_bound> lower_bounds;
887 : :
888 : : /* Statements in the loop that have data references that are candidates for a
889 : : runtime (loop versioning) misalignment check. */
890 : : auto_vec<stmt_vec_info> may_misalign_stmts;
891 : :
892 : : /* Reduction cycles detected in the loop. Used in loop-aware SLP. */
893 : : auto_vec<stmt_vec_info> reductions;
894 : :
895 : : /* All reduction chains in the loop, represented by the first
896 : : stmt in the chain. */
897 : : auto_vec<stmt_vec_info> reduction_chains;
898 : :
899 : : /* Cost vector for a single scalar iteration. */
900 : : auto_vec<stmt_info_for_cost> scalar_cost_vec;
901 : :
902 : : /* Map of IV base/step expressions to inserted name in the preheader. */
903 : : hash_map<tree_operand_hash, tree> *ivexpr_map;
904 : :
905 : : /* Map of OpenMP "omp simd array" scan variables to corresponding
906 : : rhs of the store of the initializer. */
907 : : hash_map<tree, tree> *scan_map;
908 : :
909 : : /* The unrolling factor needed to SLP the loop. In case of that pure SLP is
910 : : applied to the loop, i.e., no unrolling is needed, this is 1. */
911 : : poly_uint64 slp_unrolling_factor;
912 : :
913 : : /* The factor used to over weight those statements in an inner loop
914 : : relative to the loop being vectorized. */
915 : : unsigned int inner_loop_cost_factor;
916 : :
917 : : /* Is the loop vectorizable? */
918 : : bool vectorizable;
919 : :
920 : : /* Records whether we still have the option of vectorizing this loop
921 : : using partially-populated vectors; in other words, whether it is
922 : : still possible for one iteration of the vector loop to handle
923 : : fewer than VF scalars. */
924 : : bool can_use_partial_vectors_p;
925 : :
926 : : /* Records whether we must use niter masking for correctness reasons. */
927 : : bool must_use_partial_vectors_p;
928 : :
929 : : /* True if we've decided to use partially-populated vectors, so that
930 : : the vector loop can handle fewer than VF scalars. */
931 : : bool using_partial_vectors_p;
932 : :
933 : : /* True if we've decided to use a decrementing loop control IV that counts
934 : : scalars. This can be done for any loop that:
935 : :
936 : : (a) uses length "controls"; and
937 : : (b) can iterate more than once. */
938 : : bool using_decrementing_iv_p;
939 : :
940 : : /* True if we've decided to use output of select_vl to adjust IV of
941 : : both loop control and data reference pointer. This is only true
942 : : for single-rgroup control. */
943 : : bool using_select_vl_p;
944 : :
945 : : /* True if we've decided to use partially-populated vectors for the
946 : : epilogue of loop. */
947 : : bool epil_using_partial_vectors_p;
948 : :
949 : : /* The bias for len_load and len_store. For now, only 0 and -1 are
950 : : supported. -1 must be used when a backend does not support
951 : : len_load/len_store with a length of zero. */
952 : : signed char partial_load_store_bias;
953 : :
954 : : /* When we have grouped data accesses with gaps, we may introduce invalid
955 : : memory accesses. We peel the last iteration of the loop to prevent
956 : : this. */
957 : : bool peeling_for_gaps;
958 : :
959 : : /* When the number of iterations is not a multiple of the vector size
960 : : we need to peel off iterations at the end to form an epilogue loop. */
961 : : bool peeling_for_niter;
962 : :
963 : : /* When the loop has early breaks that we can vectorize we need to peel
964 : : the loop for the break finding loop. */
965 : : bool early_breaks;
966 : :
967 : : /* List of loop additional IV conditionals found in the loop. */
968 : : auto_vec<gcond *> conds;
969 : :
970 : : /* Main loop IV cond. */
971 : : gcond* loop_iv_cond;
972 : :
973 : : /* True if there are no loop carried data dependencies in the loop.
974 : : If loop->safelen <= 1, then this is always true, either the loop
975 : : didn't have any loop carried data dependencies, or the loop is being
976 : : vectorized guarded with some runtime alias checks, or couldn't
977 : : be vectorized at all, but then this field shouldn't be used.
978 : : For loop->safelen >= 2, the user has asserted that there are no
979 : : backward dependencies, but there still could be loop carried forward
980 : : dependencies in such loops. This flag will be false if normal
981 : : vectorizer data dependency analysis would fail or require versioning
982 : : for alias, but because of loop->safelen >= 2 it has been vectorized
983 : : even without versioning for alias. E.g. in:
984 : : #pragma omp simd
985 : : for (int i = 0; i < m; i++)
986 : : a[i] = a[i + k] * c;
987 : : (or #pragma simd or #pragma ivdep) we can vectorize this and it will
988 : : DTRT even for k > 0 && k < m, but without safelen we would not
989 : : vectorize this, so this field would be false. */
990 : : bool no_data_dependencies;
991 : :
992 : : /* Mark loops having masked stores. */
993 : : bool has_mask_store;
994 : :
995 : : /* Queued scaling factor for the scalar loop. */
996 : : profile_probability scalar_loop_scaling;
997 : :
998 : : /* If if-conversion versioned this loop before conversion, this is the
999 : : loop version without if-conversion. */
1000 : : class loop *scalar_loop;
1001 : :
1002 : : /* For loops being epilogues of already vectorized loops
1003 : : this points to the main vectorized loop. Otherwise NULL. */
1004 : : _loop_vec_info *main_loop_info;
1005 : :
1006 : : /* For loops being epilogues of already vectorized loops
1007 : : this points to the preceeding vectorized (possibly epilogue) loop.
1008 : : Otherwise NULL. */
1009 : : _loop_vec_info *orig_loop_info;
1010 : :
1011 : : /* Used to store loop_vec_infos of the epilogue of this loop during
1012 : : analysis. */
1013 : : _loop_vec_info *epilogue_vinfo;
1014 : :
1015 : : /* If this is an epilogue loop the DR advancement applied. */
1016 : : tree drs_advanced_by;
1017 : :
1018 : : /* The controlling loop IV for the current loop when vectorizing. This IV
1019 : : controls the natural exits of the loop. */
1020 : : edge vec_loop_iv_exit;
1021 : :
1022 : : /* The controlling loop IV for the epilogue loop when vectorizing. This IV
1023 : : controls the natural exits of the loop. */
1024 : : edge vec_epilogue_loop_iv_exit;
1025 : :
1026 : : /* The controlling loop IV for the scalar loop being vectorized. This IV
1027 : : controls the natural exits of the loop. */
1028 : : edge scalar_loop_iv_exit;
1029 : :
1030 : : /* Used to store the list of stores needing to be moved if doing early
1031 : : break vectorization as they would violate the scalar loop semantics if
1032 : : vectorized in their current location. These are stored in order that they
1033 : : need to be moved. */
1034 : : auto_vec<gimple *> early_break_stores;
1035 : :
1036 : : /* The final basic block where to move statements to. In the case of
1037 : : multiple exits this could be pretty far away. */
1038 : : basic_block early_break_dest_bb;
1039 : :
1040 : : /* Statements whose VUSES need updating if early break vectorization is to
1041 : : happen. */
1042 : : auto_vec<gimple*> early_break_vuses;
1043 : :
1044 : : /* Record statements that are needed to be live for early break vectorization
1045 : : but may not have an LC PHI node materialized yet in the exits. */
1046 : : auto_vec<stmt_vec_info> early_break_live_ivs;
1047 : : } *loop_vec_info;
1048 : :
1049 : : /* Access Functions. */
1050 : : #define LOOP_VINFO_LOOP(L) (L)->loop
1051 : : #define LOOP_VINFO_IV_EXIT(L) (L)->vec_loop_iv_exit
1052 : : #define LOOP_VINFO_EPILOGUE_IV_EXIT(L) (L)->vec_epilogue_loop_iv_exit
1053 : : #define LOOP_VINFO_SCALAR_IV_EXIT(L) (L)->scalar_loop_iv_exit
1054 : : #define LOOP_VINFO_BBS(L) (L)->bbs
1055 : : #define LOOP_VINFO_NBBS(L) (L)->nbbs
1056 : : #define LOOP_VINFO_NITERSM1(L) (L)->num_itersm1
1057 : : #define LOOP_VINFO_NITERS(L) (L)->num_iters
1058 : : /* Since LOOP_VINFO_NITERS and LOOP_VINFO_NITERSM1 can change after
1059 : : prologue peeling retain total unchanged scalar loop iterations for
1060 : : cost model. */
1061 : : #define LOOP_VINFO_NITERS_UNCHANGED(L) (L)->num_iters_unchanged
1062 : : #define LOOP_VINFO_NITERS_ASSUMPTIONS(L) (L)->num_iters_assumptions
1063 : : #define LOOP_VINFO_COST_MODEL_THRESHOLD(L) (L)->th
1064 : : #define LOOP_VINFO_VERSIONING_THRESHOLD(L) (L)->versioning_threshold
1065 : : #define LOOP_VINFO_VECTORIZABLE_P(L) (L)->vectorizable
1066 : : #define LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P(L) (L)->can_use_partial_vectors_p
1067 : : #define LOOP_VINFO_MUST_USE_PARTIAL_VECTORS_P(L) (L)->must_use_partial_vectors_p
1068 : : #define LOOP_VINFO_USING_PARTIAL_VECTORS_P(L) (L)->using_partial_vectors_p
1069 : : #define LOOP_VINFO_USING_DECREMENTING_IV_P(L) (L)->using_decrementing_iv_p
1070 : : #define LOOP_VINFO_USING_SELECT_VL_P(L) (L)->using_select_vl_p
1071 : : #define LOOP_VINFO_EPIL_USING_PARTIAL_VECTORS_P(L) \
1072 : : (L)->epil_using_partial_vectors_p
1073 : : #define LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS(L) (L)->partial_load_store_bias
1074 : : #define LOOP_VINFO_VECT_FACTOR(L) (L)->vectorization_factor
1075 : : #define LOOP_VINFO_MAX_VECT_FACTOR(L) (L)->max_vectorization_factor
1076 : : #define LOOP_VINFO_MASKS(L) (L)->masks
1077 : : #define LOOP_VINFO_LENS(L) (L)->lens
1078 : : #define LOOP_VINFO_MASK_SKIP_NITERS(L) (L)->mask_skip_niters
1079 : : #define LOOP_VINFO_MASK_NITERS_PFA_OFFSET(L) (L)->mask_skip_niters_pfa_offset
1080 : : #define LOOP_VINFO_RGROUP_COMPARE_TYPE(L) (L)->rgroup_compare_type
1081 : : #define LOOP_VINFO_RGROUP_IV_TYPE(L) (L)->rgroup_iv_type
1082 : : #define LOOP_VINFO_PARTIAL_VECTORS_STYLE(L) (L)->partial_vector_style
1083 : : #define LOOP_VINFO_PTR_MASK(L) (L)->ptr_mask
1084 : : #define LOOP_VINFO_LOOP_NEST(L) (L)->shared->loop_nest
1085 : : #define LOOP_VINFO_DATAREFS(L) (L)->shared->datarefs
1086 : : #define LOOP_VINFO_DDRS(L) (L)->shared->ddrs
1087 : : #define LOOP_VINFO_INT_NITERS(L) (TREE_INT_CST_LOW ((L)->num_iters))
1088 : : #define LOOP_VINFO_PEELING_FOR_ALIGNMENT(L) (L)->peeling_for_alignment
1089 : : #define LOOP_VINFO_NON_LINEAR_IV(L) (L)->nonlinear_iv
1090 : : #define LOOP_VINFO_UNALIGNED_DR(L) (L)->unaligned_dr
1091 : : #define LOOP_VINFO_MAY_MISALIGN_STMTS(L) (L)->may_misalign_stmts
1092 : : #define LOOP_VINFO_MAY_ALIAS_DDRS(L) (L)->may_alias_ddrs
1093 : : #define LOOP_VINFO_COMP_ALIAS_DDRS(L) (L)->comp_alias_ddrs
1094 : : #define LOOP_VINFO_CHECK_UNEQUAL_ADDRS(L) (L)->check_unequal_addrs
1095 : : #define LOOP_VINFO_CHECK_NONZERO(L) (L)->check_nonzero
1096 : : #define LOOP_VINFO_LOWER_BOUNDS(L) (L)->lower_bounds
1097 : : #define LOOP_VINFO_GROUPED_STORES(L) (L)->grouped_stores
1098 : : #define LOOP_VINFO_SLP_INSTANCES(L) (L)->slp_instances
1099 : : #define LOOP_VINFO_SLP_UNROLLING_FACTOR(L) (L)->slp_unrolling_factor
1100 : : #define LOOP_VINFO_REDUCTIONS(L) (L)->reductions
1101 : : #define LOOP_VINFO_REDUCTION_CHAINS(L) (L)->reduction_chains
1102 : : #define LOOP_VINFO_PEELING_FOR_GAPS(L) (L)->peeling_for_gaps
1103 : : #define LOOP_VINFO_PEELING_FOR_NITER(L) (L)->peeling_for_niter
1104 : : #define LOOP_VINFO_EARLY_BREAKS(L) (L)->early_breaks
1105 : : #define LOOP_VINFO_EARLY_BRK_STORES(L) (L)->early_break_stores
1106 : : #define LOOP_VINFO_EARLY_BREAKS_VECT_PEELED(L) \
1107 : : (single_pred ((L)->loop->latch) != (L)->vec_loop_iv_exit->src)
1108 : : #define LOOP_VINFO_EARLY_BREAKS_LIVE_IVS(L) \
1109 : : (L)->early_break_live_ivs
1110 : : #define LOOP_VINFO_EARLY_BRK_DEST_BB(L) (L)->early_break_dest_bb
1111 : : #define LOOP_VINFO_EARLY_BRK_VUSES(L) (L)->early_break_vuses
1112 : : #define LOOP_VINFO_LOOP_CONDS(L) (L)->conds
1113 : : #define LOOP_VINFO_LOOP_IV_COND(L) (L)->loop_iv_cond
