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 : #include "ssa.h"
35 :
36 : /* Used for naming of new temporaries. */
37 : enum vect_var_kind {
38 : vect_simple_var,
39 : vect_pointer_var,
40 : vect_scalar_var,
41 : vect_mask_var
42 : };
43 :
44 : /* Defines type of operation. */
45 : enum operation_type {
46 : unary_op = 1,
47 : binary_op,
48 : ternary_op
49 : };
50 :
51 : /* Define type of available alignment support. */
52 : enum dr_alignment_support {
53 : dr_unaligned_unsupported,
54 : dr_unaligned_supported,
55 : dr_explicit_realign,
56 : dr_explicit_realign_optimized,
57 : dr_aligned
58 : };
59 :
60 : /* Define type of peeling support to indicate how peeling for alignment can help
61 : make vectorization supported. */
62 : enum peeling_support {
63 : peeling_known_supported,
64 : peeling_maybe_supported,
65 : peeling_unsupported
66 : };
67 :
68 : /* Define type of def-use cross-iteration cycle. */
69 : enum vect_def_type {
70 : vect_uninitialized_def = 0,
71 : vect_constant_def = 1,
72 : vect_external_def,
73 : vect_internal_def,
74 : vect_induction_def,
75 : vect_reduction_def,
76 : vect_double_reduction_def,
77 : vect_nested_cycle,
78 : vect_first_order_recurrence,
79 : vect_condition_def,
80 : vect_unknown_def_type
81 : };
82 :
83 : /* Define operation type of linear/non-linear induction variable. */
84 : enum vect_induction_op_type {
85 : vect_step_op_add = 0,
86 : vect_step_op_neg,
87 : vect_step_op_mul,
88 : vect_step_op_shl,
89 : vect_step_op_shr
90 : };
91 :
92 : /* Define type of reduction. */
93 : enum vect_reduction_type {
94 : TREE_CODE_REDUCTION,
95 : COND_REDUCTION,
96 : INTEGER_INDUC_COND_REDUCTION,
97 : CONST_COND_REDUCTION,
98 :
99 : /* Retain a scalar phi and use a FOLD_EXTRACT_LAST within the loop
100 : to implement:
101 :
102 : for (int i = 0; i < VF; ++i)
103 : res = cond[i] ? val[i] : res; */
104 : EXTRACT_LAST_REDUCTION,
105 :
106 : /* Use a folding reduction within the loop to implement:
107 :
108 : for (int i = 0; i < VF; ++i)
109 : res = res OP val[i];
110 :
111 : (with no reassociation). */
112 : FOLD_LEFT_REDUCTION
113 : };
114 :
115 : #define VECTORIZABLE_CYCLE_DEF(D) (((D) == vect_reduction_def) \
116 : || ((D) == vect_double_reduction_def) \
117 : || ((D) == vect_nested_cycle))
118 :
119 : /* Structure to encapsulate information about a group of like
120 : instructions to be presented to the target cost model. */
121 : struct stmt_info_for_cost {
122 : int count;
123 : enum vect_cost_for_stmt kind;
124 : enum vect_cost_model_location where;
125 : stmt_vec_info stmt_info;
126 : slp_tree node;
127 : tree vectype;
128 : int misalign;
129 : };
130 :
131 : typedef vec<stmt_info_for_cost> stmt_vector_for_cost;
132 :
133 : /* Maps base addresses to an innermost_loop_behavior and the stmt it was
134 : derived from that gives the maximum known alignment for that base. */
135 : typedef hash_map<tree_operand_hash,
136 : std::pair<stmt_vec_info, innermost_loop_behavior *> >
137 : vec_base_alignments;
138 :
139 : /* Represents elements [START, START + LENGTH) of cyclical array OPS*
140 : (i.e. OPS repeated to give at least START + LENGTH elements) */
141 : struct vect_scalar_ops_slice
142 : {
143 : tree op (unsigned int i) const;
144 : bool all_same_p () const;
145 :
146 : vec<tree> *ops;
147 : unsigned int start;
148 : unsigned int length;
149 : };
150 :
151 : /* Return element I of the slice. */
152 : inline tree
153 2728362 : vect_scalar_ops_slice::op (unsigned int i) const
154 : {
155 5456724 : return (*ops)[(i + start) % ops->length ()];
156 : }
157 :
158 : /* Hash traits for vect_scalar_ops_slice. */
159 : struct vect_scalar_ops_slice_hash : typed_noop_remove<vect_scalar_ops_slice>
160 : {
161 : typedef vect_scalar_ops_slice value_type;
162 : typedef vect_scalar_ops_slice compare_type;
163 :
164 : static const bool empty_zero_p = true;
165 :
166 : static void mark_deleted (value_type &s) { s.length = ~0U; }
167 0 : static void mark_empty (value_type &s) { s.length = 0; }
168 447637 : static bool is_deleted (const value_type &s) { return s.length == ~0U; }
169 4220883 : static bool is_empty (const value_type &s) { return s.length == 0; }
170 : static hashval_t hash (const value_type &);
171 : static bool equal (const value_type &, const compare_type &);
172 : };
173 :
174 : /* Describes how we're going to vectorize an individual load or store,
175 : or a group of loads or stores. */
176 : enum vect_memory_access_type {
177 : VMAT_UNINITIALIZED,
178 :
179 : /* An access to an invariant address. This is used only for loads. */
180 : VMAT_INVARIANT,
181 :
182 : /* A simple contiguous access. */
183 : VMAT_CONTIGUOUS,
184 :
185 : /* A contiguous access that goes down in memory rather than up,
186 : with no additional permutation. This is used only for stores
187 : of invariants. */
188 : VMAT_CONTIGUOUS_DOWN,
189 :
190 : /* A simple contiguous access in which the elements need to be reversed
191 : after loading or before storing. */
192 : VMAT_CONTIGUOUS_REVERSE,
193 :
194 : /* An access that uses IFN_LOAD_LANES or IFN_STORE_LANES. */
195 : VMAT_LOAD_STORE_LANES,
196 :
197 : /* An access in which each scalar element is loaded or stored
198 : individually. */
199 : VMAT_ELEMENTWISE,
200 :
201 : /* A hybrid of VMAT_CONTIGUOUS and VMAT_ELEMENTWISE, used for grouped
202 : SLP accesses. Each unrolled iteration uses a contiguous load
203 : or store for the whole group, but the groups from separate iterations
204 : are combined in the same way as for VMAT_ELEMENTWISE. */
205 : VMAT_STRIDED_SLP,
206 :
207 : /* The access uses gather loads or scatter stores. */
208 : VMAT_GATHER_SCATTER_LEGACY,
209 : VMAT_GATHER_SCATTER_IFN,
210 : VMAT_GATHER_SCATTER_EMULATED
211 : };
212 :
213 : /* Returns whether MAT is any of the VMAT_GATHER_SCATTER_* kinds. */
214 :
215 : inline bool
216 6660484 : mat_gather_scatter_p (vect_memory_access_type mat)
217 : {
218 6660484 : return (mat == VMAT_GATHER_SCATTER_LEGACY
219 : || mat == VMAT_GATHER_SCATTER_IFN
220 6660484 : || mat == VMAT_GATHER_SCATTER_EMULATED);
221 : }
222 :
223 : /*-----------------------------------------------------------------*/
224 : /* Info on vectorized defs. */
225 : /*-----------------------------------------------------------------*/
226 : enum stmt_vec_info_type {
227 : undef_vec_info_type = 0,
228 : load_vec_info_type,
229 : store_vec_info_type,
230 : shift_vec_info_type,
231 : op_vec_info_type,
232 : call_vec_info_type,
233 : call_simd_clone_vec_info_type,
234 : assignment_vec_info_type,
235 : condition_vec_info_type,
236 : comparison_vec_info_type,
237 : reduc_vec_info_type,
238 : induc_vec_info_type,
239 : type_promotion_vec_info_type,
240 : type_demotion_vec_info_type,
241 : type_conversion_vec_info_type,
242 : cycle_phi_info_type,
243 : lc_phi_info_type,
244 : phi_info_type,
245 : recurr_info_type,
246 : loop_exit_ctrl_vec_info_type,
247 : permute_info_type
248 : };
249 :
250 : /************************************************************************
251 : SLP
252 : ************************************************************************/
253 : typedef vec<std::pair<unsigned, unsigned> > lane_permutation_t;
254 : typedef auto_vec<std::pair<unsigned, unsigned>, 16> auto_lane_permutation_t;
255 : typedef vec<unsigned> load_permutation_t;
256 : typedef auto_vec<unsigned, 16> auto_load_permutation_t;
257 :
258 3424622 : struct vect_data {
259 2133527 : virtual ~vect_data () = default;
260 : };
261 :
262 : /* Analysis data from vectorizable_simd_clone_call for
263 : call_simd_clone_vec_info_type. */
264 : struct vect_simd_clone_data : vect_data {
265 1895 : virtual ~vect_simd_clone_data () = default;
266 1418 : vect_simd_clone_data () = default;
267 477 : vect_simd_clone_data (vect_simd_clone_data &&other) = default;
268 :
269 : /* Selected SIMD clone and clone for in-branch. */
270 : cgraph_node *clone;
271 : cgraph_node *clone_inbranch;
272 :
273 : /* Selected SIMD clone's function info. First vector element
274 : is NULL_TREE, followed by a pair of trees (base + step)
275 : for linear arguments (pair of NULLs for other arguments). */
276 : auto_vec<tree> simd_clone_info;
277 : };
278 :
279 : /* Analysis data from vectorizable_load and vectorizable_store for
280 : load_vec_info_type and store_vec_info_type. */
281 : struct vect_load_store_data : vect_data {
282 1289200 : vect_load_store_data (vect_load_store_data &&other) = default;
283 2133527 : vect_load_store_data () = default;
284 3421938 : virtual ~vect_load_store_data () = default;
285 :
286 : vect_memory_access_type memory_access_type;
287 : dr_alignment_support alignment_support_scheme;
288 : int misalignment;
289 : internal_fn lanes_ifn; // VMAT_LOAD_STORE_LANES
290 : poly_int64 poffset;
291 : union {
292 : internal_fn ifn; // VMAT_GATHER_SCATTER_IFN
293 : tree decl; // VMAT_GATHER_SCATTER_DECL
294 : } gs;
295 : tree strided_offset_vectype; // VMAT_GATHER_SCATTER_IFN, originally strided
296 : /* Load/store type with larger element mode used for punning the vectype. */
297 : tree ls_type; // VMAT_GATHER_SCATTER_IFN, VMAT_STRIDED_SLP
298 : /* Load/store element type used for punning the vectype. Relevant when
299 : that is a vector type. */
300 : tree ls_eltype; // VMAT_STRIDED_SLP
301 : /* This is set to a supported offset vector type if we don't support the
302 : originally requested offset type, otherwise NULL.
303 : If nonzero there will be an additional offset conversion before
304 : the gather/scatter. */
305 : tree supported_offset_vectype; // VMAT_GATHER_SCATTER_IFN
306 : /* Similar for scale. Only nonzero if we don't support the requested
307 : scale. Then we need to multiply the offset vector before the
308 : gather/scatter. */
309 : int supported_scale; // VMAT_GATHER_SCATTER_IFN
310 : auto_vec<int> elsvals;
311 : /* True if the load requires a load permutation. */
312 : bool slp_perm; // SLP_TREE_LOAD_PERMUTATION
313 : unsigned n_perms; // SLP_TREE_LOAD_PERMUTATION
314 : unsigned n_loads; // SLP_TREE_LOAD_PERMUTATION
315 : /* Whether the load permutation is consecutive and simple. */
316 : bool subchain_p; // VMAT_STRIDED_SLP and VMAT_GATHER_SCATTER
317 : };
318 :
319 : struct match_elt_t {
320 : int v;
321 : operator bool () const = delete;
322 : match_elt_t& operator= (bool) = delete;
323 43529367 : match_elt_t& operator= (int v_) { v = v_; return *this; }
324 : };
325 :
326 : inline bool operator!= (const match_elt_t &m, const match_elt_t &m2) { return m.v != m2.v; }
327 : inline bool operator== (const match_elt_t &m, const match_elt_t &m2) { return m.v == m2.v; }
328 47056052 : inline bool operator== (const match_elt_t &m, int v) { return m.v == v; }
329 41632937 : inline bool operator!= (const match_elt_t &m, int v) { return m.v != v; }
330 :
331 : /* A computation tree of an SLP instance. Each node corresponds to a group of
332 : stmts to be packed in a SIMD stmt. */
333 : struct _slp_tree {
334 : _slp_tree ();
335 : ~_slp_tree ();
336 :
337 : void push_vec_def (gimple *def);
338 12267 : void push_vec_def (tree def) { vec_defs.quick_push (def); }
339 :
340 : /* Nodes that contain def-stmts of this node statements operands. */
341 : vec<slp_tree> children;
342 :
343 : /* A group of scalar stmts to be vectorized together. Unused when
344 : def_type is vect_external_def or vect_costant_def. */
345 : vec<stmt_vec_info> stmts;
346 : /* A group of scalar operands to be vectorized together. Unused
347 : unless def_type is vect_external_def or vect_constant_def. */
348 : vec<tree> ops;
349 : /* A set of lane indices that are live and to be code-generated from
350 : this SLP node. */
351 : vec<unsigned> live_lanes;
352 :
353 : /* The representative that should be used for analysis and
354 : code generation. NULL when code is VEC_PERM_EXPR. */
355 : stmt_vec_info representative;
356 :
357 : struct {
358 : /* SLP cycle the node resides in, or -1. */
359 : int id;
360 : /* The SLP operand index with the edge on the SLP cycle, or -1. */
361 : int reduc_idx;
362 : } cycle_info;
363 :
364 : /* Load permutation mapping outgoing vector lanes to lanes of
365 : the single DR group the load accesses. NULL if there is no
366 : permutation. Unused when this is not a load. */
367 : load_permutation_t load_permutation;
368 : /* Lane permutation of the operands scalar lanes encoded as pairs
369 : of { operand number, lane number }. The number of elements
370 : denotes the number of output lanes. Unused unless code is
371 : VEC_PERM_EXPR. */
372 : lane_permutation_t lane_permutation;
373 :
374 : tree vectype;
375 : /* Vectorized defs. */
376 : vec<tree> vec_defs;
377 : /* Insertion place for verification purposes. Only set for
378 : BB vectorization. NULL denotes region entry. */
379 : gimple *si;
380 :
381 : /* Reference count in the SLP graph. */
382 : unsigned int refcnt;
383 : /* The DEF type of this node. */
384 : enum vect_def_type def_type;
385 : /* The number of scalar lanes produced by this node. */
386 : unsigned int lanes;
387 : /* The operation of this node. Either VEC_PERM_EXPR or ERROR_MARK. */
388 : enum tree_code code;
389 : /* For gather/scatter memory operations the scale each offset element
390 : should be multiplied by before being added to the base. */
391 : int gs_scale;
392 : /* For gather/scatter memory operations the loop-invariant base value. */
393 : tree gs_base;
394 : /* Whether uses of this load or feeders of this store are suitable
395 : for load/store-lanes. */
396 : bool ldst_lanes;
397 : /* For BB vect, flag to indicate this load node should be vectorized
398 : as to avoid STLF fails because of related stores. */
399 : bool avoid_stlf_fail;
400 :
401 : /* The vertex index of this node when a full graph is built. */
402 : int vertex;
403 :
404 : /* The kind of operation as determined by analysis and optional
405 : kind specific data. */
406 : enum stmt_vec_info_type type;
407 : vect_data *data;
408 :
409 : template <class T>
410 2134945 : T& get_data (T& else_) { return data ? *static_cast <T *> (data) : else_; }
411 :
412 : /* If not NULL this is a cached failed SLP discovery attempt with
413 : the lanes that failed during SLP discovery as 'false'. This is
414 : a copy of the matches array. */
415 : match_elt_t *failed;
416 :
417 : /* Allocate from slp_tree_pool. */
418 : static void *operator new (size_t);
419 :
420 : /* Return memory to slp_tree_pool. */
421 : static void operator delete (void *, size_t);
422 :
423 : /* Linked list of nodes to release when we free the slp_tree_pool. */
424 : slp_tree next_node;
425 : slp_tree prev_node;
426 : };
427 :
428 : /* The enum describes the type of operations that an SLP instance
429 : can perform. */
430 :
431 : enum slp_instance_kind {
432 : slp_inst_kind_store,
433 : slp_inst_kind_reduc_group,
434 : slp_inst_kind_reduc_chain,
435 : slp_inst_kind_bb_reduc,
436 : slp_inst_kind_ctor,
437 : slp_inst_kind_gcond
438 : };
439 :
440 : /* SLP instance is a sequence of stmts in a loop that can be packed into
441 : SIMD stmts. */
442 : typedef class _slp_instance {
443 : public:
444 : /* The root of SLP tree. */
445 : slp_tree root;
446 :
447 : /* For vector constructors, the constructor stmt that the SLP tree is built
448 : from, NULL otherwise. */
449 : vec<stmt_vec_info> root_stmts;
450 :
451 : /* For slp_inst_kind_bb_reduc the defs that were not vectorized, NULL
452 : otherwise. */
453 : vec<tree> remain_defs;
454 :
455 : /* The group of nodes that contain loads of this SLP instance. */
456 : vec<slp_tree> loads;
457 :
458 : /* The SLP node containing the reduction PHIs. */
459 : slp_tree reduc_phis;
460 :
461 : /* Vector cost of this entry to the SLP graph. */
462 : stmt_vector_for_cost cost_vec;
463 :
464 : /* If this instance is the main entry of a subgraph the set of
465 : entries into the same subgraph, including itself. */
466 : vec<_slp_instance *> subgraph_entries;
467 :
468 : /* The type of operation the SLP instance is performing. */
469 : slp_instance_kind kind;
470 :
471 : dump_user_location_t location () const;
472 : } *slp_instance;
473 :
474 :
475 : /* Access Functions. */
476 : #define SLP_INSTANCE_TREE(S) (S)->root
477 : #define SLP_INSTANCE_LOADS(S) (S)->loads
478 : #define SLP_INSTANCE_ROOT_STMTS(S) (S)->root_stmts
479 : #define SLP_INSTANCE_REMAIN_DEFS(S) (S)->remain_defs
480 : #define SLP_INSTANCE_KIND(S) (S)->kind
481 :
482 : #define SLP_TREE_CHILDREN(S) (S)->children
483 : #define SLP_TREE_SCALAR_STMTS(S) (S)->stmts
484 : #define SLP_TREE_SCALAR_OPS(S) (S)->ops
485 : #define SLP_TREE_LIVE_LANES(S) (S)->live_lanes
486 : #define SLP_TREE_REF_COUNT(S) (S)->refcnt
487 : #define SLP_TREE_VEC_DEFS(S) (S)->vec_defs
488 : #define SLP_TREE_LOAD_PERMUTATION(S) (S)->load_permutation
489 : #define SLP_TREE_LANE_PERMUTATION(S) (S)->lane_permutation
490 : #define SLP_TREE_DEF_TYPE(S) (S)->def_type
491 : #define SLP_TREE_VECTYPE(S) (S)->vectype
492 : #define SLP_TREE_REPRESENTATIVE(S) (S)->representative
493 : #define SLP_TREE_LANES(S) (S)->lanes
494 : #define SLP_TREE_CODE(S) (S)->code
495 : #define SLP_TREE_TYPE(S) (S)->type
496 : #define SLP_TREE_GS_SCALE(S) (S)->gs_scale
497 : #define SLP_TREE_GS_BASE(S) (S)->gs_base
498 : #define SLP_TREE_REDUC_IDX(S) (S)->cycle_info.reduc_idx
499 : #define SLP_TREE_PERMUTE_P(S) ((S)->code == VEC_PERM_EXPR)
500 :
501 : inline vect_memory_access_type
502 1065398 : SLP_TREE_MEMORY_ACCESS_TYPE (slp_tree node)
503 : {
504 383396 : if (SLP_TREE_TYPE (node) == load_vec_info_type
505 365686 : || SLP_TREE_TYPE (node) == store_vec_info_type)
506 116450 : return static_cast<vect_load_store_data *> (node->data)->memory_access_type;
507 : return VMAT_UNINITIALIZED;
508 : }
509 :
510 : enum vect_partial_vector_style {
511 : vect_partial_vectors_none,
512 : vect_partial_vectors_while_ult,
513 : vect_partial_vectors_avx512,
514 : vect_partial_vectors_len
515 : };
516 :
517 : /* Key for map that records association between
518 : scalar conditions and corresponding loop mask, and
519 : is populated by vect_record_loop_mask. */
520 :
521 : struct scalar_cond_masked_key
522 : {
523 60012 : scalar_cond_masked_key (tree t, unsigned ncopies_)
524 60012 : : ncopies (ncopies_)
525 : {
526 60012 : get_cond_ops_from_tree (t);
527 : }
528 :
529 : void get_cond_ops_from_tree (tree);
530 :
531 : unsigned ncopies;
532 : bool inverted_p;
533 : tree_code code;
534 : tree op0;
535 : tree op1;
536 : };
537 :
538 : template<>
539 : struct default_hash_traits<scalar_cond_masked_key>
540 : {
541 : typedef scalar_cond_masked_key compare_type;
542 : typedef scalar_cond_masked_key value_type;
543 :
544 : static inline hashval_t
545 68923 : hash (value_type v)
546 : {
547 68923 : inchash::hash h;
548 68923 : h.add_int (v.code);
549 68923 : inchash::add_expr (v.op0, h, 0);
550 68923 : inchash::add_expr (v.op1, h, 0);
551 68923 : h.add_int (v.ncopies);
552 68923 : h.add_flag (v.inverted_p);
553 68923 : return h.end ();
554 : }
555 :
556 : static inline bool
557 10569 : equal (value_type existing, value_type candidate)
558 : {
559 10569 : return (existing.ncopies == candidate.ncopies
560 10353 : && existing.code == candidate.code
561 6429 : && existing.inverted_p == candidate.inverted_p
562 4981 : && operand_equal_p (existing.op0, candidate.op0, 0)
563 13722 : && operand_equal_p (existing.op1, candidate.op1, 0));
564 : }
565 :
566 : static const bool empty_zero_p = true;
567 :
568 : static inline void
569 0 : mark_empty (value_type &v)
570 : {
571 0 : v.ncopies = 0;
572 0 : v.inverted_p = false;
573 : }
574 :
575 : static inline bool
576 9268432 : is_empty (value_type v)
577 : {
578 9209590 : return v.ncopies == 0;
579 : }
580 :
581 : static inline void mark_deleted (value_type &) {}
582 :
583 : static inline bool is_deleted (const value_type &)
584 : {
585 : return false;
586 : }
587 :
588 51688 : static inline void remove (value_type &) {}
589 : };
590 :
591 : typedef hash_set<scalar_cond_masked_key> scalar_cond_masked_set_type;
592 :
593 : /* Key and map that records association between vector conditions and
594 : corresponding loop mask, and is populated by prepare_vec_mask. */
595 :
596 : typedef pair_hash<tree_operand_hash, tree_operand_hash> tree_cond_mask_hash;
597 : typedef hash_set<tree_cond_mask_hash> vec_cond_masked_set_type;
598 :
599 : /* Describes two objects whose addresses must be unequal for the vectorized
600 : loop to be valid. */
601 : typedef std::pair<tree, tree> vec_object_pair;
602 :
603 : /* Records that vectorization is only possible if abs (EXPR) >= MIN_VALUE.
