Line data Source code
1 : /* Vectorizer
2 : Copyright (C) 2003-2026 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 peeling support to indicate how peeling for alignment can help
60 : make vectorization supported. */
61 : enum peeling_support {
62 : peeling_known_supported,
63 : peeling_maybe_supported,
64 : peeling_unsupported
65 : };
66 :
67 : /* Define type of def-use cross-iteration cycle. */
68 : enum vect_def_type {
69 : vect_uninitialized_def = 0,
70 : vect_constant_def = 1,
71 : vect_external_def,
72 : vect_internal_def,
73 : vect_induction_def,
74 : vect_reduction_def,
75 : vect_double_reduction_def,
76 : vect_nested_cycle,
77 : vect_first_order_recurrence,
78 : vect_condition_def,
79 : vect_unknown_def_type
80 : };
81 :
82 : /* Define operation type of linear/non-linear induction variable. */
83 : enum vect_induction_op_type {
84 : vect_step_op_add = 0,
85 : vect_step_op_neg,
86 : vect_step_op_mul,
87 : vect_step_op_shl,
88 : vect_step_op_shr
89 : };
90 :
91 : /* Define type of reduction. */
92 : enum vect_reduction_type {
93 : TREE_CODE_REDUCTION,
94 : COND_REDUCTION,
95 : INTEGER_INDUC_COND_REDUCTION,
96 : CONST_COND_REDUCTION,
97 :
98 : /* Retain a scalar phi and use a FOLD_EXTRACT_LAST within the loop
99 : to implement:
100 :
101 : for (int i = 0; i < VF; ++i)
102 : res = cond[i] ? val[i] : res; */
103 : EXTRACT_LAST_REDUCTION,
104 :
105 : /* Use a folding reduction within the loop to implement:
106 :
107 : for (int i = 0; i < VF; ++i)
108 : res = res OP val[i];
109 :
110 : (with no reassociation). */
111 : FOLD_LEFT_REDUCTION
112 : };
113 :
114 : #define VECTORIZABLE_CYCLE_DEF(D) (((D) == vect_reduction_def) \
115 : || ((D) == vect_double_reduction_def) \
116 : || ((D) == vect_nested_cycle))
117 :
118 : /* Structure to encapsulate information about a group of like
119 : instructions to be presented to the target cost model. */
120 : struct stmt_info_for_cost {
121 : int count;
122 : enum vect_cost_for_stmt kind;
123 : enum vect_cost_model_location where;
124 : stmt_vec_info stmt_info;
125 : slp_tree node;
126 : tree vectype;
127 : int misalign;
128 : };
129 :
130 : typedef vec<stmt_info_for_cost> stmt_vector_for_cost;
131 :
132 : /* Maps base addresses to an innermost_loop_behavior and the stmt it was
133 : derived from that gives the maximum known alignment for that base. */
134 : typedef hash_map<tree_operand_hash,
135 : std::pair<stmt_vec_info, innermost_loop_behavior *> >
136 : vec_base_alignments;
137 :
138 : /* Represents elements [START, START + LENGTH) of cyclical array OPS*
139 : (i.e. OPS repeated to give at least START + LENGTH elements) */
140 : struct vect_scalar_ops_slice
141 : {
142 : tree op (unsigned int i) const;
143 : bool all_same_p () const;
144 :
145 : vec<tree> *ops;
146 : unsigned int start;
147 : unsigned int length;
148 : };
149 :
150 : /* Return element I of the slice. */
151 : inline tree
152 2730988 : vect_scalar_ops_slice::op (unsigned int i) const
153 : {
154 5461976 : return (*ops)[(i + start) % ops->length ()];
155 : }
156 :
157 : /* Hash traits for vect_scalar_ops_slice. */
158 : struct vect_scalar_ops_slice_hash : typed_noop_remove<vect_scalar_ops_slice>
159 : {
160 : typedef vect_scalar_ops_slice value_type;
161 : typedef vect_scalar_ops_slice compare_type;
162 :
163 : static const bool empty_zero_p = true;
164 :
165 : static void mark_deleted (value_type &s) { s.length = ~0U; }
166 0 : static void mark_empty (value_type &s) { s.length = 0; }
167 429034 : static bool is_deleted (const value_type &s) { return s.length == ~0U; }
168 4105712 : static bool is_empty (const value_type &s) { return s.length == 0; }
169 : static hashval_t hash (const value_type &);
170 : static bool equal (const value_type &, const compare_type &);
171 : };
172 :
173 : /* Describes how we're going to vectorize an individual load or store,
174 : or a group of loads or stores. */
175 : enum vect_memory_access_type {
176 : VMAT_UNINITIALIZED,
177 :
178 : /* An access to an invariant address. This is used only for loads. */
179 : VMAT_INVARIANT,
180 :
181 : /* A simple contiguous access. */
182 : VMAT_CONTIGUOUS,
183 :
184 : /* A contiguous access that goes down in memory rather than up,
185 : with no additional permutation. This is used only for stores
186 : of invariants. */
187 : VMAT_CONTIGUOUS_DOWN,
188 :
189 : /* A simple contiguous access in which the elements need to be reversed
190 : after loading or before storing. */
191 : VMAT_CONTIGUOUS_REVERSE,
192 :
193 : /* An access that uses IFN_LOAD_LANES or IFN_STORE_LANES. */
194 : VMAT_LOAD_STORE_LANES,
195 :
196 : /* An access in which each scalar element is loaded or stored
197 : individually. */
198 : VMAT_ELEMENTWISE,
199 :
200 : /* A hybrid of VMAT_CONTIGUOUS and VMAT_ELEMENTWISE, used for grouped
201 : SLP accesses. Each unrolled iteration uses a contiguous load
202 : or store for the whole group, but the groups from separate iterations
203 : are combined in the same way as for VMAT_ELEMENTWISE. */
204 : VMAT_STRIDED_SLP,
205 :
206 : /* The access uses gather loads or scatter stores. */
207 : VMAT_GATHER_SCATTER_LEGACY,
208 : VMAT_GATHER_SCATTER_IFN,
209 : VMAT_GATHER_SCATTER_EMULATED
210 : };
211 :
212 : /* Returns whether MAT is any of the VMAT_GATHER_SCATTER_* kinds. */
213 :
214 : inline bool
215 6580700 : mat_gather_scatter_p (vect_memory_access_type mat)
216 : {
217 6580700 : return (mat == VMAT_GATHER_SCATTER_LEGACY
218 : || mat == VMAT_GATHER_SCATTER_IFN
219 6580700 : || mat == VMAT_GATHER_SCATTER_EMULATED);
220 : }
221 :
222 : /*-----------------------------------------------------------------*/
223 : /* Info on vectorized defs. */
224 : /*-----------------------------------------------------------------*/
225 : enum stmt_vec_info_type {
226 : undef_vec_info_type = 0,
227 : load_vec_info_type,
228 : store_vec_info_type,
229 : shift_vec_info_type,
230 : op_vec_info_type,
231 : call_vec_info_type,
232 : call_simd_clone_vec_info_type,
233 : assignment_vec_info_type,
234 : condition_vec_info_type,
235 : comparison_vec_info_type,
236 : reduc_vec_info_type,
237 : induc_vec_info_type,
238 : type_promotion_vec_info_type,
239 : type_demotion_vec_info_type,
240 : type_conversion_vec_info_type,
241 : cycle_phi_info_type,
242 : lc_phi_info_type,
243 : phi_info_type,
244 : recurr_info_type,
245 : loop_exit_ctrl_vec_info_type,
246 : permute_info_type
247 : };
248 :
249 : /************************************************************************
250 : SLP
251 : ************************************************************************/
252 : typedef vec<std::pair<unsigned, unsigned> > lane_permutation_t;
253 : typedef auto_vec<std::pair<unsigned, unsigned>, 16> auto_lane_permutation_t;
254 : typedef vec<unsigned> load_permutation_t;
255 : typedef auto_vec<unsigned, 16> auto_load_permutation_t;
256 :
257 3377843 : struct vect_data {
258 2104472 : virtual ~vect_data () = default;
259 : };
260 :
261 : /* Analysis data from vectorizable_simd_clone_call for
262 : call_simd_clone_vec_info_type. */
263 : struct vect_simd_clone_data : vect_data {
264 1851 : virtual ~vect_simd_clone_data () = default;
265 1396 : vect_simd_clone_data () = default;
266 455 : vect_simd_clone_data (vect_simd_clone_data &&other) = default;
267 :
268 : /* Selected SIMD clone and clone for in-branch. */
269 : cgraph_node *clone;
270 : cgraph_node *clone_inbranch;
271 :
272 : /* Selected SIMD clone's function info. First vector element
273 : is NULL_TREE, followed by a pair of trees (base + step)
274 : for linear arguments (pair of NULLs for other arguments). */
275 : auto_vec<tree> simd_clone_info;
276 : };
277 :
278 : /* Analysis data from vectorizable_load and vectorizable_store for
279 : load_vec_info_type and store_vec_info_type. */
280 : struct vect_load_store_data : vect_data {
281 1271520 : vect_load_store_data (vect_load_store_data &&other) = default;
282 2104472 : vect_load_store_data () = default;
283 3375233 : virtual ~vect_load_store_data () = default;
284 :
285 : vect_memory_access_type memory_access_type;
286 : dr_alignment_support alignment_support_scheme;
287 : int misalignment;
288 : internal_fn lanes_ifn; // VMAT_LOAD_STORE_LANES
289 : poly_int64 poffset;
290 : union {
291 : internal_fn ifn; // VMAT_GATHER_SCATTER_IFN
292 : tree decl; // VMAT_GATHER_SCATTER_DECL
293 : } gs;
294 : tree strided_offset_vectype; // VMAT_GATHER_SCATTER_IFN, originally strided
295 : /* Load/store type with larger element mode used for punning the vectype. */
296 : tree ls_type; // VMAT_GATHER_SCATTER_IFN, VMAT_STRIDED_SLP
297 : /* Load/store element type used for punning the vectype. Relevant when
298 : that is a vector type. */
299 : tree ls_eltype; // VMAT_STRIDED_SLP
300 : /* This is set to a supported offset vector type if we don't support the
301 : originally requested offset type, otherwise NULL.
302 : If nonzero there will be an additional offset conversion before
303 : the gather/scatter. */
304 : tree supported_offset_vectype; // VMAT_GATHER_SCATTER_IFN
305 : /* Similar for scale. Only nonzero if we don't support the requested
306 : scale. Then we need to multiply the offset vector before the
307 : gather/scatter. */
308 : int supported_scale; // VMAT_GATHER_SCATTER_IFN
309 : auto_vec<int> elsvals;
310 : /* True if the load requires a load permutation. */
311 : bool slp_perm; // SLP_TREE_LOAD_PERMUTATION
312 : unsigned n_perms; // SLP_TREE_LOAD_PERMUTATION
313 : unsigned n_loads; // SLP_TREE_LOAD_PERMUTATION
314 : /* Whether the load permutation is consecutive and simple. */
315 : bool subchain_p; // VMAT_STRIDED_SLP and VMAT_GATHER_SCATTER
316 : };
317 :
318 : /* A computation tree of an SLP instance. Each node corresponds to a group of
319 : stmts to be packed in a SIMD stmt. */
320 : struct _slp_tree {
321 : _slp_tree ();
322 : ~_slp_tree ();
323 :
324 : void push_vec_def (gimple *def);
325 8543 : void push_vec_def (tree def) { vec_defs.quick_push (def); }
326 :
327 : /* Nodes that contain def-stmts of this node statements operands. */
328 : vec<slp_tree> children;
329 :
330 : /* A group of scalar stmts to be vectorized together. */
331 : vec<stmt_vec_info> stmts;
332 : /* A group of scalar operands to be vectorized together. */
333 : vec<tree> ops;
334 : /* A set of lane indices that are live and to be code-generated from
335 : this SLP node. */
336 : vec<unsigned> live_lanes;
337 :
338 : /* The representative that should be used for analysis and
339 : code generation. */
340 : stmt_vec_info representative;
341 :
342 : struct {
343 : /* SLP cycle the node resides in, or -1. */
344 : int id;
345 : /* The SLP operand index with the edge on the SLP cycle, or -1. */
346 : int reduc_idx;
347 : } cycle_info;
348 :
349 : /* Load permutation relative to the stores, NULL if there is no
350 : permutation. */
351 : load_permutation_t load_permutation;
352 : /* Lane permutation of the operands scalar lanes encoded as pairs
353 : of { operand number, lane number }. The number of elements
354 : denotes the number of output lanes. */
355 : lane_permutation_t lane_permutation;
356 :
357 : tree vectype;
358 : /* Vectorized defs. */
359 : vec<tree> vec_defs;
360 :
361 : /* Reference count in the SLP graph. */
362 : unsigned int refcnt;
363 : /* The maximum number of vector elements for the subtree rooted
364 : at this node. */
365 : poly_uint64 max_nunits;
366 : /* The DEF type of this node. */
367 : enum vect_def_type def_type;
368 : /* The number of scalar lanes produced by this node. */
369 : unsigned int lanes;
370 : /* The operation of this node. */
371 : enum tree_code code;
372 : /* For gather/scatter memory operations the scale each offset element
373 : should be multiplied by before being added to the base. */
374 : int gs_scale;
375 : /* For gather/scatter memory operations the loop-invariant base value. */
376 : tree gs_base;
377 : /* Whether uses of this load or feeders of this store are suitable
378 : for load/store-lanes. */
379 : bool ldst_lanes;
380 : /* For BB vect, flag to indicate this load node should be vectorized
381 : as to avoid STLF fails because of related stores. */
382 : bool avoid_stlf_fail;
383 :
384 : int vertex;
385 :
386 : /* The kind of operation as determined by analysis and optional
387 : kind specific data. */
388 : enum stmt_vec_info_type type;
389 : vect_data *data;
390 :
391 : template <class T>
392 2105868 : T& get_data (T& else_) { return data ? *static_cast <T *> (data) : else_; }
393 :
394 : /* If not NULL this is a cached failed SLP discovery attempt with
395 : the lanes that failed during SLP discovery as 'false'. This is
396 : a copy of the matches array. */
397 : bool *failed;
398 :
399 : /* Allocate from slp_tree_pool. */
400 : static void *operator new (size_t);
401 :
402 : /* Return memory to slp_tree_pool. */
403 : static void operator delete (void *, size_t);
404 :
405 : /* Linked list of nodes to release when we free the slp_tree_pool. */
406 : slp_tree next_node;
407 : slp_tree prev_node;
408 : };
409 :
410 : /* The enum describes the type of operations that an SLP instance
411 : can perform. */
412 :
413 : enum slp_instance_kind {
414 : slp_inst_kind_store,
415 : slp_inst_kind_reduc_group,
416 : slp_inst_kind_reduc_chain,
417 : slp_inst_kind_bb_reduc,
418 : slp_inst_kind_ctor,
419 : slp_inst_kind_gcond
420 : };
421 :
422 : /* SLP instance is a sequence of stmts in a loop that can be packed into
423 : SIMD stmts. */
424 : typedef class _slp_instance {
425 : public:
426 : /* The root of SLP tree. */
427 : slp_tree root;
428 :
429 : /* For vector constructors, the constructor stmt that the SLP tree is built
430 : from, NULL otherwise. */
431 : vec<stmt_vec_info> root_stmts;
432 :
433 : /* For slp_inst_kind_bb_reduc the defs that were not vectorized, NULL
434 : otherwise. */
435 : vec<tree> remain_defs;
436 :
437 : /* The group of nodes that contain loads of this SLP instance. */
438 : vec<slp_tree> loads;
439 :
440 : /* The SLP node containing the reduction PHIs. */
441 : slp_tree reduc_phis;
442 :
443 : /* Vector cost of this entry to the SLP graph. */
444 : stmt_vector_for_cost cost_vec;
445 :
446 : /* If this instance is the main entry of a subgraph the set of
447 : entries into the same subgraph, including itself. */
448 : vec<_slp_instance *> subgraph_entries;
449 :
450 : /* The type of operation the SLP instance is performing. */
451 : slp_instance_kind kind;
452 :
453 : dump_user_location_t location () const;
454 : } *slp_instance;
455 :
456 :
457 : /* Access Functions. */
458 : #define SLP_INSTANCE_TREE(S) (S)->root
459 : #define SLP_INSTANCE_LOADS(S) (S)->loads
460 : #define SLP_INSTANCE_ROOT_STMTS(S) (S)->root_stmts
461 : #define SLP_INSTANCE_REMAIN_DEFS(S) (S)->remain_defs
462 : #define SLP_INSTANCE_KIND(S) (S)->kind
463 :
464 : #define SLP_TREE_CHILDREN(S) (S)->children
465 : #define SLP_TREE_SCALAR_STMTS(S) (S)->stmts
466 : #define SLP_TREE_SCALAR_OPS(S) (S)->ops
467 : #define SLP_TREE_LIVE_LANES(S) (S)->live_lanes
468 : #define SLP_TREE_REF_COUNT(S) (S)->refcnt
469 : #define SLP_TREE_VEC_DEFS(S) (S)->vec_defs
470 : #define SLP_TREE_LOAD_PERMUTATION(S) (S)->load_permutation
471 : #define SLP_TREE_LANE_PERMUTATION(S) (S)->lane_permutation
472 : #define SLP_TREE_DEF_TYPE(S) (S)->def_type
473 : #define SLP_TREE_VECTYPE(S) (S)->vectype
474 : #define SLP_TREE_REPRESENTATIVE(S) (S)->representative
475 : #define SLP_TREE_LANES(S) (S)->lanes
476 : #define SLP_TREE_CODE(S) (S)->code
477 : #define SLP_TREE_TYPE(S) (S)->type
478 : #define SLP_TREE_GS_SCALE(S) (S)->gs_scale
479 : #define SLP_TREE_GS_BASE(S) (S)->gs_base
480 : #define SLP_TREE_REDUC_IDX(S) (S)->cycle_info.reduc_idx
481 : #define SLP_TREE_PERMUTE_P(S) ((S)->code == VEC_PERM_EXPR)
482 :
483 : inline vect_memory_access_type
484 916725 : SLP_TREE_MEMORY_ACCESS_TYPE (slp_tree node)
485 : {
486 333635 : if (SLP_TREE_TYPE (node) == load_vec_info_type
487 315988 : || SLP_TREE_TYPE (node) == store_vec_info_type)
488 115419 : return static_cast<vect_load_store_data *> (node->data)->memory_access_type;
489 : return VMAT_UNINITIALIZED;
490 : }
491 :
492 : enum vect_partial_vector_style {
493 : vect_partial_vectors_none,
494 : vect_partial_vectors_while_ult,
495 : vect_partial_vectors_avx512,
496 : vect_partial_vectors_len
497 : };
498 :
499 : /* Key for map that records association between
500 : scalar conditions and corresponding loop mask, and
501 : is populated by vect_record_loop_mask. */
502 :
503 : struct scalar_cond_masked_key
504 : {
505 63836 : scalar_cond_masked_key (tree t, unsigned ncopies_)
506 63836 : : ncopies (ncopies_)
507 : {
508 63836 : get_cond_ops_from_tree (t);
509 : }
510 :
511 : void get_cond_ops_from_tree (tree);
512 :
513 : unsigned ncopies;
514 : bool inverted_p;
515 : tree_code code;
516 : tree op0;
517 : tree op1;
518 : };
519 :
520 : template<>
521 : struct default_hash_traits<scalar_cond_masked_key>
522 : {
523 : typedef scalar_cond_masked_key compare_type;
524 : typedef scalar_cond_masked_key value_type;
525 :
526 : static inline hashval_t
527 72756 : hash (value_type v)
528 : {
529 72756 : inchash::hash h;
530 72756 : h.add_int (v.code);
531 72756 : inchash::add_expr (v.op0, h, 0);
532 72756 : inchash::add_expr (v.op1, h, 0);
533 72756 : h.add_int (v.ncopies);
534 72756 : h.add_flag (v.inverted_p);
535 72756 : return h.end ();
536 : }
537 :
538 : static inline bool
539 10791 : equal (value_type existing, value_type candidate)
540 : {
541 10791 : return (existing.ncopies == candidate.ncopies
542 10575 : && existing.code == candidate.code
543 6909 : && existing.inverted_p == candidate.inverted_p
544 5229 : && operand_equal_p (existing.op0, candidate.op0, 0)
545 13934 : && operand_equal_p (existing.op1, candidate.op1, 0));
546 : }
547 :
548 : static const bool empty_zero_p = true;
549 :
550 : static inline void
551 0 : mark_empty (value_type &v)
552 : {
553 0 : v.ncopies = 0;
554 0 : v.inverted_p = false;
555 : }
556 :
557 : static inline bool
558 9303363 : is_empty (value_type v)
559 : {
560 9240696 : return v.ncopies == 0;
561 : }
562 :
563 : static inline void mark_deleted (value_type &) {}
564 :
565 : static inline bool is_deleted (const value_type &)
566 : {
567 : return false;
568 : }
569 :
570 55525 : static inline void remove (value_type &) {}
571 : };
572 :
573 : typedef hash_set<scalar_cond_masked_key> scalar_cond_masked_set_type;
574 :
575 : /* Key and map that records association between vector conditions and
576 : corresponding loop mask, and is populated by prepare_vec_mask. */
577 :
578 : typedef pair_hash<tree_operand_hash, tree_operand_hash> tree_cond_mask_hash;
579 : typedef hash_set<tree_cond_mask_hash> vec_cond_masked_set_type;
580 :
581 : /* Describes two objects whose addresses must be unequal for the vectorized
582 : loop to be valid. */
583 : typedef std::pair<tree, tree> vec_object_pair;
584 :
585 : /* Records that vectorization is only possible if abs (EXPR) >= MIN_VALUE.