1114 : : #define LOOP_VINFO_NO_DATA_DEPENDENCIES(L) (L)->no_data_dependencies
1115 : : #define LOOP_VINFO_SCALAR_LOOP(L) (L)->scalar_loop
1116 : : #define LOOP_VINFO_SCALAR_LOOP_SCALING(L) (L)->scalar_loop_scaling
1117 : : #define LOOP_VINFO_HAS_MASK_STORE(L) (L)->has_mask_store
1118 : : #define LOOP_VINFO_SCALAR_ITERATION_COST(L) (L)->scalar_cost_vec
1119 : : #define LOOP_VINFO_MAIN_LOOP_INFO(L) (L)->main_loop_info
1120 : : #define LOOP_VINFO_ORIG_LOOP_INFO(L) (L)->orig_loop_info
1121 : : #define LOOP_VINFO_SIMD_IF_COND(L) (L)->simd_if_cond
1122 : : #define LOOP_VINFO_INNER_LOOP_COST_FACTOR(L) (L)->inner_loop_cost_factor
1123 : : #define LOOP_VINFO_INV_PATTERN_DEF_SEQ(L) (L)->inv_pattern_def_seq
1124 : : #define LOOP_VINFO_DRS_ADVANCED_BY(L) (L)->drs_advanced_by
1125 : :
1126 : : #define LOOP_VINFO_FULLY_MASKED_P(L) \
1127 : : (LOOP_VINFO_USING_PARTIAL_VECTORS_P (L) \
1128 : : && !LOOP_VINFO_MASKS (L).is_empty ())
1129 : :
1130 : : #define LOOP_VINFO_FULLY_WITH_LENGTH_P(L) \
1131 : : (LOOP_VINFO_USING_PARTIAL_VECTORS_P (L) \
1132 : : && !LOOP_VINFO_LENS (L).is_empty ())
1133 : :
1134 : : #define LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT(L) \
1135 : : ((L)->may_misalign_stmts.length () > 0)
1136 : : #define LOOP_REQUIRES_VERSIONING_FOR_ALIAS(L) \
1137 : : ((L)->comp_alias_ddrs.length () > 0 \
1138 : : || (L)->check_unequal_addrs.length () > 0 \
1139 : : || (L)->lower_bounds.length () > 0)
1140 : : #define LOOP_REQUIRES_VERSIONING_FOR_NITERS(L) \
1141 : : (LOOP_VINFO_NITERS_ASSUMPTIONS (L))
1142 : : #define LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND(L) \
1143 : : (LOOP_VINFO_SIMD_IF_COND (L))
1144 : : #define LOOP_REQUIRES_VERSIONING(L) \
1145 : : (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (L) \
1146 : : || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (L) \
1147 : : || LOOP_REQUIRES_VERSIONING_FOR_NITERS (L) \
1148 : : || LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND (L))
1149 : :
1150 : : #define LOOP_VINFO_NITERS_KNOWN_P(L) \
1151 : : (tree_fits_shwi_p ((L)->num_iters) && tree_to_shwi ((L)->num_iters) > 0)
1152 : :
1153 : : #define LOOP_VINFO_EPILOGUE_P(L) \
1154 : : (LOOP_VINFO_ORIG_LOOP_INFO (L) != NULL)
1155 : :
1156 : : #define LOOP_VINFO_ORIG_MAX_VECT_FACTOR(L) \
1157 : : (LOOP_VINFO_MAX_VECT_FACTOR (LOOP_VINFO_ORIG_LOOP_INFO (L)))
1158 : :
1159 : : /* Wrapper for loop_vec_info, for tracking success/failure, where a non-NULL
1160 : : value signifies success, and a NULL value signifies failure, supporting
1161 : : propagating an opt_problem * describing the failure back up the call
1162 : : stack. */
1163 : : typedef opt_pointer_wrapper <loop_vec_info> opt_loop_vec_info;
1164 : :
1165 : : inline loop_vec_info
1166 : 530277 : loop_vec_info_for_loop (class loop *loop)
1167 : : {
1168 : 530277 : return (loop_vec_info) loop->aux;
1169 : : }
1170 : :
1171 : : struct slp_root
1172 : : {
1173 : 1190401 : slp_root (slp_instance_kind kind_, vec<stmt_vec_info> stmts_,
1174 : 11582 : vec<stmt_vec_info> roots_, vec<tree> remain_ = vNULL)
1175 : 1190401 : : kind(kind_), stmts(stmts_), roots(roots_), remain(remain_) {}
1176 : : slp_instance_kind kind;
1177 : : vec<stmt_vec_info> stmts;
1178 : : vec<stmt_vec_info> roots;
1179 : : vec<tree> remain;
1180 : : };
1181 : :
1182 : : typedef class _bb_vec_info : public vec_info
1183 : : {
1184 : : public:
1185 : : _bb_vec_info (vec<basic_block> bbs, vec_info_shared *);
1186 : : ~_bb_vec_info ();
1187 : :
1188 : : vec<slp_root> roots;
1189 : : } *bb_vec_info;
1190 : :
1191 : : #define BB_VINFO_BBS(B) (B)->bbs
1192 : : #define BB_VINFO_NBBS(B) (B)->nbbs
1193 : : #define BB_VINFO_GROUPED_STORES(B) (B)->grouped_stores
1194 : : #define BB_VINFO_SLP_INSTANCES(B) (B)->slp_instances
1195 : : #define BB_VINFO_DATAREFS(B) (B)->shared->datarefs
1196 : : #define BB_VINFO_DDRS(B) (B)->shared->ddrs
1197 : :
1198 : : /*-----------------------------------------------------------------*/
1199 : : /* Info on vectorized defs. */
1200 : : /*-----------------------------------------------------------------*/
1201 : : enum stmt_vec_info_type {
1202 : : undef_vec_info_type = 0,
1203 : : load_vec_info_type,
1204 : : store_vec_info_type,
1205 : : shift_vec_info_type,
1206 : : op_vec_info_type,
1207 : : call_vec_info_type,
1208 : : call_simd_clone_vec_info_type,
1209 : : assignment_vec_info_type,
1210 : : condition_vec_info_type,
1211 : : comparison_vec_info_type,
1212 : : reduc_vec_info_type,
1213 : : induc_vec_info_type,
1214 : : type_promotion_vec_info_type,
1215 : : type_demotion_vec_info_type,
1216 : : type_conversion_vec_info_type,
1217 : : cycle_phi_info_type,
1218 : : lc_phi_info_type,
1219 : : phi_info_type,
1220 : : recurr_info_type,
1221 : : loop_exit_ctrl_vec_info_type
1222 : : };
1223 : :
1224 : : /* Indicates whether/how a variable is used in the scope of loop/basic
1225 : : block. */
1226 : : enum vect_relevant {
1227 : : vect_unused_in_scope = 0,
1228 : :
1229 : : /* The def is only used outside the loop. */
1230 : : vect_used_only_live,
1231 : : /* The def is in the inner loop, and the use is in the outer loop, and the
1232 : : use is a reduction stmt. */
1233 : : vect_used_in_outer_by_reduction,
1234 : : /* The def is in the inner loop, and the use is in the outer loop (and is
1235 : : not part of reduction). */
1236 : : vect_used_in_outer,
1237 : :
1238 : : /* defs that feed computations that end up (only) in a reduction. These
1239 : : defs may be used by non-reduction stmts, but eventually, any
1240 : : computations/values that are affected by these defs are used to compute
1241 : : a reduction (i.e. don't get stored to memory, for example). We use this
1242 : : to identify computations that we can change the order in which they are
1243 : : computed. */
1244 : : vect_used_by_reduction,
1245 : :
1246 : : vect_used_in_scope
1247 : : };
1248 : :
1249 : : /* The type of vectorization that can be applied to the stmt: regular loop-based
1250 : : vectorization; pure SLP - the stmt is a part of SLP instances and does not
1251 : : have uses outside SLP instances; or hybrid SLP and loop-based - the stmt is
1252 : : a part of SLP instance and also must be loop-based vectorized, since it has
1253 : : uses outside SLP sequences.
1254 : :
1255 : : In the loop context the meanings of pure and hybrid SLP are slightly
1256 : : different. By saying that pure SLP is applied to the loop, we mean that we
1257 : : exploit only intra-iteration parallelism in the loop; i.e., the loop can be
1258 : : vectorized without doing any conceptual unrolling, cause we don't pack
1259 : : together stmts from different iterations, only within a single iteration.
1260 : : Loop hybrid SLP means that we exploit both intra-iteration and
1261 : : inter-iteration parallelism (e.g., number of elements in the vector is 4
1262 : : and the slp-group-size is 2, in which case we don't have enough parallelism
1263 : : within an iteration, so we obtain the rest of the parallelism from subsequent
1264 : : iterations by unrolling the loop by 2). */
1265 : : enum slp_vect_type {
1266 : : loop_vect = 0,
1267 : : pure_slp,
1268 : : hybrid
1269 : : };
1270 : :
1271 : : /* Says whether a statement is a load, a store of a vectorized statement
1272 : : result, or a store of an invariant value. */
1273 : : enum vec_load_store_type {
1274 : : VLS_LOAD,
1275 : : VLS_STORE,
1276 : : VLS_STORE_INVARIANT
1277 : : };
1278 : :
1279 : : class dr_vec_info {
1280 : : public:
1281 : : /* The data reference itself. */
1282 : : data_reference *dr;
1283 : : /* The statement that contains the data reference. */
1284 : : stmt_vec_info stmt;
1285 : : /* The analysis group this DR belongs to when doing BB vectorization.
1286 : : DRs of the same group belong to the same conditional execution context. */
1287 : : unsigned group;
1288 : : /* The misalignment in bytes of the reference, or -1 if not known. */
1289 : : int misalignment;
1290 : : /* The byte alignment that we'd ideally like the reference to have,
1291 : : and the value that misalignment is measured against. */
1292 : : poly_uint64 target_alignment;
1293 : : /* If true the alignment of base_decl needs to be increased. */
1294 : : bool base_misaligned;
1295 : :
1296 : : /* Set by early break vectorization when this DR needs peeling for alignment
1297 : : for correctness. */
1298 : : bool safe_speculative_read_required;
1299 : :
1300 : : /* Set by early break vectorization when this DR's scalar accesses are known
1301 : : to be inbounds of a known bounds loop. */
1302 : : bool scalar_access_known_in_bounds;
1303 : :
1304 : : tree base_decl;
1305 : :
1306 : : /* Stores current vectorized loop's offset. To be added to the DR's
1307 : : offset to calculate current offset of data reference. */
1308 : : tree offset;
1309 : : };
1310 : :
1311 : : typedef struct data_reference *dr_p;
1312 : :
1313 : : class _stmt_vec_info {
1314 : : public:
1315 : :
1316 : : enum stmt_vec_info_type type;
1317 : :
1318 : : /* Indicates whether this stmts is part of a computation whose result is
1319 : : used outside the loop. */
1320 : : bool live;
1321 : :
1322 : : /* Stmt is part of some pattern (computation idiom) */
1323 : : bool in_pattern_p;
1324 : :
1325 : : /* True if the statement was created during pattern recognition as
1326 : : part of the replacement for RELATED_STMT. This implies that the
1327 : : statement isn't part of any basic block, although for convenience
1328 : : its gimple_bb is the same as for RELATED_STMT. */
1329 : : bool pattern_stmt_p;
1330 : :
1331 : : /* Is this statement vectorizable or should it be skipped in (partial)
1332 : : vectorization. */
1333 : : bool vectorizable;
1334 : :
1335 : : /* The stmt to which this info struct refers to. */
1336 : : gimple *stmt;
1337 : :
1338 : : /* The vector type to be used for the LHS of this statement. */
1339 : : tree vectype;
1340 : :
1341 : : /* The vectorized stmts. */
1342 : : vec<gimple *> vec_stmts;
1343 : :
1344 : : /* The following is relevant only for stmts that contain a non-scalar
1345 : : data-ref (array/pointer/struct access). A GIMPLE stmt is expected to have
1346 : : at most one such data-ref. */
1347 : :
1348 : : dr_vec_info dr_aux;
1349 : :
1350 : : /* Information about the data-ref relative to this loop
1351 : : nest (the loop that is being considered for vectorization). */
1352 : : innermost_loop_behavior dr_wrt_vec_loop;
1353 : :
1354 : : /* For loop PHI nodes, the base and evolution part of it. This makes sure
1355 : : this information is still available in vect_update_ivs_after_vectorizer
1356 : : where we may not be able to re-analyze the PHI nodes evolution as
1357 : : peeling for the prologue loop can make it unanalyzable. The evolution
1358 : : part is still correct after peeling, but the base may have changed from
1359 : : the version here. */
1360 : : tree loop_phi_evolution_base_unchanged;
1361 : : tree loop_phi_evolution_part;
1362 : : enum vect_induction_op_type loop_phi_evolution_type;
1363 : :
1364 : : /* Used for various bookkeeping purposes, generally holding a pointer to
1365 : : some other stmt S that is in some way "related" to this stmt.
1366 : : Current use of this field is:
1367 : : If this stmt is part of a pattern (i.e. the field 'in_pattern_p' is
1368 : : true): S is the "pattern stmt" that represents (and replaces) the
1369 : : sequence of stmts that constitutes the pattern. Similarly, the
1370 : : related_stmt of the "pattern stmt" points back to this stmt (which is
1371 : : the last stmt in the original sequence of stmts that constitutes the
1372 : : pattern). */
1373 : : stmt_vec_info related_stmt;
1374 : :
1375 : : /* Used to keep a sequence of def stmts of a pattern stmt if such exists.