604 : UNSIGNED_P is true if we can assume that abs (EXPR) == EXPR. */
605 : class vec_lower_bound {
606 : public:
607 : vec_lower_bound () {}
608 1757 : vec_lower_bound (tree e, bool u, poly_uint64 m)
609 1757 : : expr (e), unsigned_p (u), min_value (m) {}
610 :
611 : tree expr;
612 : bool unsigned_p;
613 : poly_uint64 min_value;
614 : };
615 :
616 : /* Vectorizer state shared between different analyses like vector sizes
617 : of the same CFG region. */
618 : class vec_info_shared {
619 : public:
620 : vec_info_shared();
621 : ~vec_info_shared();
622 :
623 : void save_datarefs();
624 : void check_datarefs();
625 :
626 : /* All data references. Freed by free_data_refs, so not an auto_vec. */
627 : vec<data_reference_p> datarefs;
628 : vec<data_reference> datarefs_copy;
629 :
630 : /* The loop nest in which the data dependences are computed. */
631 : auto_vec<loop_p> loop_nest;
632 :
633 : /* All data dependences. Freed by free_dependence_relations, so not
634 : an auto_vec. */
635 : vec<ddr_p> ddrs;
636 : };
637 :
638 : /* Vectorizer state common between loop and basic-block vectorization. */
639 : class vec_info {
640 : public:
641 : typedef hash_set<int_hash<machine_mode, E_VOIDmode, E_BLKmode> > mode_set;
642 : enum vec_kind { bb, loop };
643 :
644 : vec_info (vec_kind, vec_info_shared *);
645 : ~vec_info ();
646 :
647 : stmt_vec_info add_stmt (gimple *);
648 : stmt_vec_info add_pattern_stmt (gimple *, stmt_vec_info);
649 : stmt_vec_info resync_stmt_addr (gimple *);
650 : stmt_vec_info lookup_stmt (gimple *);
651 : stmt_vec_info lookup_def (tree);
652 : stmt_vec_info lookup_single_use (tree);
653 : class dr_vec_info *lookup_dr (data_reference *);
654 : void move_dr (stmt_vec_info, stmt_vec_info);
655 : void remove_stmt (stmt_vec_info);
656 : void replace_stmt (gimple_stmt_iterator *, stmt_vec_info, gimple *);
657 : void insert_on_entry (stmt_vec_info, gimple *);
658 : void insert_seq_on_entry (stmt_vec_info, gimple_seq);
659 :
660 : /* The type of vectorization. */
661 : vec_kind kind;
662 :
663 : /* Shared vectorizer state. */
664 : vec_info_shared *shared;
665 :
666 : /* The mapping of GIMPLE UID to stmt_vec_info. */
667 : vec<stmt_vec_info> stmt_vec_infos;
668 : /* Whether the above mapping is complete. */
669 : bool stmt_vec_info_ro;
670 :
671 : /* Whether we've done a transform we think OK to not update virtual
672 : SSA form. */
673 : bool any_known_not_updated_vssa;
674 :
675 : /* The SLP graph. */
676 : auto_vec<slp_instance> slp_instances;
677 :
678 : /* Maps base addresses to an innermost_loop_behavior that gives the maximum
679 : known alignment for that base. */
680 : vec_base_alignments base_alignments;
681 :
682 : /* All interleaving chains of stores, represented by the first
683 : stmt in the chain. */
684 : auto_vec<stmt_vec_info> grouped_stores;
685 :
686 : /* The set of vector modes used in the vectorized region. */
687 : mode_set used_vector_modes;
688 :
689 : /* The argument we should pass to related_vector_mode when looking up
690 : the vector mode for a scalar mode, or VOIDmode if we haven't yet
691 : made any decisions about which vector modes to use. */
692 : machine_mode vector_mode;
693 :
694 : /* The basic blocks in the vectorization region. For _loop_vec_info,
695 : the memory is internally managed, while for _bb_vec_info, it points
696 : to element space of an external auto_vec<>. This inconsistency is
697 : not a good class design pattern. TODO: improve it with an unified
698 : auto_vec<> whose lifetime is confined to vec_info object. */
699 : basic_block *bbs;
700 :
701 : /* The count of the basic blocks in the vectorization region. */
702 : unsigned int nbbs;
703 :
704 : /* Used to keep a sequence of def stmts of a pattern stmt that are loop
705 : invariant if they exists.
706 : The sequence is emitted in the loop preheader should the loop be vectorized
707 : and are reset when undoing patterns. */
708 : gimple_seq inv_pattern_def_seq;
709 :
710 : private:
711 : stmt_vec_info new_stmt_vec_info (gimple *stmt);
712 : void set_vinfo_for_stmt (gimple *, stmt_vec_info, bool = true);
713 : void free_stmt_vec_infos ();
714 : void free_stmt_vec_info (stmt_vec_info);
715 : };
716 :
717 : class _loop_vec_info;
718 : class _bb_vec_info;
719 :
720 : template<>
721 : template<>
722 : inline bool
723 381405338 : is_a_helper <_loop_vec_info *>::test (vec_info *i)
724 : {
725 380744157 : return i->kind == vec_info::loop;
726 : }
727 :
728 : template<>
729 : template<>
730 : inline bool
731 77408467 : is_a_helper <_bb_vec_info *>::test (vec_info *i)
732 : {
733 77408467 : return i->kind == vec_info::bb;
734 : }
735 :
736 : /* In general, we can divide the vector statements in a vectorized loop
737 : into related groups ("rgroups") and say that for each rgroup there is
738 : some nS such that the rgroup operates on nS values from one scalar
739 : iteration followed by nS values from the next. That is, if VF is the
740 : vectorization factor of the loop, the rgroup operates on a sequence:
741 :
742 : (1,1) (1,2) ... (1,nS) (2,1) ... (2,nS) ... (VF,1) ... (VF,nS)
743 :
744 : where (i,j) represents a scalar value with index j in a scalar
745 : iteration with index i.
746 :
747 : [ We use the term "rgroup" to emphasise that this grouping isn't
748 : necessarily the same as the grouping of statements used elsewhere.
749 : For example, if we implement a group of scalar loads using gather
750 : loads, we'll use a separate gather load for each scalar load, and
751 : thus each gather load will belong to its own rgroup. ]
752 :
753 : In general this sequence will occupy nV vectors concatenated
754 : together. If these vectors have nL lanes each, the total number
755 : of scalar values N is given by:
756 :
757 : N = nS * VF = nV * nL
758 :
759 : None of nS, VF, nV and nL are required to be a power of 2. nS and nV
760 : are compile-time constants but VF and nL can be variable (if the target
761 : supports variable-length vectors).
762 :
763 : In classical vectorization, each iteration of the vector loop would
764 : handle exactly VF iterations of the original scalar loop. However,
765 : in vector loops that are able to operate on partial vectors, a
766 : particular iteration of the vector loop might handle fewer than VF
767 : iterations of the scalar loop. The vector lanes that correspond to
768 : iterations of the scalar loop are said to be "active" and the other
769 : lanes are said to be "inactive".
770 :
771 : In such vector loops, many rgroups need to be controlled to ensure
772 : that they have no effect for the inactive lanes. Conceptually, each
773 : such rgroup needs a sequence of booleans in the same order as above,
774 : but with each (i,j) replaced by a boolean that indicates whether
775 : iteration i is active. This sequence occupies nV vector controls
776 : that again have nL lanes each. Thus the control sequence as a whole
777 : consists of VF independent booleans that are each repeated nS times.
778 :
779 : Taking mask-based approach as a partially-populated vectors example.
780 : We make the simplifying assumption that if a sequence of nV masks is
781 : suitable for one (nS,nL) pair, we can reuse it for (nS/2,nL/2) by
782 : VIEW_CONVERTing it. This holds for all current targets that support
783 : fully-masked loops. For example, suppose the scalar loop is:
784 :
785 : float *f;
786 : double *d;
787 : for (int i = 0; i < n; ++i)
788 : {
789 : f[i * 2 + 0] += 1.0f;
790 : f[i * 2 + 1] += 2.0f;
791 : d[i] += 3.0;
792 : }
793 :
794 : and suppose that vectors have 256 bits. The vectorized f accesses
795 : will belong to one rgroup and the vectorized d access to another:
796 :
797 : f rgroup: nS = 2, nV = 1, nL = 8
798 : d rgroup: nS = 1, nV = 1, nL = 4
799 : VF = 4
800 :
801 : [ In this simple example the rgroups do correspond to the normal
802 : SLP grouping scheme. ]
803 :
804 : If only the first three lanes are active, the masks we need are:
805 :
806 : f rgroup: 1 1 | 1 1 | 1 1 | 0 0
807 : d rgroup: 1 | 1 | 1 | 0
808 :
809 : Here we can use a mask calculated for f's rgroup for d's, but not
810 : vice versa.
811 :
812 : Thus for each value of nV, it is enough to provide nV masks, with the
813 : mask being calculated based on the highest nL (or, equivalently, based
814 : on the highest nS) required by any rgroup with that nV. We therefore
815 : represent the entire collection of masks as a two-level table, with the
816 : first level being indexed by nV - 1 (since nV == 0 doesn't exist) and
817 : the second being indexed by the mask index 0 <= i < nV. */
818 :
819 : /* The controls (like masks or lengths) needed by rgroups with nV vectors,
820 : according to the description above. */
821 : struct rgroup_controls {
822 : /* The largest nS for all rgroups that use these controls.
823 : For vect_partial_vectors_avx512 this is the constant nscalars_per_iter
824 : for all members of the group. */
825 : unsigned int max_nscalars_per_iter;
826 :
827 : /* For the largest nS recorded above, the loop controls divide each scalar
828 : into FACTOR equal-sized pieces. This is useful if we need to split
829 : element-based accesses into byte-based accesses.
830 : For vect_partial_vectors_avx512 this records nV instead. */
831 : unsigned int factor;
832 :
833 : /* This is a vector type with MAX_NSCALARS_PER_ITER * VF / nV elements.
834 : For mask-based controls, it is the type of the masks in CONTROLS.