586 : UNSIGNED_P is true if we can assume that abs (EXPR) == EXPR. */
587 : class vec_lower_bound {
588 : public:
589 : vec_lower_bound () {}
590 1740 : vec_lower_bound (tree e, bool u, poly_uint64 m)
591 1740 : : expr (e), unsigned_p (u), min_value (m) {}
592 :
593 : tree expr;
594 : bool unsigned_p;
595 : poly_uint64 min_value;
596 : };
597 :
598 : /* Vectorizer state shared between different analyses like vector sizes
599 : of the same CFG region. */
600 : class vec_info_shared {
601 : public:
602 : vec_info_shared();
603 : ~vec_info_shared();
604 :
605 : void save_datarefs();
606 : void check_datarefs();
607 :
608 : /* All data references. Freed by free_data_refs, so not an auto_vec. */
609 : vec<data_reference_p> datarefs;
610 : vec<data_reference> datarefs_copy;
611 :
612 : /* The loop nest in which the data dependences are computed. */
613 : auto_vec<loop_p> loop_nest;
614 :
615 : /* All data dependences. Freed by free_dependence_relations, so not
616 : an auto_vec. */
617 : vec<ddr_p> ddrs;
618 : };
619 :
620 : /* Vectorizer state common between loop and basic-block vectorization. */
621 : class vec_info {
622 : public:
623 : typedef hash_set<int_hash<machine_mode, E_VOIDmode, E_BLKmode> > mode_set;
624 : enum vec_kind { bb, loop };
625 :
626 : vec_info (vec_kind, vec_info_shared *);
627 : ~vec_info ();
628 :
629 : stmt_vec_info add_stmt (gimple *);
630 : stmt_vec_info add_pattern_stmt (gimple *, stmt_vec_info);
631 : stmt_vec_info resync_stmt_addr (gimple *);
632 : stmt_vec_info lookup_stmt (gimple *);
633 : stmt_vec_info lookup_def (tree);
634 : stmt_vec_info lookup_single_use (tree);
635 : class dr_vec_info *lookup_dr (data_reference *);
636 : void move_dr (stmt_vec_info, stmt_vec_info);
637 : void remove_stmt (stmt_vec_info);
638 : void replace_stmt (gimple_stmt_iterator *, stmt_vec_info, gimple *);
639 : void insert_on_entry (stmt_vec_info, gimple *);
640 : void insert_seq_on_entry (stmt_vec_info, gimple_seq);
641 :
642 : /* The type of vectorization. */
643 : vec_kind kind;
644 :
645 : /* Shared vectorizer state. */
646 : vec_info_shared *shared;
647 :
648 : /* The mapping of GIMPLE UID to stmt_vec_info. */
649 : vec<stmt_vec_info> stmt_vec_infos;
650 : /* Whether the above mapping is complete. */
651 : bool stmt_vec_info_ro;
652 :
653 : /* Whether we've done a transform we think OK to not update virtual
654 : SSA form. */
655 : bool any_known_not_updated_vssa;
656 :
657 : /* The SLP graph. */
658 : auto_vec<slp_instance> slp_instances;
659 :
660 : /* Maps base addresses to an innermost_loop_behavior that gives the maximum
661 : known alignment for that base. */
662 : vec_base_alignments base_alignments;
663 :
664 : /* All interleaving chains of stores, represented by the first
665 : stmt in the chain. */
666 : auto_vec<stmt_vec_info> grouped_stores;
667 :
668 : /* The set of vector modes used in the vectorized region. */
669 : mode_set used_vector_modes;
670 :
671 : /* The argument we should pass to related_vector_mode when looking up
672 : the vector mode for a scalar mode, or VOIDmode if we haven't yet
673 : made any decisions about which vector modes to use. */
674 : machine_mode vector_mode;
675 :
676 : /* The basic blocks in the vectorization region. For _loop_vec_info,
677 : the memory is internally managed, while for _bb_vec_info, it points
678 : to element space of an external auto_vec<>. This inconsistency is
679 : not a good class design pattern. TODO: improve it with an unified
680 : auto_vec<> whose lifetime is confined to vec_info object. */
681 : basic_block *bbs;
682 :
683 : /* The count of the basic blocks in the vectorization region. */
684 : unsigned int nbbs;
685 :
686 : /* Used to keep a sequence of def stmts of a pattern stmt that are loop
687 : invariant if they exists.
688 : The sequence is emitted in the loop preheader should the loop be vectorized
689 : and are reset when undoing patterns. */
690 : gimple_seq inv_pattern_def_seq;
691 :
692 : private:
693 : stmt_vec_info new_stmt_vec_info (gimple *stmt);
694 : void set_vinfo_for_stmt (gimple *, stmt_vec_info, bool = true);
695 : void free_stmt_vec_infos ();
696 : void free_stmt_vec_info (stmt_vec_info);
697 : };
698 :
699 : class _loop_vec_info;
700 : class _bb_vec_info;
701 :
702 : template<>
703 : template<>
704 : inline bool
705 383997490 : is_a_helper <_loop_vec_info *>::test (vec_info *i)
706 : {
707 383368040 : return i->kind == vec_info::loop;
708 : }
709 :
710 : template<>
711 : template<>
712 : inline bool
713 72389512 : is_a_helper <_bb_vec_info *>::test (vec_info *i)
714 : {
715 72389512 : return i->kind == vec_info::bb;
716 : }
717 :
718 : /* In general, we can divide the vector statements in a vectorized loop
719 : into related groups ("rgroups") and say that for each rgroup there is
720 : some nS such that the rgroup operates on nS values from one scalar
721 : iteration followed by nS values from the next. That is, if VF is the
722 : vectorization factor of the loop, the rgroup operates on a sequence:
723 :
724 : (1,1) (1,2) ... (1,nS) (2,1) ... (2,nS) ... (VF,1) ... (VF,nS)
725 :
726 : where (i,j) represents a scalar value with index j in a scalar
727 : iteration with index i.
728 :
729 : [ We use the term "rgroup" to emphasise that this grouping isn't
730 : necessarily the same as the grouping of statements used elsewhere.
731 : For example, if we implement a group of scalar loads using gather
732 : loads, we'll use a separate gather load for each scalar load, and
733 : thus each gather load will belong to its own rgroup. ]
734 :
735 : In general this sequence will occupy nV vectors concatenated
736 : together. If these vectors have nL lanes each, the total number
737 : of scalar values N is given by:
738 :
739 : N = nS * VF = nV * nL
740 :
741 : None of nS, VF, nV and nL are required to be a power of 2. nS and nV
742 : are compile-time constants but VF and nL can be variable (if the target
743 : supports variable-length vectors).
744 :
745 : In classical vectorization, each iteration of the vector loop would
746 : handle exactly VF iterations of the original scalar loop. However,
747 : in vector loops that are able to operate on partial vectors, a
748 : particular iteration of the vector loop might handle fewer than VF
749 : iterations of the scalar loop. The vector lanes that correspond to
750 : iterations of the scalar loop are said to be "active" and the other
751 : lanes are said to be "inactive".
752 :
753 : In such vector loops, many rgroups need to be controlled to ensure
754 : that they have no effect for the inactive lanes. Conceptually, each
755 : such rgroup needs a sequence of booleans in the same order as above,
756 : but with each (i,j) replaced by a boolean that indicates whether
757 : iteration i is active. This sequence occupies nV vector controls
758 : that again have nL lanes each. Thus the control sequence as a whole
759 : consists of VF independent booleans that are each repeated nS times.
760 :
761 : Taking mask-based approach as a partially-populated vectors example.
762 : We make the simplifying assumption that if a sequence of nV masks is
763 : suitable for one (nS,nL) pair, we can reuse it for (nS/2,nL/2) by
764 : VIEW_CONVERTing it. This holds for all current targets that support
765 : fully-masked loops. For example, suppose the scalar loop is:
766 :
767 : float *f;
768 : double *d;
769 : for (int i = 0; i < n; ++i)
770 : {
771 : f[i * 2 + 0] += 1.0f;
772 : f[i * 2 + 1] += 2.0f;
773 : d[i] += 3.0;
774 : }
775 :
776 : and suppose that vectors have 256 bits. The vectorized f accesses
777 : will belong to one rgroup and the vectorized d access to another:
778 :
779 : f rgroup: nS = 2, nV = 1, nL = 8
780 : d rgroup: nS = 1, nV = 1, nL = 4
781 : VF = 4
782 :
783 : [ In this simple example the rgroups do correspond to the normal
784 : SLP grouping scheme. ]
785 :
786 : If only the first three lanes are active, the masks we need are:
787 :
788 : f rgroup: 1 1 | 1 1 | 1 1 | 0 0
789 : d rgroup: 1 | 1 | 1 | 0
790 :
791 : Here we can use a mask calculated for f's rgroup for d's, but not
792 : vice versa.
793 :
794 : Thus for each value of nV, it is enough to provide nV masks, with the
795 : mask being calculated based on the highest nL (or, equivalently, based
796 : on the highest nS) required by any rgroup with that nV. We therefore
797 : represent the entire collection of masks as a two-level table, with the
798 : first level being indexed by nV - 1 (since nV == 0 doesn't exist) and
799 : the second being indexed by the mask index 0 <= i < nV. */
800 :
801 : /* The controls (like masks or lengths) needed by rgroups with nV vectors,
802 : according to the description above. */
803 : struct rgroup_controls {
804 : /* The largest nS for all rgroups that use these controls.
805 : For vect_partial_vectors_avx512 this is the constant nscalars_per_iter
806 : for all members of the group. */
807 : unsigned int max_nscalars_per_iter;
808 :
809 : /* For the largest nS recorded above, the loop controls divide each scalar
810 : into FACTOR equal-sized pieces. This is useful if we need to split
811 : element-based accesses into byte-based accesses.
812 : For vect_partial_vectors_avx512 this records nV instead. */
813 : unsigned int factor;
814 :
815 : /* This is a vector type with MAX_NSCALARS_PER_ITER * VF / nV elements.
816 : For mask-based controls, it is the type of the masks in CONTROLS.
817 : For length-based controls, it can be any vector type that has the
818 : specified number of elements; the type of the elements doesn't matter. */
819 : tree type;
820 :
821 : /* When there is no uniformly used LOOP_VINFO_RGROUP_COMPARE_TYPE this
822 : is the rgroup specific type used. */
823 : tree compare_type;
824 :
825 : /* A vector of nV controls, in iteration order. */
826 : vec<tree> controls;
827 :
828 : /* In case of len_load and len_store with a bias there is only one
829 : rgroup. This holds the adjusted loop length for the this rgroup. */
830 : tree bias_adjusted_ctrl;
831 : };
832 :
833 585795 : struct vec_loop_masks
834 : {
835 517543 : bool is_empty () const { return mask_set.is_empty (); }
836 :
837 : /* Set to record vectype, nvector pairs. */
838 : hash_set<pair_hash <nofree_ptr_hash <tree_node>,
839 : int_hash<unsigned, 0>>> mask_set;
840 :
841 : /* rgroup_controls used for the partial vector scheme. */
842 : auto_vec<rgroup_controls> rgc_vec;
843 : };
844 :
845 : typedef auto_vec<rgroup_controls> vec_loop_lens;
846 :
847 : typedef auto_vec<std::pair<data_reference*, tree> > drs_init_vec;
848 :
849 : /* Abstraction around info on reductions which is still in stmt_vec_info
850 : but will be duplicated or moved elsewhere. */
851 203796 : class vect_reduc_info_s
852 : {
853 : public:
854 : /* The def type of the main reduction PHI, vect_reduction_def or
855 : vect_double_reduction_def. */
856 : enum vect_def_type def_type;
857 :
858 : /* The reduction type as detected by
859 : vect_is_simple_reduction and vectorizable_reduction. */
860 : enum vect_reduction_type reduc_type;
861 :
862 : /* The original scalar reduction code, to be used in the epilogue. */
863 : code_helper reduc_code;
864 :
865 : /* A vector internal function we should use in the epilogue. */
866 : internal_fn reduc_fn;
867 :
868 : /* For loop reduction with multiple vectorized results (ncopies > 1), a
869 : lane-reducing operation participating in it may not use all of those
870 : results, this field specifies result index starting from which any
871 : following land-reducing operation would be assigned to. */
872 : unsigned int reduc_result_pos;
873 :
874 : /* Whether this represents a reduction chain. */
875 : bool is_reduc_chain;
876 :
877 : /* Whether we force a single cycle PHI during reduction vectorization. */
878 : bool force_single_cycle;
879 :
880 : /* The vector type for performing the actual reduction operation. */
881 : tree reduc_vectype;
882 :
883 : /* The vector type we should use for the final reduction in the epilogue
884 : when we reduce a mask. */
885 : tree reduc_vectype_for_mask;
886 :
887 : /* The neutral operand to use, if any. */
888 : tree neutral_op;
889 :
890 : /* For INTEGER_INDUC_COND_REDUCTION, the initial value to be used. */
891 : tree induc_cond_initial_val;
892 :
893 : /* If not NULL the value to be added to compute final reduction value. */
894 : tree reduc_epilogue_adjustment;
895 :
896 : /* If non-null, the reduction is being performed by an epilogue loop
897 : and we have decided to reuse this accumulator from the main loop. */
898 : struct vect_reusable_accumulator *reused_accumulator;
899 :
900 : /* If the vector code is performing N scalar reductions in parallel,
901 : this variable gives the initial scalar values of those N reductions. */
902 : auto_vec<tree> reduc_initial_values;
903 :
904 : /* If the vector code is performing N scalar reductions in parallel, this
905 : variable gives the vectorized code's final (scalar) result for each of
906 : those N reductions. In other words, REDUC_SCALAR_RESULTS[I] replaces
907 : the original scalar code's loop-closed SSA PHI for reduction number I. */
908 : auto_vec<tree> reduc_scalar_results;
909 : };
910 :
911 : typedef class vect_reduc_info_s *vect_reduc_info;
912 :
913 : #define VECT_REDUC_INFO_DEF_TYPE(I) ((I)->def_type)
914 : #define VECT_REDUC_INFO_TYPE(I) ((I)->reduc_type)
915 : #define VECT_REDUC_INFO_CODE(I) ((I)->reduc_code)
916 : #define VECT_REDUC_INFO_FN(I) ((I)->reduc_fn)
917 : #define VECT_REDUC_INFO_SCALAR_RESULTS(I) ((I)->reduc_scalar_results)
918 : #define VECT_REDUC_INFO_INITIAL_VALUES(I) ((I)->reduc_initial_values)
919 : #define VECT_REDUC_INFO_REUSED_ACCUMULATOR(I) ((I)->reused_accumulator)
920 : #define VECT_REDUC_INFO_INDUC_COND_INITIAL_VAL(I) ((I)->induc_cond_initial_val)
921 : #define VECT_REDUC_INFO_EPILOGUE_ADJUSTMENT(I) ((I)->reduc_epilogue_adjustment)
922 : #define VECT_REDUC_INFO_VECTYPE(I) ((I)->reduc_vectype)
923 : #define VECT_REDUC_INFO_VECTYPE_FOR_MASK(I) ((I)->reduc_vectype_for_mask)
924 : #define VECT_REDUC_INFO_FORCE_SINGLE_CYCLE(I) ((I)->force_single_cycle)
925 : #define VECT_REDUC_INFO_RESULT_POS(I) ((I)->reduc_result_pos)
926 : #define VECT_REDUC_INFO_NEUTRAL_OP(I) ((I)->neutral_op)
927 :
928 : /* Information about a reduction accumulator from the main loop that could
929 : conceivably be reused as the input to a reduction in an epilogue loop. */
930 : struct vect_reusable_accumulator {
931 : /* The final value of the accumulator, which forms the input to the
932 : reduction operation. */
933 : tree reduc_input;
934 :
935 : /* The stmt_vec_info that describes the reduction (i.e. the one for
936 : which is_reduc_info is true). */
937 : vect_reduc_info reduc_info;
938 : };
939 :
940 : /*-----------------------------------------------------------------*/
941 : /* Info on vectorized loops. */
942 : /*-----------------------------------------------------------------*/
943 : typedef class _loop_vec_info : public vec_info {
944 : public:
945 : _loop_vec_info (class loop *, vec_info_shared *);
946 : ~_loop_vec_info ();
947 :
948 : /* The loop to which this info struct refers to. */
949 : class loop *loop;
950 :
951 : /* Number of latch executions. */
952 : tree num_itersm1;
953 : /* Number of iterations. */
954 : tree num_iters;
955 : /* Number of iterations of the original loop. */
956 : tree num_iters_unchanged;
957 : /* Condition under which this loop is analyzed and versioned. */
958 : tree num_iters_assumptions;
959 :
960 : /* The cost of the vector code. */
961 : class vector_costs *vector_costs;
962 :
963 : /* The cost of the scalar code. */
964 : class vector_costs *scalar_costs;
965 :
966 : /* Threshold of number of iterations below which vectorization will not be
967 : performed. It is calculated from MIN_PROFITABLE_ITERS and
968 : param_min_vect_loop_bound. */
969 : unsigned int th;
970 :
971 : /* When applying loop versioning, the vector form should only be used
972 : if the number of scalar iterations is >= this value, on top of all
973 : the other requirements. Ignored when loop versioning is not being
974 : used. */
975 : poly_uint64 versioning_threshold;
976 :
977 : /* Unrolling factor. In case of suitable super-word parallelism
978 : it can be that no unrolling is needed, and thus this is 1. */
979 : poly_uint64 vectorization_factor;
980 :
981 : /* Gimple operand for the number of scalar iteration handed per loop
982 : iteration, and therefore how much to increment each IV by. */
983 : tree iv_increment;
984 :
985 : /* If this loop is an epilogue loop whose main loop can be skipped,
986 : MAIN_LOOP_EDGE is the edge from the main loop to this loop's
987 : preheader. SKIP_MAIN_LOOP_EDGE is then the edge that skips the
988 : main loop and goes straight to this loop's preheader.