1376 : : The sequence is attached to the original statement rather than the
1377 : : pattern statement. */
1378 : : gimple_seq pattern_def_seq;
1379 : :
1380 : : /* Selected SIMD clone's function info. First vector element
1381 : : is SIMD clone's function decl, followed by a pair of trees (base + step)
1382 : : for linear arguments (pair of NULLs for other arguments). */
1383 : : vec<tree> simd_clone_info;
1384 : :
1385 : : /* Classify the def of this stmt. */
1386 : : enum vect_def_type def_type;
1387 : :
1388 : : /* Whether the stmt is SLPed, loop-based vectorized, or both. */
1389 : : enum slp_vect_type slp_type;
1390 : :
1391 : : /* Interleaving and reduction chains info. */
1392 : : /* First element in the group. */
1393 : : stmt_vec_info first_element;
1394 : : /* Pointer to the next element in the group. */
1395 : : stmt_vec_info next_element;
1396 : : /* The size of the group. */
1397 : : unsigned int size;
1398 : : /* For stores, number of stores from this group seen. We vectorize the last
1399 : : one. */
1400 : : unsigned int store_count;
1401 : : /* For loads only, the gap from the previous load. For consecutive loads, GAP
1402 : : is 1. */
1403 : : unsigned int gap;
1404 : :
1405 : : /* The minimum negative dependence distance this stmt participates in
1406 : : or zero if none. */
1407 : : unsigned int min_neg_dist;
1408 : :
1409 : : /* Not all stmts in the loop need to be vectorized. e.g, the increment
1410 : : of the loop induction variable and computation of array indexes. relevant
1411 : : indicates whether the stmt needs to be vectorized. */
1412 : : enum vect_relevant relevant;
1413 : :
1414 : : /* For loads if this is a gather, for stores if this is a scatter. */
1415 : : bool gather_scatter_p;
1416 : :
1417 : : /* True if this is an access with loop-invariant stride. */
1418 : : bool strided_p;
1419 : :
1420 : : /* For both loads and stores. */
1421 : : unsigned simd_lane_access_p : 3;
1422 : :
1423 : : /* Classifies how the load or store is going to be implemented
1424 : : for loop vectorization. */
1425 : : vect_memory_access_type memory_access_type;
1426 : :
1427 : : /* For INTEGER_INDUC_COND_REDUCTION, the initial value to be used. */
1428 : : tree induc_cond_initial_val;
1429 : :
1430 : : /* If not NULL the value to be added to compute final reduction value. */
1431 : : tree reduc_epilogue_adjustment;
1432 : :
1433 : : /* On a reduction PHI the reduction type as detected by
1434 : : vect_is_simple_reduction and vectorizable_reduction. */
1435 : : enum vect_reduction_type reduc_type;
1436 : :
1437 : : /* The original reduction code, to be used in the epilogue. */
1438 : : code_helper reduc_code;
1439 : : /* An internal function we should use in the epilogue. */
1440 : : internal_fn reduc_fn;
1441 : :
1442 : : /* On a stmt participating in the reduction the index of the operand
1443 : : on the reduction SSA cycle. */
1444 : : int reduc_idx;
1445 : :
1446 : : /* On a reduction PHI the def returned by vect_force_simple_reduction.
1447 : : On the def returned by vect_force_simple_reduction the
1448 : : corresponding PHI. */
1449 : : stmt_vec_info reduc_def;
1450 : :
1451 : : /* The vector input type relevant for reduction vectorization. */
1452 : : tree reduc_vectype_in;
1453 : :
1454 : : /* The vector type for performing the actual reduction. */
1455 : : tree reduc_vectype;
1456 : :
1457 : : /* For loop reduction with multiple vectorized results (ncopies > 1), a
1458 : : lane-reducing operation participating in it may not use all of those
1459 : : results, this field specifies result index starting from which any
1460 : : following land-reducing operation would be assigned to. */
1461 : : unsigned int reduc_result_pos;
1462 : :
1463 : : /* If IS_REDUC_INFO is true and if the vector code is performing
1464 : : N scalar reductions in parallel, this variable gives the initial
1465 : : scalar values of those N reductions. */
1466 : : vec<tree> reduc_initial_values;
1467 : :
1468 : : /* If IS_REDUC_INFO is true and if the vector code is performing
1469 : : N scalar reductions in parallel, this variable gives the vectorized code's
1470 : : final (scalar) result for each of those N reductions. In other words,
1471 : : REDUC_SCALAR_RESULTS[I] replaces the original scalar code's loop-closed
1472 : : SSA PHI for reduction number I. */
1473 : : vec<tree> reduc_scalar_results;
1474 : :
1475 : : /* Only meaningful if IS_REDUC_INFO. If non-null, the reduction is
1476 : : being performed by an epilogue loop and we have decided to reuse
1477 : : this accumulator from the main loop. */
1478 : : vect_reusable_accumulator *reused_accumulator;
1479 : :
1480 : : /* Whether we force a single cycle PHI during reduction vectorization. */
1481 : : bool force_single_cycle;
1482 : :
1483 : : /* Whether on this stmt reduction meta is recorded. */
1484 : : bool is_reduc_info;
1485 : :
1486 : : /* If nonzero, the lhs of the statement could be truncated to this
1487 : : many bits without affecting any users of the result. */
1488 : : unsigned int min_output_precision;
1489 : :
1490 : : /* If nonzero, all non-boolean input operands have the same precision,
1491 : : and they could each be truncated to this many bits without changing
1492 : : the result. */
1493 : : unsigned int min_input_precision;
1494 : :
1495 : : /* If OPERATION_BITS is nonzero, the statement could be performed on
1496 : : an integer with the sign and number of bits given by OPERATION_SIGN
1497 : : and OPERATION_BITS without changing the result. */
1498 : : unsigned int operation_precision;
1499 : : signop operation_sign;
1500 : :
1501 : : /* If the statement produces a boolean result, this value describes
1502 : : how we should choose the associated vector type. The possible
1503 : : values are:
1504 : :
1505 : : - an integer precision N if we should use the vector mask type
1506 : : associated with N-bit integers. This is only used if all relevant
1507 : : input booleans also want the vector mask type for N-bit integers,
1508 : : or if we can convert them into that form by pattern-matching.
1509 : :
1510 : : - ~0U if we considered choosing a vector mask type but decided
1511 : : to treat the boolean as a normal integer type instead.
1512 : :
1513 : : - 0 otherwise. This means either that the operation isn't one that
1514 : : could have a vector mask type (and so should have a normal vector
1515 : : type instead) or that we simply haven't made a choice either way. */
1516 : : unsigned int mask_precision;
1517 : :
1518 : : /* True if this is only suitable for SLP vectorization. */
1519 : : bool slp_vect_only_p;
1520 : :
1521 : : /* True if this is a pattern that can only be handled by SLP
1522 : : vectorization. */
1523 : : bool slp_vect_pattern_only_p;
1524 : : };
1525 : :
1526 : : /* Information about a gather/scatter call. */
1527 : : struct gather_scatter_info {
1528 : : /* The internal function to use for the gather/scatter operation,
1529 : : or IFN_LAST if a built-in function should be used instead. */
1530 : : internal_fn ifn;
1531 : :
1532 : : /* The FUNCTION_DECL for the built-in gather/scatter function,
1533 : : or null if an internal function should be used instead. */
1534 : : tree decl;
1535 : :
1536 : : /* The loop-invariant base value. */
1537 : : tree base;
1538 : :
1539 : : /* The original scalar offset, which is a non-loop-invariant SSA_NAME. */
1540 : : tree offset;
1541 : :
1542 : : /* Each offset element should be multiplied by this amount before
1543 : : being added to the base. */
1544 : : int scale;
1545 : :
1546 : : /* The definition type for the vectorized offset. */
1547 : : enum vect_def_type offset_dt;
1548 : :
1549 : : /* The type of the vectorized offset. */
1550 : : tree offset_vectype;
1551 : :
1552 : : /* The type of the scalar elements after loading or before storing. */
1553 : : tree element_type;
1554 : :
1555 : : /* The type of the scalar elements being loaded or stored. */
1556 : : tree memory_type;
1557 : : };
1558 : :
1559 : : /* Access Functions. */
1560 : : #define STMT_VINFO_TYPE(S) (S)->type
1561 : : #define STMT_VINFO_STMT(S) (S)->stmt
1562 : : #define STMT_VINFO_RELEVANT(S) (S)->relevant
1563 : : #define STMT_VINFO_LIVE_P(S) (S)->live
1564 : : #define STMT_VINFO_VECTYPE(S) (S)->vectype
1565 : : #define STMT_VINFO_VEC_STMTS(S) (S)->vec_stmts
1566 : : #define STMT_VINFO_VECTORIZABLE(S) (S)->vectorizable
1567 : : #define STMT_VINFO_DATA_REF(S) ((S)->dr_aux.dr + 0)
1568 : : #define STMT_VINFO_GATHER_SCATTER_P(S) (S)->gather_scatter_p
1569 : : #define STMT_VINFO_STRIDED_P(S) (S)->strided_p
1570 : : #define STMT_VINFO_MEMORY_ACCESS_TYPE(S) (S)->memory_access_type
1571 : : #define STMT_VINFO_SIMD_LANE_ACCESS_P(S) (S)->simd_lane_access_p
1572 : : #define STMT_VINFO_VEC_INDUC_COND_INITIAL_VAL(S) (S)->induc_cond_initial_val
1573 : : #define STMT_VINFO_REDUC_EPILOGUE_ADJUSTMENT(S) (S)->reduc_epilogue_adjustment
1574 : : #define STMT_VINFO_REDUC_IDX(S) (S)->reduc_idx
1575 : : #define STMT_VINFO_FORCE_SINGLE_CYCLE(S) (S)->force_single_cycle
1576 : :
1577 : : #define STMT_VINFO_DR_WRT_VEC_LOOP(S) (S)->dr_wrt_vec_loop
1578 : : #define STMT_VINFO_DR_BASE_ADDRESS(S) (S)->dr_wrt_vec_loop.base_address
1579 : : #define STMT_VINFO_DR_INIT(S) (S)->dr_wrt_vec_loop.init
1580 : : #define STMT_VINFO_DR_OFFSET(S) (S)->dr_wrt_vec_loop.offset
1581 : : #define STMT_VINFO_DR_STEP(S) (S)->dr_wrt_vec_loop.step
1582 : : #define STMT_VINFO_DR_BASE_ALIGNMENT(S) (S)->dr_wrt_vec_loop.base_alignment
1583 : : #define STMT_VINFO_DR_BASE_MISALIGNMENT(S) \
1584 : : (S)->dr_wrt_vec_loop.base_misalignment
1585 : : #define STMT_VINFO_DR_OFFSET_ALIGNMENT(S) \
1586 : : (S)->dr_wrt_vec_loop.offset_alignment
1587 : : #define STMT_VINFO_DR_STEP_ALIGNMENT(S) \
1588 : : (S)->dr_wrt_vec_loop.step_alignment
1589 : :
1590 : : #define STMT_VINFO_DR_INFO(S) \
1591 : : (gcc_checking_assert ((S)->dr_aux.stmt == (S)), &(S)->dr_aux)
1592 : :
1593 : : #define STMT_VINFO_IN_PATTERN_P(S) (S)->in_pattern_p
1594 : : #define STMT_VINFO_RELATED_STMT(S) (S)->related_stmt
1595 : : #define STMT_VINFO_PATTERN_DEF_SEQ(S) (S)->pattern_def_seq
1596 : : #define STMT_VINFO_SIMD_CLONE_INFO(S) (S)->simd_clone_info
1597 : : #define STMT_VINFO_DEF_TYPE(S) (S)->def_type
1598 : : #define STMT_VINFO_GROUPED_ACCESS(S) \
1599 : : ((S)->dr_aux.dr && DR_GROUP_FIRST_ELEMENT(S))
1600 : : #define STMT_VINFO_LOOP_PHI_EVOLUTION_BASE_UNCHANGED(S) (S)->loop_phi_evolution_base_unchanged
1601 : : #define STMT_VINFO_LOOP_PHI_EVOLUTION_PART(S) (S)->loop_phi_evolution_part
1602 : : #define STMT_VINFO_LOOP_PHI_EVOLUTION_TYPE(S) (S)->loop_phi_evolution_type
1603 : : #define STMT_VINFO_MIN_NEG_DIST(S) (S)->min_neg_dist
1604 : : #define STMT_VINFO_REDUC_TYPE(S) (S)->reduc_type
1605 : : #define STMT_VINFO_REDUC_CODE(S) (S)->reduc_code
1606 : : #define STMT_VINFO_REDUC_FN(S) (S)->reduc_fn
1607 : : #define STMT_VINFO_REDUC_DEF(S) (S)->reduc_def
1608 : : #define STMT_VINFO_REDUC_VECTYPE(S) (S)->reduc_vectype
1609 : : #define STMT_VINFO_REDUC_VECTYPE_IN(S) (S)->reduc_vectype_in
1610 : : #define STMT_VINFO_SLP_VECT_ONLY(S) (S)->slp_vect_only_p
1611 : : #define STMT_VINFO_SLP_VECT_ONLY_PATTERN(S) (S)->slp_vect_pattern_only_p
1612 : :
1613 : : #define DR_GROUP_FIRST_ELEMENT(S) \
1614 : : (gcc_checking_assert ((S)->dr_aux.dr), (S)->first_element)
1615 : : #define DR_GROUP_NEXT_ELEMENT(S) \
1616 : : (gcc_checking_assert ((S)->dr_aux.dr), (S)->next_element)
1617 : : #define DR_GROUP_SIZE(S) \
1618 : : (gcc_checking_assert ((S)->dr_aux.dr), (S)->size)
1619 : : #define DR_GROUP_STORE_COUNT(S) \
1620 : : (gcc_checking_assert ((S)->dr_aux.dr), (S)->store_count)
1621 : : #define DR_GROUP_GAP(S) \
1622 : : (gcc_checking_assert ((S)->dr_aux.dr), (S)->gap)
1623 : :
1624 : : #define REDUC_GROUP_FIRST_ELEMENT(S) \
1625 : : (gcc_checking_assert (!(S)->dr_aux.dr), (S)->first_element)
1626 : : #define REDUC_GROUP_NEXT_ELEMENT(S) \
1627 : : (gcc_checking_assert (!(S)->dr_aux.dr), (S)->next_element)
1628 : : #define REDUC_GROUP_SIZE(S) \
1629 : : (gcc_checking_assert (!(S)->dr_aux.dr), (S)->size)
1630 : :
1631 : : #define STMT_VINFO_RELEVANT_P(S) ((S)->relevant != vect_unused_in_scope)
1632 : :
1633 : : #define HYBRID_SLP_STMT(S) ((S)->slp_type == hybrid)
1634 : : #define PURE_SLP_STMT(S) ((S)->slp_type == pure_slp)
1635 : : #define STMT_SLP_TYPE(S) (S)->slp_type
1636 : :
1637 : : /* Contains the scalar or vector costs for a vec_info. */
1638 : : class vector_costs
1639 : : {
1640 : : public:
1641 : : vector_costs (vec_info *, bool);
1642 : 0 : virtual ~vector_costs () {}
1643 : :
1644 : : /* Update the costs in response to adding COUNT copies of a statement.