835 : For length-based controls, it can be any vector type that has the
836 : specified number of elements; the type of the elements doesn't matter. */
837 : tree type;
838 :
839 : /* When there is no uniformly used LOOP_VINFO_RGROUP_COMPARE_TYPE this
840 : is the rgroup specific type used. */
841 : tree compare_type;
842 :
843 : /* A vector of nV controls, in iteration order. */
844 : vec<tree> controls;
845 :
846 : /* In case of len_load and len_store with a bias there is only one
847 : rgroup. This holds the adjusted loop length for the this rgroup. */
848 : tree bias_adjusted_ctrl;
849 : };
850 :
851 591676 : struct vec_loop_masks
852 : {
853 529232 : bool is_empty () const { return mask_set.is_empty (); }
854 :
855 : /* Set to record vectype, nvector pairs. */
856 : hash_set<pair_hash <nofree_ptr_hash <tree_node>,
857 : int_hash<unsigned, 0>>> mask_set;
858 :
859 : /* rgroup_controls used for the partial vector scheme. */
860 : auto_vec<rgroup_controls> rgc_vec;
861 : };
862 :
863 : typedef auto_vec<rgroup_controls> vec_loop_lens;
864 :
865 : typedef auto_vec<std::pair<data_reference*, tree> > drs_init_vec;
866 :
867 : /* Abstraction around info on reductions which is still in stmt_vec_info
868 : but will be duplicated or moved elsewhere. */
869 208104 : class vect_reduc_info_s
870 : {
871 : public:
872 : /* The def type of the main reduction PHI, vect_reduction_def or
873 : vect_double_reduction_def. */
874 : enum vect_def_type def_type;
875 :
876 : /* The reduction type as detected by
877 : vect_is_simple_reduction and vectorizable_reduction. */
878 : enum vect_reduction_type reduc_type;
879 :
880 : /* The original scalar reduction code, to be used in the epilogue. */
881 : code_helper reduc_code;
882 :
883 : /* A vector internal function we should use in the epilogue. */
884 : internal_fn reduc_fn;
885 :
886 : /* For loop reduction with multiple vectorized results (ncopies > 1), a
887 : lane-reducing operation participating in it may not use all of those
888 : results, this field specifies result index starting from which any
889 : following land-reducing operation would be assigned to. */
890 : unsigned int reduc_result_pos;
891 :
892 : /* Whether this represents a reduction chain. */
893 : bool is_reduc_chain;
894 :
895 : /* Whether we force a single cycle PHI during reduction vectorization. */
896 : bool force_single_cycle;
897 :
898 : /* The vector type for performing the actual reduction operation. */
899 : tree reduc_vectype;
900 :
901 : /* The vector type we should use for the final reduction in the epilogue
902 : when we reduce a mask. */
903 : tree reduc_vectype_for_mask;
904 :
905 : /* The neutral operand to use, if any. */
906 : tree neutral_op;
907 :
908 : /* For INTEGER_INDUC_COND_REDUCTION, the initial value to be used. */
909 : tree induc_cond_initial_val;
910 :
911 : /* If not NULL the value to be added to compute final reduction value. */
912 : tree reduc_epilogue_adjustment;
913 :
914 : /* If non-null, the reduction is being performed by an epilogue loop
915 : and we have decided to reuse this accumulator from the main loop. */
916 : struct vect_reusable_accumulator *reused_accumulator;
917 :
918 : /* If the vector code is performing N scalar reductions in parallel,
919 : this variable gives the initial scalar values of those N reductions. */
920 : auto_vec<tree> reduc_initial_values;
921 :
922 : /* If the vector code is performing N scalar reductions in parallel, this
923 : variable gives the vectorized code's final (scalar) result for each of
924 : those N reductions. In other words, REDUC_SCALAR_RESULTS[I] replaces
925 : the original scalar code's loop-closed SSA PHI for reduction number I. */
926 : auto_vec<tree> reduc_scalar_results;
927 : };
928 :
929 : typedef class vect_reduc_info_s *vect_reduc_info;
930 :
931 : #define VECT_REDUC_INFO_DEF_TYPE(I) ((I)->def_type)
932 : #define VECT_REDUC_INFO_TYPE(I) ((I)->reduc_type)
933 : #define VECT_REDUC_INFO_CODE(I) ((I)->reduc_code)
934 : #define VECT_REDUC_INFO_FN(I) ((I)->reduc_fn)
935 : #define VECT_REDUC_INFO_SCALAR_RESULTS(I) ((I)->reduc_scalar_results)
936 : #define VECT_REDUC_INFO_INITIAL_VALUES(I) ((I)->reduc_initial_values)
937 : #define VECT_REDUC_INFO_REUSED_ACCUMULATOR(I) ((I)->reused_accumulator)
938 : #define VECT_REDUC_INFO_INDUC_COND_INITIAL_VAL(I) ((I)->induc_cond_initial_val)
939 : #define VECT_REDUC_INFO_EPILOGUE_ADJUSTMENT(I) ((I)->reduc_epilogue_adjustment)
940 : #define VECT_REDUC_INFO_VECTYPE(I) ((I)->reduc_vectype)
941 : #define VECT_REDUC_INFO_VECTYPE_FOR_MASK(I) ((I)->reduc_vectype_for_mask)
942 : #define VECT_REDUC_INFO_FORCE_SINGLE_CYCLE(I) ((I)->force_single_cycle)
943 : #define VECT_REDUC_INFO_RESULT_POS(I) ((I)->reduc_result_pos)
944 : #define VECT_REDUC_INFO_NEUTRAL_OP(I) ((I)->neutral_op)
945 :
946 : /* Information about a reduction accumulator from the main loop that could
947 : conceivably be reused as the input to a reduction in an epilogue loop. */
948 : struct vect_reusable_accumulator {
949 : /* The final value of the accumulator, which forms the input to the
950 : reduction operation. */
951 : tree reduc_input;
952 :
953 : /* The stmt_vec_info that describes the reduction (i.e. the one for
954 : which is_reduc_info is true). */
955 : vect_reduc_info reduc_info;
956 : };
957 :
958 : /*-----------------------------------------------------------------*/
959 : /* Info on vectorized loops. */
960 : /*-----------------------------------------------------------------*/
961 : typedef class _loop_vec_info : public vec_info {
962 : public:
963 : _loop_vec_info (class loop *, vec_info_shared *);
964 : ~_loop_vec_info ();
965 :
966 : /* The loop to which this info struct refers to. */
967 : class loop *loop;
968 :
969 : /* Number of latch executions. */
970 : tree num_itersm1;
971 : /* Number of iterations. */
972 : tree num_iters;
973 : /* Number of iterations of the original loop. */
974 : tree num_iters_unchanged;
975 : /* Condition under which this loop is analyzed and versioned. */
976 : tree num_iters_assumptions;
977 :
978 : /* The cost of the vector code. */
979 : class vector_costs *vector_costs;
980 :
981 : /* The cost of the scalar code. */
982 : class vector_costs *scalar_costs;
983 :
984 : /* Threshold of number of iterations below which vectorization will not be
985 : performed. It is calculated from MIN_PROFITABLE_ITERS and
986 : param_min_vect_loop_bound. */
987 : unsigned int th;
988 :
989 : /* When applying loop versioning, the vector form should only be used
990 : if the number of scalar iterations is >= this value, on top of all
991 : the other requirements. Ignored when loop versioning is not being
992 : used. */
993 : poly_uint64 versioning_threshold;
994 :
995 : /* Unrolling factor. In case of suitable super-word parallelism
996 : it can be that no unrolling is needed, and thus this is 1. */
997 : poly_uint64 vectorization_factor;
998 :
999 : /* Gimple operand for the number of scalar iteration handed per loop
1000 : iteration, and therefore how much to increment each IV by. */
1001 : tree iv_increment;
1002 :
1003 : /* If this loop is an epilogue loop whose main loop can be skipped,
1004 : MAIN_LOOP_EDGE is the edge from the main loop to this loop's
1005 : preheader. SKIP_MAIN_LOOP_EDGE is then the edge that skips the
1006 : main loop and goes straight to this loop's preheader.
1007 :
1008 : Both fields are null otherwise. */
1009 : edge main_loop_edge;
1010 : edge skip_main_loop_edge;
1011 :
1012 : /* If this loop is an epilogue loop that might be skipped after executing
1013 : the main loop, this edge is the one that skips the epilogue. */
1014 : edge skip_this_loop_edge;
1015 :
1016 : /* Reduction descriptors of this loop. Referenced to from SLP nodes
1017 : by index. */
1018 : auto_vec<vect_reduc_info> reduc_infos;
1019 :
1020 : /* The vectorized form of a standard reduction replaces the original
1021 : scalar code's final result (a loop-closed SSA PHI) with the result
1022 : of a vector-to-scalar reduction operation. After vectorization,
1023 : this variable maps these vector-to-scalar results to information
1024 : about the reductions that generated them. */
1025 : hash_map<tree, vect_reusable_accumulator> reusable_accumulators;
1026 :
1027 : /* The number of times that the target suggested we unroll the vector loop
1028 : in order to promote more ILP. This value will be used to re-analyze the
1029 : loop for vectorization and if successful the value will be folded into
1030 : vectorization_factor (and therefore exactly divides
1031 : vectorization_factor). */
1032 : unsigned int suggested_unroll_factor;
1033 :
1034 : /* Maximum runtime vectorization factor, or MAX_VECTORIZATION_FACTOR
1035 : if there is no particular limit. */
1036 : unsigned HOST_WIDE_INT max_vectorization_factor;
1037 :
1038 : /* The masks that a fully-masked loop should use to avoid operating
1039 : on inactive scalars. */
1040 : vec_loop_masks masks;
1041 :
1042 : /* The lengths that a loop with length should use to avoid operating
1043 : on inactive scalars. */
1044 : vec_loop_lens lens;
1045 :
1046 : /* Set of scalar conditions that have loop mask applied. */
1047 : scalar_cond_masked_set_type scalar_cond_masked_set;
1048 :
1049 : /* Set of vector conditions that have loop mask applied. */
1050 : vec_cond_masked_set_type vec_cond_masked_set;
1051 :
1052 : /* If we are using a loop mask to align memory addresses, this variable
1053 : contains the number of vector elements that we should skip in the
1054 : first iteration of the vector loop (i.e. the number of leading
1055 : elements that should be false in the first mask). */
1056 : tree mask_skip_niters;
1057 :
1058 : /* If we are using a loop mask to align memory addresses and we're in an
1059 : early break loop then this variable contains the number of elements that
1060 : were skipped during the initial iteration of the loop. */
1061 : tree mask_skip_niters_pfa_offset;
1062 :
1063 : /* The type that the loop control IV should be converted to before
1064 : testing which of the VF scalars are active and inactive.
1065 : Only meaningful if LOOP_VINFO_USING_PARTIAL_VECTORS_P. */
1066 : tree rgroup_compare_type;
1067 :
1068 : /* For #pragma omp simd if (x) loops the x expression. If constant 0,
1069 : the loop should not be vectorized, if constant non-zero, simd_if_cond
1070 : shouldn't be set and loop vectorized normally, if SSA_NAME, the loop
1071 : should be versioned on that condition, using scalar loop if the condition
1072 : is false and vectorized loop otherwise. */
1073 : tree simd_if_cond;
1074 :
1075 : /* The type that the vector loop control IV should have when
1076 : LOOP_VINFO_USING_PARTIAL_VECTORS_P is true. */
1077 : tree rgroup_iv_type;
1078 :
1079 : /* The style used for implementing partial vectors when
1080 : LOOP_VINFO_USING_PARTIAL_VECTORS_P is true. */
1081 : vect_partial_vector_style partial_vector_style;
1082 :
1083 : /* Unknown DRs according to which loop was peeled. */
1084 : class dr_vec_info *unaligned_dr;
1085 :
1086 : /* peeling_for_alignment indicates whether peeling for alignment will take
1087 : place, and what the peeling factor should be:
1088 : peeling_for_alignment = X means:
1089 : If X=0: Peeling for alignment will not be applied.
1090 : If X>0: Peel first X iterations.
1091 : If X=-1: Generate a runtime test to calculate the number of iterations
1092 : to be peeled, using the dataref recorded in the field
1093 : unaligned_dr. */
1094 : int peeling_for_alignment;
1095 :
1096 : /* The mask used to check the alignment of pointers or arrays. */
1097 : poly_uint64 ptr_mask;
1098 :
1099 : /* The maximum speculative read amount in VLA modes for runtime check. */
1100 : poly_uint64 max_spec_read_amount;
1101 :
1102 : /* Indicates whether the loop has any non-linear IV. */
1103 : bool nonlinear_iv;
1104 :
1105 : /* Data Dependence Relations defining address ranges that are candidates
1106 : for a run-time aliasing check. */
1107 : auto_vec<ddr_p> may_alias_ddrs;
1108 :
1109 : /* Data Dependence Relations defining address ranges together with segment
1110 : lengths from which the run-time aliasing check is built. */
1111 : auto_vec<dr_with_seg_len_pair_t> comp_alias_ddrs;
1112 :
1113 : /* Check that the addresses of each pair of objects is unequal. */
1114 : auto_vec<vec_object_pair> check_unequal_addrs;
1115 :
1116 : /* List of values that are required to be nonzero. This is used to check
1117 : whether things like "x[i * n] += 1;" are safe and eventually gets added
1118 : to the checks for lower bounds below. */
1119 : auto_vec<tree> check_nonzero;
1120 :
1121 : /* List of values that need to be checked for a minimum value. */
1122 : auto_vec<vec_lower_bound> lower_bounds;
1123 :
1124 : /* Statements in the loop that have data references that are candidates for a
1125 : runtime (loop versioning) misalignment check. */
1126 : auto_vec<stmt_vec_info> may_misalign_stmts;
1127 :
1128 : /* Reduction cycles detected in the loop. Used in loop-aware SLP. */
1129 : auto_vec<stmt_vec_info> reductions;
1130 :
1131 : /* Defs that could not be analyzed such as OMP SIMD calls without
1132 : a LHS. */
1133 : auto_vec<stmt_vec_info> alternate_defs;
1134 :
1135 : /* Cost vector for a single scalar iteration. */
1136 : auto_vec<stmt_info_for_cost> scalar_cost_vec;
1137 :
1138 : /* Map of IV base/step expressions to inserted name in the preheader. */
1139 : hash_map<tree_operand_hash, tree> *ivexpr_map;
1140 :
1141 : /* Map of OpenMP "omp simd array" scan variables to corresponding
1142 : rhs of the store of the initializer. */
1143 : hash_map<tree, tree> *scan_map;
1144 :
1145 : /* The factor used to over weight those statements in an inner loop
1146 : relative to the loop being vectorized. */
1147 : unsigned int inner_loop_cost_factor;
1148 :
1149 : /* Is the loop vectorizable? */
1150 : bool vectorizable;
1151 :
1152 : /* Records whether we still have the option of vectorizing this loop
1153 : using partially-populated vectors; in other words, whether it is
1154 : still possible for one iteration of the vector loop to handle
1155 : fewer than VF scalars. */
1156 : bool can_use_partial_vectors_p;
1157 :
1158 : /* Records whether we must use niter masking for correctness reasons. */
1159 : bool must_use_partial_vectors_p;
1160 :
1161 : /* True if we've decided to use partially-populated vectors, so that
1162 : the vector loop can handle fewer than VF scalars. */
1163 : bool using_partial_vectors_p;
1164 :
1165 : /* True if we've decided to use a decrementing loop control IV that counts
1166 : scalars. This can be done for any loop that:
1167 :
1168 : (a) uses length "controls"; and
1169 : (b) can iterate more than once. */
1170 : bool using_decrementing_iv_p;
1171 :
1172 : /* True if we've decided to use output of select_vl to adjust IV of
1173 : both loop control and data reference pointer. This is only true
1174 : for single-rgroup control. */
1175 : bool using_select_vl_p;
1176 :
1177 : /* True if we've decided to use peeling with versioning together, which allows
1178 : unaligned unsupported data refs to be uniformly aligned after a certain
1179 : amount of peeling (mutual alignment). Otherwise, we use versioning alone
1180 : so these data refs must be already aligned to a power-of-two boundary
1181 : without peeling. */
1182 : bool allow_mutual_alignment;
1183 :
1184 : /* The bias for len_load and len_store. For now, only 0 and -1 are
1185 : supported. -1 must be used when a backend does not support
1186 : len_load/len_store with a length of zero. */
1187 : signed char partial_load_store_bias;
1188 :
1189 : /* When we have grouped data accesses with gaps, we may introduce invalid
1190 : memory accesses. We peel the last iteration of the loop to prevent
1191 : this. */
1192 : bool peeling_for_gaps;
1193 :
1194 : /* When the number of iterations is not a multiple of the vector size
1195 : we need to peel off iterations at the end to form an epilogue loop. */
1196 : bool peeling_for_niter;
1197 :
1198 : /* When the loop has early breaks that we can vectorize we need to peel
1199 : the loop for the break finding loop. */
1200 : bool early_breaks;
1201 :
1202 : /* List of loop additional IV conditionals found in the loop. */
1203 : auto_vec<gcond *> conds;
1204 :
1205 : /* Main loop IV cond. */
1206 : gcond* loop_iv_cond;
1207 :
1208 : /* True if we have an unroll factor requested by the user through pragma GCC
1209 : unroll. */
1210 : bool user_unroll;
1211 :
1212 : /* True if there are no loop carried data dependencies in the loop.
1213 : If loop->safelen <= 1, then this is always true, either the loop
1214 : didn't have any loop carried data dependencies, or the loop is being
1215 : vectorized guarded with some runtime alias checks, or couldn't
1216 : be vectorized at all, but then this field shouldn't be used.
1217 : For loop->safelen >= 2, the user has asserted that there are no
1218 : backward dependencies, but there still could be loop carried forward
1219 : dependencies in such loops. This flag will be false if normal
1220 : vectorizer data dependency analysis would fail or require versioning
1221 : for alias, but because of loop->safelen >= 2 it has been vectorized
1222 : even without versioning for alias. E.g. in:
1223 : #pragma omp simd
1224 : for (int i = 0; i < m; i++)
1225 : a[i] = a[i + k] * c;
1226 : (or #pragma simd or #pragma ivdep) we can vectorize this and it will
1227 : DTRT even for k > 0 && k < m, but without safelen we would not
1228 : vectorize this, so this field would be false. */
1229 : bool no_data_dependencies;
1230 :
1231 : /* Mark loops having masked stores. */
1232 : bool has_mask_store;
1233 :
1234 : /* Queued scaling factor for the scalar loop. */
1235 : profile_probability scalar_loop_scaling;
1236 :
1237 : /* If if-conversion versioned this loop before conversion, this is the
1238 : loop version without if-conversion. */
1239 : class loop *scalar_loop;
1240 :
1241 : /* For loops being epilogues of already vectorized loops
1242 : this points to the main vectorized loop. Otherwise NULL. */
1243 : _loop_vec_info *main_loop_info;
1244 :
1245 : /* For loops being epilogues of already vectorized loops
1246 : this points to the preceding vectorized (possibly epilogue) loop.
1247 : Otherwise NULL. */
1248 : _loop_vec_info *orig_loop_info;
1249 :
1250 : /* Used to store loop_vec_infos of the epilogue of this loop during
1251 : analysis. */
1252 : _loop_vec_info *epilogue_vinfo;
1253 :
1254 : /* If this is an epilogue loop the DR advancement applied. */
1255 : tree drs_advanced_by;
1256 :
1257 : /* The controlling loop exit for the current loop when vectorizing.
1258 : For counted loops, this IV controls the natural exits of the loop. */
1259 : edge vec_loop_main_exit;
1260 :
1261 : /* The controlling loop exit for the epilogue loop when vectorizing.
1262 : For counted loops, this IV controls the natural exits of the loop. */
1263 : edge vec_epilogue_loop_main_exit;
1264 :
1265 : /* The controlling loop exit for the scalar loop being vectorized.