989 :
990 : Both fields are null otherwise. */
991 : edge main_loop_edge;
992 : edge skip_main_loop_edge;
993 :
994 : /* If this loop is an epilogue loop that might be skipped after executing
995 : the main loop, this edge is the one that skips the epilogue. */
996 : edge skip_this_loop_edge;
997 :
998 : /* Reduction descriptors of this loop. Referenced to from SLP nodes
999 : by index. */
1000 : auto_vec<vect_reduc_info> reduc_infos;
1001 :
1002 : /* The vectorized form of a standard reduction replaces the original
1003 : scalar code's final result (a loop-closed SSA PHI) with the result
1004 : of a vector-to-scalar reduction operation. After vectorization,
1005 : this variable maps these vector-to-scalar results to information
1006 : about the reductions that generated them. */
1007 : hash_map<tree, vect_reusable_accumulator> reusable_accumulators;
1008 :
1009 : /* The number of times that the target suggested we unroll the vector loop
1010 : in order to promote more ILP. This value will be used to re-analyze the
1011 : loop for vectorization and if successful the value will be folded into
1012 : vectorization_factor (and therefore exactly divides
1013 : vectorization_factor). */
1014 : unsigned int suggested_unroll_factor;
1015 :
1016 : /* Maximum runtime vectorization factor, or MAX_VECTORIZATION_FACTOR
1017 : if there is no particular limit. */
1018 : unsigned HOST_WIDE_INT max_vectorization_factor;
1019 :
1020 : /* The masks that a fully-masked loop should use to avoid operating
1021 : on inactive scalars. */
1022 : vec_loop_masks masks;
1023 :
1024 : /* The lengths that a loop with length should use to avoid operating
1025 : on inactive scalars. */
1026 : vec_loop_lens lens;
1027 :
1028 : /* Set of scalar conditions that have loop mask applied. */
1029 : scalar_cond_masked_set_type scalar_cond_masked_set;
1030 :
1031 : /* Set of vector conditions that have loop mask applied. */
1032 : vec_cond_masked_set_type vec_cond_masked_set;
1033 :
1034 : /* If we are using a loop mask to align memory addresses, this variable
1035 : contains the number of vector elements that we should skip in the
1036 : first iteration of the vector loop (i.e. the number of leading
1037 : elements that should be false in the first mask). */
1038 : tree mask_skip_niters;
1039 :
1040 : /* If we are using a loop mask to align memory addresses and we're in an
1041 : early break loop then this variable contains the number of elements that
1042 : were skipped during the initial iteration of the loop. */
1043 : tree mask_skip_niters_pfa_offset;
1044 :
1045 : /* The type that the loop control IV should be converted to before
1046 : testing which of the VF scalars are active and inactive.
1047 : Only meaningful if LOOP_VINFO_USING_PARTIAL_VECTORS_P. */
1048 : tree rgroup_compare_type;
1049 :
1050 : /* For #pragma omp simd if (x) loops the x expression. If constant 0,
1051 : the loop should not be vectorized, if constant non-zero, simd_if_cond
1052 : shouldn't be set and loop vectorized normally, if SSA_NAME, the loop
1053 : should be versioned on that condition, using scalar loop if the condition
1054 : is false and vectorized loop otherwise. */
1055 : tree simd_if_cond;
1056 :
1057 : /* The type that the vector loop control IV should have when
1058 : LOOP_VINFO_USING_PARTIAL_VECTORS_P is true. */
1059 : tree rgroup_iv_type;
1060 :
1061 : /* The style used for implementing partial vectors when
1062 : LOOP_VINFO_USING_PARTIAL_VECTORS_P is true. */
1063 : vect_partial_vector_style partial_vector_style;
1064 :
1065 : /* Unknown DRs according to which loop was peeled. */
1066 : class dr_vec_info *unaligned_dr;
1067 :
1068 : /* peeling_for_alignment indicates whether peeling for alignment will take
1069 : place, and what the peeling factor should be:
1070 : peeling_for_alignment = X means:
1071 : If X=0: Peeling for alignment will not be applied.
1072 : If X>0: Peel first X iterations.
1073 : If X=-1: Generate a runtime test to calculate the number of iterations
1074 : to be peeled, using the dataref recorded in the field
1075 : unaligned_dr. */
1076 : int peeling_for_alignment;
1077 :
1078 : /* The mask used to check the alignment of pointers or arrays. */
1079 : poly_uint64 ptr_mask;
1080 :
1081 : /* The maximum speculative read amount in VLA modes for runtime check. */
1082 : poly_uint64 max_spec_read_amount;
1083 :
1084 : /* Indicates whether the loop has any non-linear IV. */
1085 : bool nonlinear_iv;
1086 :
1087 : /* Data Dependence Relations defining address ranges that are candidates
1088 : for a run-time aliasing check. */
1089 : auto_vec<ddr_p> may_alias_ddrs;
1090 :
1091 : /* Data Dependence Relations defining address ranges together with segment
1092 : lengths from which the run-time aliasing check is built. */
1093 : auto_vec<dr_with_seg_len_pair_t> comp_alias_ddrs;
1094 :
1095 : /* Check that the addresses of each pair of objects is unequal. */
1096 : auto_vec<vec_object_pair> check_unequal_addrs;
1097 :
1098 : /* List of values that are required to be nonzero. This is used to check
1099 : whether things like "x[i * n] += 1;" are safe and eventually gets added
1100 : to the checks for lower bounds below. */
1101 : auto_vec<tree> check_nonzero;
1102 :
1103 : /* List of values that need to be checked for a minimum value. */
1104 : auto_vec<vec_lower_bound> lower_bounds;
1105 :
1106 : /* Statements in the loop that have data references that are candidates for a
1107 : runtime (loop versioning) misalignment check. */
1108 : auto_vec<stmt_vec_info> may_misalign_stmts;
1109 :
1110 : /* Reduction cycles detected in the loop. Used in loop-aware SLP. */
1111 : auto_vec<stmt_vec_info> reductions;
1112 :
1113 : /* Defs that could not be analyzed such as OMP SIMD calls without
1114 : a LHS. */
1115 : auto_vec<stmt_vec_info> alternate_defs;
1116 :
1117 : /* Cost vector for a single scalar iteration. */
1118 : auto_vec<stmt_info_for_cost> scalar_cost_vec;
1119 :
1120 : /* Map of IV base/step expressions to inserted name in the preheader. */
1121 : hash_map<tree_operand_hash, tree> *ivexpr_map;
1122 :
1123 : /* Map of OpenMP "omp simd array" scan variables to corresponding
1124 : rhs of the store of the initializer. */
1125 : hash_map<tree, tree> *scan_map;
1126 :
1127 : /* The factor used to over weight those statements in an inner loop
1128 : relative to the loop being vectorized. */
1129 : unsigned int inner_loop_cost_factor;
1130 :
1131 : /* Is the loop vectorizable? */
1132 : bool vectorizable;
1133 :
1134 : /* Records whether we still have the option of vectorizing this loop
1135 : using partially-populated vectors; in other words, whether it is
1136 : still possible for one iteration of the vector loop to handle
1137 : fewer than VF scalars. */
1138 : bool can_use_partial_vectors_p;
1139 :
1140 : /* Records whether we must use niter masking for correctness reasons. */
1141 : bool must_use_partial_vectors_p;
1142 :
1143 : /* True if we've decided to use partially-populated vectors, so that
1144 : the vector loop can handle fewer than VF scalars. */
1145 : bool using_partial_vectors_p;
1146 :
1147 : /* True if we've decided to use a decrementing loop control IV that counts
1148 : scalars. This can be done for any loop that:
1149 :
1150 : (a) uses length "controls"; and
1151 : (b) can iterate more than once. */
1152 : bool using_decrementing_iv_p;
1153 :
1154 : /* True if we've decided to use output of select_vl to adjust IV of
1155 : both loop control and data reference pointer. This is only true
1156 : for single-rgroup control. */
1157 : bool using_select_vl_p;
1158 :
1159 : /* True if we've decided to use peeling with versioning together, which allows
1160 : unaligned unsupported data refs to be uniformly aligned after a certain
1161 : amount of peeling (mutual alignment). Otherwise, we use versioning alone
1162 : so these data refs must be already aligned to a power-of-two boundary
1163 : without peeling. */
1164 : bool allow_mutual_alignment;
1165 :
1166 : /* The bias for len_load and len_store. For now, only 0 and -1 are
1167 : supported. -1 must be used when a backend does not support
1168 : len_load/len_store with a length of zero. */
1169 : signed char partial_load_store_bias;
1170 :
1171 : /* When we have grouped data accesses with gaps, we may introduce invalid
1172 : memory accesses. We peel the last iteration of the loop to prevent
1173 : this. */
1174 : bool peeling_for_gaps;
1175 :
1176 : /* When the number of iterations is not a multiple of the vector size
1177 : we need to peel off iterations at the end to form an epilogue loop. */
1178 : bool peeling_for_niter;
1179 :
1180 : /* When the loop has early breaks that we can vectorize we need to peel
1181 : the loop for the break finding loop. */
1182 : bool early_breaks;
1183 :
1184 : /* List of loop additional IV conditionals found in the loop. */
1185 : auto_vec<gcond *> conds;
1186 :
1187 : /* Main loop IV cond. */
1188 : gcond* loop_iv_cond;
1189 :
1190 : /* True if we have an unroll factor requested by the user through pragma GCC
1191 : unroll. */
1192 : bool user_unroll;
1193 :
1194 : /* True if there are no loop carried data dependencies in the loop.
1195 : If loop->safelen <= 1, then this is always true, either the loop
1196 : didn't have any loop carried data dependencies, or the loop is being
1197 : vectorized guarded with some runtime alias checks, or couldn't
1198 : be vectorized at all, but then this field shouldn't be used.
1199 : For loop->safelen >= 2, the user has asserted that there are no
1200 : backward dependencies, but there still could be loop carried forward
1201 : dependencies in such loops. This flag will be false if normal
1202 : vectorizer data dependency analysis would fail or require versioning
1203 : for alias, but because of loop->safelen >= 2 it has been vectorized
1204 : even without versioning for alias. E.g. in:
1205 : #pragma omp simd
1206 : for (int i = 0; i < m; i++)
1207 : a[i] = a[i + k] * c;
1208 : (or #pragma simd or #pragma ivdep) we can vectorize this and it will
1209 : DTRT even for k > 0 && k < m, but without safelen we would not
1210 : vectorize this, so this field would be false. */
1211 : bool no_data_dependencies;
1212 :
1213 : /* Mark loops having masked stores. */
1214 : bool has_mask_store;
1215 :
1216 : /* Queued scaling factor for the scalar loop. */
1217 : profile_probability scalar_loop_scaling;
1218 :
1219 : /* If if-conversion versioned this loop before conversion, this is the
1220 : loop version without if-conversion. */
1221 : class loop *scalar_loop;
1222 :
1223 : /* For loops being epilogues of already vectorized loops
1224 : this points to the main vectorized loop. Otherwise NULL. */
1225 : _loop_vec_info *main_loop_info;
1226 :
1227 : /* For loops being epilogues of already vectorized loops
1228 : this points to the preceding vectorized (possibly epilogue) loop.
1229 : Otherwise NULL. */
1230 : _loop_vec_info *orig_loop_info;
1231 :
1232 : /* Used to store loop_vec_infos of the epilogue of this loop during
1233 : analysis. */
1234 : _loop_vec_info *epilogue_vinfo;
1235 :
1236 : /* If this is an epilogue loop the DR advancement applied. */
1237 : tree drs_advanced_by;
1238 :
1239 : /* The controlling loop exit for the current loop when vectorizing.
1240 : For counted loops, this IV controls the natural exits of the loop. */
1241 : edge vec_loop_main_exit;
1242 :
1243 : /* The controlling loop exit for the epilogue loop when vectorizing.
1244 : For counted loops, this IV controls the natural exits of the loop. */
1245 : edge vec_epilogue_loop_main_exit;
1246 :
1247 : /* The controlling loop exit for the scalar loop being vectorized.