1645 : :
1646 : : - WHERE specifies whether the cost occurs in the loop prologue,
1647 : : the loop body, or the loop epilogue.
1648 : : - KIND is the kind of statement, which is always meaningful.
1649 : : - STMT_INFO or NODE, if nonnull, describe the statement that will be
1650 : : vectorized.
1651 : : - VECTYPE, if nonnull, is the vector type that the vectorized
1652 : : statement will operate on. Note that this should be used in
1653 : : preference to STMT_VINFO_VECTYPE (STMT_INFO) since the latter
1654 : : is not correct for SLP.
1655 : : - for unaligned_load and unaligned_store statements, MISALIGN is
1656 : : the byte misalignment of the load or store relative to the target's
1657 : : preferred alignment for VECTYPE, or DR_MISALIGNMENT_UNKNOWN
1658 : : if the misalignment is not known.
1659 : :
1660 : : Return the calculated cost as well as recording it. The return
1661 : : value is used for dumping purposes. */
1662 : : virtual unsigned int add_stmt_cost (int count, vect_cost_for_stmt kind,
1663 : : stmt_vec_info stmt_info,
1664 : : slp_tree node,
1665 : : tree vectype, int misalign,
1666 : : vect_cost_model_location where);
1667 : :
1668 : : /* Finish calculating the cost of the code. The results can be
1669 : : read back using the functions below.
1670 : :
1671 : : If the costs describe vector code, SCALAR_COSTS gives the costs
1672 : : of the corresponding scalar code, otherwise it is null. */
1673 : : virtual void finish_cost (const vector_costs *scalar_costs);
1674 : :
1675 : : /* The costs in THIS and OTHER both describe ways of vectorizing
1676 : : a main loop. Return true if the costs described by THIS are
1677 : : cheaper than the costs described by OTHER. Return false if any
1678 : : of the following are true:
1679 : :
1680 : : - THIS and OTHER are of equal cost
1681 : : - OTHER is better than THIS
1682 : : - we can't be sure about the relative costs of THIS and OTHER. */
1683 : : virtual bool better_main_loop_than_p (const vector_costs *other) const;
1684 : :
1685 : : /* Likewise, but the costs in THIS and OTHER both describe ways of
1686 : : vectorizing an epilogue loop of MAIN_LOOP. */
1687 : : virtual bool better_epilogue_loop_than_p (const vector_costs *other,
1688 : : loop_vec_info main_loop) const;
1689 : :
1690 : : unsigned int prologue_cost () const;
1691 : : unsigned int body_cost () const;
1692 : : unsigned int epilogue_cost () const;
1693 : : unsigned int outside_cost () const;
1694 : : unsigned int total_cost () const;
1695 : : unsigned int suggested_unroll_factor () const;
1696 : : machine_mode suggested_epilogue_mode () const;
1697 : :
1698 : : protected:
1699 : : unsigned int record_stmt_cost (stmt_vec_info, vect_cost_model_location,
1700 : : unsigned int);
1701 : : unsigned int adjust_cost_for_freq (stmt_vec_info, vect_cost_model_location,
1702 : : unsigned int);
1703 : : int compare_inside_loop_cost (const vector_costs *) const;
1704 : : int compare_outside_loop_cost (const vector_costs *) const;
1705 : :
1706 : : /* The region of code that we're considering vectorizing. */
1707 : : vec_info *m_vinfo;
1708 : :
1709 : : /* True if we're costing the scalar code, false if we're costing
1710 : : the vector code. */
1711 : : bool m_costing_for_scalar;
1712 : :
1713 : : /* The costs of the three regions, indexed by vect_cost_model_location. */
1714 : : unsigned int m_costs[3];
1715 : :
1716 : : /* The suggested unrolling factor determined at finish_cost. */
1717 : : unsigned int m_suggested_unroll_factor;
1718 : :
1719 : : /* The suggested mode to be used for a vectorized epilogue or VOIDmode,
1720 : : determined at finish_cost. */
1721 : : machine_mode m_suggested_epilogue_mode;
1722 : :
1723 : : /* True if finish_cost has been called. */
1724 : : bool m_finished;
1725 : : };
1726 : :
1727 : : /* Create costs for VINFO. COSTING_FOR_SCALAR is true if the costs
1728 : : are for scalar code, false if they are for vector code. */
1729 : :
1730 : : inline
1731 : 2056426 : vector_costs::vector_costs (vec_info *vinfo, bool costing_for_scalar)
1732 : 2056426 : : m_vinfo (vinfo),
1733 : 2056426 : m_costing_for_scalar (costing_for_scalar),
1734 : 2056426 : m_costs (),
1735 : 2056426 : m_suggested_unroll_factor(1),
1736 : 2056426 : m_suggested_epilogue_mode(VOIDmode),
1737 : 2056426 : m_finished (false)
1738 : : {
1739 : : }
1740 : :
1741 : : /* Return the cost of the prologue code (in abstract units). */
1742 : :
1743 : : inline unsigned int
1744 : 889778 : vector_costs::prologue_cost () const
1745 : : {
1746 : 889778 : gcc_checking_assert (m_finished);
1747 : 889778 : return m_costs[vect_prologue];
1748 : : }
1749 : :
1750 : : /* Return the cost of the body code (in abstract units). */
1751 : :
1752 : : inline unsigned int
1753 : 1622478 : vector_costs::body_cost () const
1754 : : {
1755 : 1622478 : gcc_checking_assert (m_finished);
1756 : 1622478 : return m_costs[vect_body];
1757 : : }
1758 : :
1759 : : /* Return the cost of the epilogue code (in abstract units). */
1760 : :
1761 : : inline unsigned int
1762 : 889778 : vector_costs::epilogue_cost () const
1763 : : {
1764 : 889778 : gcc_checking_assert (m_finished);
1765 : 889778 : return m_costs[vect_epilogue];
1766 : : }
1767 : :
1768 : : /* Return the cost of the prologue and epilogue code (in abstract units). */
1769 : :
1770 : : inline unsigned int
1771 : 78539 : vector_costs::outside_cost () const
1772 : : {
1773 : 78539 : return prologue_cost () + epilogue_cost ();
1774 : : }
1775 : :
1776 : : /* Return the cost of the prologue, body and epilogue code
1777 : : (in abstract units). */
1778 : :
1779 : : inline unsigned int
1780 : 78539 : vector_costs::total_cost () const
1781 : : {
1782 : 78539 : return body_cost () + outside_cost ();
1783 : : }
1784 : :
1785 : : /* Return the suggested unroll factor. */
1786 : :
1787 : : inline unsigned int
1788 : 78539 : vector_costs::suggested_unroll_factor () const
1789 : : {
1790 : 78539 : gcc_checking_assert (m_finished);
1791 : 78539 : return m_suggested_unroll_factor;
1792 : : }
1793 : :
1794 : : /* Return the suggested epilogue mode. */
1795 : :
1796 : : inline machine_mode
1797 : 12032 : vector_costs::suggested_epilogue_mode () const
1798 : : {
1799 : 12032 : gcc_checking_assert (m_finished);
1800 : 12032 : return m_suggested_epilogue_mode;
1801 : : }
1802 : :
1803 : : #define VECT_MAX_COST 1000
1804 : :
1805 : : /* The maximum number of intermediate steps required in multi-step type
1806 : : conversion. */
1807 : : #define MAX_INTERM_CVT_STEPS 3
1808 : :
1809 : : #define MAX_VECTORIZATION_FACTOR INT_MAX
1810 : :
1811 : : /* Nonzero if TYPE represents a (scalar) boolean type or type
1812 : : in the middle-end compatible with it (unsigned precision 1 integral
1813 : : types). Used to determine which types should be vectorized as
1814 : : VECTOR_BOOLEAN_TYPE_P. */
1815 : :
1816 : : #define VECT_SCALAR_BOOLEAN_TYPE_P(TYPE) \
1817 : : (TREE_CODE (TYPE) == BOOLEAN_TYPE \
1818 : : || ((TREE_CODE (TYPE) == INTEGER_TYPE \
1819 : : || TREE_CODE (TYPE) == ENUMERAL_TYPE) \
1820 : : && TYPE_PRECISION (TYPE) == 1 \
1821 : : && TYPE_UNSIGNED (TYPE)))
1822 : :
1823 : : inline bool
1824 : 9699824 : nested_in_vect_loop_p (class loop *loop, stmt_vec_info stmt_info)
1825 : : {
1826 : 9699824 : return (loop->inner
1827 : 7727821 : && (loop->inner == (gimple_bb (stmt_info->stmt))->loop_father));
1828 : : }
1829 : :
1830 : : /* PHI is either a scalar reduction phi or a scalar induction phi.