1266 : For counted loops, this IV controls the natural exits of the loop. */
1267 : edge scalar_loop_main_exit;
1268 :
1269 : /* Indicate if the multiple exit loop has any side-effects that require it to
1270 : have a scalar epilogue. */
1271 : bool early_break_needs_epilogue;
1272 :
1273 : /* Used to store the list of stores needing to be moved if doing early
1274 : break vectorization as they would violate the scalar loop semantics if
1275 : vectorized in their current location. These are stored in order that they
1276 : need to be moved. */
1277 : auto_vec<gimple *> early_break_stores;
1278 :
1279 : /* The final basic block where to move statements to. In the case of
1280 : multiple exits this could be pretty far away. */
1281 : basic_block early_break_dest_bb;
1282 :
1283 : /* Statements whose VUSES need updating if early break vectorization is to
1284 : happen. */
1285 : auto_vec<gimple*> early_break_vuses;
1286 :
1287 : /* The IV adjustment value for inductions that needs to be materialized
1288 : inside the relevant exit blocks in order to adjust for early break. */
1289 : tree early_break_niters_var;
1290 :
1291 : /* The type of the variable to be used to create the scalar IV for early break
1292 : loops. */
1293 : tree early_break_iv_type;
1294 :
1295 : /* Record statements that are needed to be live for early break vectorization
1296 : but may not have an LC PHI node materialized yet in the exits. */
1297 : auto_vec<stmt_vec_info> early_break_live_ivs;
1298 : } *loop_vec_info;
1299 :
1300 : /* Access Functions. */
1301 : #define LOOP_VINFO_LOOP(L) (L)->loop
1302 : #define LOOP_VINFO_MAIN_EXIT(L) (L)->vec_loop_main_exit
1303 : #define LOOP_VINFO_EPILOGUE_MAIN_EXIT(L) (L)->vec_epilogue_loop_main_exit
1304 : #define LOOP_VINFO_SCALAR_MAIN_EXIT(L) (L)->scalar_loop_main_exit
1305 : #define LOOP_VINFO_BBS(L) (L)->bbs
1306 : #define LOOP_VINFO_NBBS(L) (L)->nbbs
1307 : #define LOOP_VINFO_NITERSM1(L) (L)->num_itersm1
1308 : #define LOOP_VINFO_NITERS(L) (L)->num_iters
1309 : #define LOOP_VINFO_NITERS_UNCOUNTED_P(L) (LOOP_VINFO_NITERS (L) \
1310 : == chrec_dont_know)
1311 : /* Since LOOP_VINFO_NITERS and LOOP_VINFO_NITERSM1 can change after
1312 : prologue peeling retain total unchanged scalar loop iterations for
1313 : cost model. */
1314 : #define LOOP_VINFO_NITERS_UNCHANGED(L) (L)->num_iters_unchanged
1315 : #define LOOP_VINFO_NITERS_ASSUMPTIONS(L) (L)->num_iters_assumptions
1316 : #define LOOP_VINFO_COST_MODEL_THRESHOLD(L) (L)->th
1317 : #define LOOP_VINFO_VERSIONING_THRESHOLD(L) (L)->versioning_threshold
1318 : #define LOOP_VINFO_VECTORIZABLE_P(L) (L)->vectorizable
1319 : #define LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P(L) (L)->can_use_partial_vectors_p
1320 : #define LOOP_VINFO_MUST_USE_PARTIAL_VECTORS_P(L) (L)->must_use_partial_vectors_p
1321 : #define LOOP_VINFO_USING_PARTIAL_VECTORS_P(L) (L)->using_partial_vectors_p
1322 : #define LOOP_VINFO_USING_DECREMENTING_IV_P(L) (L)->using_decrementing_iv_p
1323 : #define LOOP_VINFO_USING_SELECT_VL_P(L) (L)->using_select_vl_p
1324 : #define LOOP_VINFO_ALLOW_MUTUAL_ALIGNMENT(L) (L)->allow_mutual_alignment
1325 : #define LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS(L) (L)->partial_load_store_bias
1326 : #define LOOP_VINFO_VECT_FACTOR(L) (L)->vectorization_factor
1327 : #define LOOP_VINFO_IV_INCREMENT(L) (L)->iv_increment
1328 : #define LOOP_VINFO_IV_INCREMENT_INVARIANT_P(L) \
1329 : (!LOOP_VINFO_USING_SELECT_VL_P (L))
1330 : #define LOOP_VINFO_MAX_VECT_FACTOR(L) (L)->max_vectorization_factor
1331 : #define LOOP_VINFO_MASKS(L) (L)->masks
1332 : #define LOOP_VINFO_LENS(L) (L)->lens
1333 : #define LOOP_VINFO_MASK_SKIP_NITERS(L) (L)->mask_skip_niters
1334 : #define LOOP_VINFO_MASK_NITERS_PFA_OFFSET(L) (L)->mask_skip_niters_pfa_offset
1335 : #define LOOP_VINFO_RGROUP_COMPARE_TYPE(L) (L)->rgroup_compare_type
1336 : #define LOOP_VINFO_RGROUP_IV_TYPE(L) (L)->rgroup_iv_type
1337 : #define LOOP_VINFO_PARTIAL_VECTORS_STYLE(L) (L)->partial_vector_style
1338 : #define LOOP_VINFO_PTR_MASK(L) (L)->ptr_mask
1339 : #define LOOP_VINFO_MAX_SPEC_READ_AMOUNT(L) (L)->max_spec_read_amount
1340 : #define LOOP_VINFO_LOOP_NEST(L) (L)->shared->loop_nest
1341 : #define LOOP_VINFO_DATAREFS(L) (L)->shared->datarefs
1342 : #define LOOP_VINFO_DDRS(L) (L)->shared->ddrs
1343 : #define LOOP_VINFO_INT_NITERS(L) (TREE_INT_CST_LOW ((L)->num_iters))
1344 : #define LOOP_VINFO_PEELING_FOR_ALIGNMENT(L) (L)->peeling_for_alignment
1345 : #define LOOP_VINFO_NON_LINEAR_IV(L) (L)->nonlinear_iv
1346 : #define LOOP_VINFO_UNALIGNED_DR(L) (L)->unaligned_dr
1347 : #define LOOP_VINFO_MAY_MISALIGN_STMTS(L) (L)->may_misalign_stmts
1348 : #define LOOP_VINFO_MAY_ALIAS_DDRS(L) (L)->may_alias_ddrs
1349 : #define LOOP_VINFO_COMP_ALIAS_DDRS(L) (L)->comp_alias_ddrs
1350 : #define LOOP_VINFO_CHECK_UNEQUAL_ADDRS(L) (L)->check_unequal_addrs
1351 : #define LOOP_VINFO_CHECK_NONZERO(L) (L)->check_nonzero
1352 : #define LOOP_VINFO_LOWER_BOUNDS(L) (L)->lower_bounds
1353 : #define LOOP_VINFO_USER_UNROLL(L) (L)->user_unroll
1354 : #define LOOP_VINFO_GROUPED_STORES(L) (L)->grouped_stores
1355 : #define LOOP_VINFO_SLP_INSTANCES(L) (L)->slp_instances
1356 : #define LOOP_VINFO_REDUCTIONS(L) (L)->reductions
1357 : #define LOOP_VINFO_PEELING_FOR_GAPS(L) (L)->peeling_for_gaps
1358 : #define LOOP_VINFO_PEELING_FOR_NITER(L) (L)->peeling_for_niter
1359 : #define LOOP_VINFO_EARLY_BREAKS(L) (L)->early_breaks
1360 : #define LOOP_VINFO_EARLY_BRK_NEEDS_EPILOG(L) (L)->early_break_needs_epilogue
1361 : #define LOOP_VINFO_EARLY_BRK_STORES(L) (L)->early_break_stores
1362 : #define LOOP_VINFO_EARLY_BREAKS_VECT_PEELED(L) \
1363 : ((single_pred ((L)->loop->latch) != (L)->vec_loop_main_exit->src) \
1364 : || LOOP_VINFO_NITERS_UNCOUNTED_P (L))
1365 : #define LOOP_VINFO_EARLY_BREAKS_LIVE_IVS(L) \
1366 : (L)->early_break_live_ivs
1367 : #define LOOP_VINFO_EARLY_BRK_DEST_BB(L) (L)->early_break_dest_bb
1368 : #define LOOP_VINFO_EARLY_BRK_VUSES(L) (L)->early_break_vuses
1369 : #define LOOP_VINFO_EARLY_BRK_NITERS_VAR(L) (L)->early_break_niters_var
1370 : #define LOOP_VINFO_EARLY_BRK_IV_TYPE(L) (L)->early_break_iv_type
1371 : #define LOOP_VINFO_LOOP_CONDS(L) (L)->conds
1372 : #define LOOP_VINFO_LOOP_IV_COND(L) (L)->loop_iv_cond
1373 : #define LOOP_VINFO_NO_DATA_DEPENDENCIES(L) (L)->no_data_dependencies
1374 : #define LOOP_VINFO_SCALAR_LOOP(L) (L)->scalar_loop
1375 : #define LOOP_VINFO_SCALAR_LOOP_SCALING(L) (L)->scalar_loop_scaling
1376 : #define LOOP_VINFO_HAS_MASK_STORE(L) (L)->has_mask_store
1377 : #define LOOP_VINFO_SCALAR_ITERATION_COST(L) (L)->scalar_cost_vec
1378 : #define LOOP_VINFO_MAIN_LOOP_INFO(L) (L)->main_loop_info
1379 : #define LOOP_VINFO_ORIG_LOOP_INFO(L) (L)->orig_loop_info
1380 : #define LOOP_VINFO_SIMD_IF_COND(L) (L)->simd_if_cond
1381 : #define LOOP_VINFO_INNER_LOOP_COST_FACTOR(L) (L)->inner_loop_cost_factor
1382 : #define LOOP_VINFO_INV_PATTERN_DEF_SEQ(L) (L)->inv_pattern_def_seq
1383 : #define LOOP_VINFO_DRS_ADVANCED_BY(L) (L)->drs_advanced_by
1384 : #define LOOP_VINFO_ALTERNATE_DEFS(L) (L)->alternate_defs
1385 :
1386 : #define LOOP_VINFO_FULLY_MASKED_P(L) \
1387 : (LOOP_VINFO_USING_PARTIAL_VECTORS_P (L) \
1388 : && !LOOP_VINFO_MASKS (L).is_empty ())
1389 :
1390 : #define LOOP_VINFO_FULLY_WITH_LENGTH_P(L) \
1391 : (LOOP_VINFO_USING_PARTIAL_VECTORS_P (L) \
1392 : && !LOOP_VINFO_LENS (L).is_empty ())
1393 :
1394 : #define LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT(L) \
1395 : ((L)->may_misalign_stmts.length () > 0)
1396 : #define LOOP_REQUIRES_VERSIONING_FOR_SPEC_READ(L) \
1397 : (maybe_gt ((L)->max_spec_read_amount, 0U))
1398 : #define LOOP_REQUIRES_VERSIONING_FOR_ALIAS(L) \
1399 : ((L)->comp_alias_ddrs.length () > 0 \
1400 : || (L)->check_unequal_addrs.length () > 0 \
1401 : || (L)->lower_bounds.length () > 0)
1402 : #define LOOP_REQUIRES_VERSIONING_FOR_NITERS(L) \
1403 : (LOOP_VINFO_NITERS_ASSUMPTIONS (L))
1404 : #define LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND(L) \
1405 : (LOOP_VINFO_SIMD_IF_COND (L))
1406 : #define LOOP_REQUIRES_VERSIONING(L) \
1407 : (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (L) \
1408 : || LOOP_REQUIRES_VERSIONING_FOR_SPEC_READ (L) \
1409 : || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (L) \
1410 : || LOOP_REQUIRES_VERSIONING_FOR_NITERS (L) \
1411 : || LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND (L))
1412 :
1413 : #define LOOP_VINFO_USE_VERSIONING_WITHOUT_PEELING(L) \
1414 : ((L)->may_misalign_stmts.length () > 0 \
1415 : && !LOOP_VINFO_ALLOW_MUTUAL_ALIGNMENT (L))
1416 :
1417 : #define LOOP_VINFO_NITERS_KNOWN_P(L) \
1418 : (tree_fits_shwi_p ((L)->num_iters) && tree_to_shwi ((L)->num_iters) > 0)
1419 :
1420 : #define LOOP_VINFO_EPILOGUE_P(L) \
1421 : (LOOP_VINFO_ORIG_LOOP_INFO (L) != NULL)
1422 :
1423 : #define LOOP_VINFO_ORIG_MAX_VECT_FACTOR(L) \
1424 : (LOOP_VINFO_MAX_VECT_FACTOR (LOOP_VINFO_ORIG_LOOP_INFO (L)))
1425 :
1426 : /* Wrapper for loop_vec_info, for tracking success/failure, where a non-NULL
1427 : value signifies success, and a NULL value signifies failure, supporting
1428 : propagating an opt_problem * describing the failure back up the call
1429 : stack. */
1430 : typedef opt_pointer_wrapper <loop_vec_info> opt_loop_vec_info;
1431 :
1432 : inline loop_vec_info
1433 543591 : loop_vec_info_for_loop (class loop *loop)
1434 : {
1435 543591 : return (loop_vec_info) loop->aux;
1436 : }
1437 :
1438 : struct slp_root
1439 : {
1440 1313013 : slp_root (slp_instance_kind kind_, vec<stmt_vec_info> stmts_,
1441 14484 : vec<stmt_vec_info> roots_, vec<tree> remain_ = vNULL)
1442 1313013 : : kind(kind_), stmts(stmts_), roots(roots_), remain(remain_) {}
1443 : slp_instance_kind kind;
1444 : vec<stmt_vec_info> stmts;
1445 : vec<stmt_vec_info> roots;
1446 : vec<tree> remain;
1447 : };
1448 :
1449 : typedef class _bb_vec_info : public vec_info
1450 : {
1451 : public:
1452 : _bb_vec_info (vec<basic_block> bbs, vec_info_shared *);
1453 : ~_bb_vec_info ();
1454 :
1455 : vec<slp_root> roots;
1456 : } *bb_vec_info;
1457 :
1458 : #define BB_VINFO_BBS(B) (B)->bbs
1459 : #define BB_VINFO_NBBS(B) (B)->nbbs
1460 : #define BB_VINFO_GROUPED_STORES(B) (B)->grouped_stores
1461 : #define BB_VINFO_SLP_INSTANCES(B) (B)->slp_instances
1462 : #define BB_VINFO_DATAREFS(B) (B)->shared->datarefs
1463 : #define BB_VINFO_DDRS(B) (B)->shared->ddrs
1464 :
1465 : /* Indicates whether/how a variable is used in the scope of loop/basic
1466 : block. */
1467 : enum vect_relevant {
1468 : vect_unused_in_scope = 0,
1469 :
1470 : /* The def is only used outside the loop. */
1471 : vect_used_only_live,
1472 : /* The def is in the inner loop, and the use is in the outer loop, and the
1473 : use is a reduction stmt. */
1474 : vect_used_in_outer_by_reduction,
1475 : /* The def is in the inner loop, and the use is in the outer loop (and is
1476 : not part of reduction). */
1477 : vect_used_in_outer,
1478 :
1479 : /* defs that feed computations that end up (only) in a reduction. These
1480 : defs may be used by non-reduction stmts, but eventually, any
1481 : computations/values that are affected by these defs are used to compute
1482 : a reduction (i.e. don't get stored to memory, for example). We use this
1483 : to identify computations that we can change the order in which they are
1484 : computed. */
1485 : vect_used_by_reduction,
1486 :
1487 : vect_used_in_scope
1488 : };
1489 :
1490 : /* The type of vectorization. pure_slp means the stmt is covered by the
1491 : SLP graph, not_vect means it is not. This is mostly used by BB
1492 : vectorization. */
1493 : enum slp_vect_type {
1494 : not_vect = 0,
1495 : pure_slp,
1496 : };
1497 :
1498 : /* Says whether a statement is a load, a store of a vectorized statement
1499 : result, or a store of an invariant value. */
1500 : enum vec_load_store_type {
1501 : VLS_LOAD,
1502 : VLS_STORE,
1503 : VLS_STORE_INVARIANT
1504 : };
1505 :
1506 : class dr_vec_info {
1507 : public:
1508 : /* The data reference itself. */
1509 : data_reference *dr;
1510 : /* The statement that contains the data reference. */
1511 : stmt_vec_info stmt;
1512 : /* The analysis group this DR belongs to when doing BB vectorization.
1513 : DRs of the same group belong to the same conditional execution context. */
1514 : unsigned group;
1515 : /* The misalignment in bytes of the reference, or -1 if not known. */
1516 : int misalignment;
1517 : /* The byte alignment that we'd ideally like the reference to have,
1518 : and the value that misalignment is measured against. */
1519 : poly_uint64 target_alignment;
1520 : /* If true the alignment of base_decl needs to be increased. */
1521 : bool base_misaligned;
1522 :
1523 : /* Set by early break vectorization when this DR needs peeling for alignment
1524 : for correctness. */
1525 : bool safe_speculative_read_required;
1526 :
1527 : /* Set by early break vectorization when this DR's scalar accesses are known
1528 : to be inbounds of a known bounds loop. */
1529 : bool scalar_access_known_in_bounds;
1530 :
1531 : tree base_decl;
1532 :
1533 : /* Stores current vectorized loop's offset. To be added to the DR's
1534 : offset to calculate current offset of data reference. */
1535 : tree offset;
1536 : };
1537 :
1538 : typedef struct data_reference *dr_p;
1539 :
1540 : class _stmt_vec_info {
1541 : public:
1542 :
1543 : /* Indicates whether this stmts is part of a computation whose result is
1544 : used outside the loop. */
1545 : bool live;
1546 :
1547 : /* Stmt is part of some pattern (computation idiom) */
1548 : bool in_pattern_p;
1549 :
1550 : /* True if the statement was created during pattern recognition as
1551 : part of the replacement for RELATED_STMT. This implies that the
1552 : statement isn't part of any basic block, although for convenience
1553 : its gimple_bb is the same as for RELATED_STMT. */
1554 : bool pattern_stmt_p;
1555 :
1556 : /* Is this statement vectorizable or should it be skipped in (partial)
1557 : vectorization. */
1558 : bool vectorizable;
1559 :
1560 : /* The stmt to which this info struct refers to. */
1561 : gimple *stmt;
1562 :
1563 : /* The vector type to be used for the LHS of this statement. */
1564 : tree vectype;
1565 :
1566 : /* The following is relevant only for stmts that contain a non-scalar
1567 : data-ref (array/pointer/struct access). A GIMPLE stmt is expected to have
1568 : at most one such data-ref. */
1569 :
1570 : dr_vec_info dr_aux;
1571 :
1572 : /* Information about the data-ref relative to this loop
1573 : nest (the loop that is being considered for vectorization). */
1574 : innermost_loop_behavior dr_wrt_vec_loop;
1575 :
1576 : /* For loop PHI nodes, the base and evolution part of it. This makes sure
1577 : this information is still available in vect_update_ivs_after_vectorizer
1578 : where we may not be able to re-analyze the PHI nodes evolution as
1579 : peeling for the prologue loop can make it unanalyzable. The evolution
1580 : part is still correct after peeling, but the base may have changed from
1581 : the version here. */
1582 : tree loop_phi_evolution_base_unchanged;
1583 : tree loop_phi_evolution_part;
1584 : enum vect_induction_op_type loop_phi_evolution_type;
1585 :
1586 : /* Used for various bookkeeping purposes, generally holding a pointer to
1587 : some other stmt S that is in some way "related" to this stmt.