1248 : For counted loops, this IV controls the natural exits of the loop. */
1249 : edge scalar_loop_main_exit;
1250 :
1251 : /* Indicate if the multiple exit loop has any side-effects that require it to
1252 : have a scalar epilogue. */
1253 : bool early_break_needs_epilogue;
1254 :
1255 : /* Used to store the list of stores needing to be moved if doing early
1256 : break vectorization as they would violate the scalar loop semantics if
1257 : vectorized in their current location. These are stored in order that they
1258 : need to be moved. */
1259 : auto_vec<gimple *> early_break_stores;
1260 :
1261 : /* The final basic block where to move statements to. In the case of
1262 : multiple exits this could be pretty far away. */
1263 : basic_block early_break_dest_bb;
1264 :
1265 : /* Statements whose VUSES need updating if early break vectorization is to
1266 : happen. */
1267 : auto_vec<gimple*> early_break_vuses;
1268 :
1269 : /* The IV adjustment value for inductions that needs to be materialized
1270 : inside the relevant exit blocks in order to adjust for early break. */
1271 : tree early_break_niters_var;
1272 :
1273 : /* The type of the variable to be used to create the scalar IV for early break
1274 : loops. */
1275 : tree early_break_iv_type;
1276 :
1277 : /* Record statements that are needed to be live for early break vectorization
1278 : but may not have an LC PHI node materialized yet in the exits. */
1279 : auto_vec<stmt_vec_info> early_break_live_ivs;
1280 : } *loop_vec_info;
1281 :
1282 : /* Access Functions. */
1283 : #define LOOP_VINFO_LOOP(L) (L)->loop
1284 : #define LOOP_VINFO_MAIN_EXIT(L) (L)->vec_loop_main_exit
1285 : #define LOOP_VINFO_EPILOGUE_MAIN_EXIT(L) (L)->vec_epilogue_loop_main_exit
1286 : #define LOOP_VINFO_SCALAR_MAIN_EXIT(L) (L)->scalar_loop_main_exit
1287 : #define LOOP_VINFO_BBS(L) (L)->bbs
1288 : #define LOOP_VINFO_NBBS(L) (L)->nbbs
1289 : #define LOOP_VINFO_NITERSM1(L) (L)->num_itersm1
1290 : #define LOOP_VINFO_NITERS(L) (L)->num_iters
1291 : #define LOOP_VINFO_NITERS_UNCOUNTED_P(L) (LOOP_VINFO_NITERS (L) \
1292 : == chrec_dont_know)
1293 : /* Since LOOP_VINFO_NITERS and LOOP_VINFO_NITERSM1 can change after
1294 : prologue peeling retain total unchanged scalar loop iterations for
1295 : cost model. */
1296 : #define LOOP_VINFO_NITERS_UNCHANGED(L) (L)->num_iters_unchanged
1297 : #define LOOP_VINFO_NITERS_ASSUMPTIONS(L) (L)->num_iters_assumptions
1298 : #define LOOP_VINFO_COST_MODEL_THRESHOLD(L) (L)->th
1299 : #define LOOP_VINFO_VERSIONING_THRESHOLD(L) (L)->versioning_threshold
1300 : #define LOOP_VINFO_VECTORIZABLE_P(L) (L)->vectorizable
1301 : #define LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P(L) (L)->can_use_partial_vectors_p
1302 : #define LOOP_VINFO_MUST_USE_PARTIAL_VECTORS_P(L) (L)->must_use_partial_vectors_p
1303 : #define LOOP_VINFO_USING_PARTIAL_VECTORS_P(L) (L)->using_partial_vectors_p
1304 : #define LOOP_VINFO_USING_DECREMENTING_IV_P(L) (L)->using_decrementing_iv_p
1305 : #define LOOP_VINFO_USING_SELECT_VL_P(L) (L)->using_select_vl_p
1306 : #define LOOP_VINFO_ALLOW_MUTUAL_ALIGNMENT(L) (L)->allow_mutual_alignment
1307 : #define LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS(L) (L)->partial_load_store_bias
1308 : #define LOOP_VINFO_VECT_FACTOR(L) (L)->vectorization_factor
1309 : #define LOOP_VINFO_IV_INCREMENT(L) (L)->iv_increment
1310 : #define LOOP_VINFO_IV_INCREMENT_INVARIANT_P(L) \
1311 : (!LOOP_VINFO_USING_SELECT_VL_P (L))
1312 : #define LOOP_VINFO_MAX_VECT_FACTOR(L) (L)->max_vectorization_factor
1313 : #define LOOP_VINFO_MASKS(L) (L)->masks
1314 : #define LOOP_VINFO_LENS(L) (L)->lens
1315 : #define LOOP_VINFO_MASK_SKIP_NITERS(L) (L)->mask_skip_niters
1316 : #define LOOP_VINFO_MASK_NITERS_PFA_OFFSET(L) (L)->mask_skip_niters_pfa_offset
1317 : #define LOOP_VINFO_RGROUP_COMPARE_TYPE(L) (L)->rgroup_compare_type
1318 : #define LOOP_VINFO_RGROUP_IV_TYPE(L) (L)->rgroup_iv_type
1319 : #define LOOP_VINFO_PARTIAL_VECTORS_STYLE(L) (L)->partial_vector_style
1320 : #define LOOP_VINFO_PTR_MASK(L) (L)->ptr_mask
1321 : #define LOOP_VINFO_MAX_SPEC_READ_AMOUNT(L) (L)->max_spec_read_amount
1322 : #define LOOP_VINFO_LOOP_NEST(L) (L)->shared->loop_nest
1323 : #define LOOP_VINFO_DATAREFS(L) (L)->shared->datarefs
1324 : #define LOOP_VINFO_DDRS(L) (L)->shared->ddrs
1325 : #define LOOP_VINFO_INT_NITERS(L) (TREE_INT_CST_LOW ((L)->num_iters))
1326 : #define LOOP_VINFO_PEELING_FOR_ALIGNMENT(L) (L)->peeling_for_alignment
1327 : #define LOOP_VINFO_NON_LINEAR_IV(L) (L)->nonlinear_iv
1328 : #define LOOP_VINFO_UNALIGNED_DR(L) (L)->unaligned_dr
1329 : #define LOOP_VINFO_MAY_MISALIGN_STMTS(L) (L)->may_misalign_stmts
1330 : #define LOOP_VINFO_MAY_ALIAS_DDRS(L) (L)->may_alias_ddrs
1331 : #define LOOP_VINFO_COMP_ALIAS_DDRS(L) (L)->comp_alias_ddrs
1332 : #define LOOP_VINFO_CHECK_UNEQUAL_ADDRS(L) (L)->check_unequal_addrs
1333 : #define LOOP_VINFO_CHECK_NONZERO(L) (L)->check_nonzero
1334 : #define LOOP_VINFO_LOWER_BOUNDS(L) (L)->lower_bounds
1335 : #define LOOP_VINFO_USER_UNROLL(L) (L)->user_unroll
1336 : #define LOOP_VINFO_GROUPED_STORES(L) (L)->grouped_stores
1337 : #define LOOP_VINFO_SLP_INSTANCES(L) (L)->slp_instances
1338 : #define LOOP_VINFO_REDUCTIONS(L) (L)->reductions
1339 : #define LOOP_VINFO_PEELING_FOR_GAPS(L) (L)->peeling_for_gaps
1340 : #define LOOP_VINFO_PEELING_FOR_NITER(L) (L)->peeling_for_niter
1341 : #define LOOP_VINFO_EARLY_BREAKS(L) (L)->early_breaks
1342 : #define LOOP_VINFO_EARLY_BRK_NEEDS_EPILOG(L) (L)->early_break_needs_epilogue
1343 : #define LOOP_VINFO_EARLY_BRK_STORES(L) (L)->early_break_stores
1344 : #define LOOP_VINFO_EARLY_BREAKS_VECT_PEELED(L) \
1345 : ((single_pred ((L)->loop->latch) != (L)->vec_loop_main_exit->src) \
1346 : || LOOP_VINFO_NITERS_UNCOUNTED_P (L))
1347 : #define LOOP_VINFO_EARLY_BREAKS_LIVE_IVS(L) \
1348 : (L)->early_break_live_ivs
1349 : #define LOOP_VINFO_EARLY_BRK_DEST_BB(L) (L)->early_break_dest_bb
1350 : #define LOOP_VINFO_EARLY_BRK_VUSES(L) (L)->early_break_vuses
1351 : #define LOOP_VINFO_EARLY_BRK_NITERS_VAR(L) (L)->early_break_niters_var
1352 : #define LOOP_VINFO_EARLY_BRK_IV_TYPE(L) (L)->early_break_iv_type
1353 : #define LOOP_VINFO_LOOP_CONDS(L) (L)->conds
1354 : #define LOOP_VINFO_LOOP_IV_COND(L) (L)->loop_iv_cond
1355 : #define LOOP_VINFO_NO_DATA_DEPENDENCIES(L) (L)->no_data_dependencies
1356 : #define LOOP_VINFO_SCALAR_LOOP(L) (L)->scalar_loop
1357 : #define LOOP_VINFO_SCALAR_LOOP_SCALING(L) (L)->scalar_loop_scaling
1358 : #define LOOP_VINFO_HAS_MASK_STORE(L) (L)->has_mask_store
1359 : #define LOOP_VINFO_SCALAR_ITERATION_COST(L) (L)->scalar_cost_vec
1360 : #define LOOP_VINFO_MAIN_LOOP_INFO(L) (L)->main_loop_info
1361 : #define LOOP_VINFO_ORIG_LOOP_INFO(L) (L)->orig_loop_info
1362 : #define LOOP_VINFO_SIMD_IF_COND(L) (L)->simd_if_cond
1363 : #define LOOP_VINFO_INNER_LOOP_COST_FACTOR(L) (L)->inner_loop_cost_factor
1364 : #define LOOP_VINFO_INV_PATTERN_DEF_SEQ(L) (L)->inv_pattern_def_seq
1365 : #define LOOP_VINFO_DRS_ADVANCED_BY(L) (L)->drs_advanced_by
1366 : #define LOOP_VINFO_ALTERNATE_DEFS(L) (L)->alternate_defs
1367 :
1368 : #define LOOP_VINFO_FULLY_MASKED_P(L) \
1369 : (LOOP_VINFO_USING_PARTIAL_VECTORS_P (L) \
1370 : && !LOOP_VINFO_MASKS (L).is_empty ())
1371 :
1372 : #define LOOP_VINFO_FULLY_WITH_LENGTH_P(L) \
1373 : (LOOP_VINFO_USING_PARTIAL_VECTORS_P (L) \
1374 : && !LOOP_VINFO_LENS (L).is_empty ())
1375 :
1376 : #define LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT(L) \
1377 : ((L)->may_misalign_stmts.length () > 0)
1378 : #define LOOP_REQUIRES_VERSIONING_FOR_SPEC_READ(L) \
1379 : (maybe_gt ((L)->max_spec_read_amount, 0U))
1380 : #define LOOP_REQUIRES_VERSIONING_FOR_ALIAS(L) \
1381 : ((L)->comp_alias_ddrs.length () > 0 \
1382 : || (L)->check_unequal_addrs.length () > 0 \
1383 : || (L)->lower_bounds.length () > 0)
1384 : #define LOOP_REQUIRES_VERSIONING_FOR_NITERS(L) \
1385 : (LOOP_VINFO_NITERS_ASSUMPTIONS (L))
1386 : #define LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND(L) \
1387 : (LOOP_VINFO_SIMD_IF_COND (L))
1388 : #define LOOP_REQUIRES_VERSIONING(L) \
1389 : (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (L) \
1390 : || LOOP_REQUIRES_VERSIONING_FOR_SPEC_READ (L) \
1391 : || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (L) \
1392 : || LOOP_REQUIRES_VERSIONING_FOR_NITERS (L) \
1393 : || LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND (L))
1394 :
1395 : #define LOOP_VINFO_USE_VERSIONING_WITHOUT_PEELING(L) \
1396 : ((L)->may_misalign_stmts.length () > 0 \
1397 : && !LOOP_VINFO_ALLOW_MUTUAL_ALIGNMENT (L))
1398 :
1399 : #define LOOP_VINFO_NITERS_KNOWN_P(L) \
1400 : (tree_fits_shwi_p ((L)->num_iters) && tree_to_shwi ((L)->num_iters) > 0)
1401 :
1402 : #define LOOP_VINFO_EPILOGUE_P(L) \
1403 : (LOOP_VINFO_ORIG_LOOP_INFO (L) != NULL)
1404 :
1405 : #define LOOP_VINFO_ORIG_MAX_VECT_FACTOR(L) \
1406 : (LOOP_VINFO_MAX_VECT_FACTOR (LOOP_VINFO_ORIG_LOOP_INFO (L)))
1407 :
1408 : /* Wrapper for loop_vec_info, for tracking success/failure, where a non-NULL
1409 : value signifies success, and a NULL value signifies failure, supporting
1410 : propagating an opt_problem * describing the failure back up the call
1411 : stack. */
1412 : typedef opt_pointer_wrapper <loop_vec_info> opt_loop_vec_info;
1413 :
1414 : inline loop_vec_info
1415 541975 : loop_vec_info_for_loop (class loop *loop)
1416 : {
1417 541975 : return (loop_vec_info) loop->aux;
1418 : }
1419 :
1420 : struct slp_root
1421 : {
1422 1294328 : slp_root (slp_instance_kind kind_, vec<stmt_vec_info> stmts_,
1423 14399 : vec<stmt_vec_info> roots_, vec<tree> remain_ = vNULL)
1424 1294328 : : kind(kind_), stmts(stmts_), roots(roots_), remain(remain_) {}
1425 : slp_instance_kind kind;
1426 : vec<stmt_vec_info> stmts;
1427 : vec<stmt_vec_info> roots;
1428 : vec<tree> remain;
1429 : };
1430 :
1431 : typedef class _bb_vec_info : public vec_info
1432 : {
1433 : public:
1434 : _bb_vec_info (vec<basic_block> bbs, vec_info_shared *);
1435 : ~_bb_vec_info ();
1436 :
1437 : vec<slp_root> roots;
1438 : } *bb_vec_info;
1439 :
1440 : #define BB_VINFO_BBS(B) (B)->bbs
1441 : #define BB_VINFO_NBBS(B) (B)->nbbs
1442 : #define BB_VINFO_GROUPED_STORES(B) (B)->grouped_stores
1443 : #define BB_VINFO_SLP_INSTANCES(B) (B)->slp_instances
1444 : #define BB_VINFO_DATAREFS(B) (B)->shared->datarefs
1445 : #define BB_VINFO_DDRS(B) (B)->shared->ddrs
1446 :
1447 : /* Indicates whether/how a variable is used in the scope of loop/basic
1448 : block. */
1449 : enum vect_relevant {
1450 : vect_unused_in_scope = 0,
1451 :
1452 : /* The def is only used outside the loop. */
1453 : vect_used_only_live,
1454 : /* The def is in the inner loop, and the use is in the outer loop, and the
1455 : use is a reduction stmt. */
1456 : vect_used_in_outer_by_reduction,
1457 : /* The def is in the inner loop, and the use is in the outer loop (and is
1458 : not part of reduction). */
1459 : vect_used_in_outer,
1460 :
1461 : /* defs that feed computations that end up (only) in a reduction. These
1462 : defs may be used by non-reduction stmts, but eventually, any
1463 : computations/values that are affected by these defs are used to compute
1464 : a reduction (i.e. don't get stored to memory, for example). We use this
1465 : to identify computations that we can change the order in which they are
1466 : computed. */
1467 : vect_used_by_reduction,
1468 :
1469 : vect_used_in_scope
1470 : };
1471 :
1472 : /* The type of vectorization. pure_slp means the stmt is covered by the
1473 : SLP graph, not_vect means it is not. This is mostly used by BB
1474 : vectorization. */
1475 : enum slp_vect_type {
1476 : not_vect = 0,
1477 : pure_slp,
1478 : };
1479 :
1480 : /* Says whether a statement is a load, a store of a vectorized statement
1481 : result, or a store of an invariant value. */
1482 : enum vec_load_store_type {
1483 : VLS_LOAD,
1484 : VLS_STORE,
1485 : VLS_STORE_INVARIANT
1486 : };
1487 :
1488 : class dr_vec_info {
1489 : public:
1490 : /* The data reference itself. */
1491 : data_reference *dr;
1492 : /* The statement that contains the data reference. */
1493 : stmt_vec_info stmt;
1494 : /* The analysis group this DR belongs to when doing BB vectorization.
1495 : DRs of the same group belong to the same conditional execution context. */
1496 : unsigned group;
1497 : /* The misalignment in bytes of the reference, or -1 if not known. */
1498 : int misalignment;
1499 : /* The byte alignment that we'd ideally like the reference to have,
1500 : and the value that misalignment is measured against. */
1501 : poly_uint64 target_alignment;
1502 : /* If true the alignment of base_decl needs to be increased. */
1503 : bool base_misaligned;
1504 :
1505 : /* Set by early break vectorization when this DR needs peeling for alignment
1506 : for correctness. */
1507 : bool safe_speculative_read_required;
1508 :
1509 : /* Set by early break vectorization when this DR's scalar accesses are known
1510 : to be inbounds of a known bounds loop. */
1511 : bool scalar_access_known_in_bounds;
1512 :
1513 : tree base_decl;
1514 :
1515 : /* Stores current vectorized loop's offset. To be added to the DR's
1516 : offset to calculate current offset of data reference. */
1517 : tree offset;
1518 : };
1519 :
1520 : typedef struct data_reference *dr_p;
1521 :
1522 : class _stmt_vec_info {
1523 : public:
1524 :
1525 : /* Indicates whether this stmts is part of a computation whose result is
1526 : used outside the loop. */
1527 : bool live;
1528 :
1529 : /* Stmt is part of some pattern (computation idiom) */
1530 : bool in_pattern_p;
1531 :
1532 : /* True if the statement was created during pattern recognition as
1533 : part of the replacement for RELATED_STMT. This implies that the
1534 : statement isn't part of any basic block, although for convenience
1535 : its gimple_bb is the same as for RELATED_STMT. */
1536 : bool pattern_stmt_p;
1537 :
1538 : /* Is this statement vectorizable or should it be skipped in (partial)
1539 : vectorization. */
1540 : bool vectorizable;
1541 :
1542 : /* The stmt to which this info struct refers to. */
1543 : gimple *stmt;
1544 :
1545 : /* The vector type to be used for the LHS of this statement. */
1546 : tree vectype;
1547 :
1548 : /* The following is relevant only for stmts that contain a non-scalar
1549 : data-ref (array/pointer/struct access). A GIMPLE stmt is expected to have
1550 : at most one such data-ref. */
1551 :
1552 : dr_vec_info dr_aux;
1553 :
1554 : /* Information about the data-ref relative to this loop
1555 : nest (the loop that is being considered for vectorization). */
1556 : innermost_loop_behavior dr_wrt_vec_loop;
1557 :
1558 : /* For loop PHI nodes, the base and evolution part of it. This makes sure
1559 : this information is still available in vect_update_ivs_after_vectorizer
1560 : where we may not be able to re-analyze the PHI nodes evolution as
1561 : peeling for the prologue loop can make it unanalyzable. The evolution
1562 : part is still correct after peeling, but the base may have changed from
1563 : the version here. */
1564 : tree loop_phi_evolution_base_unchanged;
1565 : tree loop_phi_evolution_part;
1566 : enum vect_induction_op_type loop_phi_evolution_type;
1567 :
1568 : /* Used for various bookkeeping purposes, generally holding a pointer to
1569 : some other stmt S that is in some way "related" to this stmt.
1570 : Current use of this field is:
1571 : If this stmt is part of a pattern (i.e. the field 'in_pattern_p' is
1572 : true): S is the "pattern stmt" that represents (and replaces) the
1573 : sequence of stmts that constitutes the pattern. Similarly, the
1574 : related_stmt of the "pattern stmt" points back to this stmt (which is
1575 : the last stmt in the original sequence of stmts that constitutes the
1576 : pattern). */
1577 : stmt_vec_info related_stmt;
1578 :
1579 : /* Used to keep a sequence of def stmts of a pattern stmt if such exists.