1831 : : Return the initial value of the variable on entry to the containing
1832 : : loop. */
1833 : :
1834 : : inline tree
1835 : 34410 : vect_phi_initial_value (gphi *phi)
1836 : : {
1837 : 34410 : basic_block bb = gimple_bb (phi);
1838 : 34410 : edge pe = loop_preheader_edge (bb->loop_father);
1839 : 34410 : gcc_assert (pe->dest == bb);
1840 : 34410 : return PHI_ARG_DEF_FROM_EDGE (phi, pe);
1841 : : }
1842 : :
1843 : : /* Return true if STMT_INFO should produce a vector mask type rather than
1844 : : a normal nonmask type. */
1845 : :
1846 : : inline bool
1847 : 5866512 : vect_use_mask_type_p (stmt_vec_info stmt_info)
1848 : : {
1849 : 5866512 : return stmt_info->mask_precision && stmt_info->mask_precision != ~0U;
1850 : : }
1851 : :
1852 : : /* Return TRUE if a statement represented by STMT_INFO is a part of a
1853 : : pattern. */
1854 : :
1855 : : inline bool
1856 : 124969708 : is_pattern_stmt_p (stmt_vec_info stmt_info)
1857 : : {
1858 : 77663115 : return stmt_info->pattern_stmt_p;
1859 : : }
1860 : :
1861 : : /* If STMT_INFO is a pattern statement, return the statement that it
1862 : : replaces, otherwise return STMT_INFO itself. */
1863 : :
1864 : : inline stmt_vec_info
1865 : 51974997 : vect_orig_stmt (stmt_vec_info stmt_info)
1866 : : {
1867 : 39296719 : if (is_pattern_stmt_p (stmt_info))
1868 : 3209076 : return STMT_VINFO_RELATED_STMT (stmt_info);
1869 : : return stmt_info;
1870 : : }
1871 : :
1872 : : /* Return the later statement between STMT1_INFO and STMT2_INFO. */
1873 : :
1874 : : inline stmt_vec_info
1875 : 6232924 : get_later_stmt (stmt_vec_info stmt1_info, stmt_vec_info stmt2_info)
1876 : : {
1877 : 6232924 : gimple *stmt1 = vect_orig_stmt (stmt1_info)->stmt;
1878 : 6232924 : gimple *stmt2 = vect_orig_stmt (stmt2_info)->stmt;
1879 : 6232924 : if (gimple_bb (stmt1) == gimple_bb (stmt2))
1880 : : {
1881 : 6205418 : if (gimple_uid (stmt1) > gimple_uid (stmt2))
1882 : : return stmt1_info;
1883 : : else
1884 : : return stmt2_info;
1885 : : }
1886 : : /* ??? We should be really calling this function only with stmts
1887 : : in the same BB but we can recover if there's a domination
1888 : : relationship between them. */
1889 : 27506 : else if (dominated_by_p (CDI_DOMINATORS,
1890 : 27506 : gimple_bb (stmt1), gimple_bb (stmt2)))
1891 : : return stmt1_info;
1892 : 8534 : else if (dominated_by_p (CDI_DOMINATORS,
1893 : 8534 : gimple_bb (stmt2), gimple_bb (stmt1)))
1894 : : return stmt2_info;
1895 : 0 : gcc_unreachable ();
1896 : : }
1897 : :
1898 : : /* If STMT_INFO has been replaced by a pattern statement, return the
1899 : : replacement statement, otherwise return STMT_INFO itself. */
1900 : :
1901 : : inline stmt_vec_info
1902 : 53121115 : vect_stmt_to_vectorize (stmt_vec_info stmt_info)
1903 : : {
1904 : 53121114 : if (STMT_VINFO_IN_PATTERN_P (stmt_info))
1905 : 1788953 : return STMT_VINFO_RELATED_STMT (stmt_info);
1906 : : return stmt_info;
1907 : : }
1908 : :
1909 : : /* Return true if BB is a loop header. */
1910 : :
1911 : : inline bool
1912 : 1386077 : is_loop_header_bb_p (basic_block bb)
1913 : : {
1914 : 1386079 : if (bb == (bb->loop_father)->header)
1915 : 1375399 : return true;
1916 : :
1917 : : return false;
1918 : : }
1919 : :
1920 : : /* Return pow2 (X). */
1921 : :
1922 : : inline int
1923 : : vect_pow2 (int x)
1924 : : {
1925 : : int i, res = 1;
1926 : :
1927 : : for (i = 0; i < x; i++)
1928 : : res *= 2;
1929 : :
1930 : : return res;
1931 : : }
1932 : :
1933 : : /* Alias targetm.vectorize.builtin_vectorization_cost. */
1934 : :
1935 : : inline int
1936 : 10043100 : builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost,
1937 : : tree vectype, int misalign)
1938 : : {
1939 : 10043100 : return targetm.vectorize.builtin_vectorization_cost (type_of_cost,
1940 : : vectype, misalign);
1941 : : }
1942 : :
1943 : : /* Get cost by calling cost target builtin. */
1944 : :
1945 : : inline
1946 : 26 : int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost)
1947 : : {
1948 : 47278 : return builtin_vectorization_cost (type_of_cost, NULL, 0);
1949 : : }
1950 : :
1951 : : /* Alias targetm.vectorize.init_cost. */
1952 : :
1953 : : inline vector_costs *
1954 : 2056426 : init_cost (vec_info *vinfo, bool costing_for_scalar)
1955 : : {
1956 : 2056426 : return targetm.vectorize.create_costs (vinfo, costing_for_scalar);
1957 : : }
1958 : :
1959 : : extern void dump_stmt_cost (FILE *, int, enum vect_cost_for_stmt,
1960 : : stmt_vec_info, slp_tree, tree, int, unsigned,
1961 : : enum vect_cost_model_location);
1962 : :
1963 : : /* Dump and add costs. */
1964 : :
1965 : : inline unsigned
1966 : 6961783 : add_stmt_cost (vector_costs *costs, int count,
1967 : : enum vect_cost_for_stmt kind,
1968 : : stmt_vec_info stmt_info, slp_tree node,
1969 : : tree vectype, int misalign,
1970 : : enum vect_cost_model_location where)
1971 : : {
1972 : 6961783 : unsigned cost = costs->add_stmt_cost (count, kind, stmt_info, node, vectype,
1973 : : misalign, where);
1974 : 6961783 : if (dump_file && (dump_flags & TDF_DETAILS))
1975 : 228739 : dump_stmt_cost (dump_file, count, kind, stmt_info, node, vectype, misalign,
1976 : : cost, where);
1977 : 6961783 : return cost;
1978 : : }
1979 : :
1980 : : inline unsigned
1981 : 56212 : add_stmt_cost (vector_costs *costs, int count, enum vect_cost_for_stmt kind,
1982 : : enum vect_cost_model_location where)
1983 : : {
1984 : 56212 : gcc_assert (kind == cond_branch_taken || kind == cond_branch_not_taken
1985 : : || kind == scalar_stmt);
1986 : 56212 : return add_stmt_cost (costs, count, kind, NULL, NULL, NULL_TREE, 0, where);
1987 : : }
1988 : :
1989 : : inline unsigned
1990 : 4268934 : add_stmt_cost (vector_costs *costs, stmt_info_for_cost *i)
1991 : : {
1992 : 4268934 : return add_stmt_cost (costs, i->count, i->kind, i->stmt_info, i->node,
1993 : 4268934 : i->vectype, i->misalign, i->where);
1994 : : }
1995 : :
1996 : : inline void
1997 : 455883 : add_stmt_costs (vector_costs *costs, stmt_vector_for_cost *cost_vec)
1998 : : {
1999 : 455883 : stmt_info_for_cost *cost;
2000 : 455883 : unsigned i;
2001 : 2879003 : FOR_EACH_VEC_ELT (*cost_vec, i, cost)
2002 : 2423120 : add_stmt_cost (costs, cost->count, cost->kind, cost->stmt_info,
2003 : : cost->node, cost->vectype, cost->misalign, cost->where);
2004 : 455883 : }
2005 : :
2006 : : /*-----------------------------------------------------------------*/
2007 : : /* Info on data references alignment. */
2008 : : /*-----------------------------------------------------------------*/
2009 : : #define DR_MISALIGNMENT_UNKNOWN (-1)
2010 : : #define DR_MISALIGNMENT_UNINITIALIZED (-2)
2011 : :
2012 : : inline void
2013 : 2571518 : set_dr_misalignment (dr_vec_info *dr_info, int val)
2014 : : {
2015 : 2571518 : dr_info->misalignment = val;
2016 : : }
2017 : :
2018 : : extern int dr_misalignment (dr_vec_info *dr_info, tree vectype,
2019 : : poly_int64 offset = 0);
2020 : :
2021 : : #define SET_DR_MISALIGNMENT(DR, VAL) set_dr_misalignment (DR, VAL)
2022 : :
2023 : : /* Only defined once DR_MISALIGNMENT is defined. */
2024 : : inline const poly_uint64
2025 : 6622151 : dr_target_alignment (dr_vec_info *dr_info)
2026 : : {
2027 : 6622151 : if (STMT_VINFO_GROUPED_ACCESS (dr_info->stmt))
2028 : 4900473 : dr_info = STMT_VINFO_DR_INFO (DR_GROUP_FIRST_ELEMENT (dr_info->stmt));
2029 : 6622151 : return dr_info->target_alignment;
2030 : : }
2031 : : #define DR_TARGET_ALIGNMENT(DR) dr_target_alignment (DR)
2032 : : #define DR_SCALAR_KNOWN_BOUNDS(DR) (DR)->scalar_access_known_in_bounds
2033 : :
2034 : : /* Return if the stmt_vec_info requires peeling for alignment. */
2035 : : inline bool
2036 : 5080138 : dr_safe_speculative_read_required (stmt_vec_info stmt_info)
2037 : : {
2038 : 5080138 : dr_vec_info *dr_info;
2039 : 5080138 : if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
2040 : 2557802 : dr_info = STMT_VINFO_DR_INFO (DR_GROUP_FIRST_ELEMENT (stmt_info));
2041 : : else
2042 : 2522336 : dr_info = STMT_VINFO_DR_INFO (stmt_info);
2043 : :
2044 : 5080138 : return dr_info->safe_speculative_read_required;
2045 : : }
2046 : :
2047 : : /* Set the safe_speculative_read_required for the the stmt_vec_info, if group
2048 : : access then set on the fist element otherwise set on DR directly. */
2049 : : inline void
2050 : 190275 : dr_set_safe_speculative_read_required (stmt_vec_info stmt_info,
2051 : : bool requires_alignment)
2052 : : {
2053 : 190275 : dr_vec_info *dr_info;
2054 : 190275 : if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
2055 : 53642 : dr_info = STMT_VINFO_DR_INFO (DR_GROUP_FIRST_ELEMENT (stmt_info));
2056 : : else
2057 : 136633 : dr_info = STMT_VINFO_DR_INFO (stmt_info);
2058 : :
2059 : 190275 : dr_info->safe_speculative_read_required = requires_alignment;
2060 : 190275 : }
2061 : :
2062 : : inline void
2063 : 1557240 : set_dr_target_alignment (dr_vec_info *dr_info, poly_uint64 val)
2064 : : {
2065 : 1557240 : dr_info->target_alignment = val;
2066 : : }
2067 : : #define SET_DR_TARGET_ALIGNMENT(DR, VAL) set_dr_target_alignment (DR, VAL)
2068 : :
2069 : : /* Return true if data access DR_INFO is aligned to the targets
2070 : : preferred alignment for VECTYPE (which may be less than a full vector). */
2071 : :
2072 : : inline bool
2073 : 308272 : aligned_access_p (dr_vec_info *dr_info, tree vectype)
2074 : : {
2075 : 308272 : return (dr_misalignment (dr_info, vectype) == 0);
2076 : : }
2077 : :
2078 : : /* Return TRUE if the (mis-)alignment of the data access is known with
2079 : : respect to the targets preferred alignment for VECTYPE, and FALSE
2080 : : otherwise. */
2081 : :
2082 : : inline bool
2083 : 1805365 : known_alignment_for_access_p (dr_vec_info *dr_info, tree vectype)
2084 : : {
2085 : 1618982 : return (dr_misalignment (dr_info, vectype) != DR_MISALIGNMENT_UNKNOWN);
2086 : : }
2087 : :
2088 : : /* Return the minimum alignment in bytes that the vectorized version
2089 : : of DR_INFO is guaranteed to have. */
2090 : :
2091 : : inline unsigned int
2092 : 228741 : vect_known_alignment_in_bytes (dr_vec_info *dr_info, tree vectype,
2093 : : poly_int64 offset = 0)
2094 : : {
2095 : 228741 : int misalignment = dr_misalignment (dr_info, vectype, offset);
2096 : 228741 : if (misalignment == DR_MISALIGNMENT_UNKNOWN)
2097 : 109071 : return TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr_info->dr)));
2098 : 119670 : else if (misalignment == 0)
2099 : 90325 : return known_alignment (DR_TARGET_ALIGNMENT (dr_info));
2100 : 29345 : return misalignment & -misalignment;
2101 : : }
2102 : :
2103 : : /* Return the behavior of DR_INFO with respect to the vectorization context
2104 : : (which for outer loop vectorization might not be the behavior recorded
2105 : : in DR_INFO itself). */
2106 : :
2107 : : inline innermost_loop_behavior *
2108 : 6138844 : vect_dr_behavior (vec_info *vinfo, dr_vec_info *dr_info)
2109 : : {
2110 : 6138844 : stmt_vec_info stmt_info = dr_info->stmt;
2111 : 6138844 : loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo);
2112 : 1979662 : if (loop_vinfo == NULL
2113 : 1979662 : || !nested_in_vect_loop_p (LOOP_VINFO_LOOP (loop_vinfo), stmt_info))
2114 : 6134473 : return &DR_INNERMOST (dr_info->dr);
2115 : : else
2116 : 4371 : return &STMT_VINFO_DR_WRT_VEC_LOOP (stmt_info);
2117 : : }
2118 : :
2119 : : /* Return the offset calculated by adding the offset of this DR_INFO to the
2120 : : corresponding data_reference's offset. If CHECK_OUTER then use
2121 : : vect_dr_behavior to select the appropriate data_reference to use. */
2122 : :
2123 : : inline tree
2124 : 720641 : get_dr_vinfo_offset (vec_info *vinfo,
2125 : : dr_vec_info *dr_info, bool check_outer = false)
2126 : : {
2127 : 720641 : innermost_loop_behavior *base;
2128 : 720641 : if (check_outer)
2129 : 681004 : base = vect_dr_behavior (vinfo, dr_info);
2130 : : else
2131 : 39637 : base = &dr_info->dr->innermost;
2132 : :
2133 : 720641 : tree offset = base->offset;
2134 : :
2135 : 720641 : if (!dr_info->offset)
2136 : : return offset;
2137 : :
2138 : 18335 : offset = fold_convert (sizetype, offset);
2139 : 18335 : return fold_build2 (PLUS_EXPR, TREE_TYPE (dr_info->offset), offset,
2140 : : dr_info->offset);
2141 : : }
2142 : :
2143 : :
2144 : : /* Return the vect cost model for LOOP. */
2145 : : inline enum vect_cost_model
2146 : 1872786 : loop_cost_model (loop_p loop)
2147 : : {
2148 : 1872786 : if (loop != NULL
2149 : 1140386 : && loop->force_vectorize
2150 : 70698 : && flag_simd_cost_model != VECT_COST_MODEL_DEFAULT)
2151 : : return flag_simd_cost_model;
2152 : 1802088 : return flag_vect_cost_model;
2153 : : }
2154 : :
2155 : : /* Return true if the vect cost model is unlimited. */
2156 : : inline bool
2157 : 1332205 : unlimited_cost_model (loop_p loop)
2158 : : {
2159 : 1332205 : return loop_cost_model (loop) == VECT_COST_MODEL_UNLIMITED;
2160 : : }
2161 : :
2162 : : /* Return true if the loop described by LOOP_VINFO is fully-masked and
2163 : : if the first iteration should use a partial mask in order to achieve
2164 : : alignment. */
2165 : :
2166 : : inline bool
2167 : 203819 : vect_use_loop_mask_for_alignment_p (loop_vec_info loop_vinfo)
2168 : : {
2169 : : /* With early break vectorization we don't know whether the accesses will stay
2170 : : inside the loop or not. TODO: The early break adjustment code can be
2171 : : implemented the same way as vectorizable_linear_induction. However we
2172 : : can't test this today so reject it. */
2173 : 46 : return (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)
2174 : 46 : && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo)
2175 : 203823 : && !(LOOP_VINFO_NON_LINEAR_IV (loop_vinfo)
2176 : 0 : && LOOP_VINFO_EARLY_BREAKS (loop_vinfo)));
2177 : : }
2178 : :
2179 : : /* Return the number of vectors of type VECTYPE that are needed to get
2180 : : NUNITS elements. NUNITS should be based on the vectorization factor,
2181 : : so it is always a known multiple of the number of elements in VECTYPE. */
2182 : :
2183 : : inline unsigned int
2184 : 3765540 : vect_get_num_vectors (poly_uint64 nunits, tree vectype)
2185 : : {
2186 : 3765540 : return exact_div (nunits, TYPE_VECTOR_SUBPARTS (vectype)).to_constant ();
2187 : : }
2188 : :
2189 : : /* Return the number of vectors in the context of vectorization region VINFO,
2190 : : needed for a group of statements, whose size is specified by lanes of NODE,
2191 : : if NULL, it is 1. The statements are supposed to be interleaved together
2192 : : with no gap, and all operate on vectors of type VECTYPE, if NULL, the
2193 : : vectype of NODE is used. */
2194 : :
2195 : : inline unsigned int
2196 : 3765540 : vect_get_num_copies (vec_info *vinfo, slp_tree node, tree vectype = NULL)
2197 : : {
2198 : 3765540 : poly_uint64 vf;
2199 : :
2200 : 3797000 : if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
2201 : 1593085 : vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
2202 : : else
2203 : : vf = 1;
2204 : :
2205 : 3765540 : if (node)
2206 : : {
2207 : 3749810 : vf *= SLP_TREE_LANES (node);
2208 : 3749810 : if (!vectype)
2209 : 3736878 : vectype = SLP_TREE_VECTYPE (node);
2210 : : }
2211 : :
2212 : 3765540 : return vect_get_num_vectors (vf, vectype);
2213 : : }
2214 : :
2215 : : /* Return the number of copies needed for loop vectorization when
2216 : : a statement operates on vectors of type VECTYPE. This is the
2217 : : vectorization factor divided by the number of elements in
2218 : : VECTYPE and is always known at compile time. */
2219 : :
2220 : : inline unsigned int
2221 : 15730 : vect_get_num_copies (loop_vec_info loop_vinfo, tree vectype)
2222 : : {
2223 : 15730 : return vect_get_num_copies (loop_vinfo, NULL, vectype);
2224 : : }
2225 : :
2226 : : /* Update maximum unit count *MAX_NUNITS so that it accounts for
2227 : : NUNITS. *MAX_NUNITS can be 1 if we haven't yet recorded anything. */
2228 : :
2229 : : inline void
2230 : 11832802 : vect_update_max_nunits (poly_uint64 *max_nunits, poly_uint64 nunits)
2231 : : {
2232 : : /* All unit counts have the form vec_info::vector_size * X for some
2233 : : rational X, so two unit sizes must have a common multiple.