1588 : Current use of this field is:
1589 : If this stmt is part of a pattern (i.e. the field 'in_pattern_p' is
1590 : true): S is the "pattern stmt" that represents (and replaces) the
1591 : sequence of stmts that constitutes the pattern. Similarly, the
1592 : related_stmt of the "pattern stmt" points back to this stmt (which is
1593 : the last stmt in the original sequence of stmts that constitutes the
1594 : pattern). */
1595 : stmt_vec_info related_stmt;
1596 :
1597 : /* Used to keep a sequence of def stmts of a pattern stmt if such exists.
1598 : The sequence is attached to the original statement rather than the
1599 : pattern statement. */
1600 : gimple_seq pattern_def_seq;
1601 :
1602 : /* Classify the def of this stmt. */
1603 : enum vect_def_type def_type;
1604 :
1605 : /* Whether the stmt is SLPed, loop-based vectorized, or both. */
1606 : enum slp_vect_type slp_type;
1607 :
1608 : /* Interleaving chains info. */
1609 : /* First element in the group. */
1610 : stmt_vec_info first_element;
1611 : /* Pointer to the next element in the group. */
1612 : stmt_vec_info next_element;
1613 : /* The size of the group. */
1614 : unsigned int size;
1615 : /* For loads only, the gap from the previous load. For consecutive loads, GAP
1616 : is 1. */
1617 : unsigned int gap;
1618 :
1619 : /* The minimum negative dependence distance this stmt participates in
1620 : or zero if none. */
1621 : unsigned int min_neg_dist;
1622 :
1623 : /* Not all stmts in the loop need to be vectorized. e.g, the increment
1624 : of the loop induction variable and computation of array indexes. relevant
1625 : indicates whether the stmt needs to be vectorized. */
1626 : enum vect_relevant relevant;
1627 :
1628 : /* For loads if this is a gather, for stores if this is a scatter. */
1629 : bool gather_scatter_p;
1630 :
1631 : /* True if this is an access with loop-invariant stride. */
1632 : bool strided_p;
1633 :
1634 : /* For both loads and stores. */
1635 : unsigned simd_lane_access_p : 3;
1636 :
1637 : /* On a reduction PHI the reduction type as detected by
1638 : vect_is_simple_reduction. */
1639 : enum vect_reduction_type reduc_type;
1640 :
1641 : /* On a reduction PHI, the original reduction code as detected by
1642 : vect_is_simple_reduction. */
1643 : code_helper reduc_code;
1644 :
1645 : /* On a stmt participating in a reduction the index of the operand
1646 : on the reduction SSA cycle. */
1647 : int reduc_idx;
1648 :
1649 : /* On a reduction PHI the def returned by vect_is_simple_reduction.
1650 : On the def returned by vect_is_simple_reduction the corresponding PHI. */
1651 : stmt_vec_info reduc_def;
1652 :
1653 : /* If nonzero, the lhs of the statement could be truncated to this
1654 : many bits without affecting any users of the result. */
1655 : unsigned int min_output_precision;
1656 :
1657 : /* If nonzero, all non-boolean input operands have the same precision,
1658 : and they could each be truncated to this many bits without changing
1659 : the result. */
1660 : unsigned int min_input_precision;
1661 :
1662 : /* If OPERATION_BITS is nonzero, the statement could be performed on
1663 : an integer with the sign and number of bits given by OPERATION_SIGN
1664 : and OPERATION_BITS without changing the result. */
1665 : unsigned int operation_precision;
1666 : signop operation_sign;
1667 :
1668 : /* If the statement produces a boolean result, this value describes
1669 : how we should choose the associated vector type. The possible
1670 : values are:
1671 :
1672 : - an integer precision N if we should use the vector mask type
1673 : associated with N-bit integers. This is only used if all relevant
1674 : input booleans also want the vector mask type for N-bit integers,
1675 : or if we can convert them into that form by pattern-matching.
1676 :
1677 : - ~0U if we considered choosing a vector mask type but decided
1678 : to treat the boolean as a normal integer type instead.
1679 :
1680 : - 0 otherwise. This means either that the operation isn't one that
1681 : could have a vector mask type (and so should have a normal vector
1682 : type instead) or that we simply haven't made a choice either way. */
1683 : unsigned int mask_precision;
1684 :
1685 : /* True if this is only suitable for SLP vectorization. */
1686 : bool slp_vect_only_p;
1687 : };
1688 :
1689 : /* Information about a gather/scatter call. */
1690 : struct gather_scatter_info {
1691 : /* The internal function to use for the gather/scatter operation,
1692 : or IFN_LAST if a built-in function should be used instead. */
1693 : internal_fn ifn;
1694 :
1695 : /* The FUNCTION_DECL for the built-in gather/scatter function,
1696 : or null if an internal function should be used instead. */
1697 : tree decl;
1698 :
1699 : /* The loop-invariant base value. */
1700 : tree base;
1701 :
1702 : /* The TBBA alias pointer the value of which determines the alignment
1703 : of the scalar accesses. */
1704 : tree alias_ptr;
1705 :
1706 : /* The original scalar offset, which is a non-loop-invariant SSA_NAME. */
1707 : tree offset;
1708 :
1709 : /* Each offset element should be multiplied by this amount before
1710 : being added to the base. */
1711 : int scale;
1712 :
1713 : /* The type of the vectorized offset. */
1714 : tree offset_vectype;
1715 :
1716 : /* The type of the scalar elements after loading or before storing. */
1717 : tree element_type;
1718 :
1719 : /* The type of the scalar elements being loaded or stored. */
1720 : tree memory_type;
1721 : };
1722 :
1723 : /* Access Functions. */
1724 : #define STMT_VINFO_STMT(S) (S)->stmt
1725 : #define STMT_VINFO_RELEVANT(S) (S)->relevant
1726 : #define STMT_VINFO_LIVE_P(S) (S)->live
1727 : #define STMT_VINFO_VECTYPE(S) (S)->vectype
1728 : #define STMT_VINFO_VECTORIZABLE(S) (S)->vectorizable
1729 : #define STMT_VINFO_DATA_REF(S) ((S)->dr_aux.dr + 0)
1730 : #define STMT_VINFO_GATHER_SCATTER_P(S) (S)->gather_scatter_p
1731 : #define STMT_VINFO_STRIDED_P(S) (S)->strided_p
1732 : #define STMT_VINFO_SIMD_LANE_ACCESS_P(S) (S)->simd_lane_access_p
1733 : #define STMT_VINFO_REDUC_IDX(S) (S)->reduc_idx
1734 :
1735 : #define STMT_VINFO_DR_WRT_VEC_LOOP(S) (S)->dr_wrt_vec_loop
1736 : #define STMT_VINFO_DR_BASE_ADDRESS(S) (S)->dr_wrt_vec_loop.base_address
1737 : #define STMT_VINFO_DR_INIT(S) (S)->dr_wrt_vec_loop.init
1738 : #define STMT_VINFO_DR_OFFSET(S) (S)->dr_wrt_vec_loop.offset
1739 : #define STMT_VINFO_DR_STEP(S) (S)->dr_wrt_vec_loop.step
1740 : #define STMT_VINFO_DR_BASE_ALIGNMENT(S) (S)->dr_wrt_vec_loop.base_alignment
1741 : #define STMT_VINFO_DR_BASE_MISALIGNMENT(S) \
1742 : (S)->dr_wrt_vec_loop.base_misalignment
1743 : #define STMT_VINFO_DR_OFFSET_ALIGNMENT(S) \
1744 : (S)->dr_wrt_vec_loop.offset_alignment
1745 : #define STMT_VINFO_DR_STEP_ALIGNMENT(S) \
1746 : (S)->dr_wrt_vec_loop.step_alignment
1747 :
1748 : #define STMT_VINFO_DR_INFO(S) \
1749 : (gcc_checking_assert ((S)->dr_aux.stmt == (S)), &(S)->dr_aux)
1750 :
1751 : #define STMT_VINFO_IN_PATTERN_P(S) (S)->in_pattern_p
1752 : #define STMT_VINFO_RELATED_STMT(S) (S)->related_stmt
1753 : #define STMT_VINFO_PATTERN_DEF_SEQ(S) (S)->pattern_def_seq
1754 : #define STMT_VINFO_DEF_TYPE(S) (S)->def_type
1755 : #define STMT_VINFO_GROUPED_ACCESS(S) \
1756 : ((S)->dr_aux.dr && DR_GROUP_FIRST_ELEMENT(S))
1757 : #define STMT_VINFO_LOOP_PHI_EVOLUTION_BASE_UNCHANGED(S) (S)->loop_phi_evolution_base_unchanged
1758 : #define STMT_VINFO_LOOP_PHI_EVOLUTION_PART(S) (S)->loop_phi_evolution_part
1759 : #define STMT_VINFO_LOOP_PHI_EVOLUTION_TYPE(S) (S)->loop_phi_evolution_type
1760 : #define STMT_VINFO_MIN_NEG_DIST(S) (S)->min_neg_dist
1761 : #define STMT_VINFO_REDUC_TYPE(S) (S)->reduc_type
1762 : #define STMT_VINFO_REDUC_CODE(S) (S)->reduc_code
1763 : #define STMT_VINFO_REDUC_DEF(S) (S)->reduc_def
1764 : #define STMT_VINFO_SLP_VECT_ONLY(S) (S)->slp_vect_only_p
1765 : #define STMT_VINFO_REDUC_VECTYPE_IN(S) (S)->reduc_vectype_in
1766 :
1767 : #define DR_GROUP_FIRST_ELEMENT(S) \
1768 : (gcc_checking_assert ((S)->dr_aux.dr), (S)->first_element)
1769 : #define DR_GROUP_NEXT_ELEMENT(S) \
1770 : (gcc_checking_assert ((S)->dr_aux.dr), (S)->next_element)
1771 : #define DR_GROUP_SIZE(S) \
1772 : (gcc_checking_assert ((S)->dr_aux.dr), (S)->size)
1773 : #define DR_GROUP_GAP(S) \
1774 : (gcc_checking_assert ((S)->dr_aux.dr), (S)->gap)
1775 :
1776 : #define STMT_VINFO_RELEVANT_P(S) ((S)->relevant != vect_unused_in_scope)
1777 :
1778 : #define PURE_SLP_STMT(S) ((S)->slp_type == pure_slp)
1779 : #define STMT_SLP_TYPE(S) (S)->slp_type
1780 :
1781 :
1782 : /* Contains the scalar or vector costs for a vec_info. */
1783 : class vector_costs
1784 : {
1785 : public:
1786 : vector_costs (vec_info *, bool);
1787 0 : virtual ~vector_costs () {}
1788 :
1789 : /* Update the costs in response to adding COUNT copies of a statement.
1790 :
1791 : - WHERE specifies whether the cost occurs in the loop prologue,
1792 : the loop body, or the loop epilogue.
1793 : - KIND is the kind of statement, which is always meaningful.
1794 : - STMT_INFO or NODE, if nonnull, describe the statement that will be
1795 : vectorized.
1796 : - VECTYPE, if nonnull, is the vector type that the vectorized
1797 : statement will operate on. Note that this should be used in
1798 : preference to STMT_VINFO_VECTYPE (STMT_INFO) since the latter
1799 : is not correct for SLP.
1800 : - for unaligned_load and unaligned_store statements, MISALIGN is
1801 : the byte misalignment of the load or store relative to the target's
1802 : preferred alignment for VECTYPE, or DR_MISALIGNMENT_UNKNOWN
1803 : if the misalignment is not known.
1804 :
1805 : Return the calculated cost as well as recording it. The return
1806 : value is used for dumping purposes. */
1807 : virtual unsigned int add_stmt_cost (int count, vect_cost_for_stmt kind,
1808 : stmt_vec_info stmt_info,
1809 : slp_tree node,
1810 : tree vectype, int misalign,
1811 : vect_cost_model_location where);
1812 :
1813 : /* Update the costs in response to adding costs in V which are all from
1814 : vectorizing NODE to the respective part. */
1815 : virtual unsigned int add_slp_cost (slp_tree node,
1816 : const array_slice<stmt_info_for_cost> &v);
1817 :
1818 : /* Finish calculating the cost of the code. The results can be
1819 : read back using the functions below.
1820 :
1821 : If the costs describe vector code, SCALAR_COSTS gives the costs
1822 : of the corresponding scalar code, otherwise it is null. */
1823 : virtual void finish_cost (const vector_costs *scalar_costs);
1824 :
1825 : /* The costs in THIS and OTHER both describe ways of vectorizing
1826 : a main loop. Return true if the costs described by THIS are
1827 : cheaper than the costs described by OTHER. Return false if any
1828 : of the following are true:
1829 :
1830 : - THIS and OTHER are of equal cost
1831 : - OTHER is better than THIS
1832 : - we can't be sure about the relative costs of THIS and OTHER. */
1833 : virtual bool better_main_loop_than_p (const vector_costs *other) const;
1834 :
1835 : /* Likewise, but the costs in THIS and OTHER both describe ways of
1836 : vectorizing an epilogue loop of MAIN_LOOP. */
1837 : virtual bool better_epilogue_loop_than_p (const vector_costs *other,
1838 : loop_vec_info main_loop) const;
1839 :
1840 : unsigned int prologue_cost () const;
1841 : unsigned int body_cost () const;
1842 : unsigned int epilogue_cost () const;
1843 : unsigned int outside_cost () const;
1844 : unsigned int total_cost () const;
1845 :
1846 : unsigned int suggested_unroll_factor () const;
1847 : machine_mode suggested_epilogue_mode (int &masked) const;
1848 :
1849 32751 : vec_info *vinfo () const { return m_vinfo; }
1850 7924079 : bool costing_for_scalar () const { return m_costing_for_scalar; }
1851 :
1852 : protected:
1853 : unsigned int record_stmt_cost (stmt_vec_info, vect_cost_model_location,
1854 : unsigned int);
1855 : unsigned int adjust_cost_for_freq (stmt_vec_info, vect_cost_model_location,
1856 : unsigned int);
1857 : int compare_inside_loop_cost (const vector_costs *) const;
1858 : int compare_outside_loop_cost (const vector_costs *) const;
1859 :
1860 : /* The region of code that we're considering vectorizing. */
1861 : vec_info *m_vinfo;
1862 :
1863 : /* True if we're costing the scalar code, false if we're costing
1864 : the vector code. */
1865 : bool m_costing_for_scalar;
1866 :
1867 : /* The costs of the three regions, indexed by vect_cost_model_location. */
1868 : unsigned int m_costs[3];
1869 :
1870 : /* The suggested unrolling factor determined at finish_cost. */
1871 : unsigned int m_suggested_unroll_factor;
1872 :
1873 : /* The suggested mode to be used for a vectorized epilogue or VOIDmode,
1874 : determined at finish_cost. m_masked_epilogue specifies whether the
1875 : epilogue should use masked vectorization, regardless of the
1876 : --param vect-partial-vector-usage default. If -1 then the
1877 : --param setting takes precedence. If the user explicitly specified
1878 : --param vect-partial-vector-usage then that takes precedence. */
1879 : machine_mode m_suggested_epilogue_mode;
1880 : int m_masked_epilogue;
1881 :
1882 : /* True if finish_cost has been called. */
1883 : bool m_finished;
1884 : };
1885 :
1886 : /* Create costs for VINFO. COSTING_FOR_SCALAR is true if the costs
1887 : are for scalar code, false if they are for vector code. */
1888 :
1889 : inline
1890 2255000 : vector_costs::vector_costs (vec_info *vinfo, bool costing_for_scalar)
1891 2255000 : : m_vinfo (vinfo),
1892 2255000 : m_costing_for_scalar (costing_for_scalar),
1893 2255000 : m_costs (),
1894 2255000 : m_suggested_unroll_factor(1),
1895 2255000 : m_suggested_epilogue_mode(VOIDmode),
1896 2255000 : m_masked_epilogue (-1),
1897 2255000 : m_finished (false)
1898 : {
1899 : }
1900 :
1901 : /* Return the cost of the prologue code (in abstract units). */
1902 :
1903 : inline unsigned int
1904 1379882 : vector_costs::prologue_cost () const
1905 : {
1906 1379882 : gcc_checking_assert (m_finished);
1907 1379882 : return m_costs[vect_prologue];
1908 : }
1909 :
1910 : /* Return the cost of the body code (in abstract units). */
1911 :
1912 : inline unsigned int
1913 2143984 : vector_costs::body_cost () const
1914 : {
1915 2143984 : gcc_checking_assert (m_finished);
1916 2143984 : return m_costs[vect_body];
1917 : }
1918 :
1919 : /* Return the cost of the epilogue code (in abstract units). */
1920 :
1921 : inline unsigned int
1922 1379882 : vector_costs::epilogue_cost () const
1923 : {
1924 1379882 : gcc_checking_assert (m_finished);
1925 1379882 : return m_costs[vect_epilogue];
1926 : }
1927 :
1928 : /* Return the cost of the prologue and epilogue code (in abstract units). */
1929 :
1930 : inline unsigned int
1931 510306 : vector_costs::outside_cost () const
1932 : {
1933 510306 : return prologue_cost () + epilogue_cost ();
1934 : }
1935 :
1936 : /* Return the cost of the prologue, body and epilogue code
1937 : (in abstract units). */
1938 :
1939 : inline unsigned int
1940 126040 : vector_costs::total_cost () const
1941 : {
1942 126040 : return body_cost () + outside_cost ();
1943 : }
1944 :
1945 : /* Return the suggested unroll factor. */
1946 :
1947 : inline unsigned int
1948 125650 : vector_costs::suggested_unroll_factor () const
1949 : {
1950 125650 : gcc_checking_assert (m_finished);
1951 125650 : return m_suggested_unroll_factor;
1952 : }
1953 :
1954 : /* Return the suggested epilogue mode. */
1955 :
1956 : inline machine_mode
1957 14520 : vector_costs::suggested_epilogue_mode (int &masked_p) const
1958 : {
1959 14520 : gcc_checking_assert (m_finished);
1960 14520 : masked_p = m_masked_epilogue;
1961 14520 : return m_suggested_epilogue_mode;
1962 : }
1963 :
1964 : #define VECT_MAX_COST 1000
1965 :
1966 : /* The maximum number of intermediate steps required in multi-step type
1967 : conversion. */
1968 : #define MAX_INTERM_CVT_STEPS 3
1969 :
1970 : #define MAX_VECTORIZATION_FACTOR INT_MAX
1971 :
1972 : /* Nonzero if TYPE represents a (scalar) boolean type or type
1973 : in the middle-end compatible with it (unsigned precision 1 integral
1974 : types). Used to determine which types should be vectorized as
1975 : VECTOR_BOOLEAN_TYPE_P. */
1976 :
1977 : #define VECT_SCALAR_BOOLEAN_TYPE_P(TYPE) \
1978 : (TREE_CODE (TYPE) == BOOLEAN_TYPE \
1979 : || ((TREE_CODE (TYPE) == INTEGER_TYPE \
1980 : || TREE_CODE (TYPE) == ENUMERAL_TYPE) \
1981 : && TYPE_PRECISION (TYPE) == 1 \
1982 : && TYPE_UNSIGNED (TYPE)))
1983 :
1984 : inline bool
1985 11921116 : nested_in_vect_loop_p (class loop *loop, stmt_vec_info stmt_info)
1986 : {
1987 11921116 : return (loop->inner
1988 9592004 : && (loop->inner == (gimple_bb (stmt_info->stmt))->loop_father));
1989 : }
1990 :
1991 : /* PHI is either a scalar reduction phi or a scalar induction phi.