1580 : The sequence is attached to the original statement rather than the
1581 : pattern statement. */
1582 : gimple_seq pattern_def_seq;
1583 :
1584 : /* Classify the def of this stmt. */
1585 : enum vect_def_type def_type;
1586 :
1587 : /* Whether the stmt is SLPed, loop-based vectorized, or both. */
1588 : enum slp_vect_type slp_type;
1589 :
1590 : /* Interleaving chains info. */
1591 : /* First element in the group. */
1592 : stmt_vec_info first_element;
1593 : /* Pointer to the next element in the group. */
1594 : stmt_vec_info next_element;
1595 : /* The size of the group. */
1596 : unsigned int size;
1597 : /* For loads only, the gap from the previous load. For consecutive loads, GAP
1598 : is 1. */
1599 : unsigned int gap;
1600 :
1601 : /* The minimum negative dependence distance this stmt participates in
1602 : or zero if none. */
1603 : unsigned int min_neg_dist;
1604 :
1605 : /* Not all stmts in the loop need to be vectorized. e.g, the increment
1606 : of the loop induction variable and computation of array indexes. relevant
1607 : indicates whether the stmt needs to be vectorized. */
1608 : enum vect_relevant relevant;
1609 :
1610 : /* For loads if this is a gather, for stores if this is a scatter. */
1611 : bool gather_scatter_p;
1612 :
1613 : /* True if this is an access with loop-invariant stride. */
1614 : bool strided_p;
1615 :
1616 : /* For both loads and stores. */
1617 : unsigned simd_lane_access_p : 3;
1618 :
1619 : /* On a reduction PHI the reduction type as detected by
1620 : vect_is_simple_reduction. */
1621 : enum vect_reduction_type reduc_type;
1622 :
1623 : /* On a reduction PHI, the original reduction code as detected by
1624 : vect_is_simple_reduction. */
1625 : code_helper reduc_code;
1626 :
1627 : /* On a stmt participating in a reduction the index of the operand
1628 : on the reduction SSA cycle. */
1629 : int reduc_idx;
1630 :
1631 : /* On a reduction PHI the def returned by vect_is_simple_reduction.
1632 : On the def returned by vect_is_simple_reduction the corresponding PHI. */
1633 : stmt_vec_info reduc_def;
1634 :
1635 : /* If nonzero, the lhs of the statement could be truncated to this
1636 : many bits without affecting any users of the result. */
1637 : unsigned int min_output_precision;
1638 :
1639 : /* If nonzero, all non-boolean input operands have the same precision,
1640 : and they could each be truncated to this many bits without changing
1641 : the result. */
1642 : unsigned int min_input_precision;
1643 :
1644 : /* If OPERATION_BITS is nonzero, the statement could be performed on
1645 : an integer with the sign and number of bits given by OPERATION_SIGN
1646 : and OPERATION_BITS without changing the result. */
1647 : unsigned int operation_precision;
1648 : signop operation_sign;
1649 :
1650 : /* If the statement produces a boolean result, this value describes
1651 : how we should choose the associated vector type. The possible
1652 : values are:
1653 :
1654 : - an integer precision N if we should use the vector mask type
1655 : associated with N-bit integers. This is only used if all relevant
1656 : input booleans also want the vector mask type for N-bit integers,
1657 : or if we can convert them into that form by pattern-matching.
1658 :
1659 : - ~0U if we considered choosing a vector mask type but decided
1660 : to treat the boolean as a normal integer type instead.
1661 :
1662 : - 0 otherwise. This means either that the operation isn't one that
1663 : could have a vector mask type (and so should have a normal vector
1664 : type instead) or that we simply haven't made a choice either way. */
1665 : unsigned int mask_precision;
1666 :
1667 : /* True if this is only suitable for SLP vectorization. */
1668 : bool slp_vect_only_p;
1669 : };
1670 :
1671 : /* Information about a gather/scatter call. */
1672 : struct gather_scatter_info {
1673 : /* The internal function to use for the gather/scatter operation,
1674 : or IFN_LAST if a built-in function should be used instead. */
1675 : internal_fn ifn;
1676 :
1677 : /* The FUNCTION_DECL for the built-in gather/scatter function,
1678 : or null if an internal function should be used instead. */
1679 : tree decl;
1680 :
1681 : /* The loop-invariant base value. */
1682 : tree base;
1683 :
1684 : /* The TBBA alias pointer the value of which determines the alignment
1685 : of the scalar accesses. */
1686 : tree alias_ptr;
1687 :
1688 : /* The original scalar offset, which is a non-loop-invariant SSA_NAME. */
1689 : tree offset;
1690 :
1691 : /* Each offset element should be multiplied by this amount before
1692 : being added to the base. */
1693 : int scale;
1694 :
1695 : /* The type of the vectorized offset. */
1696 : tree offset_vectype;
1697 :
1698 : /* The type of the scalar elements after loading or before storing. */
1699 : tree element_type;
1700 :
1701 : /* The type of the scalar elements being loaded or stored. */
1702 : tree memory_type;
1703 : };
1704 :
1705 : /* Access Functions. */
1706 : #define STMT_VINFO_STMT(S) (S)->stmt
1707 : #define STMT_VINFO_RELEVANT(S) (S)->relevant
1708 : #define STMT_VINFO_LIVE_P(S) (S)->live
1709 : #define STMT_VINFO_VECTYPE(S) (S)->vectype
1710 : #define STMT_VINFO_VECTORIZABLE(S) (S)->vectorizable
1711 : #define STMT_VINFO_DATA_REF(S) ((S)->dr_aux.dr + 0)
1712 : #define STMT_VINFO_GATHER_SCATTER_P(S) (S)->gather_scatter_p
1713 : #define STMT_VINFO_STRIDED_P(S) (S)->strided_p
1714 : #define STMT_VINFO_SIMD_LANE_ACCESS_P(S) (S)->simd_lane_access_p
1715 : #define STMT_VINFO_REDUC_IDX(S) (S)->reduc_idx
1716 :
1717 : #define STMT_VINFO_DR_WRT_VEC_LOOP(S) (S)->dr_wrt_vec_loop
1718 : #define STMT_VINFO_DR_BASE_ADDRESS(S) (S)->dr_wrt_vec_loop.base_address
1719 : #define STMT_VINFO_DR_INIT(S) (S)->dr_wrt_vec_loop.init
1720 : #define STMT_VINFO_DR_OFFSET(S) (S)->dr_wrt_vec_loop.offset
1721 : #define STMT_VINFO_DR_STEP(S) (S)->dr_wrt_vec_loop.step
1722 : #define STMT_VINFO_DR_BASE_ALIGNMENT(S) (S)->dr_wrt_vec_loop.base_alignment
1723 : #define STMT_VINFO_DR_BASE_MISALIGNMENT(S) \
1724 : (S)->dr_wrt_vec_loop.base_misalignment
1725 : #define STMT_VINFO_DR_OFFSET_ALIGNMENT(S) \
1726 : (S)->dr_wrt_vec_loop.offset_alignment
1727 : #define STMT_VINFO_DR_STEP_ALIGNMENT(S) \
1728 : (S)->dr_wrt_vec_loop.step_alignment
1729 :
1730 : #define STMT_VINFO_DR_INFO(S) \
1731 : (gcc_checking_assert ((S)->dr_aux.stmt == (S)), &(S)->dr_aux)
1732 :
1733 : #define STMT_VINFO_IN_PATTERN_P(S) (S)->in_pattern_p
1734 : #define STMT_VINFO_RELATED_STMT(S) (S)->related_stmt
1735 : #define STMT_VINFO_PATTERN_DEF_SEQ(S) (S)->pattern_def_seq
1736 : #define STMT_VINFO_DEF_TYPE(S) (S)->def_type
1737 : #define STMT_VINFO_GROUPED_ACCESS(S) \
1738 : ((S)->dr_aux.dr && DR_GROUP_FIRST_ELEMENT(S))
1739 : #define STMT_VINFO_LOOP_PHI_EVOLUTION_BASE_UNCHANGED(S) (S)->loop_phi_evolution_base_unchanged
1740 : #define STMT_VINFO_LOOP_PHI_EVOLUTION_PART(S) (S)->loop_phi_evolution_part
1741 : #define STMT_VINFO_LOOP_PHI_EVOLUTION_TYPE(S) (S)->loop_phi_evolution_type
1742 : #define STMT_VINFO_MIN_NEG_DIST(S) (S)->min_neg_dist
1743 : #define STMT_VINFO_REDUC_TYPE(S) (S)->reduc_type
1744 : #define STMT_VINFO_REDUC_CODE(S) (S)->reduc_code
1745 : #define STMT_VINFO_REDUC_DEF(S) (S)->reduc_def
1746 : #define STMT_VINFO_SLP_VECT_ONLY(S) (S)->slp_vect_only_p
1747 : #define STMT_VINFO_REDUC_VECTYPE_IN(S) (S)->reduc_vectype_in
1748 :
1749 : #define DR_GROUP_FIRST_ELEMENT(S) \
1750 : (gcc_checking_assert ((S)->dr_aux.dr), (S)->first_element)
1751 : #define DR_GROUP_NEXT_ELEMENT(S) \
1752 : (gcc_checking_assert ((S)->dr_aux.dr), (S)->next_element)
1753 : #define DR_GROUP_SIZE(S) \
1754 : (gcc_checking_assert ((S)->dr_aux.dr), (S)->size)
1755 : #define DR_GROUP_GAP(S) \
1756 : (gcc_checking_assert ((S)->dr_aux.dr), (S)->gap)
1757 :
1758 : #define STMT_VINFO_RELEVANT_P(S) ((S)->relevant != vect_unused_in_scope)
1759 :
1760 : #define PURE_SLP_STMT(S) ((S)->slp_type == pure_slp)
1761 : #define STMT_SLP_TYPE(S) (S)->slp_type
1762 :
1763 :
1764 : /* Contains the scalar or vector costs for a vec_info. */
1765 : class vector_costs
1766 : {
1767 : public:
1768 : vector_costs (vec_info *, bool);
1769 0 : virtual ~vector_costs () {}
1770 :
1771 : /* Update the costs in response to adding COUNT copies of a statement.
1772 :
1773 : - WHERE specifies whether the cost occurs in the loop prologue,
1774 : the loop body, or the loop epilogue.
1775 : - KIND is the kind of statement, which is always meaningful.
1776 : - STMT_INFO or NODE, if nonnull, describe the statement that will be
1777 : vectorized.
1778 : - VECTYPE, if nonnull, is the vector type that the vectorized
1779 : statement will operate on. Note that this should be used in
1780 : preference to STMT_VINFO_VECTYPE (STMT_INFO) since the latter
1781 : is not correct for SLP.
1782 : - for unaligned_load and unaligned_store statements, MISALIGN is
1783 : the byte misalignment of the load or store relative to the target's
1784 : preferred alignment for VECTYPE, or DR_MISALIGNMENT_UNKNOWN
1785 : if the misalignment is not known.
1786 :
1787 : Return the calculated cost as well as recording it. The return
1788 : value is used for dumping purposes. */
1789 : virtual unsigned int add_stmt_cost (int count, vect_cost_for_stmt kind,
1790 : stmt_vec_info stmt_info,
1791 : slp_tree node,
1792 : tree vectype, int misalign,
1793 : vect_cost_model_location where);
1794 :
1795 : /* Update the costs in response to adding costs in V which are all from
1796 : vectorizing NODE to the respective part. */
1797 : virtual unsigned int add_slp_cost (slp_tree node,
1798 : const array_slice<stmt_info_for_cost> &v);
1799 :
1800 : /* Finish calculating the cost of the code. The results can be
1801 : read back using the functions below.
1802 :
1803 : If the costs describe vector code, SCALAR_COSTS gives the costs
1804 : of the corresponding scalar code, otherwise it is null. */
1805 : virtual void finish_cost (const vector_costs *scalar_costs);
1806 :
1807 : /* The costs in THIS and OTHER both describe ways of vectorizing
1808 : a main loop. Return true if the costs described by THIS are
1809 : cheaper than the costs described by OTHER. Return false if any
1810 : of the following are true:
1811 :
1812 : - THIS and OTHER are of equal cost
1813 : - OTHER is better than THIS
1814 : - we can't be sure about the relative costs of THIS and OTHER. */
1815 : virtual bool better_main_loop_than_p (const vector_costs *other) const;
1816 :
1817 : /* Likewise, but the costs in THIS and OTHER both describe ways of
1818 : vectorizing an epilogue loop of MAIN_LOOP. */
1819 : virtual bool better_epilogue_loop_than_p (const vector_costs *other,
1820 : loop_vec_info main_loop) const;
1821 :
1822 : unsigned int prologue_cost () const;
1823 : unsigned int body_cost () const;
1824 : unsigned int epilogue_cost () const;
1825 : unsigned int outside_cost () const;
1826 : unsigned int total_cost () const;
1827 :
1828 : unsigned int suggested_unroll_factor () const;
1829 : machine_mode suggested_epilogue_mode (int &masked) const;
1830 :
1831 32512 : vec_info *vinfo () const { return m_vinfo; }
1832 7613488 : bool costing_for_scalar () const { return m_costing_for_scalar; }
1833 :
1834 : protected:
1835 : unsigned int record_stmt_cost (stmt_vec_info, vect_cost_model_location,
1836 : unsigned int);
1837 : unsigned int adjust_cost_for_freq (stmt_vec_info, vect_cost_model_location,
1838 : unsigned int);
1839 : int compare_inside_loop_cost (const vector_costs *) const;
1840 : int compare_outside_loop_cost (const vector_costs *) const;
1841 :
1842 : /* The region of code that we're considering vectorizing. */
1843 : vec_info *m_vinfo;
1844 :
1845 : /* True if we're costing the scalar code, false if we're costing
1846 : the vector code. */
1847 : bool m_costing_for_scalar;
1848 :
1849 : /* The costs of the three regions, indexed by vect_cost_model_location. */
1850 : unsigned int m_costs[3];
1851 :
1852 : /* The suggested unrolling factor determined at finish_cost. */
1853 : unsigned int m_suggested_unroll_factor;
1854 :
1855 : /* The suggested mode to be used for a vectorized epilogue or VOIDmode,
1856 : determined at finish_cost. m_masked_epilogue specifies whether the
1857 : epilogue should use masked vectorization, regardless of the
1858 : --param vect-partial-vector-usage default. If -1 then the
1859 : --param setting takes precedence. If the user explicitly specified
1860 : --param vect-partial-vector-usage then that takes precedence. */
1861 : machine_mode m_suggested_epilogue_mode;
1862 : int m_masked_epilogue;
1863 :
1864 : /* True if finish_cost has been called. */
1865 : bool m_finished;
1866 : };
1867 :
1868 : /* Create costs for VINFO. COSTING_FOR_SCALAR is true if the costs
1869 : are for scalar code, false if they are for vector code. */
1870 :
1871 : inline
1872 2145564 : vector_costs::vector_costs (vec_info *vinfo, bool costing_for_scalar)
1873 2145564 : : m_vinfo (vinfo),
1874 2145564 : m_costing_for_scalar (costing_for_scalar),
1875 2145564 : m_costs (),
1876 2145564 : m_suggested_unroll_factor(1),
1877 2145564 : m_suggested_epilogue_mode(VOIDmode),
1878 2145564 : m_masked_epilogue (-1),
1879 2145564 : m_finished (false)
1880 : {
1881 : }
1882 :
1883 : /* Return the cost of the prologue code (in abstract units). */
1884 :
1885 : inline unsigned int
1886 1299049 : vector_costs::prologue_cost () const
1887 : {
1888 1299049 : gcc_checking_assert (m_finished);
1889 1299049 : return m_costs[vect_prologue];
1890 : }
1891 :
1892 : /* Return the cost of the body code (in abstract units). */
1893 :
1894 : inline unsigned int
1895 2036057 : vector_costs::body_cost () const
1896 : {
1897 2036057 : gcc_checking_assert (m_finished);
1898 2036057 : return m_costs[vect_body];
1899 : }
1900 :
1901 : /* Return the cost of the epilogue code (in abstract units). */
1902 :
1903 : inline unsigned int
1904 1299049 : vector_costs::epilogue_cost () const
1905 : {
1906 1299049 : gcc_checking_assert (m_finished);
1907 1299049 : return m_costs[vect_epilogue];
1908 : }
1909 :
1910 : /* Return the cost of the prologue and epilogue code (in abstract units). */
1911 :
1912 : inline unsigned int
1913 498492 : vector_costs::outside_cost () const
1914 : {
1915 498492 : return prologue_cost () + epilogue_cost ();
1916 : }
1917 :
1918 : /* Return the cost of the prologue, body and epilogue code
1919 : (in abstract units). */
1920 :
1921 : inline unsigned int
1922 125267 : vector_costs::total_cost () const
1923 : {
1924 125267 : return body_cost () + outside_cost ();
1925 : }
1926 :
1927 : /* Return the suggested unroll factor. */
1928 :
1929 : inline unsigned int
1930 124877 : vector_costs::suggested_unroll_factor () const
1931 : {
1932 124877 : gcc_checking_assert (m_finished);
1933 124877 : return m_suggested_unroll_factor;
1934 : }
1935 :
1936 : /* Return the suggested epilogue mode. */
1937 :
1938 : inline machine_mode
1939 14437 : vector_costs::suggested_epilogue_mode (int &masked_p) const
1940 : {
1941 14437 : gcc_checking_assert (m_finished);
1942 14437 : masked_p = m_masked_epilogue;
1943 14437 : return m_suggested_epilogue_mode;
1944 : }
1945 :
1946 : #define VECT_MAX_COST 1000
1947 :
1948 : /* The maximum number of intermediate steps required in multi-step type
1949 : conversion. */
1950 : #define MAX_INTERM_CVT_STEPS 3
1951 :
1952 : #define MAX_VECTORIZATION_FACTOR INT_MAX
1953 :
1954 : /* Nonzero if TYPE represents a (scalar) boolean type or type
1955 : in the middle-end compatible with it (unsigned precision 1 integral
1956 : types). Used to determine which types should be vectorized as
1957 : VECTOR_BOOLEAN_TYPE_P. */
1958 :
1959 : #define VECT_SCALAR_BOOLEAN_TYPE_P(TYPE) \
1960 : (TREE_CODE (TYPE) == BOOLEAN_TYPE \
1961 : || ((TREE_CODE (TYPE) == INTEGER_TYPE \
1962 : || TREE_CODE (TYPE) == ENUMERAL_TYPE) \
1963 : && TYPE_PRECISION (TYPE) == 1 \
1964 : && TYPE_UNSIGNED (TYPE)))
1965 :
1966 : inline bool
1967 11759309 : nested_in_vect_loop_p (class loop *loop, stmt_vec_info stmt_info)
1968 : {
1969 11759309 : return (loop->inner
1970 9460099 : && (loop->inner == (gimple_bb (stmt_info->stmt))->loop_father));
1971 : }
1972 :
1973 : /* PHI is either a scalar reduction phi or a scalar induction phi.