2234 : : Everything is a multiple of the initial value of 1. */
2235 : 8073617 : *max_nunits = force_common_multiple (*max_nunits, nunits);
2236 : : }
2237 : :
2238 : : /* Update maximum unit count *MAX_NUNITS so that it accounts for
2239 : : the number of units in vector type VECTYPE. *MAX_NUNITS can be 1
2240 : : if we haven't yet recorded any vector types. */
2241 : :
2242 : : inline void
2243 : 7678079 : vect_update_max_nunits (poly_uint64 *max_nunits, tree vectype)
2244 : : {
2245 : 7678079 : vect_update_max_nunits (max_nunits, TYPE_VECTOR_SUBPARTS (vectype));
2246 : 7678079 : }
2247 : :
2248 : : /* Return the vectorization factor that should be used for costing
2249 : : purposes while vectorizing the loop described by LOOP_VINFO.
2250 : : Pick a reasonable estimate if the vectorization factor isn't
2251 : : known at compile time. */
2252 : :
2253 : : inline unsigned int
2254 : 930820 : vect_vf_for_cost (loop_vec_info loop_vinfo)
2255 : : {
2256 : 930820 : return estimated_poly_value (LOOP_VINFO_VECT_FACTOR (loop_vinfo));
2257 : : }
2258 : :
2259 : : /* Estimate the number of elements in VEC_TYPE for costing purposes.
2260 : : Pick a reasonable estimate if the exact number isn't known at
2261 : : compile time. */
2262 : :
2263 : : inline unsigned int
2264 : 30774 : vect_nunits_for_cost (tree vec_type)
2265 : : {
2266 : 30774 : return estimated_poly_value (TYPE_VECTOR_SUBPARTS (vec_type));
2267 : : }
2268 : :
2269 : : /* Return the maximum possible vectorization factor for LOOP_VINFO. */
2270 : :
2271 : : inline unsigned HOST_WIDE_INT
2272 : 31578 : vect_max_vf (loop_vec_info loop_vinfo)
2273 : : {
2274 : 31578 : unsigned HOST_WIDE_INT vf;
2275 : 31578 : if (LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (&vf))
2276 : 31578 : return vf;
2277 : : return MAX_VECTORIZATION_FACTOR;
2278 : : }
2279 : :
2280 : : /* Return the size of the value accessed by unvectorized data reference
2281 : : DR_INFO. This is only valid once STMT_VINFO_VECTYPE has been calculated
2282 : : for the associated gimple statement, since that guarantees that DR_INFO
2283 : : accesses either a scalar or a scalar equivalent. ("Scalar equivalent"
2284 : : here includes things like V1SI, which can be vectorized in the same way
2285 : : as a plain SI.) */
2286 : :
2287 : : inline unsigned int
2288 : 2520919 : vect_get_scalar_dr_size (dr_vec_info *dr_info)
2289 : : {
2290 : 2520919 : return tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_info->dr))));
2291 : : }
2292 : :
2293 : : /* Return true if LOOP_VINFO requires a runtime check for whether the
2294 : : vector loop is profitable. */
2295 : :
2296 : : inline bool
2297 : 62467 : vect_apply_runtime_profitability_check_p (loop_vec_info loop_vinfo)
2298 : : {
2299 : 62467 : unsigned int th = LOOP_VINFO_COST_MODEL_THRESHOLD (loop_vinfo);
2300 : 34056 : return (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
2301 : 62467 : && th >= vect_vf_for_cost (loop_vinfo));
2302 : : }
2303 : :
2304 : : /* Return true if CODE is a lane-reducing opcode. */
2305 : :
2306 : : inline bool
2307 : 422357 : lane_reducing_op_p (code_helper code)
2308 : : {
2309 : 422357 : return code == DOT_PROD_EXPR || code == WIDEN_SUM_EXPR || code == SAD_EXPR;
2310 : : }
2311 : :
2312 : : /* Return true if STMT is a lane-reducing statement. */
2313 : :
2314 : : inline bool
2315 : 508524 : lane_reducing_stmt_p (gimple *stmt)
2316 : : {
2317 : 508524 : if (auto *assign = dyn_cast <gassign *> (stmt))
2318 : 383216 : return lane_reducing_op_p (gimple_assign_rhs_code (assign));
2319 : : return false;
2320 : : }
2321 : :
2322 : : /* Source location + hotness information. */
2323 : : extern dump_user_location_t vect_location;
2324 : :
2325 : : /* A macro for calling:
2326 : : dump_begin_scope (MSG, vect_location);
2327 : : via an RAII object, thus printing "=== MSG ===\n" to the dumpfile etc,
2328 : : and then calling
2329 : : dump_end_scope ();
2330 : : once the object goes out of scope, thus capturing the nesting of
2331 : : the scopes.
2332 : :
2333 : : These scopes affect dump messages within them: dump messages at the
2334 : : top level implicitly default to MSG_PRIORITY_USER_FACING, whereas those
2335 : : in a nested scope implicitly default to MSG_PRIORITY_INTERNALS. */
2336 : :
2337 : : #define DUMP_VECT_SCOPE(MSG) \
2338 : : AUTO_DUMP_SCOPE (MSG, vect_location)
2339 : :
2340 : : /* A sentinel class for ensuring that the "vect_location" global gets
2341 : : reset at the end of a scope.
2342 : :
2343 : : The "vect_location" global is used during dumping and contains a
2344 : : location_t, which could contain references to a tree block via the
2345 : : ad-hoc data. This data is used for tracking inlining information,
2346 : : but it's not a GC root; it's simply assumed that such locations never
2347 : : get accessed if the blocks are optimized away.
2348 : :
2349 : : Hence we need to ensure that such locations are purged at the end
2350 : : of any operations using them (e.g. via this class). */
2351 : :
2352 : : class auto_purge_vect_location
2353 : : {
2354 : : public:
2355 : : ~auto_purge_vect_location ();
2356 : : };
2357 : :
2358 : : /*-----------------------------------------------------------------*/
2359 : : /* Function prototypes. */
2360 : : /*-----------------------------------------------------------------*/
2361 : :
2362 : : /* Simple loop peeling and versioning utilities for vectorizer's purposes -
2363 : : in tree-vect-loop-manip.cc. */
2364 : : extern void vect_set_loop_condition (class loop *, edge, loop_vec_info,
2365 : : tree, tree, tree, bool);
2366 : : extern bool slpeel_can_duplicate_loop_p (const class loop *, const_edge,
2367 : : const_edge);
2368 : : class loop *slpeel_tree_duplicate_loop_to_edge_cfg (class loop *, edge,
2369 : : class loop *, edge,
2370 : : edge, edge *, bool = true,
2371 : : vec<basic_block> * = NULL);
2372 : : class loop *vect_loop_versioning (loop_vec_info, gimple *);
2373 : : extern class loop *vect_do_peeling (loop_vec_info, tree, tree,
2374 : : tree *, tree *, tree *, int, bool, bool,
2375 : : tree *);
2376 : : extern tree vect_get_main_loop_result (loop_vec_info, tree, tree);
2377 : : extern void vect_prepare_for_masked_peels (loop_vec_info);
2378 : : extern dump_user_location_t find_loop_location (class loop *);
2379 : : extern bool vect_can_advance_ivs_p (loop_vec_info);
2380 : : extern void vect_update_inits_of_drs (loop_vec_info, tree, tree_code);
2381 : : extern edge vec_init_loop_exit_info (class loop *);
2382 : : extern void vect_iv_increment_position (edge, gimple_stmt_iterator *, bool *);
2383 : :
2384 : : /* In tree-vect-stmts.cc. */
2385 : : extern tree get_related_vectype_for_scalar_type (machine_mode, tree,
2386 : : poly_uint64 = 0);
2387 : : extern tree get_vectype_for_scalar_type (vec_info *, tree, unsigned int = 0);
2388 : : extern tree get_vectype_for_scalar_type (vec_info *, tree, slp_tree);
2389 : : extern tree get_mask_type_for_scalar_type (vec_info *, tree, unsigned int = 0);
2390 : : extern tree get_mask_type_for_scalar_type (vec_info *, tree, slp_tree);
2391 : : extern tree get_same_sized_vectype (tree, tree);
2392 : : extern bool vect_chooses_same_modes_p (vec_info *, machine_mode);
2393 : : extern bool vect_get_loop_mask_type (loop_vec_info);
2394 : : extern bool vect_is_simple_use (tree, vec_info *, enum vect_def_type *,
2395 : : stmt_vec_info * = NULL, gimple ** = NULL);
2396 : : extern bool vect_is_simple_use (tree, vec_info *, enum vect_def_type *,
2397 : : tree *, stmt_vec_info * = NULL,
2398 : : gimple ** = NULL);
2399 : : extern bool vect_is_simple_use (vec_info *, stmt_vec_info, slp_tree,
2400 : : unsigned, tree *, slp_tree *,
2401 : : enum vect_def_type *,
2402 : : tree *, stmt_vec_info * = NULL);
2403 : : extern bool vect_maybe_update_slp_op_vectype (slp_tree, tree);
2404 : : extern tree perm_mask_for_reverse (tree);
2405 : : extern bool supportable_widening_operation (vec_info*, code_helper,
2406 : : stmt_vec_info, tree, tree,
2407 : : code_helper*, code_helper*,
2408 : : int*, vec<tree> *);
2409 : : extern bool supportable_narrowing_operation (code_helper, tree, tree,
2410 : : code_helper *, int *,
2411 : : vec<tree> *);
2412 : : extern bool supportable_indirect_convert_operation (code_helper,
2413 : : tree, tree,
2414 : : vec<std::pair<tree, tree_code> > &,
2415 : : tree = NULL_TREE,
2416 : : slp_tree = NULL);
2417 : : extern int compare_step_with_zero (vec_info *, stmt_vec_info);
2418 : :
2419 : : extern unsigned record_stmt_cost (stmt_vector_for_cost *, int,
2420 : : enum vect_cost_for_stmt, stmt_vec_info,
2421 : : tree, int, enum vect_cost_model_location);
2422 : : extern unsigned record_stmt_cost (stmt_vector_for_cost *, int,
2423 : : enum vect_cost_for_stmt, slp_tree,
2424 : : tree, int, enum vect_cost_model_location);
2425 : : extern unsigned record_stmt_cost (stmt_vector_for_cost *, int,
2426 : : enum vect_cost_for_stmt,
2427 : : enum vect_cost_model_location);
2428 : : extern unsigned record_stmt_cost (stmt_vector_for_cost *, int,
2429 : : enum vect_cost_for_stmt, stmt_vec_info,
2430 : : slp_tree, tree, int,
2431 : : enum vect_cost_model_location);
2432 : :
2433 : : /* Overload of record_stmt_cost with VECTYPE derived from STMT_INFO. */
2434 : :
2435 : : inline unsigned
2436 : 3298294 : record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count,
2437 : : enum vect_cost_for_stmt kind, stmt_vec_info stmt_info,
2438 : : int misalign, enum vect_cost_model_location where)
2439 : : {
2440 : 3138962 : return record_stmt_cost (body_cost_vec, count, kind, stmt_info,
2441 : 1314454 : STMT_VINFO_VECTYPE (stmt_info), misalign, where);
2442 : : }
2443 : :
2444 : : /* Overload of record_stmt_cost with VECTYPE derived from SLP node. */
2445 : :
2446 : : inline unsigned
2447 : 914561 : record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count,
2448 : : enum vect_cost_for_stmt kind, slp_tree node,
2449 : : int misalign, enum vect_cost_model_location where)
2450 : : {
2451 : 884637 : return record_stmt_cost (body_cost_vec, count, kind, node,
2452 : : SLP_TREE_VECTYPE (node), misalign, where);
2453 : : }
2454 : :
2455 : : extern void vect_finish_replace_stmt (vec_info *, stmt_vec_info, gimple *);
2456 : : extern void vect_finish_stmt_generation (vec_info *, stmt_vec_info, gimple *,
2457 : : gimple_stmt_iterator *);
2458 : : extern opt_result vect_mark_stmts_to_be_vectorized (loop_vec_info, bool *);
2459 : : extern tree vect_get_store_rhs (stmt_vec_info);
2460 : : void vect_get_vec_defs_for_operand (vec_info *vinfo, stmt_vec_info, unsigned,
2461 : : tree op, vec<tree> *, tree = NULL);
2462 : : void vect_get_vec_defs (vec_info *, stmt_vec_info, slp_tree, unsigned,
2463 : : tree, vec<tree> *,
2464 : : tree = NULL, vec<tree> * = NULL,
2465 : : tree = NULL, vec<tree> * = NULL,
2466 : : tree = NULL, vec<tree> * = NULL);
2467 : : void vect_get_vec_defs (vec_info *, stmt_vec_info, slp_tree, unsigned,
2468 : : tree, tree, vec<tree> *,
2469 : : tree = NULL, tree = NULL, vec<tree> * = NULL,
2470 : : tree = NULL, tree = NULL, vec<tree> * = NULL,
2471 : : tree = NULL, tree = NULL, vec<tree> * = NULL);
2472 : : extern tree vect_init_vector (vec_info *, stmt_vec_info, tree, tree,
2473 : : gimple_stmt_iterator *);
2474 : : extern tree vect_get_slp_vect_def (slp_tree, unsigned);
2475 : : extern bool vect_transform_stmt (vec_info *, stmt_vec_info,
2476 : : gimple_stmt_iterator *,
2477 : : slp_tree, slp_instance);
2478 : : extern void vect_remove_stores (vec_info *, stmt_vec_info);
2479 : : extern bool vect_nop_conversion_p (stmt_vec_info);
2480 : : extern opt_result vect_analyze_stmt (vec_info *, stmt_vec_info, bool *,
2481 : : slp_tree,
2482 : : slp_instance, stmt_vector_for_cost *);