1992 : Return the initial value of the variable on entry to the containing
1993 : loop. */
1994 :
1995 : inline tree
1996 34583 : vect_phi_initial_value (gphi *phi)
1997 : {
1998 34583 : basic_block bb = gimple_bb (phi);
1999 34583 : edge pe = loop_preheader_edge (bb->loop_father);
2000 34583 : gcc_assert (pe->dest == bb);
2001 34583 : return PHI_ARG_DEF_FROM_EDGE (phi, pe);
2002 : }
2003 :
2004 : /* Return true if STMT_INFO should produce a vector mask type rather than
2005 : a normal nonmask type. */
2006 :
2007 : inline bool
2008 7520996 : vect_use_mask_type_p (stmt_vec_info stmt_info)
2009 : {
2010 7520996 : return stmt_info->mask_precision && stmt_info->mask_precision != ~0U;
2011 : }
2012 :
2013 : /* Return TRUE if a statement represented by STMT_INFO is a part of a
2014 : pattern. */
2015 :
2016 : inline bool
2017 132598199 : is_pattern_stmt_p (stmt_vec_info stmt_info)
2018 : {
2019 85344181 : return stmt_info->pattern_stmt_p;
2020 : }
2021 :
2022 : /* If STMT_INFO is a pattern statement, return the statement that it
2023 : replaces, otherwise return STMT_INFO itself. */
2024 :
2025 : inline stmt_vec_info
2026 51752760 : vect_orig_stmt (stmt_vec_info stmt_info)
2027 : {
2028 39235375 : if (is_pattern_stmt_p (stmt_info))
2029 3849423 : return STMT_VINFO_RELATED_STMT (stmt_info);
2030 : return stmt_info;
2031 : }
2032 :
2033 : /* Return the later statement between STMT1_INFO and STMT2_INFO. */
2034 :
2035 : inline stmt_vec_info
2036 6098597 : get_later_stmt (stmt_vec_info stmt1_info, stmt_vec_info stmt2_info)
2037 : {
2038 6098597 : gimple *stmt1 = vect_orig_stmt (stmt1_info)->stmt;
2039 6098597 : gimple *stmt2 = vect_orig_stmt (stmt2_info)->stmt;
2040 6098597 : if (gimple_bb (stmt1) == gimple_bb (stmt2))
2041 : {
2042 6080037 : if (gimple_uid (stmt1) > gimple_uid (stmt2))
2043 : return stmt1_info;
2044 : else
2045 1252848 : return stmt2_info;
2046 : }
2047 : /* ??? We should be really calling this function only with stmts
2048 : in the same BB but we can recover if there's a domination
2049 : relationship between them. */
2050 18560 : else if (dominated_by_p (CDI_DOMINATORS,
2051 18560 : gimple_bb (stmt1), gimple_bb (stmt2)))
2052 : return stmt1_info;
2053 7553 : else if (dominated_by_p (CDI_DOMINATORS,
2054 7553 : gimple_bb (stmt2), gimple_bb (stmt1)))
2055 : return stmt2_info;
2056 0 : gcc_unreachable ();
2057 : }
2058 :
2059 : /* If STMT_INFO has been replaced by a pattern statement, return the
2060 : replacement statement, otherwise return STMT_INFO itself. */
2061 :
2062 : inline stmt_vec_info
2063 54062030 : vect_stmt_to_vectorize (stmt_vec_info stmt_info)
2064 : {
2065 54062030 : if (STMT_VINFO_IN_PATTERN_P (stmt_info))
2066 1655517 : return STMT_VINFO_RELATED_STMT (stmt_info);
2067 : return stmt_info;
2068 : }
2069 :
2070 : /* Return true if BB is a loop header. */
2071 :
2072 : inline bool
2073 1566040 : is_loop_header_bb_p (basic_block bb)
2074 : {
2075 1566040 : if (bb == (bb->loop_father)->header)
2076 1555236 : return true;
2077 :
2078 : return false;
2079 : }
2080 :
2081 : /* Return pow2 (X). */
2082 :
2083 : inline int
2084 : vect_pow2 (int x)
2085 : {
2086 : int i, res = 1;
2087 :
2088 : for (i = 0; i < x; i++)
2089 : res *= 2;
2090 :
2091 : return res;
2092 : }
2093 :
2094 : /* Alias targetm.vectorize.builtin_vectorization_cost. */
2095 :
2096 : inline int
2097 9386258 : builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost,
2098 : tree vectype, int misalign)
2099 : {
2100 9318299 : return targetm.vectorize.builtin_vectorization_cost (type_of_cost,
2101 : vectype, misalign);
2102 : }
2103 :
2104 : /* Get cost by calling cost target builtin. */
2105 :
2106 : inline
2107 155 : int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost)
2108 : {
2109 67804 : return builtin_vectorization_cost (type_of_cost, NULL, 0);
2110 : }
2111 :
2112 : /* Alias targetm.vectorize.init_cost. */
2113 :
2114 : inline vector_costs *
2115 2255000 : init_cost (vec_info *vinfo, bool costing_for_scalar)
2116 : {
2117 2255000 : return targetm.vectorize.create_costs (vinfo, costing_for_scalar);
2118 : }
2119 :
2120 : extern void dump_stmt_cost (FILE *, int, enum vect_cost_for_stmt,
2121 : stmt_vec_info, slp_tree, tree, int, unsigned,
2122 : enum vect_cost_model_location);
2123 :
2124 : /* Dump and add costs. */
2125 :
2126 : inline unsigned
2127 7924079 : add_stmt_cost (vector_costs *costs, int count,
2128 : enum vect_cost_for_stmt kind,
2129 : stmt_vec_info stmt_info, slp_tree node,
2130 : tree vectype, int misalign,
2131 : enum vect_cost_model_location where)
2132 : {
2133 : /* Even though a vector type might be set on stmt do not pass that on when
2134 : costing the scalar IL. A SLP node shouldn't have been recorded. */
2135 7924079 : if (costs->costing_for_scalar ())
2136 : {
2137 4065427 : vectype = NULL_TREE;
2138 4065427 : gcc_checking_assert (node == NULL);
2139 : }
2140 7924079 : unsigned cost = costs->add_stmt_cost (count, kind, stmt_info, node, vectype,
2141 : misalign, where);
2142 7924079 : if (dump_file && (dump_flags & TDF_DETAILS))
2143 222428 : dump_stmt_cost (dump_file, count, kind, stmt_info, node, vectype, misalign,
2144 : cost, where);
2145 7924079 : return cost;
2146 : }
2147 :
2148 : inline unsigned
2149 83142 : add_stmt_cost (vector_costs *costs, int count, enum vect_cost_for_stmt kind,
2150 : enum vect_cost_model_location where)
2151 : {
2152 83142 : gcc_assert (kind == cond_branch_taken || kind == cond_branch_not_taken
2153 : || kind == scalar_stmt);
2154 83142 : return add_stmt_cost (costs, count, kind, NULL, NULL, NULL_TREE, 0, where);
2155 : }
2156 :
2157 : inline unsigned
2158 2288638 : add_stmt_cost (vector_costs *costs, stmt_info_for_cost *i)
2159 : {
2160 2288638 : return add_stmt_cost (costs, i->count, i->kind, i->stmt_info, i->node,
2161 2288638 : i->vectype, i->misalign, i->where);
2162 : }
2163 :
2164 : inline void
2165 374587 : add_stmt_costs (vector_costs *costs, stmt_vector_for_cost *cost_vec)
2166 : {
2167 374587 : stmt_info_for_cost *cost;
2168 374587 : unsigned i;
2169 2151376 : FOR_EACH_VEC_ELT (*cost_vec, i, cost)
2170 1776789 : add_stmt_cost (costs, cost->count, cost->kind, cost->stmt_info,
2171 : cost->node, cost->vectype, cost->misalign, cost->where);
2172 374587 : }
2173 :
2174 : /*-----------------------------------------------------------------*/
2175 : /* Info on data references alignment. */
2176 : /*-----------------------------------------------------------------*/
2177 : #define DR_MISALIGNMENT_UNKNOWN (-1)
2178 : #define DR_MISALIGNMENT_UNINITIALIZED (-2)
2179 :
2180 : inline void
2181 2665643 : set_dr_misalignment (dr_vec_info *dr_info, int val)
2182 : {
2183 2665643 : dr_info->misalignment = val;
2184 : }
2185 :
2186 : extern int dr_misalignment (dr_vec_info *dr_info, tree vectype,
2187 : poly_int64 offset = 0);
2188 :
2189 : #define SET_DR_MISALIGNMENT(DR, VAL) set_dr_misalignment (DR, VAL)
2190 :
2191 : /* Only defined once DR_MISALIGNMENT is defined. */
2192 : inline const poly_uint64
2193 8226291 : dr_target_alignment (dr_vec_info *dr_info)
2194 : {
2195 8226291 : if (STMT_VINFO_GROUPED_ACCESS (dr_info->stmt))
2196 6063871 : dr_info = STMT_VINFO_DR_INFO (DR_GROUP_FIRST_ELEMENT (dr_info->stmt));
2197 8226291 : return dr_info->target_alignment;
2198 : }
2199 : #define DR_TARGET_ALIGNMENT(DR) dr_target_alignment (DR)
2200 : #define DR_SCALAR_KNOWN_BOUNDS(DR) (DR)->scalar_access_known_in_bounds
2201 :
2202 : /* Return if the stmt_vec_info requires peeling for alignment. */
2203 : inline bool
2204 4644856 : dr_safe_speculative_read_required (stmt_vec_info stmt_info)
2205 : {
2206 4644856 : dr_vec_info *dr_info;
2207 4644856 : if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
2208 1713793 : dr_info = STMT_VINFO_DR_INFO (DR_GROUP_FIRST_ELEMENT (stmt_info));
2209 : else
2210 2931063 : dr_info = STMT_VINFO_DR_INFO (stmt_info);
2211 :
2212 4644856 : return dr_info->safe_speculative_read_required;
2213 : }
2214 :
2215 : /* Set the safe_speculative_read_required for the stmt_vec_info, if group
2216 : access then set on the fist element otherwise set on DR directly. */
2217 : inline void
2218 236891 : dr_set_safe_speculative_read_required (stmt_vec_info stmt_info,
2219 : bool requires_alignment)
2220 : {
2221 236891 : dr_vec_info *dr_info;
2222 236891 : if (STMT_VINFO_GROUPED_ACCESS (stmt_info))
2223 67856 : dr_info = STMT_VINFO_DR_INFO (DR_GROUP_FIRST_ELEMENT (stmt_info));
2224 : else
2225 169035 : dr_info = STMT_VINFO_DR_INFO (stmt_info);
2226 :
2227 236891 : dr_info->safe_speculative_read_required = requires_alignment;
2228 236891 : }
2229 :
2230 : inline void
2231 1641627 : set_dr_target_alignment (dr_vec_info *dr_info, poly_uint64 val)
2232 : {
2233 1641627 : dr_info->target_alignment = val;
2234 : }
2235 : #define SET_DR_TARGET_ALIGNMENT(DR, VAL) set_dr_target_alignment (DR, VAL)
2236 :
2237 : /* Return true if data access DR_INFO is aligned to the targets
2238 : preferred alignment for VECTYPE (which may be less than a full vector). */
2239 :
2240 : inline bool
2241 395334 : aligned_access_p (dr_vec_info *dr_info, tree vectype)
2242 : {
2243 395334 : return (dr_misalignment (dr_info, vectype) == 0);
2244 : }
2245 :
2246 : /* Return TRUE if the (mis-)alignment of the data access is known with
2247 : respect to the targets preferred alignment for VECTYPE, and FALSE
2248 : otherwise. */
2249 :
2250 : inline bool
2251 2390033 : known_alignment_for_access_p (dr_vec_info *dr_info, tree vectype)
2252 : {
2253 2140429 : return (dr_misalignment (dr_info, vectype) != DR_MISALIGNMENT_UNKNOWN);
2254 : }
2255 :
2256 : /* Return the minimum alignment in bytes that the vectorized version
2257 : of DR_INFO is guaranteed to have. */
2258 :
2259 : inline unsigned int
2260 283064 : vect_known_alignment_in_bytes (dr_vec_info *dr_info, tree vectype,
2261 : poly_int64 offset = 0)
2262 : {
2263 283064 : int misalignment = dr_misalignment (dr_info, vectype, offset);
2264 283064 : if (misalignment == DR_MISALIGNMENT_UNKNOWN)
2265 140215 : return TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr_info->dr)));
2266 142849 : else if (misalignment == 0)
2267 99173 : return known_alignment (DR_TARGET_ALIGNMENT (dr_info));
2268 43676 : return misalignment & -misalignment;
2269 : }
2270 :
2271 : /* Return the behavior of DR_INFO with respect to the vectorization context
2272 : (which for outer loop vectorization might not be the behavior recorded
2273 : in DR_INFO itself). */
2274 :
2275 : inline innermost_loop_behavior *
2276 5911956 : vect_dr_behavior (vec_info *vinfo, dr_vec_info *dr_info)
2277 : {
2278 5911956 : stmt_vec_info stmt_info = dr_info->stmt;
2279 5911956 : loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo);
2280 2337524 : if (loop_vinfo == NULL
2281 2337524 : || !nested_in_vect_loop_p (LOOP_VINFO_LOOP (loop_vinfo), stmt_info))
2282 5907614 : return &DR_INNERMOST (dr_info->dr);
2283 : else
2284 4342 : return &STMT_VINFO_DR_WRT_VEC_LOOP (stmt_info);
2285 : }
2286 :
2287 : /* Return the offset calculated by adding the offset of this DR_INFO to the
2288 : corresponding data_reference's offset. If CHECK_OUTER then use
2289 : vect_dr_behavior to select the appropriate data_reference to use. */
2290 :
2291 : inline tree
2292 735778 : get_dr_vinfo_offset (vec_info *vinfo,
2293 : dr_vec_info *dr_info, bool check_outer = false)
2294 : {
2295 735778 : innermost_loop_behavior *base;
2296 735778 : if (check_outer)
2297 695873 : base = vect_dr_behavior (vinfo, dr_info);
2298 : else
2299 39905 : base = &dr_info->dr->innermost;
2300 :
2301 735778 : tree offset = base->offset;
2302 :
2303 735778 : if (!dr_info->offset)
2304 : return offset;
2305 :
2306 1160 : offset = fold_convert (sizetype, offset);
2307 1160 : return fold_build2 (PLUS_EXPR, TREE_TYPE (dr_info->offset), offset,
2308 : dr_info->offset);
2309 : }
2310 :
2311 :
2312 : /* Return the vect cost model for LOOP. */
2313 : inline enum vect_cost_model
2314 2494897 : loop_cost_model (loop_p loop)
2315 : {
2316 2494897 : if (loop != NULL
2317 1804335 : && loop->force_vectorize
2318 79691 : && flag_simd_cost_model != VECT_COST_MODEL_DEFAULT)
2319 : return flag_simd_cost_model;
2320 2415206 : return flag_vect_cost_model;
2321 : }
2322 :
2323 : /* Return true if the vect cost model is unlimited. */
2324 : inline bool
2325 1735332 : unlimited_cost_model (loop_p loop)
2326 : {
2327 1735332 : return loop_cost_model (loop) == VECT_COST_MODEL_UNLIMITED;
2328 : }
2329 :
2330 : /* Return true if the loop described by LOOP_VINFO is fully-masked and
2331 : if the first iteration should use a partial mask in order to achieve
2332 : alignment. */
2333 :
2334 : inline bool
2335 268092 : vect_use_loop_mask_for_alignment_p (loop_vec_info loop_vinfo)
2336 : {
2337 : /* With early break vectorization we don't know whether the accesses will stay
2338 : inside the loop or not. TODO: The early break adjustment code can be
2339 : implemented the same way as vectorizable_linear_induction. However we
2340 : can't test this today so reject it. */
2341 98 : return (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)
2342 98 : && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo)
2343 268100 : && !(LOOP_VINFO_NON_LINEAR_IV (loop_vinfo)
2344 0 : && LOOP_VINFO_EARLY_BREAKS (loop_vinfo)));
2345 : }
2346 :
2347 : /* Return the number of vectors of type VECTYPE that are needed to get
2348 : NUNITS elements. NUNITS should be based on the vectorization factor,
2349 : so it is always a known multiple of the number of elements in VECTYPE. */
2350 :
2351 : inline unsigned int
2352 5641903 : vect_get_num_vectors (poly_uint64 nunits, tree vectype)
2353 : {
2354 5641903 : return exact_div (nunits, TYPE_VECTOR_SUBPARTS (vectype)).to_constant ();
2355 : }
2356 :
2357 : /* Return the number of vectors in the context of vectorization region VINFO,
2358 : needed for a group of statements and a vector type as specified by NODE. */
2359 :
2360 : inline unsigned int
2361 5641068 : vect_get_num_copies (vec_info *vinfo, slp_tree node)
2362 : {
2363 5641068 : poly_uint64 vf;
2364 :
2365 5641068 : if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
2366 3035733 : vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo);
2367 : else
2368 : vf = 1;
2369 :
2370 5641068 : vf *= SLP_TREE_LANES (node);
2371 5641068 : tree vectype = SLP_TREE_VECTYPE (node);
2372 :
2373 5641068 : return vect_get_num_vectors (vf, vectype);
2374 : }
2375 :
2376 : /* Return the vectorization factor that should be used for costing
2377 : purposes while vectorizing the loop described by LOOP_VINFO.