1974 : Return the initial value of the variable on entry to the containing
1975 : loop. */
1976 :
1977 : inline tree
1978 34224 : vect_phi_initial_value (gphi *phi)
1979 : {
1980 34224 : basic_block bb = gimple_bb (phi);
1981 34224 : edge pe = loop_preheader_edge (bb->loop_father);
1982 34224 : gcc_assert (pe->dest == bb);
1983 34224 : return PHI_ARG_DEF_FROM_EDGE (phi, pe);
1984 : }
1985 :
1986 : /* Return true if STMT_INFO should produce a vector mask type rather than
1987 : a normal nonmask type. */
1988 :
1989 : inline bool
1990 7397827 : vect_use_mask_type_p (stmt_vec_info stmt_info)
1991 : {
1992 7397827 : return stmt_info->mask_precision && stmt_info->mask_precision != ~0U;
1993 : }
1994 :
1995 : /* Return TRUE if a statement represented by STMT_INFO is a part of a
1996 : pattern. */
1997 :
1998 : inline bool
1999 129712481 : is_pattern_stmt_p (stmt_vec_info stmt_info)
2000 : {
2001 82870784 : return stmt_info->pattern_stmt_p;
2002 : }
2003 :
2004 : /* If STMT_INFO is a pattern statement, return the statement that it
2005 : replaces, otherwise return STMT_INFO itself. */
2006 :
2007 : inline stmt_vec_info
2008 51195842 : vect_orig_stmt (stmt_vec_info stmt_info)
2009 : {
2010 38573107 : if (is_pattern_stmt_p (stmt_info))
2011 3526766 : return STMT_VINFO_RELATED_STMT (stmt_info);
2012 : return stmt_info;
2013 : }
2014 :
2015 : /* Return the later statement between STMT1_INFO and STMT2_INFO. */
2016 :
2017 : inline stmt_vec_info
2018 6173097 : get_later_stmt (stmt_vec_info stmt1_info, stmt_vec_info stmt2_info)
2019 : {
2020 6173097 : gimple *stmt1 = vect_orig_stmt (stmt1_info)->stmt;
2021 6173097 : gimple *stmt2 = vect_orig_stmt (stmt2_info)->stmt;
2022 6173097 : if (gimple_bb (stmt1) == gimple_bb (stmt2))
2023 : {
2024 6147753 : if (gimple_uid (stmt1) > gimple_uid (stmt2))
2025 : return stmt1_info;
2026 : else
2027 : return stmt2_info;
2028 : }
2029 : /* ??? We should be really calling this function only with stmts
2030 : in the same BB but we can recover if there's a domination
2031 : relationship between them. */
2032 25344 : else if (dominated_by_p (CDI_DOMINATORS,
2033 25344 : gimple_bb (stmt1), gimple_bb (stmt2)))
2034 : return stmt1_info;
2035 9042 : else if (dominated_by_p (CDI_DOMINATORS,
2036 9042 : gimple_bb (stmt2), gimple_bb (stmt1)))
2037 : return stmt2_info;
2038 0 : gcc_unreachable ();
2039 : }
2040 :
2041 : /* If STMT_INFO has been replaced by a pattern statement, return the
2042 : replacement statement, otherwise return STMT_INFO itself. */
2043 :
2044 : inline stmt_vec_info
2045 52832078 : vect_stmt_to_vectorize (stmt_vec_info stmt_info)
2046 : {
2047 52832078 : if (STMT_VINFO_IN_PATTERN_P (stmt_info))
2048 1623753 : return STMT_VINFO_RELATED_STMT (stmt_info);
2049 : return stmt_info;
2050 : }
2051 :
2052 : /* Return true if BB is a loop header. */
2053 :
2054 : inline bool
2055 1551123 : is_loop_header_bb_p (basic_block bb)
2056 : {
2057 1551123 : if (bb == (bb->loop_father)->header)
2058 1540333 : return true;
2059 :
2060 : return false;
2061 : }
2062 :
2063 : /* Return pow2 (X). */
2064 :
2065 : inline int
2066 : vect_pow2 (int x)
2067 : {
2068 : int i, res = 1;
2069 :
2070 : for (i = 0; i < x; i++)
2071 : res *= 2;
2072 :
2073 : return res;
2074 : }
2075 :
2076 : /* Alias targetm.vectorize.builtin_vectorization_cost. */
2077 :
2078 : inline int
2079 9272284 : builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost,
2080 : tree vectype, int misalign)
2081 : {
2082 9204649 : return targetm.vectorize.builtin_vectorization_cost (type_of_cost,
2083 : vectype, misalign);
2084 : }
2085 :
2086 : /* Get cost by calling cost target builtin. */
2087 :
2088 : inline
2089 155 : int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost)
2090 : {
2091 67480 : return builtin_vectorization_cost (type_of_cost, NULL, 0);
2092 : }
2093 :
2094 : /* Alias targetm.vectorize.init_cost. */
2095 :
2096 : inline vector_costs *
2097 2145564 : init_cost (vec_info *vinfo, bool costing_for_scalar)
2098 : {
2099 2145564 : return targetm.vectorize.create_costs (vinfo, costing_for_scalar);
2100 : }
2101 :
2102 : extern void dump_stmt_cost (FILE *, int, enum vect_cost_for_stmt,
2103 : stmt_vec_info, slp_tree, tree, int, unsigned,
2104 : enum vect_cost_model_location);
2105 :
2106 : /* Dump and add costs. */
2107 :
2108 : inline unsigned
2109 7613488 : add_stmt_cost (vector_costs *costs, int count,
2110 : enum vect_cost_for_stmt kind,
2111 : stmt_vec_info stmt_info, slp_tree node,
2112 : tree vectype, int misalign,
2113 : enum vect_cost_model_location where)
2114 : {
2115 : /* Even though a vector type might be set on stmt do not pass that on when
2116 : costing the scalar IL. A SLP node shouldn't have been recorded. */
2117 7613488 : if (costs->costing_for_scalar ())
2118 : {
2119 3954908 : vectype = NULL_TREE;
2120 3954908 : gcc_checking_assert (node == NULL);
2121 : }
2122 7613488 : unsigned cost = costs->add_stmt_cost (count, kind, stmt_info, node, vectype,
2123 : misalign, where);
2124 7613488 : if (dump_file && (dump_flags & TDF_DETAILS))
2125 220212 : dump_stmt_cost (dump_file, count, kind, stmt_info, node, vectype, misalign,
2126 : cost, where);
2127 7613488 : return cost;
2128 : }
2129 :
2130 : inline unsigned
2131 82698 : add_stmt_cost (vector_costs *costs, int count, enum vect_cost_for_stmt kind,
2132 : enum vect_cost_model_location where)
2133 : {
2134 82698 : gcc_assert (kind == cond_branch_taken || kind == cond_branch_not_taken
2135 : || kind == scalar_stmt);
2136 82698 : return add_stmt_cost (costs, count, kind, NULL, NULL, NULL_TREE, 0, where);
2137 : }
2138 :
2139 : inline unsigned
2140 2189762 : add_stmt_cost (vector_costs *costs, stmt_info_for_cost *i)
2141 : {
2142 2189762 : return add_stmt_cost (costs, i->count, i->kind, i->stmt_info, i->node,
2143 2189762 : i->vectype, i->misalign, i->where);
2144 : }
2145 :
2146 : inline void
2147 371401 : add_stmt_costs (vector_costs *costs, stmt_vector_for_cost *cost_vec)
2148 : {
2149 371401 : stmt_info_for_cost *cost;
2150 371401 : unsigned i;
2151 2136547 : FOR_EACH_VEC_ELT (*cost_vec, i, cost)
2152 1765146 : add_stmt_cost (costs, cost->count, cost->kind, cost->stmt_info,
2153 : cost->node, cost->vectype, cost->misalign, cost->where);
2154 371401 : }
2155 :
2156 : /*-----------------------------------------------------------------*/
2157 : /* Info on data references alignment. */
2158 : /*-----------------------------------------------------------------*/
2159 : #define DR_MISALIGNMENT_UNKNOWN (-1)
2160 : #define DR_MISALIGNMENT_UNINITIALIZED (-2)
2161 :
2162 : inline void
2163 2626342 : set_dr_misalignment (dr_vec_info *dr_info, int val)
2164 : {
2165 2626342 : dr_info->misalignment = val;
2166 : }
2167 :
2168 : extern int dr_misalignment (dr_vec_info *dr_info, tree vectype,
2169 : poly_int64 offset = 0);
2170 :
2171 : #define SET_DR_MISALIGNMENT(DR, VAL) set_dr_misalignment (DR, VAL)
2172 :
2173 : /* Only defined once DR_MISALIGNMENT is defined. */
2174 : inline const poly_uint64
2175 8082002 : dr_target_alignment (dr_vec_info *dr_info)
2176 : {
2177 8082002 : if (STMT_VINFO_GROUPED_ACCESS (dr_info->stmt))
2178 5946239 : dr_info = STMT_VINFO_DR_INFO (DR_GROUP_FIRST_ELEMENT (dr_info->stmt));
2179 8082002 : return dr_info->target_alignment;
2180 : }
2181 : #define DR_TARGET_ALIGNMENT(DR) dr_target_alignment (DR)
2182 : #define DR_SCALAR_KNOWN_BOUNDS(DR) (DR)->scalar_access_known_in_bounds
2183 :
2184 : /* Return if the stmt_vec_info requires peeling for alignment. */
2185 : inline bool
2186 4562164 : dr_safe_speculative_read_required (stmt_vec_info stmt_info)
2187 : {
2188 4562164 : dr_vec_info *dr_info;
2189 4562164 : if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
2190 1687990 : dr_info = STMT_VINFO_DR_INFO (DR_GROUP_FIRST_ELEMENT (stmt_info));
2191 : else
2192 2874174 : dr_info = STMT_VINFO_DR_INFO (stmt_info);
2193 :
2194 4562164 : return dr_info->safe_speculative_read_required;
2195 : }
2196 :
2197 : /* Set the safe_speculative_read_required for the stmt_vec_info, if group
2198 : access then set on the fist element otherwise set on DR directly. */
2199 : inline void
2200 234882 : dr_set_safe_speculative_read_required (stmt_vec_info stmt_info,
2201 : bool requires_alignment)
2202 : {
2203 234882 : dr_vec_info *dr_info;
2204 234882 : if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
2205 68102 : dr_info = STMT_VINFO_DR_INFO (DR_GROUP_FIRST_ELEMENT (stmt_info));
2206 : else
2207 166780 : dr_info = STMT_VINFO_DR_INFO (stmt_info);
2208 :
2209 234882 : dr_info->safe_speculative_read_required = requires_alignment;
2210 234882 : }
2211 :
2212 : inline void
2213 1615254 : set_dr_target_alignment (dr_vec_info *dr_info, poly_uint64 val)
2214 : {
2215 1615254 : dr_info->target_alignment = val;
2216 : }
2217 : #define SET_DR_TARGET_ALIGNMENT(DR, VAL) set_dr_target_alignment (DR, VAL)
2218 :
2219 : /* Return true if data access DR_INFO is aligned to the targets
2220 : preferred alignment for VECTYPE (which may be less than a full vector). */
2221 :
2222 : inline bool
2223 387633 : aligned_access_p (dr_vec_info *dr_info, tree vectype)
2224 : {
2225 387633 : return (dr_misalignment (dr_info, vectype) == 0);
2226 : }
2227 :
2228 : /* Return TRUE if the (mis-)alignment of the data access is known with
2229 : respect to the targets preferred alignment for VECTYPE, and FALSE
2230 : otherwise. */
2231 :
2232 : inline bool
2233 2324288 : known_alignment_for_access_p (dr_vec_info *dr_info, tree vectype)
2234 : {
2235 2082610 : return (dr_misalignment (dr_info, vectype) != DR_MISALIGNMENT_UNKNOWN);
2236 : }
2237 :
2238 : /* Return the minimum alignment in bytes that the vectorized version
2239 : of DR_INFO is guaranteed to have. */
2240 :
2241 : inline unsigned int
2242 277672 : vect_known_alignment_in_bytes (dr_vec_info *dr_info, tree vectype,
2243 : poly_int64 offset = 0)
2244 : {
2245 277672 : int misalignment = dr_misalignment (dr_info, vectype, offset);
2246 277672 : if (misalignment == DR_MISALIGNMENT_UNKNOWN)
2247 134486 : return TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr_info->dr)));
2248 143186 : else if (misalignment == 0)
2249 100024 : return known_alignment (DR_TARGET_ALIGNMENT (dr_info));
2250 43162 : return misalignment & -misalignment;
2251 : }
2252 :
2253 : /* Return the behavior of DR_INFO with respect to the vectorization context
2254 : (which for outer loop vectorization might not be the behavior recorded
2255 : in DR_INFO itself). */
2256 :
2257 : inline innermost_loop_behavior *
2258 5831123 : vect_dr_behavior (vec_info *vinfo, dr_vec_info *dr_info)
2259 : {
2260 5831123 : stmt_vec_info stmt_info = dr_info->stmt;
2261 5831123 : loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo);
2262 2307484 : if (loop_vinfo == NULL
2263 2307484 : || !nested_in_vect_loop_p (LOOP_VINFO_LOOP (loop_vinfo), stmt_info))
2264 5826830 : return &DR_INNERMOST (dr_info->dr);
2265 : else
2266 4293 : return &STMT_VINFO_DR_WRT_VEC_LOOP (stmt_info);
2267 : }
2268 :
2269 : /* Return the offset calculated by adding the offset of this DR_INFO to the
2270 : corresponding data_reference's offset. If CHECK_OUTER then use
2271 : vect_dr_behavior to select the appropriate data_reference to use. */
2272 :
2273 : inline tree
2274 727468 : get_dr_vinfo_offset (vec_info *vinfo,
2275 : dr_vec_info *dr_info, bool check_outer = false)
2276 : {
2277 727468 : innermost_loop_behavior *base;
2278 727468 : if (check_outer)
2279 687783 : base = vect_dr_behavior (vinfo, dr_info);
2280 : else
2281 39685 : base = &dr_info->dr->innermost;
2282 :
2283 727468 : tree offset = base->offset;
2284 :
2285 727468 : if (!dr_info->offset)
2286 : return offset;
2287 :
2288 1155 : offset = fold_convert (sizetype, offset);
2289 1155 : return fold_build2 (PLUS_EXPR, TREE_TYPE (dr_info->offset), offset,
2290 : dr_info->offset);
2291 : }
2292 :
2293 :
2294 : /* Return the vect cost model for LOOP. */
2295 : inline enum vect_cost_model
2296 2442313 : loop_cost_model (loop_p loop)
2297 : {
2298 2442313 : if (loop != NULL
2299 1771508 : && loop->force_vectorize
2300 77652 : && flag_simd_cost_model != VECT_COST_MODEL_DEFAULT)
2301 : return flag_simd_cost_model;
2302 2364661 : return flag_vect_cost_model;
2303 : }
2304 :
2305 : /* Return true if the vect cost model is unlimited. */
2306 : inline bool
2307 1697569 : unlimited_cost_model (loop_p loop)
2308 : {
2309 1697569 : return loop_cost_model (loop) == VECT_COST_MODEL_UNLIMITED;
2310 : }
2311 :
2312 : /* Return true if the loop described by LOOP_VINFO is fully-masked and
2313 : if the first iteration should use a partial mask in order to achieve
2314 : alignment. */
2315 :
2316 : inline bool
2317 266454 : vect_use_loop_mask_for_alignment_p (loop_vec_info loop_vinfo)
2318 : {
2319 : /* With early break vectorization we don't know whether the accesses will stay
2320 : inside the loop or not. TODO: The early break adjustment code can be
2321 : implemented the same way as vectorizable_linear_induction. However we
2322 : can't test this today so reject it. */
2323 85 : return (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)
2324 85 : && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo)
2325 266458 : && !(LOOP_VINFO_NON_LINEAR_IV (loop_vinfo)
2326 0 : && LOOP_VINFO_EARLY_BREAKS (loop_vinfo)));
2327 : }
2328 :
2329 : /* Return the number of vectors of type VECTYPE that are needed to get
2330 : NUNITS elements. NUNITS should be based on the vectorization factor,
2331 : so it is always a known multiple of the number of elements in VECTYPE. */
2332 :
2333 : inline unsigned int
2334 7352219 : vect_get_num_vectors (poly_uint64 nunits, tree vectype)
2335 : {
2336 7352219 : return exact_div (nunits, TYPE_VECTOR_SUBPARTS (vectype)).to_constant ();
2337 : }
2338 :
2339 : /* Return the number of vectors in the context of vectorization region VINFO,
2340 : needed for a group of statements and a vector type as specified by NODE. */
2341 :
2342 : inline unsigned int
2343 7351406 : vect_get_num_copies (vec_info *vinfo, slp_tree node)
2344 : {
2345 7351406 : poly_uint64 vf;
2346 :
2347 7351406 : if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
2348 3361873 : vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
2349 : else
2350 : vf = 1;
2351 :
2352 7351406 : vf *= SLP_TREE_LANES (node);
2353 7351406 : tree vectype = SLP_TREE_VECTYPE (node);
2354 :
2355 7351406 : return vect_get_num_vectors (vf, vectype);
2356 : }
2357 :
2358 : /* Update maximum unit count *MAX_NUNITS so that it accounts for
2359 : NUNITS. *MAX_NUNITS can be 1 if we haven't yet recorded anything. */
2360 :
2361 : inline void
2362 9726399 : vect_update_max_nunits (poly_uint64 *max_nunits, poly_uint64 nunits)
2363 : {
2364 : /* All unit counts have the form vec_info::vector_size * X for some
2365 : rational X, so two unit sizes must have a common multiple.