2483 : : extern void vect_get_load_cost (vec_info *, stmt_vec_info, slp_tree, int,
2484 : : dr_alignment_support, int, bool,
2485 : : unsigned int *, unsigned int *,
2486 : : stmt_vector_for_cost *,
2487 : : stmt_vector_for_cost *, bool);
2488 : : extern void vect_get_store_cost (vec_info *, stmt_vec_info, slp_tree, int,
2489 : : dr_alignment_support, int,
2490 : : unsigned int *, stmt_vector_for_cost *);
2491 : : extern bool vect_supportable_shift (vec_info *, enum tree_code, tree);
2492 : : extern tree vect_gen_perm_mask_any (tree, const vec_perm_indices &);
2493 : : extern tree vect_gen_perm_mask_checked (tree, const vec_perm_indices &);
2494 : : extern void optimize_mask_stores (class loop*);
2495 : : extern tree vect_gen_while (gimple_seq *, tree, tree, tree,
2496 : : const char * = nullptr);
2497 : : extern tree vect_gen_while_not (gimple_seq *, tree, tree, tree);
2498 : : extern opt_result vect_get_vector_types_for_stmt (vec_info *,
2499 : : stmt_vec_info, tree *,
2500 : : tree *, unsigned int = 0);
2501 : : extern opt_tree vect_get_mask_type_for_stmt (stmt_vec_info, unsigned int = 0);
2502 : :
2503 : : /* In tree-if-conv.cc. */
2504 : : extern bool ref_within_array_bound (gimple *, tree);
2505 : :
2506 : : /* In tree-vect-data-refs.cc. */
2507 : : extern bool vect_can_force_dr_alignment_p (const_tree, poly_uint64);
2508 : : extern enum dr_alignment_support vect_supportable_dr_alignment
2509 : : (vec_info *, dr_vec_info *, tree, int);
2510 : : extern tree vect_get_smallest_scalar_type (stmt_vec_info, tree);
2511 : : extern opt_result vect_analyze_data_ref_dependences (loop_vec_info, unsigned int *);
2512 : : extern bool vect_slp_analyze_instance_dependence (vec_info *, slp_instance);
2513 : : extern opt_result vect_enhance_data_refs_alignment (loop_vec_info);
2514 : : extern opt_result vect_analyze_data_refs_alignment (loop_vec_info);
2515 : : extern bool vect_slp_analyze_instance_alignment (vec_info *, slp_instance);
2516 : : extern opt_result vect_analyze_data_ref_accesses (vec_info *, vec<int> *);
2517 : : extern opt_result vect_prune_runtime_alias_test_list (loop_vec_info);
2518 : : extern bool vect_gather_scatter_fn_p (vec_info *, bool, bool, tree, tree,
2519 : : tree, int, internal_fn *, tree *,
2520 : : vec<int> * = nullptr);
2521 : : extern bool vect_check_gather_scatter (stmt_vec_info, loop_vec_info,
2522 : : gather_scatter_info *,
2523 : : vec<int> * = nullptr);
2524 : : extern opt_result vect_find_stmt_data_reference (loop_p, gimple *,
2525 : : vec<data_reference_p> *,
2526 : : vec<int> *, int);
2527 : : extern opt_result vect_analyze_data_refs (vec_info *, poly_uint64 *, bool *);
2528 : : extern void vect_record_base_alignments (vec_info *);
2529 : : extern tree vect_create_data_ref_ptr (vec_info *,
2530 : : stmt_vec_info, tree, class loop *, tree,
2531 : : tree *, gimple_stmt_iterator *,
2532 : : gimple **, bool,
2533 : : tree = NULL_TREE);
2534 : : extern tree bump_vector_ptr (vec_info *, tree, gimple *, gimple_stmt_iterator *,
2535 : : stmt_vec_info, tree);
2536 : : extern void vect_copy_ref_info (tree, tree);
2537 : : extern tree vect_create_destination_var (tree, tree);
2538 : : extern bool vect_grouped_store_supported (tree, unsigned HOST_WIDE_INT);
2539 : : extern internal_fn vect_store_lanes_supported (tree, unsigned HOST_WIDE_INT, bool);
2540 : : extern bool vect_grouped_load_supported (tree, bool, unsigned HOST_WIDE_INT);
2541 : : extern internal_fn vect_load_lanes_supported (tree, unsigned HOST_WIDE_INT,
2542 : : bool, vec<int> * = nullptr);
2543 : : extern void vect_permute_store_chain (vec_info *, vec<tree> &,
2544 : : unsigned int, stmt_vec_info,
2545 : : gimple_stmt_iterator *, vec<tree> *);
2546 : : extern tree vect_setup_realignment (vec_info *,
2547 : : stmt_vec_info, gimple_stmt_iterator *,
2548 : : tree *, enum dr_alignment_support, tree,
2549 : : class loop **);
2550 : : extern void vect_transform_grouped_load (vec_info *, stmt_vec_info, vec<tree>,
2551 : : int, gimple_stmt_iterator *);
2552 : : extern void vect_record_grouped_load_vectors (vec_info *,
2553 : : stmt_vec_info, vec<tree>);
2554 : : extern tree vect_get_new_vect_var (tree, enum vect_var_kind, const char *);
2555 : : extern tree vect_get_new_ssa_name (tree, enum vect_var_kind,
2556 : : const char * = NULL);
2557 : : extern tree vect_create_addr_base_for_vector_ref (vec_info *,
2558 : : stmt_vec_info, gimple_seq *,
2559 : : tree);
2560 : :
2561 : : /* In tree-vect-loop.cc. */
2562 : : extern tree neutral_op_for_reduction (tree, code_helper, tree, bool = true);
2563 : : extern widest_int vect_iv_limit_for_partial_vectors (loop_vec_info loop_vinfo);
2564 : : bool vect_rgroup_iv_might_wrap_p (loop_vec_info, rgroup_controls *);
2565 : : /* Used in tree-vect-loop-manip.cc */
2566 : : extern opt_result vect_determine_partial_vectors_and_peeling (loop_vec_info);
2567 : : /* Used in gimple-loop-interchange.c and tree-parloops.cc. */
2568 : : extern bool check_reduction_path (dump_user_location_t, loop_p, gphi *, tree,
2569 : : enum tree_code);
2570 : : extern bool needs_fold_left_reduction_p (tree, code_helper);
2571 : : /* Drive for loop analysis stage. */
2572 : : extern opt_loop_vec_info vect_analyze_loop (class loop *, gimple *,
2573 : : vec_info_shared *);
2574 : : extern tree vect_build_loop_niters (loop_vec_info, bool * = NULL);
2575 : : extern void vect_gen_vector_loop_niters (loop_vec_info, tree, tree *,
2576 : : tree *, bool);
2577 : : extern tree vect_halve_mask_nunits (tree, machine_mode);
2578 : : extern tree vect_double_mask_nunits (tree, machine_mode);
2579 : : extern void vect_record_loop_mask (loop_vec_info, vec_loop_masks *,
2580 : : unsigned int, tree, tree);
2581 : : extern tree vect_get_loop_mask (loop_vec_info, gimple_stmt_iterator *,
2582 : : vec_loop_masks *,
2583 : : unsigned int, tree, unsigned int);
2584 : : extern void vect_record_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
2585 : : tree, unsigned int);
2586 : : extern tree vect_get_loop_len (loop_vec_info, gimple_stmt_iterator *,
2587 : : vec_loop_lens *, unsigned int, tree,
2588 : : unsigned int, unsigned int);
2589 : : extern tree vect_gen_loop_len_mask (loop_vec_info, gimple_stmt_iterator *,
2590 : : gimple_stmt_iterator *, vec_loop_lens *,
2591 : : unsigned int, tree, tree, unsigned int,
2592 : : unsigned int);
2593 : : extern gimple_seq vect_gen_len (tree, tree, tree, tree);
2594 : : extern stmt_vec_info info_for_reduction (vec_info *, stmt_vec_info);
2595 : : extern bool reduction_fn_for_scalar_code (code_helper, internal_fn *);
2596 : :
2597 : : /* Drive for loop transformation stage. */
2598 : : extern class loop *vect_transform_loop (loop_vec_info, gimple *);
2599 : 913446 : struct vect_loop_form_info
2600 : : {
2601 : : tree number_of_iterations;
2602 : : tree number_of_iterationsm1;
2603 : : tree assumptions;
2604 : : auto_vec<gcond *> conds;
2605 : : gcond *inner_loop_cond;
2606 : : edge loop_exit;
2607 : : };
2608 : : extern opt_result vect_analyze_loop_form (class loop *, gimple *,
2609 : : vect_loop_form_info *);
2610 : : extern loop_vec_info vect_create_loop_vinfo (class loop *, vec_info_shared *,
2611 : : const vect_loop_form_info *,
2612 : : loop_vec_info = nullptr);
2613 : : extern bool vectorizable_live_operation (vec_info *, stmt_vec_info,
2614 : : slp_tree, slp_instance, int,
2615 : : bool, stmt_vector_for_cost *);
2616 : : extern bool vectorizable_lane_reducing (loop_vec_info, stmt_vec_info,
2617 : : slp_tree, stmt_vector_for_cost *);
2618 : : extern bool vectorizable_reduction (loop_vec_info, stmt_vec_info,
2619 : : slp_tree, slp_instance,
2620 : : stmt_vector_for_cost *);
2621 : : extern bool vectorizable_induction (loop_vec_info, stmt_vec_info,
2622 : : gimple **, slp_tree,
2623 : : stmt_vector_for_cost *);
2624 : : extern bool vect_transform_reduction (loop_vec_info, stmt_vec_info,
2625 : : gimple_stmt_iterator *,
2626 : : slp_tree);
2627 : : extern bool vect_transform_cycle_phi (loop_vec_info, stmt_vec_info,
2628 : : slp_tree, slp_instance);
2629 : : extern bool vectorizable_lc_phi (loop_vec_info, stmt_vec_info, slp_tree);
2630 : : extern bool vect_transform_lc_phi (loop_vec_info, stmt_vec_info, slp_tree);
2631 : : extern bool vectorizable_phi (vec_info *, stmt_vec_info, gimple **, slp_tree,
2632 : : stmt_vector_for_cost *);
2633 : : extern bool vectorizable_recurr (loop_vec_info, stmt_vec_info,
2634 : : gimple **, slp_tree, stmt_vector_for_cost *);
2635 : : extern bool vectorizable_early_exit (vec_info *, stmt_vec_info,
2636 : : gimple_stmt_iterator *, gimple **,
2637 : : slp_tree, stmt_vector_for_cost *);
2638 : : extern bool vect_emulated_vector_p (tree);
2639 : : extern bool vect_can_vectorize_without_simd_p (tree_code);
2640 : : extern bool vect_can_vectorize_without_simd_p (code_helper);
2641 : : extern int vect_get_known_peeling_cost (loop_vec_info, int, int *,
2642 : : stmt_vector_for_cost *,
2643 : : stmt_vector_for_cost *,
2644 : : stmt_vector_for_cost *);
2645 : : extern tree cse_and_gimplify_to_preheader (loop_vec_info, tree);
2646 : :
2647 : : /* Nonlinear induction. */
2648 : : extern tree vect_peel_nonlinear_iv_init (gimple_seq*, tree, tree,
2649 : : tree, enum vect_induction_op_type);
2650 : :
2651 : : /* In tree-vect-slp.cc. */
2652 : : extern void vect_slp_init (void);
2653 : : extern void vect_slp_fini (void);
2654 : : extern void vect_free_slp_instance (slp_instance);
2655 : : extern bool vect_transform_slp_perm_load (vec_info *, slp_tree, const vec<tree> &,
2656 : : gimple_stmt_iterator *, poly_uint64,
2657 : : bool, unsigned *,
2658 : : unsigned * = nullptr, bool = false);
2659 : : extern bool vect_slp_analyze_operations (vec_info *);
2660 : : extern void vect_schedule_slp (vec_info *, const vec<slp_instance> &);
2661 : : extern opt_result vect_analyze_slp (vec_info *, unsigned, bool);
2662 : : extern bool vect_make_slp_decision (loop_vec_info);
2663 : : extern void vect_detect_hybrid_slp (loop_vec_info);
2664 : : extern void vect_optimize_slp (vec_info *);
2665 : : extern void vect_gather_slp_loads (vec_info *);
2666 : : extern tree vect_get_slp_scalar_def (slp_tree, unsigned);
2667 : : extern void vect_get_slp_defs (slp_tree, vec<tree> *);
2668 : : extern void vect_get_slp_defs (vec_info *, slp_tree, vec<vec<tree> > *,
2669 : : unsigned n = -1U);
2670 : : extern bool vect_slp_if_converted_bb (basic_block bb, loop_p orig_loop);
2671 : : extern bool vect_slp_function (function *);
2672 : : extern stmt_vec_info vect_find_last_scalar_stmt_in_slp (slp_tree);
2673 : : extern stmt_vec_info vect_find_first_scalar_stmt_in_slp (slp_tree);
2674 : : extern bool is_simple_and_all_uses_invariant (stmt_vec_info, loop_vec_info);
2675 : : extern bool can_duplicate_and_interleave_p (vec_info *, unsigned int, tree,
2676 : : unsigned int * = NULL,
2677 : : tree * = NULL, tree * = NULL);
2678 : : extern void duplicate_and_interleave (vec_info *, gimple_seq *, tree,
2679 : : const vec<tree> &, unsigned int, vec<tree> &);
2680 : : extern int vect_get_place_in_interleaving_chain (stmt_vec_info, stmt_vec_info);
2681 : : extern slp_tree vect_create_new_slp_node (unsigned, tree_code);
2682 : : extern void vect_free_slp_tree (slp_tree);
2683 : : extern bool compatible_calls_p (gcall *, gcall *);
2684 : : extern int vect_slp_child_index_for_operand (const gimple *, int op, bool);
2685 : :
2686 : : extern tree prepare_vec_mask (loop_vec_info, tree, tree, tree,
2687 : : gimple_stmt_iterator *);
2688 : : extern tree vect_get_mask_load_else (int, tree);
2689 : :
2690 : : /* In tree-vect-patterns.cc. */
2691 : : extern void
2692 : : vect_mark_pattern_stmts (vec_info *, stmt_vec_info, gimple *, tree);
2693 : : extern bool vect_get_range_info (tree, wide_int*, wide_int*);
2694 : :
2695 : : /* Pattern recognition functions.