2378 : Pick a reasonable estimate if the vectorization factor isn't
2379 : known at compile time. */
2380 :
2381 : inline unsigned int
2382 1446292 : vect_vf_for_cost (loop_vec_info loop_vinfo)
2383 : {
2384 1446292 : return estimated_poly_value (LOOP_VINFO_VECT_FACTOR (loop_vinfo));
2385 : }
2386 :
2387 : /* Estimate the number of elements in VEC_TYPE for costing purposes.
2388 : Pick a reasonable estimate if the exact number isn't known at
2389 : compile time. */
2390 :
2391 : inline unsigned int
2392 57738 : vect_nunits_for_cost (tree vec_type)
2393 : {
2394 57738 : return estimated_poly_value (TYPE_VECTOR_SUBPARTS (vec_type));
2395 : }
2396 :
2397 : /* Return the maximum possible vectorization factor for LOOP_VINFO. */
2398 :
2399 : inline unsigned HOST_WIDE_INT
2400 106011 : vect_max_vf (loop_vec_info loop_vinfo)
2401 : {
2402 106011 : unsigned HOST_WIDE_INT vf;
2403 106011 : if (LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (&vf))
2404 106011 : return vf;
2405 : return MAX_VECTORIZATION_FACTOR;
2406 : }
2407 :
2408 : /* Return the size of the value accessed by unvectorized data reference
2409 : DR_INFO. This is only valid once STMT_VINFO_VECTYPE has been calculated
2410 : for the associated gimple statement, since that guarantees that DR_INFO
2411 : accesses either a scalar or a scalar equivalent. ("Scalar equivalent"
2412 : here includes things like V1SI, which can be vectorized in the same way
2413 : as a plain SI.) */
2414 :
2415 : inline unsigned int
2416 2032695 : vect_get_scalar_dr_size (dr_vec_info *dr_info)
2417 : {
2418 2032695 : return tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_info->dr))));
2419 : }
2420 :
2421 : /* Return true if LOOP_VINFO requires a runtime check for whether the
2422 : vector loop is profitable. */
2423 :
2424 : inline bool
2425 72018 : vect_apply_runtime_profitability_check_p (loop_vec_info loop_vinfo)
2426 : {
2427 72018 : unsigned int th = LOOP_VINFO_COST_MODEL_THRESHOLD (loop_vinfo);
2428 37895 : return (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
2429 72018 : && th >= vect_vf_for_cost (loop_vinfo));
2430 : }
2431 :
2432 : /* Return true if CODE is a lane-reducing opcode. */
2433 :
2434 : inline bool
2435 396249 : lane_reducing_op_p (code_helper code)
2436 : {
2437 396249 : return code == DOT_PROD_EXPR || code == WIDEN_SUM_EXPR || code == SAD_EXPR;
2438 : }
2439 :
2440 : /* Return true if STMT is a lane-reducing statement. */
2441 :
2442 : inline bool
2443 484392 : lane_reducing_stmt_p (gimple *stmt)
2444 : {
2445 484392 : if (auto *assign = dyn_cast <gassign *> (stmt))
2446 338321 : return lane_reducing_op_p (gimple_assign_rhs_code (assign));
2447 : return false;
2448 : }
2449 :
2450 : /* Source location + hotness information. */
2451 : extern dump_user_location_t vect_location;
2452 :
2453 : /* A macro for calling:
2454 : dump_begin_scope (MSG, vect_location);
2455 : via an RAII object, thus printing "=== MSG ===\n" to the dumpfile etc,
2456 : and then calling
2457 : dump_end_scope ();
2458 : once the object goes out of scope, thus capturing the nesting of
2459 : the scopes.
2460 :
2461 : These scopes affect dump messages within them: dump messages at the
2462 : top level implicitly default to MSG_PRIORITY_USER_FACING, whereas those
2463 : in a nested scope implicitly default to MSG_PRIORITY_INTERNALS. */
2464 :
2465 : #define DUMP_VECT_SCOPE(MSG) \
2466 : AUTO_DUMP_SCOPE (MSG, vect_location)
2467 :
2468 : /* A sentinel class for ensuring that the "vect_location" global gets
2469 : reset at the end of a scope.
2470 :
2471 : The "vect_location" global is used during dumping and contains a
2472 : location_t, which could contain references to a tree block via the
2473 : ad-hoc data. This data is used for tracking inlining information,
2474 : but it's not a GC root; it's simply assumed that such locations never
2475 : get accessed if the blocks are optimized away.
2476 :
2477 : Hence we need to ensure that such locations are purged at the end
2478 : of any operations using them (e.g. via this class). */
2479 :
2480 : class auto_purge_vect_location
2481 : {
2482 : public:
2483 : ~auto_purge_vect_location ();
2484 : };
2485 :
2486 : /*-----------------------------------------------------------------*/
2487 : /* Function prototypes. */
2488 : /*-----------------------------------------------------------------*/
2489 :
2490 : /* Simple loop peeling and versioning utilities for vectorizer's purposes -
2491 : in tree-vect-loop-manip.cc. */
2492 : extern void vect_set_loop_condition (class loop *, edge, loop_vec_info,
2493 : tree, tree, tree, bool);
2494 : extern bool slpeel_can_duplicate_loop_p (const class loop *, const_edge,
2495 : const_edge);
2496 : class loop *slpeel_tree_duplicate_loop_to_edge_cfg (class loop *, edge,
2497 : class loop *, edge,
2498 : edge, edge *, bool = true,
2499 : vec<basic_block> * = NULL,
2500 : bool = false, bool = false,
2501 : bool = true);
2502 : class loop *vect_loop_versioning (loop_vec_info, gimple *);
2503 : extern class loop *vect_do_peeling (loop_vec_info, tree, tree,
2504 : tree *, tree *, tree *, int, bool, bool,
2505 : tree *);
2506 : extern tree vect_get_main_loop_result (loop_vec_info, tree, tree);
2507 : extern void vect_prepare_for_masked_peels (loop_vec_info);
2508 : extern dump_user_location_t find_loop_location (class loop *);
2509 : extern bool vect_can_advance_ivs_p (loop_vec_info);
2510 : extern void vect_update_inits_of_drs (loop_vec_info, tree, tree_code);
2511 : extern edge vec_init_loop_exit_info (class loop *);
2512 : extern void vect_iv_increment_position (edge, gimple_stmt_iterator *, bool *);
2513 :
2514 : /* In tree-vect-stmts.cc. */
2515 : extern tree get_related_vectype_for_scalar_type (machine_mode, tree,
2516 : poly_uint64 = 0);
2517 : extern tree get_vectype_for_scalar_type (vec_info *, tree, unsigned int = 0);
2518 : extern tree get_vectype_for_scalar_type (vec_info *, tree, slp_tree);
2519 : extern tree get_mask_type_for_scalar_type (vec_info *, tree, unsigned int = 0);
2520 : extern tree get_mask_type_for_scalar_type (vec_info *, tree, slp_tree);
2521 : extern tree get_same_sized_vectype (tree, tree);
2522 : extern bool vect_chooses_same_modes_p (vec_info *, machine_mode);
2523 : extern bool vect_chooses_same_modes_p (machine_mode, machine_mode);
2524 : extern bool vect_get_loop_mask_type (loop_vec_info);
2525 : extern bool vect_is_simple_use (tree, vec_info *, enum vect_def_type *,
2526 : stmt_vec_info * = NULL, gimple ** = NULL);
2527 : extern bool vect_is_simple_use (vec_info *, slp_tree,
2528 : unsigned, tree *, slp_tree *,
2529 : enum vect_def_type *,
2530 : tree *, stmt_vec_info * = NULL);
2531 : extern bool vect_is_simple_use (vec_info *, slp_tree,
2532 : unsigned, slp_tree *,
2533 : enum vect_def_type *, tree *);
2534 : extern bool vect_maybe_update_slp_op_vectype (slp_tree, tree);
2535 : extern tree perm_mask_for_reverse (tree);
2536 : extern bool supportable_widening_operation (code_helper, tree, tree, bool,
2537 : code_helper*, code_helper*,
2538 : int*, vec<tree> *);
2539 : extern bool supportable_narrowing_operation (code_helper, tree, tree,
2540 : code_helper *, int *,
2541 : vec<tree> *);
2542 : extern bool supportable_indirect_convert_operation (code_helper,
2543 : tree, tree,
2544 : vec<std::pair<tree, tree_code> > &,
2545 : slp_tree = NULL);
2546 : extern int compare_step_with_zero (vec_info *, stmt_vec_info);
2547 :
2548 : extern unsigned record_stmt_cost (stmt_vector_for_cost *, int,
2549 : enum vect_cost_for_stmt, stmt_vec_info,
2550 : tree, int, enum vect_cost_model_location);
2551 : extern unsigned record_stmt_cost (stmt_vector_for_cost *, int,
2552 : enum vect_cost_for_stmt, slp_tree,
2553 : tree, int, enum vect_cost_model_location);
2554 : extern unsigned record_stmt_cost (stmt_vector_for_cost *, int,
2555 : enum vect_cost_for_stmt,
2556 : enum vect_cost_model_location);
2557 : extern unsigned record_stmt_cost (stmt_vector_for_cost *, int,
2558 : enum vect_cost_for_stmt, stmt_vec_info,
2559 : slp_tree, tree, int,
2560 : enum vect_cost_model_location);
2561 :
2562 : /* Overload of record_stmt_cost with VECTYPE derived from STMT_INFO. */
2563 :
2564 : inline unsigned
2565 1845037 : record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count,
2566 : enum vect_cost_for_stmt kind, stmt_vec_info stmt_info,
2567 : int misalign, enum vect_cost_model_location where)
2568 : {
2569 1844472 : return record_stmt_cost (body_cost_vec, count, kind, stmt_info,
2570 : STMT_VINFO_VECTYPE (stmt_info), misalign, where);
2571 : }
2572 :
2573 : /* Overload of record_stmt_cost with VECTYPE derived from SLP node. */
2574 :
2575 : inline unsigned
2576 1579460 : record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count,
2577 : enum vect_cost_for_stmt kind, slp_tree node,
2578 : int misalign, enum vect_cost_model_location where)
2579 : {
2580 1369117 : return record_stmt_cost (body_cost_vec, count, kind, node,
2581 : SLP_TREE_VECTYPE (node), misalign, where);
2582 : }
2583 :
2584 : extern void vect_finish_replace_stmt (vec_info *, stmt_vec_info, gimple *);
2585 : extern void vect_finish_stmt_generation (vec_info *, stmt_vec_info, gimple *,
2586 : gimple_stmt_iterator *);
2587 : extern opt_result vect_mark_stmts_to_be_vectorized (loop_vec_info, bool *);
2588 : extern tree vect_get_store_rhs (stmt_vec_info);
2589 : void vect_get_vec_defs (vec_info *, slp_tree,
2590 : bool, vec<tree> *,
2591 : bool = false, vec<tree> * = NULL,
2592 : bool = false, vec<tree> * = NULL,
2593 : bool = false, vec<tree> * = NULL);
2594 : extern tree vect_init_vector (vec_info *, stmt_vec_info, tree, tree,
2595 : gimple_stmt_iterator *);
2596 : extern tree vect_get_slp_vect_def (slp_tree, unsigned);
2597 : extern void vect_transform_stmt (vec_info *, stmt_vec_info,
2598 : gimple_stmt_iterator *,
2599 : slp_tree, slp_instance);
2600 : extern void vect_remove_stores (vec_info *, stmt_vec_info);
2601 : extern bool vect_nop_conversion_p (stmt_vec_info);
2602 : extern opt_result vect_analyze_stmt (vec_info *, slp_tree,
2603 : slp_instance, stmt_vector_for_cost *);
2604 : extern void vect_get_load_cost (vec_info *, stmt_vec_info, slp_tree, int,
2605 : dr_alignment_support, int, bool,
2606 : unsigned int *, unsigned int *,
2607 : stmt_vector_for_cost *,
2608 : stmt_vector_for_cost *, bool);
2609 : extern void vect_get_store_cost (vec_info *, stmt_vec_info, slp_tree, int,
2610 : dr_alignment_support, int,
2611 : unsigned int *, stmt_vector_for_cost *);
2612 : extern bool vect_supportable_shift (vec_info *, enum tree_code, tree);
2613 : extern tree vect_gen_perm_mask_any (tree, const vec_perm_indices &);
2614 : extern tree vect_gen_perm_mask_checked (tree, const vec_perm_indices &);
2615 : extern void optimize_mask_stores (class loop*);
2616 : extern tree vect_gen_while (gimple_seq *, tree, tree, tree,
2617 : const char * = nullptr);
2618 : extern tree vect_gen_while_not (gimple_seq *, tree, tree, tree);
2619 : extern opt_result vect_get_vector_types_for_stmt (vec_info *,
2620 : stmt_vec_info, tree *,
2621 : unsigned int = 0);
2622 : extern opt_tree vect_get_mask_type_for_stmt (stmt_vec_info, unsigned int = 0);
2623 :
2624 : /* In tree-if-conv.cc. */
2625 : extern bool ref_within_array_bound (gimple *, tree);
2626 :
2627 : /* In tree-vect-data-refs.cc. */
2628 : extern bool vect_can_force_dr_alignment_p (const_tree, poly_uint64);
2629 : extern enum dr_alignment_support vect_supportable_dr_alignment
2630 : (vec_info *, dr_vec_info *, tree, int,
2631 : bool = false);
2632 : extern opt_result vect_analyze_data_ref_dependences (loop_vec_info, unsigned int *);
2633 : extern bool vect_slp_analyze_instance_dependence (vec_info *, slp_instance);
2634 : extern opt_result vect_enhance_data_refs_alignment (loop_vec_info);
2635 : extern void vect_analyze_data_refs_alignment (loop_vec_info);
2636 : extern bool vect_slp_analyze_instance_alignment (vec_info *, slp_instance);
2637 : extern opt_result vect_analyze_data_ref_accesses (vec_info *, vec<int> *);
2638 : extern opt_result vect_prune_runtime_alias_test_list (loop_vec_info);
2639 : extern bool vect_gather_scatter_fn_p (vec_info *, bool, bool, tree, tree,
2640 : tree, int, int *, internal_fn *, tree *,
2641 : tree *, vec<int> * = nullptr);
2642 : extern bool vect_check_gather_scatter (stmt_vec_info, tree,
2643 : loop_vec_info, gather_scatter_info *,
2644 : vec<int> * = nullptr);
2645 : extern void vect_describe_gather_scatter_call (stmt_vec_info,
2646 : gather_scatter_info *);
2647 : extern opt_result vect_find_stmt_data_reference (loop_p, gimple *,
2648 : vec<data_reference_p> *,
2649 : vec<int> *, int);
2650 : extern opt_result vect_analyze_data_refs (vec_info *, bool *);
2651 : extern void vect_record_base_alignments (vec_info *);
2652 : extern tree vect_create_data_ref_ptr (vec_info *,
2653 : stmt_vec_info, tree, class loop *, tree,
2654 : tree *, gimple_stmt_iterator *,
2655 : gimple **, bool,
2656 : tree = NULL_TREE);
2657 : extern tree bump_vector_ptr (vec_info *, tree, gimple_stmt_iterator *,
2658 : stmt_vec_info, tree);
2659 : extern void vect_copy_ref_info (tree, tree);
2660 : extern tree vect_create_destination_var (tree, tree);
2661 : extern bool vect_grouped_store_supported (tree, unsigned HOST_WIDE_INT);
2662 : extern internal_fn vect_store_lanes_supported (tree, unsigned HOST_WIDE_INT, bool);
2663 : extern bool vect_grouped_load_supported (tree, bool, unsigned HOST_WIDE_INT);
2664 : extern internal_fn vect_load_lanes_supported (tree, unsigned HOST_WIDE_INT,
2665 : bool, vec<int> * = nullptr);
2666 : extern tree vect_setup_realignment (vec_info *,
2667 : stmt_vec_info, tree, gimple_stmt_iterator *,
2668 : tree *, enum dr_alignment_support, tree,
2669 : class loop **);
2670 : extern tree vect_get_new_vect_var (tree, enum vect_var_kind, const char *);
2671 : extern tree vect_get_new_ssa_name (tree, enum vect_var_kind,
2672 : const char * = NULL);
2673 : extern tree vect_create_addr_base_for_vector_ref (vec_info *,
2674 : stmt_vec_info, gimple_seq *,
2675 : tree);
2676 :
2677 : /* In tree-vect-loop.cc. */
2678 : extern tree neutral_op_for_reduction (tree, code_helper, tree, bool = true);
2679 : extern widest_int vect_iv_limit_for_partial_vectors (loop_vec_info loop_vinfo);
2680 : bool vect_rgroup_iv_might_wrap_p (loop_vec_info, rgroup_controls *);
2681 : /* Used in gimple-loop-interchange.c and tree-parloops.cc. */
2682 : extern bool check_reduction_path (dump_user_location_t, loop_p, gphi *, tree,
2683 : enum tree_code);
2684 : extern bool needs_fold_left_reduction_p (tree, code_helper);
2685 : /* Drive for loop analysis stage. */
2686 : extern opt_loop_vec_info vect_analyze_loop (class loop *, gimple *,
2687 : vec_info_shared *);
2688 : extern tree vect_build_loop_niters (loop_vec_info, bool * = NULL);
2689 : extern void vect_gen_vector_loop_niters (loop_vec_info, tree, tree *,
2690 : tree *, bool);
2691 : extern tree vect_get_loop_iv_increment (loop_vec_info);
2692 : extern tree vect_halve_mask_nunits (tree, machine_mode);
2693 : extern tree vect_double_mask_nunits (tree, machine_mode);
2694 : extern void vect_record_loop_mask (loop_vec_info, vec_loop_masks *,
2695 : unsigned int, tree, tree);