2366 : Everything is a multiple of the initial value of 1. */
2367 4256195 : *max_nunits = force_common_multiple (*max_nunits, nunits);
2368 : }
2369 :
2370 : /* Update maximum unit count *MAX_NUNITS so that it accounts for
2371 : the number of units in vector type VECTYPE. *MAX_NUNITS can be 1
2372 : if we haven't yet recorded any vector types. */
2373 :
2374 : inline void
2375 5470204 : vect_update_max_nunits (poly_uint64 *max_nunits, tree vectype)
2376 : {
2377 5470204 : vect_update_max_nunits (max_nunits, TYPE_VECTOR_SUBPARTS (vectype));
2378 5470204 : }
2379 :
2380 : /* Return the vectorization factor that should be used for costing
2381 : purposes while vectorizing the loop described by LOOP_VINFO.
2382 : Pick a reasonable estimate if the vectorization factor isn't
2383 : known at compile time. */
2384 :
2385 : inline unsigned int
2386 1299355 : vect_vf_for_cost (loop_vec_info loop_vinfo)
2387 : {
2388 1299355 : return estimated_poly_value (LOOP_VINFO_VECT_FACTOR (loop_vinfo));
2389 : }
2390 :
2391 : /* Estimate the number of elements in VEC_TYPE for costing purposes.
2392 : Pick a reasonable estimate if the exact number isn't known at
2393 : compile time. */
2394 :
2395 : inline unsigned int
2396 31861 : vect_nunits_for_cost (tree vec_type)
2397 : {
2398 31861 : return estimated_poly_value (TYPE_VECTOR_SUBPARTS (vec_type));
2399 : }
2400 :
2401 : /* Return the maximum possible vectorization factor for LOOP_VINFO. */
2402 :
2403 : inline unsigned HOST_WIDE_INT
2404 105635 : vect_max_vf (loop_vec_info loop_vinfo)
2405 : {
2406 105635 : unsigned HOST_WIDE_INT vf;
2407 105635 : if (LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (&vf))
2408 105635 : return vf;
2409 : return MAX_VECTORIZATION_FACTOR;
2410 : }
2411 :
2412 : /* Return the size of the value accessed by unvectorized data reference
2413 : DR_INFO. This is only valid once STMT_VINFO_VECTYPE has been calculated
2414 : for the associated gimple statement, since that guarantees that DR_INFO
2415 : accesses either a scalar or a scalar equivalent. ("Scalar equivalent"
2416 : here includes things like V1SI, which can be vectorized in the same way
2417 : as a plain SI.) */
2418 :
2419 : inline unsigned int
2420 1997034 : vect_get_scalar_dr_size (dr_vec_info *dr_info)
2421 : {
2422 1997034 : return tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_info->dr))));
2423 : }
2424 :
2425 : /* Return true if LOOP_VINFO requires a runtime check for whether the
2426 : vector loop is profitable. */
2427 :
2428 : inline bool
2429 71574 : vect_apply_runtime_profitability_check_p (loop_vec_info loop_vinfo)
2430 : {
2431 71574 : unsigned int th = LOOP_VINFO_COST_MODEL_THRESHOLD (loop_vinfo);
2432 37766 : return (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
2433 71574 : && th >= vect_vf_for_cost (loop_vinfo));
2434 : }
2435 :
2436 : /* Return true if CODE is a lane-reducing opcode. */
2437 :
2438 : inline bool
2439 387239 : lane_reducing_op_p (code_helper code)
2440 : {
2441 387239 : return code == DOT_PROD_EXPR || code == WIDEN_SUM_EXPR || code == SAD_EXPR;
2442 : }
2443 :
2444 : /* Return true if STMT is a lane-reducing statement. */
2445 :
2446 : inline bool
2447 469701 : lane_reducing_stmt_p (gimple *stmt)
2448 : {
2449 469701 : if (auto *assign = dyn_cast <gassign *> (stmt))
2450 327749 : return lane_reducing_op_p (gimple_assign_rhs_code (assign));
2451 : return false;
2452 : }
2453 :
2454 : /* Source location + hotness information. */
2455 : extern dump_user_location_t vect_location;
2456 :
2457 : /* A macro for calling:
2458 : dump_begin_scope (MSG, vect_location);
2459 : via an RAII object, thus printing "=== MSG ===\n" to the dumpfile etc,
2460 : and then calling
2461 : dump_end_scope ();
2462 : once the object goes out of scope, thus capturing the nesting of
2463 : the scopes.
2464 :
2465 : These scopes affect dump messages within them: dump messages at the
2466 : top level implicitly default to MSG_PRIORITY_USER_FACING, whereas those
2467 : in a nested scope implicitly default to MSG_PRIORITY_INTERNALS. */
2468 :
2469 : #define DUMP_VECT_SCOPE(MSG) \
2470 : AUTO_DUMP_SCOPE (MSG, vect_location)
2471 :
2472 : /* A sentinel class for ensuring that the "vect_location" global gets
2473 : reset at the end of a scope.
2474 :
2475 : The "vect_location" global is used during dumping and contains a
2476 : location_t, which could contain references to a tree block via the
2477 : ad-hoc data. This data is used for tracking inlining information,
2478 : but it's not a GC root; it's simply assumed that such locations never
2479 : get accessed if the blocks are optimized away.
2480 :
2481 : Hence we need to ensure that such locations are purged at the end
2482 : of any operations using them (e.g. via this class). */
2483 :
2484 : class auto_purge_vect_location
2485 : {
2486 : public:
2487 : ~auto_purge_vect_location ();
2488 : };
2489 :
2490 : /*-----------------------------------------------------------------*/
2491 : /* Function prototypes. */
2492 : /*-----------------------------------------------------------------*/
2493 :
2494 : /* Simple loop peeling and versioning utilities for vectorizer's purposes -
2495 : in tree-vect-loop-manip.cc. */
2496 : extern void vect_set_loop_condition (class loop *, edge, loop_vec_info,
2497 : tree, tree, tree, bool);
2498 : extern bool slpeel_can_duplicate_loop_p (const class loop *, const_edge,
2499 : const_edge);
2500 : class loop *slpeel_tree_duplicate_loop_to_edge_cfg (class loop *, edge,
2501 : class loop *, edge,
2502 : edge, edge *, bool = true,
2503 : vec<basic_block> * = NULL,
2504 : bool = false, bool = false,
2505 : bool = true);
2506 : class loop *vect_loop_versioning (loop_vec_info, gimple *);
2507 : extern class loop *vect_do_peeling (loop_vec_info, tree, tree,
2508 : tree *, tree *, tree *, int, bool, bool,
2509 : tree *);
2510 : extern tree vect_get_main_loop_result (loop_vec_info, tree, tree);
2511 : extern void vect_prepare_for_masked_peels (loop_vec_info);
2512 : extern dump_user_location_t find_loop_location (class loop *);
2513 : extern bool vect_can_advance_ivs_p (loop_vec_info);
2514 : extern void vect_update_inits_of_drs (loop_vec_info, tree, tree_code);
2515 : extern edge vec_init_loop_exit_info (class loop *);
2516 : extern void vect_iv_increment_position (edge, gimple_stmt_iterator *, bool *);
2517 :
2518 : /* In tree-vect-stmts.cc. */
2519 : extern tree get_related_vectype_for_scalar_type (machine_mode, tree,
2520 : poly_uint64 = 0);
2521 : extern tree get_vectype_for_scalar_type (vec_info *, tree, unsigned int = 0);
2522 : extern tree get_vectype_for_scalar_type (vec_info *, tree, slp_tree);
2523 : extern tree get_mask_type_for_scalar_type (vec_info *, tree, unsigned int = 0);
2524 : extern tree get_mask_type_for_scalar_type (vec_info *, tree, slp_tree);
2525 : extern tree get_same_sized_vectype (tree, tree);
2526 : extern bool vect_chooses_same_modes_p (vec_info *, machine_mode);
2527 : extern bool vect_chooses_same_modes_p (machine_mode, machine_mode);
2528 : extern bool vect_get_loop_mask_type (loop_vec_info);
2529 : extern bool vect_is_simple_use (tree, vec_info *, enum vect_def_type *,
2530 : stmt_vec_info * = NULL, gimple ** = NULL);
2531 : extern bool vect_is_simple_use (vec_info *, slp_tree,
2532 : unsigned, tree *, slp_tree *,
2533 : enum vect_def_type *,
2534 : tree *, stmt_vec_info * = NULL);
2535 : extern bool vect_maybe_update_slp_op_vectype (slp_tree, tree);
2536 : extern tree perm_mask_for_reverse (tree);
2537 : extern bool supportable_widening_operation (code_helper, tree, tree, bool,
2538 : code_helper*, code_helper*,
2539 : int*, vec<tree> *);
2540 : extern bool supportable_narrowing_operation (code_helper, tree, tree,
2541 : code_helper *, int *,
2542 : vec<tree> *);
2543 : extern bool supportable_indirect_convert_operation (code_helper,
2544 : tree, tree,
2545 : vec<std::pair<tree, tree_code> > &,
2546 : tree = NULL_TREE,
2547 : slp_tree = NULL);
2548 : extern int compare_step_with_zero (vec_info *, stmt_vec_info);
2549 :
2550 : extern unsigned record_stmt_cost (stmt_vector_for_cost *, int,
2551 : enum vect_cost_for_stmt, stmt_vec_info,
2552 : tree, int, enum vect_cost_model_location);
2553 : extern unsigned record_stmt_cost (stmt_vector_for_cost *, int,
2554 : enum vect_cost_for_stmt, slp_tree,
2555 : tree, int, enum vect_cost_model_location);
2556 : extern unsigned record_stmt_cost (stmt_vector_for_cost *, int,
2557 : enum vect_cost_for_stmt,
2558 : enum vect_cost_model_location);
2559 : extern unsigned record_stmt_cost (stmt_vector_for_cost *, int,
2560 : enum vect_cost_for_stmt, stmt_vec_info,
2561 : slp_tree, tree, int,
2562 : enum vect_cost_model_location);
2563 :
2564 : /* Overload of record_stmt_cost with VECTYPE derived from STMT_INFO. */
2565 :
2566 : inline unsigned
2567 1806315 : record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count,
2568 : enum vect_cost_for_stmt kind, stmt_vec_info stmt_info,
2569 : int misalign, enum vect_cost_model_location where)
2570 : {
2571 1805750 : return record_stmt_cost (body_cost_vec, count, kind, stmt_info,
2572 1805750 : STMT_VINFO_VECTYPE (stmt_info), misalign, where);
2573 : }
2574 :
2575 : /* Overload of record_stmt_cost with VECTYPE derived from SLP node. */
2576 :
2577 : inline unsigned
2578 1641370 : record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count,
2579 : enum vect_cost_for_stmt kind, slp_tree node,
2580 : int misalign, enum vect_cost_model_location where)
2581 : {
2582 1436001 : return record_stmt_cost (body_cost_vec, count, kind, node,
2583 71610 : SLP_TREE_VECTYPE (node), misalign, where);
2584 : }
2585 :
2586 : extern void vect_finish_replace_stmt (vec_info *, stmt_vec_info, gimple *);
2587 : extern void vect_finish_stmt_generation (vec_info *, stmt_vec_info, gimple *,
2588 : gimple_stmt_iterator *);
2589 : extern opt_result vect_mark_stmts_to_be_vectorized (loop_vec_info, bool *);
2590 : extern tree vect_get_store_rhs (stmt_vec_info);
2591 : void vect_get_vec_defs (vec_info *, slp_tree,
2592 : tree, vec<tree> *,
2593 : tree = NULL, vec<tree> * = NULL,
2594 : tree = NULL, vec<tree> * = NULL,
2595 : tree = NULL, vec<tree> * = NULL);
2596 : extern tree vect_init_vector (vec_info *, stmt_vec_info, tree, tree,
2597 : gimple_stmt_iterator *);
2598 : extern tree vect_get_slp_vect_def (slp_tree, unsigned);
2599 : extern bool vect_transform_stmt (vec_info *, stmt_vec_info,
2600 : gimple_stmt_iterator *,
2601 : slp_tree, slp_instance);
2602 : extern void vect_remove_stores (vec_info *, stmt_vec_info);
2603 : extern bool vect_nop_conversion_p (stmt_vec_info);
2604 : extern opt_result vect_analyze_stmt (vec_info *, slp_tree,
2605 : slp_instance, stmt_vector_for_cost *);
2606 : extern void vect_get_load_cost (vec_info *, stmt_vec_info, slp_tree, int,
2607 : dr_alignment_support, int, bool,
2608 : unsigned int *, unsigned int *,
2609 : stmt_vector_for_cost *,
2610 : stmt_vector_for_cost *, bool);
2611 : extern void vect_get_store_cost (vec_info *, stmt_vec_info, slp_tree, int,
2612 : dr_alignment_support, int,
2613 : unsigned int *, stmt_vector_for_cost *);
2614 : extern bool vect_supportable_shift (vec_info *, enum tree_code, tree);
2615 : extern tree vect_gen_perm_mask_any (tree, const vec_perm_indices &);
2616 : extern tree vect_gen_perm_mask_checked (tree, const vec_perm_indices &);
2617 : extern void optimize_mask_stores (class loop*);
2618 : extern tree vect_gen_while (gimple_seq *, tree, tree, tree,
2619 : const char * = nullptr);
2620 : extern tree vect_gen_while_not (gimple_seq *, tree, tree, tree);
2621 : extern opt_result vect_get_vector_types_for_stmt (vec_info *,
2622 : stmt_vec_info, tree *,
2623 : tree *, unsigned int = 0);
2624 : extern opt_tree vect_get_mask_type_for_stmt (stmt_vec_info, unsigned int = 0);
2625 :
2626 : /* In tree-if-conv.cc. */
2627 : extern bool ref_within_array_bound (gimple *, tree);
2628 :
2629 : /* In tree-vect-data-refs.cc. */
2630 : extern bool vect_can_force_dr_alignment_p (const_tree, poly_uint64);
2631 : extern enum dr_alignment_support vect_supportable_dr_alignment
2632 : (vec_info *, dr_vec_info *, tree, int,
2633 : bool = false);
2634 : extern tree vect_get_smallest_scalar_type (stmt_vec_info, tree);
2635 : extern opt_result vect_analyze_data_ref_dependences (loop_vec_info, unsigned int *);
2636 : extern bool vect_slp_analyze_instance_dependence (vec_info *, slp_instance);
2637 : extern opt_result vect_enhance_data_refs_alignment (loop_vec_info);
2638 : extern void vect_analyze_data_refs_alignment (loop_vec_info);
2639 : extern bool vect_slp_analyze_instance_alignment (vec_info *, slp_instance);
2640 : extern opt_result vect_analyze_data_ref_accesses (vec_info *, vec<int> *);
2641 : extern opt_result vect_prune_runtime_alias_test_list (loop_vec_info);
2642 : extern bool vect_gather_scatter_fn_p (vec_info *, bool, bool, tree, tree,
2643 : tree, int, int *, internal_fn *, tree *,
2644 : tree *, vec<int> * = nullptr);
2645 : extern bool vect_check_gather_scatter (stmt_vec_info, tree,
2646 : loop_vec_info, gather_scatter_info *,
2647 : vec<int> * = nullptr);
2648 : extern void vect_describe_gather_scatter_call (stmt_vec_info,
2649 : gather_scatter_info *);
2650 : extern opt_result vect_find_stmt_data_reference (loop_p, gimple *,
2651 : vec<data_reference_p> *,
2652 : vec<int> *, int);
2653 : extern opt_result vect_analyze_data_refs (vec_info *, bool *);
2654 : extern void vect_record_base_alignments (vec_info *);
2655 : extern tree vect_create_data_ref_ptr (vec_info *,
2656 : stmt_vec_info, tree, class loop *, tree,
2657 : tree *, gimple_stmt_iterator *,
2658 : gimple **, bool,
2659 : tree = NULL_TREE);
2660 : extern tree bump_vector_ptr (vec_info *, tree, gimple_stmt_iterator *,
2661 : stmt_vec_info, tree);
2662 : extern void vect_copy_ref_info (tree, tree);
2663 : extern tree vect_create_destination_var (tree, tree);
2664 : extern bool vect_grouped_store_supported (tree, unsigned HOST_WIDE_INT);
2665 : extern internal_fn vect_store_lanes_supported (tree, unsigned HOST_WIDE_INT, bool);
2666 : extern bool vect_grouped_load_supported (tree, bool, unsigned HOST_WIDE_INT);
2667 : extern internal_fn vect_load_lanes_supported (tree, unsigned HOST_WIDE_INT,
2668 : bool, vec<int> * = nullptr);
2669 : extern tree vect_setup_realignment (vec_info *,
2670 : stmt_vec_info, tree, gimple_stmt_iterator *,
2671 : tree *, enum dr_alignment_support, tree,
2672 : class loop **);
2673 : extern tree vect_get_new_vect_var (tree, enum vect_var_kind, const char *);
2674 : extern tree vect_get_new_ssa_name (tree, enum vect_var_kind,
2675 : const char * = NULL);
2676 : extern tree vect_create_addr_base_for_vector_ref (vec_info *,
2677 : stmt_vec_info, gimple_seq *,
2678 : tree);
2679 :
2680 : /* In tree-vect-loop.cc. */
2681 : extern tree neutral_op_for_reduction (tree, code_helper, tree, bool = true);
2682 : extern widest_int vect_iv_limit_for_partial_vectors (loop_vec_info loop_vinfo);
2683 : bool vect_rgroup_iv_might_wrap_p (loop_vec_info, rgroup_controls *);