2696 : : Additional pattern recognition functions can (and will) be added
2697 : : in the future. */
2698 : : void vect_pattern_recog (vec_info *);
2699 : :
2700 : : /* In tree-vectorizer.cc. */
2701 : : unsigned vectorize_loops (void);
2702 : : void vect_free_loop_info_assumptions (class loop *);
2703 : : gimple *vect_loop_vectorized_call (class loop *, gcond **cond = NULL);
2704 : : bool vect_stmt_dominates_stmt_p (gimple *, gimple *);
2705 : :
2706 : : /* SLP Pattern matcher types, tree-vect-slp-patterns.cc. */
2707 : :
2708 : : /* Forward declaration of possible two operands operation that can be matched
2709 : : by the complex numbers pattern matchers. */
2710 : : enum _complex_operation : unsigned;
2711 : :
2712 : : /* All possible load permute values that could result from the partial data-flow
2713 : : analysis. */
2714 : : typedef enum _complex_perm_kinds {
2715 : : PERM_UNKNOWN,
2716 : : PERM_EVENODD,
2717 : : PERM_ODDEVEN,
2718 : : PERM_ODDODD,
2719 : : PERM_EVENEVEN,
2720 : : /* Can be combined with any other PERM values. */
2721 : : PERM_TOP
2722 : : } complex_perm_kinds_t;
2723 : :
2724 : : /* Cache from nodes to the load permutation they represent. */
2725 : : typedef hash_map <slp_tree, complex_perm_kinds_t>
2726 : : slp_tree_to_load_perm_map_t;
2727 : :
2728 : : /* Cache from nodes pair to being compatible or not. */
2729 : : typedef pair_hash <nofree_ptr_hash <_slp_tree>,
2730 : : nofree_ptr_hash <_slp_tree>> slp_node_hash;
2731 : : typedef hash_map <slp_node_hash, bool> slp_compat_nodes_map_t;
2732 : :
2733 : :
2734 : : /* Vector pattern matcher base class. All SLP pattern matchers must inherit
2735 : : from this type. */
2736 : :
2737 : : class vect_pattern
2738 : : {
2739 : : protected:
2740 : : /* The number of arguments that the IFN requires. */
2741 : : unsigned m_num_args;
2742 : :
2743 : : /* The internal function that will be used when a pattern is created. */
2744 : : internal_fn m_ifn;
2745 : :
2746 : : /* The current node being inspected. */
2747 : : slp_tree *m_node;
2748 : :
2749 : : /* The list of operands to be the children for the node produced when the
2750 : : internal function is created. */
2751 : : vec<slp_tree> m_ops;
2752 : :
2753 : : /* Default constructor where NODE is the root of the tree to inspect. */
2754 : 860 : vect_pattern (slp_tree *node, vec<slp_tree> *m_ops, internal_fn ifn)
2755 : 860 : {
2756 : 860 : this->m_ifn = ifn;
2757 : 860 : this->m_node = node;
2758 : 860 : this->m_ops.create (0);
2759 : 860 : if (m_ops)
2760 : 20 : this->m_ops.safe_splice (*m_ops);
2761 : : }
2762 : :
2763 : : public:
2764 : :
2765 : : /* Create a new instance of the pattern matcher class of the given type. */
2766 : : static vect_pattern* recognize (slp_tree_to_load_perm_map_t *,
2767 : : slp_compat_nodes_map_t *, slp_tree *);
2768 : :
2769 : : /* Build the pattern from the data collected so far. */
2770 : : virtual void build (vec_info *) = 0;
2771 : :
2772 : : /* Default destructor. */
2773 : : virtual ~vect_pattern ()
2774 : : {
2775 : : this->m_ops.release ();
2776 : : }
2777 : : };
2778 : :
2779 : : /* Function pointer to create a new pattern matcher from a generic type. */
2780 : : typedef vect_pattern* (*vect_pattern_decl_t) (slp_tree_to_load_perm_map_t *,
2781 : : slp_compat_nodes_map_t *,
2782 : : slp_tree *);
2783 : :
2784 : : /* List of supported pattern matchers. */
2785 : : extern vect_pattern_decl_t slp_patterns[];
2786 : :
2787 : : /* Number of supported pattern matchers. */
2788 : : extern size_t num__slp_patterns;
2789 : :
2790 : : /* ----------------------------------------------------------------------
2791 : : Target support routines
2792 : : -----------------------------------------------------------------------
2793 : : The following routines are provided to simplify costing decisions in
2794 : : target code. Please add more as needed. */
2795 : :
2796 : : /* Return true if an operaton of kind KIND for STMT_INFO represents
2797 : : the extraction of an element from a vector in preparation for
2798 : : storing the element to memory. */
2799 : : inline bool
2800 : : vect_is_store_elt_extraction (vect_cost_for_stmt kind, stmt_vec_info stmt_info)
2801 : : {
2802 : : return (kind == vec_to_scalar
2803 : : && STMT_VINFO_DATA_REF (stmt_info)
2804 : : && DR_IS_WRITE (STMT_VINFO_DATA_REF (stmt_info)));
2805 : : }
2806 : :
2807 : : /* Return true if STMT_INFO represents part of a reduction. */
2808 : : inline bool
2809 : : vect_is_reduction (stmt_vec_info stmt_info)
2810 : : {
2811 : : return STMT_VINFO_REDUC_IDX (stmt_info) >= 0;
2812 : : }
2813 : :
2814 : : /* Returns the memory acccess type being used to vectorize the statement. If
2815 : : SLP this is read from NODE, otherwise it's read from the STMT_VINFO. */
2816 : :
2817 : : inline vect_memory_access_type
2818 : : vect_mem_access_type (stmt_vec_info stmt_info, slp_tree node)
2819 : : {
2820 : : if (node)
2821 : : return SLP_TREE_MEMORY_ACCESS_TYPE (node);
2822 : : else
2823 : : return STMT_VINFO_MEMORY_ACCESS_TYPE (stmt_info);
2824 : : }
2825 : :
2826 : : /* If STMT_INFO describes a reduction, return the vect_reduction_type
2827 : : of the reduction it describes, otherwise return -1. */
2828 : : inline int
2829 : : vect_reduc_type (vec_info *vinfo, stmt_vec_info stmt_info)
2830 : : {
2831 : : if (loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo))
2832 : : if (STMT_VINFO_REDUC_DEF (stmt_info))
2833 : : {
2834 : : stmt_vec_info reduc_info = info_for_reduction (loop_vinfo, stmt_info);
2835 : : return int (STMT_VINFO_REDUC_TYPE (reduc_info));
2836 : : }
2837 : : return -1;
2838 : : }
2839 : :
2840 : : /* If STMT_INFO is a COND_EXPR that includes an embedded comparison, return the
2841 : : scalar type of the values being compared. Return null otherwise. */
2842 : : inline tree
2843 : : vect_embedded_comparison_type (stmt_vec_info stmt_info)
2844 : : {
2845 : : if (auto *assign = dyn_cast<gassign *> (stmt_info->stmt))
2846 : : if (gimple_assign_rhs_code (assign) == COND_EXPR)
2847 : : {
2848 : : tree cond = gimple_assign_rhs1 (assign);
2849 : : if (COMPARISON_CLASS_P (cond))
2850 : : return TREE_TYPE (TREE_OPERAND (cond, 0));
2851 : : }
2852 : : return NULL_TREE;
2853 : : }
2854 : :
2855 : : /* If STMT_INFO is a comparison or contains an embedded comparison, return the
2856 : : scalar type of the values being compared. Return null otherwise. */
2857 : : inline tree
2858 : : vect_comparison_type (stmt_vec_info stmt_info)
2859 : : {
2860 : : if (auto *assign = dyn_cast<gassign *> (stmt_info->stmt))
2861 : : if (TREE_CODE_CLASS (gimple_assign_rhs_code (assign)) == tcc_comparison)
2862 : : return TREE_TYPE (gimple_assign_rhs1 (assign));
2863 : : return vect_embedded_comparison_type (stmt_info);
2864 : : }
2865 : :
2866 : : /* Return true if STMT_INFO extends the result of a load. */
2867 : : inline bool
2868 : : vect_is_extending_load (class vec_info *vinfo, stmt_vec_info stmt_info)
2869 : : {
2870 : : /* Although this is quite large for an inline function, this part
2871 : : at least should be inline. */
2872 : : gassign *assign = dyn_cast <gassign *> (stmt_info->stmt);
2873 : : if (!assign || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (assign)))
2874 : : return false;
2875 : :
2876 : : tree rhs = gimple_assign_rhs1 (stmt_info->stmt);
2877 : : tree lhs_type = TREE_TYPE (gimple_assign_lhs (assign));
2878 : : tree rhs_type = TREE_TYPE (rhs);
2879 : : if (!INTEGRAL_TYPE_P (lhs_type)
2880 : : || !INTEGRAL_TYPE_P (rhs_type)
2881 : : || TYPE_PRECISION (lhs_type) <= TYPE_PRECISION (rhs_type))
2882 : : return false;
2883 : :
2884 : : stmt_vec_info def_stmt_info = vinfo->lookup_def (rhs);
2885 : : return (def_stmt_info
2886 : : && STMT_VINFO_DATA_REF (def_stmt_info)
2887 : : && DR_IS_READ (STMT_VINFO_DATA_REF (def_stmt_info)));
2888 : : }
2889 : :
2890 : : /* Return true if STMT_INFO is an integer truncation. */
2891 : : inline bool
2892 : : vect_is_integer_truncation (stmt_vec_info stmt_info)
2893 : : {
2894 : : gassign *assign = dyn_cast <gassign *> (stmt_info->stmt);
2895 : : if (!assign || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (assign)))
2896 : : return false;
2897 : :
2898 : : tree lhs_type = TREE_TYPE (gimple_assign_lhs (assign));
2899 : : tree rhs_type = TREE_TYPE (gimple_assign_rhs1 (assign));
2900 : : return (INTEGRAL_TYPE_P (lhs_type)
2901 : : && INTEGRAL_TYPE_P (rhs_type)
2902 : : && TYPE_PRECISION (lhs_type) < TYPE_PRECISION (rhs_type));
2903 : : }
2904 : :
2905 : : /* Build a GIMPLE_ASSIGN or GIMPLE_CALL with the tree_code,
2906 : : or internal_fn contained in ch, respectively. */
2907 : : gimple * vect_gimple_build (tree, code_helper, tree, tree = NULL_TREE);
2908 : : #endif /* GCC_TREE_VECTORIZER_H */
|