2696 : extern tree vect_get_loop_mask (loop_vec_info, gimple_stmt_iterator *,
2697 : vec_loop_masks *,
2698 : unsigned int, tree, unsigned int);
2699 : extern void vect_record_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
2700 : tree, unsigned int);
2701 : extern tree vect_get_loop_len (loop_vec_info, gimple_stmt_iterator *,
2702 : vec_loop_lens *, unsigned int, tree,
2703 : unsigned int, unsigned int, bool);
2704 : extern tree vect_gen_loop_len_mask (loop_vec_info, gimple_stmt_iterator *,
2705 : gimple_stmt_iterator *, vec_loop_lens *,
2706 : unsigned int, tree, tree, unsigned int,
2707 : unsigned int);
2708 : extern gimple_seq vect_gen_len (tree, tree, tree, tree);
2709 : extern vect_reduc_info info_for_reduction (loop_vec_info, slp_tree);
2710 : extern bool reduction_fn_for_scalar_code (code_helper, internal_fn *);
2711 : extern unsigned vect_min_prec_for_max_niters (loop_vec_info, unsigned int);
2712 : /* Drive for loop transformation stage. */
2713 : extern class loop *vect_transform_loop (loop_vec_info, gimple *);
2714 930744 : struct vect_loop_form_info
2715 : {
2716 : tree number_of_iterations;
2717 : tree number_of_iterationsm1;
2718 : tree assumptions;
2719 : auto_vec<gcond *> conds;
2720 : gcond *inner_loop_cond;
2721 : edge loop_exit;
2722 : };
2723 : extern opt_result vect_analyze_loop_form (class loop *, gimple *,
2724 : vect_loop_form_info *);
2725 : extern loop_vec_info vect_create_loop_vinfo (class loop *, vec_info_shared *,
2726 : const vect_loop_form_info *,
2727 : loop_vec_info = nullptr);
2728 : extern bool vectorizable_live_operation (vec_info *, stmt_vec_info,
2729 : slp_tree, slp_instance, int,
2730 : bool, stmt_vector_for_cost *);
2731 : extern bool vectorizable_lane_reducing (loop_vec_info, stmt_vec_info,
2732 : slp_tree, stmt_vector_for_cost *);
2733 : extern bool vectorizable_reduction (loop_vec_info, stmt_vec_info,
2734 : slp_tree, slp_instance,
2735 : stmt_vector_for_cost *);
2736 : extern bool vectorizable_induction (loop_vec_info, stmt_vec_info,
2737 : slp_tree, stmt_vector_for_cost *);
2738 : extern bool vect_transform_reduction (loop_vec_info, stmt_vec_info,
2739 : gimple_stmt_iterator *,
2740 : slp_tree);
2741 : extern bool vect_transform_cycle_phi (loop_vec_info, stmt_vec_info,
2742 : slp_tree, slp_instance);
2743 : extern bool vectorizable_lc_phi (loop_vec_info, stmt_vec_info, slp_tree);
2744 : extern bool vect_transform_lc_phi (loop_vec_info, stmt_vec_info, slp_tree);
2745 : extern bool vectorizable_phi (bb_vec_info, stmt_vec_info, slp_tree,
2746 : stmt_vector_for_cost *);
2747 : extern bool vectorizable_recurr (loop_vec_info, stmt_vec_info,
2748 : slp_tree, stmt_vector_for_cost *);
2749 : extern bool vectorizable_early_exit (loop_vec_info, stmt_vec_info,
2750 : gimple_stmt_iterator *,
2751 : slp_tree, stmt_vector_for_cost *);
2752 : extern bool vect_emulated_vector_p (tree);
2753 : extern bool vect_can_vectorize_without_simd_p (tree_code);
2754 : extern bool vect_can_vectorize_without_simd_p (code_helper);
2755 : extern int vect_get_known_peeling_cost (loop_vec_info, int);
2756 : extern tree cse_and_gimplify_to_preheader (loop_vec_info, tree);
2757 :
2758 : /* Nonlinear induction. */
2759 : extern tree vect_peel_nonlinear_iv_init (gimple_seq*, tree, tree,
2760 : tree, enum vect_induction_op_type,
2761 : bool);
2762 :
2763 : /* In tree-vect-slp.cc. */
2764 : extern void vect_slp_init (void);
2765 : extern void vect_slp_fini (void);
2766 : extern void vect_free_slp_instance (slp_instance);
2767 : extern bool vect_transform_slp_perm_load (vec_info *, slp_tree, const vec<tree> &,
2768 : gimple_stmt_iterator *, poly_uint64,
2769 : bool, unsigned *,
2770 : unsigned * = nullptr, bool = false);
2771 : extern bool vectorizable_slp_permutation (vec_info *, gimple_stmt_iterator *,
2772 : slp_tree, stmt_vector_for_cost *);
2773 : extern bool vect_slp_analyze_operations (vec_info *);
2774 : extern bool vect_schedule_slp (vec_info *, vec<slp_instance> &, bool);
2775 : extern opt_result vect_analyze_slp (vec_info *, unsigned, bool);
2776 : extern bool vect_make_slp_decision (loop_vec_info);
2777 : extern void vect_optimize_slp (vec_info *);
2778 : extern void vect_gather_slp_loads (vec_info *);
2779 : extern tree vect_get_slp_scalar_def (slp_tree, unsigned);
2780 : extern void vect_get_slp_defs (slp_tree, vec<tree> *);
2781 : extern void vect_get_slp_defs (vec_info *, slp_tree, vec<vec<tree> > *,
2782 : unsigned n = -1U);
2783 : extern bool vect_slp_if_converted_bb (basic_block bb, loop_p orig_loop);
2784 : extern bool vect_slp_function (function *);
2785 : extern stmt_vec_info vect_find_last_scalar_stmt_in_slp (slp_tree);
2786 : extern stmt_vec_info vect_find_first_scalar_stmt_in_slp (slp_tree);
2787 : extern bool is_simple_and_all_uses_invariant (stmt_vec_info, loop_vec_info);
2788 : extern bool can_duplicate_and_interleave_p (vec_info *, unsigned int, tree,
2789 : unsigned int * = NULL,
2790 : tree * = NULL, tree * = NULL);
2791 : extern void duplicate_and_interleave (vec_info *, gimple_seq *, tree,
2792 : const vec<tree> &, unsigned int, vec<tree> &);
2793 : extern int vect_get_place_in_interleaving_chain (stmt_vec_info, stmt_vec_info);
2794 : extern slp_tree vect_create_new_slp_node (unsigned, tree_code);
2795 : extern void vect_free_slp_tree (slp_tree);
2796 : extern bool compatible_calls_p (gcall *, gcall *, bool);
2797 : extern int vect_slp_child_index_for_operand (const stmt_vec_info, int op);
2798 :
2799 : extern tree prepare_vec_mask (loop_vec_info, tree, tree, tree,
2800 : gimple_stmt_iterator *);
2801 : extern tree vect_get_mask_load_else (int, tree);
2802 : extern bool vect_load_perm_consecutive_p (slp_tree, unsigned = UINT_MAX);
2803 : extern bool vect_get_num_copies_for_invariant (vec_info *, slp_tree,
2804 : unsigned *, unsigned *);
2805 :
2806 : /* In tree-vect-patterns.cc. */
2807 : extern void
2808 : vect_mark_pattern_stmts (vec_info *, stmt_vec_info, gimple *, tree);
2809 : extern bool vect_get_range_info (tree, wide_int*, wide_int*);
2810 :
2811 : /* Pattern recognition functions.
2812 : Additional pattern recognition functions can (and will) be added
2813 : in the future. */
2814 : void vect_pattern_recog (vec_info *);
2815 :
2816 : /* In tree-vectorizer.cc. */
2817 : unsigned vectorize_loops (void);
2818 : void vect_free_loop_info_assumptions (class loop *);
2819 : gimple *vect_loop_vectorized_call (class loop *, gcond **cond = NULL);
2820 : bool vect_stmt_dominates_stmt_p (gimple *, gimple *);
2821 :
2822 : /* SLP Pattern matcher types, tree-vect-slp-patterns.cc. */
2823 :
2824 : /* Forward declaration of possible two operands operation that can be matched
2825 : by the complex numbers pattern matchers. */
2826 : enum _complex_operation : unsigned;
2827 :
2828 : /* All possible load permute values that could result from the partial data-flow
2829 : analysis. */
2830 : typedef enum _complex_perm_kinds {
2831 : PERM_UNKNOWN,
2832 : PERM_EVENODD,
2833 : PERM_ODDEVEN,
2834 : PERM_ODDODD,
2835 : PERM_EVENEVEN,
2836 : /* Can be combined with any other PERM values. */
2837 : PERM_TOP
2838 : } complex_perm_kinds_t;
2839 :
2840 : /* Cache from nodes to the load permutation they represent. */
2841 : typedef hash_map <slp_tree, complex_perm_kinds_t>
2842 : slp_tree_to_load_perm_map_t;
2843 :
2844 : /* Cache from nodes pair to being compatible or not. */
2845 : typedef pair_hash <nofree_ptr_hash <_slp_tree>,
2846 : nofree_ptr_hash <_slp_tree>> slp_node_hash;
2847 : typedef hash_map <slp_node_hash, bool> slp_compat_nodes_map_t;
2848 :
2849 :
2850 : /* Vector pattern matcher base class. All SLP pattern matchers must inherit
2851 : from this type. */
2852 :
2853 : class vect_pattern
2854 : {
2855 : protected:
2856 : /* The number of arguments that the IFN requires. */
2857 : unsigned m_num_args;
2858 :
2859 : /* The internal function that will be used when a pattern is created. */
2860 : internal_fn m_ifn;
2861 :
2862 : /* The current node being inspected. */
2863 : slp_tree *m_node;
2864 :
2865 : /* The list of operands to be the children for the node produced when the
2866 : internal function is created. */
2867 : vec<slp_tree> m_ops;
2868 :
2869 : /* Default constructor where NODE is the root of the tree to inspect. */
2870 1114 : vect_pattern (slp_tree *node, vec<slp_tree> *m_ops, internal_fn ifn)
2871 32 : {
2872 1114 : this->m_ifn = ifn;
2873 1114 : this->m_node = node;
2874 1114 : this->m_ops.create (0);
2875 1114 : if (m_ops)
2876 32 : this->m_ops.safe_splice (*m_ops);
2877 : }
2878 :
2879 : public:
2880 :
2881 : /* Create a new instance of the pattern matcher class of the given type. */
2882 : static vect_pattern* recognize (slp_tree_to_load_perm_map_t *,
2883 : slp_compat_nodes_map_t *, slp_tree *);
2884 :
2885 : /* Build the pattern from the data collected so far. */
2886 : virtual void build (vec_info *) = 0;
2887 :
2888 : /* Default destructor. */
2889 : virtual ~vect_pattern ()
2890 : {
2891 : this->m_ops.release ();
2892 : }
2893 : };
2894 :
2895 : /* Function pointer to create a new pattern matcher from a generic type. */
2896 : typedef vect_pattern* (*vect_pattern_decl_t) (slp_tree_to_load_perm_map_t *,
2897 : slp_compat_nodes_map_t *,
2898 : slp_tree *);
2899 :
2900 : /* List of supported pattern matchers. */
2901 : extern vect_pattern_decl_t slp_patterns[];
2902 :
2903 : /* Number of supported pattern matchers. */
2904 : extern size_t num__slp_patterns;
2905 :
2906 : /* ----------------------------------------------------------------------
2907 : Target support routines
2908 : -----------------------------------------------------------------------
2909 : The following routines are provided to simplify costing decisions in
2910 : target code. Please add more as needed. */
2911 :
2912 : /* Return true if an operation of kind KIND for STMT_INFO represents
2913 : the extraction of an element from a vector in preparation for
2914 : storing the element to memory. */
2915 : inline bool
2916 : vect_is_store_elt_extraction (vect_cost_for_stmt kind, stmt_vec_info stmt_info)
2917 : {
2918 : return (kind == vec_to_scalar
2919 : && STMT_VINFO_DATA_REF (stmt_info)
2920 : && DR_IS_WRITE (STMT_VINFO_DATA_REF (stmt_info)));
2921 : }
2922 :
2923 : /* Return true if STMT_INFO represents part of a reduction. */
2924 : inline bool
2925 50543794 : vect_is_reduction (stmt_vec_info stmt_info)
2926 : {
2927 50543794 : return STMT_VINFO_REDUC_IDX (stmt_info) != -1;
2928 : }
2929 :
2930 : /* Return true if SLP_NODE represents part of a reduction. */
2931 : inline bool
2932 318431 : vect_is_reduction (slp_tree slp_node)
2933 : {
2934 318431 : return SLP_TREE_REDUC_IDX (slp_node) != -1;
2935 : }
2936 :
2937 : /* If STMT_INFO describes a reduction, return the vect_reduction_type
2938 : of the reduction it describes, otherwise return -1. */
2939 : inline int
2940 55895 : vect_reduc_type (vec_info *vinfo, slp_tree node)
2941 : {
2942 55895 : if (loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo))
2943 : {
2944 55895 : vect_reduc_info reduc_info = info_for_reduction (loop_vinfo, node);
2945 55895 : if (reduc_info)
2946 52942 : return int (VECT_REDUC_INFO_TYPE (reduc_info));
2947 : }
2948 : return -1;
2949 : }
2950 :
2951 : /* If STMT_INFO is a COND_EXPR that includes an embedded comparison, return the
2952 : scalar type of the values being compared. Return null otherwise. */
2953 : inline tree
2954 0 : vect_embedded_comparison_type (stmt_vec_info stmt_info)
2955 : {
2956 0 : if (auto *assign = dyn_cast<gassign *> (stmt_info->stmt))
2957 0 : if (gimple_assign_rhs_code (assign) == COND_EXPR)
2958 : {
2959 0 : tree cond = gimple_assign_rhs1 (assign);
2960 0 : if (COMPARISON_CLASS_P (cond))
2961 0 : return TREE_TYPE (TREE_OPERAND (cond, 0));
2962 : }
2963 : return NULL_TREE;
2964 : }
2965 :
2966 : /* If STMT_INFO is a comparison or contains an embedded comparison, return the
2967 : scalar type of the values being compared. Return null otherwise. */
2968 : inline tree
2969 67651 : vect_comparison_type (stmt_vec_info stmt_info)
2970 : {
2971 67651 : if (auto *assign = dyn_cast<gassign *> (stmt_info->stmt))
2972 67651 : if (TREE_CODE_CLASS (gimple_assign_rhs_code (assign)) == tcc_comparison)
2973 67651 : return TREE_TYPE (gimple_assign_rhs1 (assign));
2974 0 : return vect_embedded_comparison_type (stmt_info);
2975 : }
2976 :
2977 : /* Return true if STMT_INFO extends the result of a load. */
2978 : inline bool
2979 : vect_is_extending_load (class vec_info *vinfo, stmt_vec_info stmt_info)
2980 : {
2981 : /* Although this is quite large for an inline function, this part
2982 : at least should be inline. */
2983 : gassign *assign = dyn_cast <gassign *> (stmt_info->stmt);
2984 : if (!assign || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (assign)))
2985 : return false;
2986 :
2987 : tree rhs = gimple_assign_rhs1 (stmt_info->stmt);
2988 : tree lhs_type = TREE_TYPE (gimple_assign_lhs (assign));
2989 : tree rhs_type = TREE_TYPE (rhs);
2990 : if (!INTEGRAL_TYPE_P (lhs_type)
2991 : || !INTEGRAL_TYPE_P (rhs_type)
2992 : || TYPE_PRECISION (lhs_type) <= TYPE_PRECISION (rhs_type))
2993 : return false;
2994 :
2995 : stmt_vec_info def_stmt_info = vinfo->lookup_def (rhs);
2996 : return (def_stmt_info
2997 : && STMT_VINFO_DATA_REF (def_stmt_info)
2998 : && DR_IS_READ (STMT_VINFO_DATA_REF (def_stmt_info)));
2999 : }
3000 :
3001 : /* Return true if STMT_INFO truncates the input of a store. */
3002 : inline bool
3003 : vect_is_truncating_store (class vec_info *vinfo, stmt_vec_info stmt_info)
3004 : {
3005 : /* Although this is quite large for an inline function, this part
3006 : at least should be inline. */
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 rhs = gimple_assign_rhs1 (stmt_info->stmt);
3012 : tree lhs = gimple_assign_lhs (assign);
3013 : tree lhs_type = TREE_TYPE (lhs);
3014 : tree rhs_type = TREE_TYPE (rhs);
3015 : if (!INTEGRAL_TYPE_P (lhs_type) || !INTEGRAL_TYPE_P (rhs_type)
3016 : || TYPE_PRECISION (lhs_type) >= TYPE_PRECISION (rhs_type))
3017 : return false;
3018 :
3019 : gimple *use_stmt;
3020 : use_operand_p use_p;
3021 : if (!single_imm_use (lhs, &use_p, &use_stmt))
3022 : return false;
3023 :
3024 : stmt_vec_info use_stmt_info = vinfo->lookup_stmt (use_stmt);
3025 : return (use_stmt_info && STMT_VINFO_DATA_REF (use_stmt_info)
3026 : && DR_IS_WRITE (STMT_VINFO_DATA_REF (use_stmt_info)));
3027 : }
3028 :
3029 : /* Return true if STMT_INFO is an integer truncation. */
3030 : inline bool
3031 : vect_is_integer_truncation (stmt_vec_info stmt_info)
3032 : {
3033 : gassign *assign = dyn_cast <gassign *> (stmt_info->stmt);
3034 : if (!assign || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (assign)))
3035 : return false;
3036 :
3037 : tree lhs_type = TREE_TYPE (gimple_assign_lhs (assign));
3038 : tree rhs_type = TREE_TYPE (gimple_assign_rhs1 (assign));
3039 : return (INTEGRAL_TYPE_P (lhs_type)
3040 : && INTEGRAL_TYPE_P (rhs_type)
3041 : && TYPE_PRECISION (lhs_type) < TYPE_PRECISION (rhs_type));
3042 : }
3043 :
3044 : /* Build a GIMPLE_ASSIGN or GIMPLE_CALL with the tree_code,
3045 : or internal_fn contained in ch, respectively. */
3046 : gimple * vect_gimple_build (tree, code_helper, tree, tree = NULL_TREE);
3047 : #endif /* GCC_TREE_VECTORIZER_H */
|