2684 : /* Used in gimple-loop-interchange.c and tree-parloops.cc. */
2685 : extern bool check_reduction_path (dump_user_location_t, loop_p, gphi *, tree,
2686 : enum tree_code);
2687 : extern bool needs_fold_left_reduction_p (tree, code_helper);
2688 : /* Drive for loop analysis stage. */
2689 : extern opt_loop_vec_info vect_analyze_loop (class loop *, gimple *,
2690 : vec_info_shared *);
2691 : extern tree vect_build_loop_niters (loop_vec_info, bool * = NULL);
2692 : extern void vect_gen_vector_loop_niters (loop_vec_info, tree, tree *,
2693 : tree *, bool);
2694 : extern tree vect_get_loop_iv_increment (loop_vec_info);
2695 : extern tree vect_halve_mask_nunits (tree, machine_mode);
2696 : extern tree vect_double_mask_nunits (tree, machine_mode);
2697 : extern void vect_record_loop_mask (loop_vec_info, vec_loop_masks *,
2698 : unsigned int, tree, tree);
2699 : extern tree vect_get_loop_mask (loop_vec_info, gimple_stmt_iterator *,
2700 : vec_loop_masks *,
2701 : unsigned int, tree, unsigned int);
2702 : extern void vect_record_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
2703 : tree, unsigned int);
2704 : extern tree vect_get_loop_len (loop_vec_info, gimple_stmt_iterator *,
2705 : vec_loop_lens *, unsigned int, tree,
2706 : unsigned int, unsigned int, bool);
2707 : extern tree vect_gen_loop_len_mask (loop_vec_info, gimple_stmt_iterator *,
2708 : gimple_stmt_iterator *, vec_loop_lens *,
2709 : unsigned int, tree, tree, unsigned int,
2710 : unsigned int);
2711 : extern gimple_seq vect_gen_len (tree, tree, tree, tree);
2712 : extern vect_reduc_info info_for_reduction (loop_vec_info, slp_tree);
2713 : extern bool reduction_fn_for_scalar_code (code_helper, internal_fn *);
2714 : extern unsigned vect_min_prec_for_max_niters (loop_vec_info, unsigned int);
2715 : /* Drive for loop transformation stage. */
2716 : extern class loop *vect_transform_loop (loop_vec_info, gimple *);
2717 925042 : struct vect_loop_form_info
2718 : {
2719 : tree number_of_iterations;
2720 : tree number_of_iterationsm1;
2721 : tree assumptions;
2722 : auto_vec<gcond *> conds;
2723 : gcond *inner_loop_cond;
2724 : edge loop_exit;
2725 : };
2726 : extern opt_result vect_analyze_loop_form (class loop *, gimple *,
2727 : vect_loop_form_info *);
2728 : extern loop_vec_info vect_create_loop_vinfo (class loop *, vec_info_shared *,
2729 : const vect_loop_form_info *,
2730 : loop_vec_info = nullptr);
2731 : extern bool vectorizable_live_operation (vec_info *, stmt_vec_info,
2732 : slp_tree, slp_instance, int,
2733 : bool, stmt_vector_for_cost *);
2734 : extern bool vectorizable_lane_reducing (loop_vec_info, stmt_vec_info,
2735 : slp_tree, stmt_vector_for_cost *);
2736 : extern bool vectorizable_reduction (loop_vec_info, stmt_vec_info,
2737 : slp_tree, slp_instance,
2738 : stmt_vector_for_cost *);
2739 : extern bool vectorizable_induction (loop_vec_info, stmt_vec_info,
2740 : slp_tree, stmt_vector_for_cost *);
2741 : extern bool vect_transform_reduction (loop_vec_info, stmt_vec_info,
2742 : gimple_stmt_iterator *,
2743 : slp_tree);
2744 : extern bool vect_transform_cycle_phi (loop_vec_info, stmt_vec_info,
2745 : slp_tree, slp_instance);
2746 : extern bool vectorizable_lc_phi (loop_vec_info, stmt_vec_info, slp_tree);
2747 : extern bool vect_transform_lc_phi (loop_vec_info, stmt_vec_info, slp_tree);
2748 : extern bool vectorizable_phi (bb_vec_info, stmt_vec_info, slp_tree,
2749 : stmt_vector_for_cost *);
2750 : extern bool vectorizable_recurr (loop_vec_info, stmt_vec_info,
2751 : slp_tree, stmt_vector_for_cost *);
2752 : extern bool vectorizable_early_exit (loop_vec_info, stmt_vec_info,
2753 : gimple_stmt_iterator *,
2754 : slp_tree, stmt_vector_for_cost *);
2755 : extern bool vect_emulated_vector_p (tree);
2756 : extern bool vect_can_vectorize_without_simd_p (tree_code);
2757 : extern bool vect_can_vectorize_without_simd_p (code_helper);
2758 : extern int vect_get_known_peeling_cost (loop_vec_info, int);
2759 : extern tree cse_and_gimplify_to_preheader (loop_vec_info, tree);
2760 :
2761 : /* Nonlinear induction. */
2762 : extern tree vect_peel_nonlinear_iv_init (gimple_seq*, tree, tree,
2763 : tree, enum vect_induction_op_type,
2764 : bool);
2765 :
2766 : /* In tree-vect-slp.cc. */
2767 : extern void vect_slp_init (void);
2768 : extern void vect_slp_fini (void);
2769 : extern void vect_free_slp_instance (slp_instance);
2770 : extern bool vect_transform_slp_perm_load (vec_info *, slp_tree, const vec<tree> &,
2771 : gimple_stmt_iterator *, poly_uint64,
2772 : bool, unsigned *,
2773 : unsigned * = nullptr, bool = false);
2774 : extern bool vectorizable_slp_permutation (vec_info *, gimple_stmt_iterator *,
2775 : slp_tree, stmt_vector_for_cost *);
2776 : extern bool vect_slp_analyze_operations (vec_info *);
2777 : extern void vect_schedule_slp (vec_info *, const vec<slp_instance> &);
2778 : extern opt_result vect_analyze_slp (vec_info *, unsigned, bool);
2779 : extern bool vect_make_slp_decision (loop_vec_info);
2780 : extern bool vect_detect_hybrid_slp (loop_vec_info);
2781 : extern void vect_optimize_slp (vec_info *);
2782 : extern void vect_gather_slp_loads (vec_info *);
2783 : extern tree vect_get_slp_scalar_def (slp_tree, unsigned);
2784 : extern void vect_get_slp_defs (slp_tree, vec<tree> *);
2785 : extern void vect_get_slp_defs (vec_info *, slp_tree, vec<vec<tree> > *,
2786 : unsigned n = -1U);
2787 : extern bool vect_slp_if_converted_bb (basic_block bb, loop_p orig_loop);
2788 : extern bool vect_slp_function (function *);
2789 : extern stmt_vec_info vect_find_last_scalar_stmt_in_slp (slp_tree);
2790 : extern stmt_vec_info vect_find_first_scalar_stmt_in_slp (slp_tree);
2791 : extern bool is_simple_and_all_uses_invariant (stmt_vec_info, loop_vec_info);
2792 : extern bool can_duplicate_and_interleave_p (vec_info *, unsigned int, tree,
2793 : unsigned int * = NULL,
2794 : tree * = NULL, tree * = NULL);
2795 : extern void duplicate_and_interleave (vec_info *, gimple_seq *, tree,
2796 : const vec<tree> &, unsigned int, vec<tree> &);
2797 : extern int vect_get_place_in_interleaving_chain (stmt_vec_info, stmt_vec_info);
2798 : extern slp_tree vect_create_new_slp_node (unsigned, tree_code);
2799 : extern void vect_free_slp_tree (slp_tree);
2800 : extern bool compatible_calls_p (gcall *, gcall *, bool);
2801 : extern int vect_slp_child_index_for_operand (const stmt_vec_info, int op);
2802 :
2803 : extern tree prepare_vec_mask (loop_vec_info, tree, tree, tree,
2804 : gimple_stmt_iterator *);
2805 : extern tree vect_get_mask_load_else (int, tree);
2806 : extern bool vect_load_perm_consecutive_p (slp_tree, unsigned = UINT_MAX);
2807 :
2808 : /* In tree-vect-patterns.cc. */
2809 : extern void
2810 : vect_mark_pattern_stmts (vec_info *, stmt_vec_info, gimple *, tree);
2811 : extern bool vect_get_range_info (tree, wide_int*, wide_int*);
2812 :
2813 : /* Pattern recognition functions.
2814 : Additional pattern recognition functions can (and will) be added
2815 : in the future. */
2816 : void vect_pattern_recog (vec_info *);
2817 :
2818 : /* In tree-vectorizer.cc. */
2819 : unsigned vectorize_loops (void);
2820 : void vect_free_loop_info_assumptions (class loop *);
2821 : gimple *vect_loop_vectorized_call (class loop *, gcond **cond = NULL);
2822 : bool vect_stmt_dominates_stmt_p (gimple *, gimple *);
2823 :
2824 : /* SLP Pattern matcher types, tree-vect-slp-patterns.cc. */
2825 :
2826 : /* Forward declaration of possible two operands operation that can be matched
2827 : by the complex numbers pattern matchers. */
2828 : enum _complex_operation : unsigned;
2829 :
2830 : /* All possible load permute values that could result from the partial data-flow
2831 : analysis. */
2832 : typedef enum _complex_perm_kinds {
2833 : PERM_UNKNOWN,
2834 : PERM_EVENODD,
2835 : PERM_ODDEVEN,
2836 : PERM_ODDODD,
2837 : PERM_EVENEVEN,
2838 : /* Can be combined with any other PERM values. */
2839 : PERM_TOP
2840 : } complex_perm_kinds_t;
2841 :
2842 : /* Cache from nodes to the load permutation they represent. */
2843 : typedef hash_map <slp_tree, complex_perm_kinds_t>
2844 : slp_tree_to_load_perm_map_t;
2845 :
2846 : /* Cache from nodes pair to being compatible or not. */
2847 : typedef pair_hash <nofree_ptr_hash <_slp_tree>,
2848 : nofree_ptr_hash <_slp_tree>> slp_node_hash;
2849 : typedef hash_map <slp_node_hash, bool> slp_compat_nodes_map_t;
2850 :
2851 :
2852 : /* Vector pattern matcher base class. All SLP pattern matchers must inherit
2853 : from this type. */
2854 :
2855 : class vect_pattern
2856 : {
2857 : protected:
2858 : /* The number of arguments that the IFN requires. */
2859 : unsigned m_num_args;
2860 :
2861 : /* The internal function that will be used when a pattern is created. */
2862 : internal_fn m_ifn;
2863 :
2864 : /* The current node being inspected. */
2865 : slp_tree *m_node;
2866 :
2867 : /* The list of operands to be the children for the node produced when the
2868 : internal function is created. */
2869 : vec<slp_tree> m_ops;
2870 :
2871 : /* Default constructor where NODE is the root of the tree to inspect. */
2872 1107 : vect_pattern (slp_tree *node, vec<slp_tree> *m_ops, internal_fn ifn)
2873 1107 : {
2874 1107 : this->m_ifn = ifn;
2875 1107 : this->m_node = node;
2876 1107 : this->m_ops.create (0);
2877 1107 : if (m_ops)
2878 32 : this->m_ops.safe_splice (*m_ops);
2879 : }
2880 :
2881 : public:
2882 :
2883 : /* Create a new instance of the pattern matcher class of the given type. */
2884 : static vect_pattern* recognize (slp_tree_to_load_perm_map_t *,
2885 : slp_compat_nodes_map_t *, slp_tree *);
2886 :
2887 : /* Build the pattern from the data collected so far. */
2888 : virtual void build (vec_info *) = 0;
2889 :
2890 : /* Default destructor. */
2891 : virtual ~vect_pattern ()
2892 : {
2893 : this->m_ops.release ();
2894 : }
2895 : };
2896 :
2897 : /* Function pointer to create a new pattern matcher from a generic type. */
2898 : typedef vect_pattern* (*vect_pattern_decl_t) (slp_tree_to_load_perm_map_t *,
2899 : slp_compat_nodes_map_t *,
2900 : slp_tree *);
2901 :
2902 : /* List of supported pattern matchers. */
2903 : extern vect_pattern_decl_t slp_patterns[];
2904 :
2905 : /* Number of supported pattern matchers. */
2906 : extern size_t num__slp_patterns;
2907 :
2908 : /* ----------------------------------------------------------------------
2909 : Target support routines
2910 : -----------------------------------------------------------------------
2911 : The following routines are provided to simplify costing decisions in
2912 : target code. Please add more as needed. */
2913 :
2914 : /* Return true if an operation of kind KIND for STMT_INFO represents
2915 : the extraction of an element from a vector in preparation for
2916 : storing the element to memory. */
2917 : inline bool
2918 : vect_is_store_elt_extraction (vect_cost_for_stmt kind, stmt_vec_info stmt_info)
2919 : {
2920 : return (kind == vec_to_scalar
2921 : && STMT_VINFO_DATA_REF (stmt_info)
2922 : && DR_IS_WRITE (STMT_VINFO_DATA_REF (stmt_info)));
2923 : }
2924 :
2925 : /* Return true if STMT_INFO represents part of a reduction. */
2926 : inline bool
2927 49155155 : vect_is_reduction (stmt_vec_info stmt_info)
2928 : {
2929 49155155 : return STMT_VINFO_REDUC_IDX (stmt_info) != -1;
2930 : }
2931 :
2932 : /* Return true if SLP_NODE represents part of a reduction. */
2933 : inline bool
2934 267143 : vect_is_reduction (slp_tree slp_node)
2935 : {
2936 267143 : return SLP_TREE_REDUC_IDX (slp_node) != -1;
2937 : }
2938 :
2939 : /* If STMT_INFO describes a reduction, return the vect_reduction_type
2940 : of the reduction it describes, otherwise return -1. */
2941 : inline int
2942 48 : vect_reduc_type (vec_info *vinfo, slp_tree node)
2943 : {
2944 48 : if (loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo))
2945 : {
2946 48 : vect_reduc_info reduc_info = info_for_reduction (loop_vinfo, node);
2947 48 : if (reduc_info)
2948 45 : return int (VECT_REDUC_INFO_TYPE (reduc_info));
2949 : }
2950 : return -1;
2951 : }
2952 :
2953 : /* If STMT_INFO is a COND_EXPR that includes an embedded comparison, return the
2954 : scalar type of the values being compared. Return null otherwise. */
2955 : inline tree
2956 0 : vect_embedded_comparison_type (stmt_vec_info stmt_info)
2957 : {
2958 0 : if (auto *assign = dyn_cast<gassign *> (stmt_info->stmt))
2959 0 : if (gimple_assign_rhs_code (assign) == COND_EXPR)
2960 : {
2961 0 : tree cond = gimple_assign_rhs1 (assign);
2962 0 : if (COMPARISON_CLASS_P (cond))
2963 0 : return TREE_TYPE (TREE_OPERAND (cond, 0));
2964 : }
2965 : return NULL_TREE;
2966 : }
2967 :
2968 : /* If STMT_INFO is a comparison or contains an embedded comparison, return the
2969 : scalar type of the values being compared. Return null otherwise. */
2970 : inline tree
2971 67181 : vect_comparison_type (stmt_vec_info stmt_info)
2972 : {
2973 67181 : if (auto *assign = dyn_cast<gassign *> (stmt_info->stmt))
2974 67181 : if (TREE_CODE_CLASS (gimple_assign_rhs_code (assign)) == tcc_comparison)
2975 67181 : return TREE_TYPE (gimple_assign_rhs1 (assign));
2976 0 : return vect_embedded_comparison_type (stmt_info);
2977 : }
2978 :
2979 : /* Return true if STMT_INFO extends the result of a load. */
2980 : inline bool
2981 : vect_is_extending_load (class vec_info *vinfo, stmt_vec_info stmt_info)
2982 : {
2983 : /* Although this is quite large for an inline function, this part
2984 : at least should be inline. */
2985 : gassign *assign = dyn_cast <gassign *> (stmt_info->stmt);
2986 : if (!assign || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (assign)))
2987 : return false;
2988 :
2989 : tree rhs = gimple_assign_rhs1 (stmt_info->stmt);
2990 : tree lhs_type = TREE_TYPE (gimple_assign_lhs (assign));
2991 : tree rhs_type = TREE_TYPE (rhs);
2992 : if (!INTEGRAL_TYPE_P (lhs_type)
2993 : || !INTEGRAL_TYPE_P (rhs_type)
2994 : || TYPE_PRECISION (lhs_type) <= TYPE_PRECISION (rhs_type))
2995 : return false;
2996 :
2997 : stmt_vec_info def_stmt_info = vinfo->lookup_def (rhs);
2998 : return (def_stmt_info
2999 : && STMT_VINFO_DATA_REF (def_stmt_info)
3000 : && DR_IS_READ (STMT_VINFO_DATA_REF (def_stmt_info)));
3001 : }
3002 :
3003 : /* Return true if STMT_INFO is an integer truncation. */
3004 : inline bool
3005 : vect_is_integer_truncation (stmt_vec_info stmt_info)
3006 : {
3007 : gassign *assign = dyn_cast <gassign *> (stmt_info->stmt);
3008 : if (!assign || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (assign)))
3009 : return false;
3010 :
3011 : tree lhs_type = TREE_TYPE (gimple_assign_lhs (assign));
3012 : tree rhs_type = TREE_TYPE (gimple_assign_rhs1 (assign));
3013 : return (INTEGRAL_TYPE_P (lhs_type)
3014 : && INTEGRAL_TYPE_P (rhs_type)
3015 : && TYPE_PRECISION (lhs_type) < TYPE_PRECISION (rhs_type));
3016 : }
3017 :
3018 : /* Build a GIMPLE_ASSIGN or GIMPLE_CALL with the tree_code,
3019 : or internal_fn contained in ch, respectively. */
3020 : gimple * vect_gimple_build (tree, code_helper, tree, tree = NULL_TREE);
3021 : #endif /* GCC_TREE_VECTORIZER_H */
|
