Daily bump.
[official-gcc.git] / gcc / tree-vectorizer.h
blob2f6e1e268fb07e9de065ff9c45af87546e565d66
1 /* Vectorizer
2 Copyright (C) 2003-2021 Free Software Foundation, Inc.
3 Contributed by Dorit Naishlos <dorit@il.ibm.com>
5 This file is part of GCC.
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.
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.
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/>. */
21 #ifndef GCC_TREE_VECTORIZER_H
22 #define GCC_TREE_VECTORIZER_H
24 typedef class _stmt_vec_info *stmt_vec_info;
26 #include "tree-data-ref.h"
27 #include "tree-hash-traits.h"
28 #include "target.h"
29 #include "internal-fn.h"
30 #include "tree-ssa-operands.h"
31 #include "gimple-match.h"
33 /* Used for naming of new temporaries. */
34 enum vect_var_kind {
35 vect_simple_var,
36 vect_pointer_var,
37 vect_scalar_var,
38 vect_mask_var
41 /* Defines type of operation. */
42 enum operation_type {
43 unary_op = 1,
44 binary_op,
45 ternary_op
48 /* Define type of available alignment support. */
49 enum dr_alignment_support {
50 dr_unaligned_unsupported,
51 dr_unaligned_supported,
52 dr_explicit_realign,
53 dr_explicit_realign_optimized,
54 dr_aligned
57 /* Define type of def-use cross-iteration cycle. */
58 enum vect_def_type {
59 vect_uninitialized_def = 0,
60 vect_constant_def = 1,
61 vect_external_def,
62 vect_internal_def,
63 vect_induction_def,
64 vect_reduction_def,
65 vect_double_reduction_def,
66 vect_nested_cycle,
67 vect_unknown_def_type
70 /* Define type of reduction. */
71 enum vect_reduction_type {
72 TREE_CODE_REDUCTION,
73 COND_REDUCTION,
74 INTEGER_INDUC_COND_REDUCTION,
75 CONST_COND_REDUCTION,
77 /* Retain a scalar phi and use a FOLD_EXTRACT_LAST within the loop
78 to implement:
80 for (int i = 0; i < VF; ++i)
81 res = cond[i] ? val[i] : res; */
82 EXTRACT_LAST_REDUCTION,
84 /* Use a folding reduction within the loop to implement:
86 for (int i = 0; i < VF; ++i)
87 res = res OP val[i];
89 (with no reassocation). */
90 FOLD_LEFT_REDUCTION
93 #define VECTORIZABLE_CYCLE_DEF(D) (((D) == vect_reduction_def) \
94 || ((D) == vect_double_reduction_def) \
95 || ((D) == vect_nested_cycle))
97 /* Structure to encapsulate information about a group of like
98 instructions to be presented to the target cost model. */
99 struct stmt_info_for_cost {
100 int count;
101 enum vect_cost_for_stmt kind;
102 enum vect_cost_model_location where;
103 stmt_vec_info stmt_info;
104 tree vectype;
105 int misalign;
108 typedef vec<stmt_info_for_cost> stmt_vector_for_cost;
110 /* Maps base addresses to an innermost_loop_behavior and the stmt it was
111 derived from that gives the maximum known alignment for that base. */
112 typedef hash_map<tree_operand_hash,
113 std::pair<stmt_vec_info, innermost_loop_behavior *> >
114 vec_base_alignments;
116 /************************************************************************
118 ************************************************************************/
119 typedef struct _slp_tree *slp_tree;
120 typedef vec<std::pair<unsigned, unsigned> > lane_permutation_t;
121 typedef vec<unsigned> load_permutation_t;
123 /* A computation tree of an SLP instance. Each node corresponds to a group of
124 stmts to be packed in a SIMD stmt. */
125 struct _slp_tree {
126 _slp_tree ();
127 ~_slp_tree ();
129 /* Nodes that contain def-stmts of this node statements operands. */
130 vec<slp_tree> children;
132 /* A group of scalar stmts to be vectorized together. */
133 vec<stmt_vec_info> stmts;
134 /* A group of scalar operands to be vectorized together. */
135 vec<tree> ops;
136 /* The representative that should be used for analysis and
137 code generation. */
138 stmt_vec_info representative;
140 /* Load permutation relative to the stores, NULL if there is no
141 permutation. */
142 load_permutation_t load_permutation;
143 /* Lane permutation of the operands scalar lanes encoded as pairs
144 of { operand number, lane number }. The number of elements
145 denotes the number of output lanes. */
146 lane_permutation_t lane_permutation;
148 tree vectype;
149 /* Vectorized stmt/s. */
150 vec<gimple *> vec_stmts;
151 vec<tree> vec_defs;
152 /* Number of vector stmts that are created to replace the group of scalar
153 stmts. It is calculated during the transformation phase as the number of
154 scalar elements in one scalar iteration (GROUP_SIZE) multiplied by VF
155 divided by vector size. */
156 unsigned int vec_stmts_size;
158 /* Reference count in the SLP graph. */
159 unsigned int refcnt;
160 /* The maximum number of vector elements for the subtree rooted
161 at this node. */
162 poly_uint64 max_nunits;
163 /* The DEF type of this node. */
164 enum vect_def_type def_type;
165 /* The number of scalar lanes produced by this node. */
166 unsigned int lanes;
167 /* The operation of this node. */
168 enum tree_code code;
170 int vertex;
172 /* If not NULL this is a cached failed SLP discovery attempt with
173 the lanes that failed during SLP discovery as 'false'. This is
174 a copy of the matches array. */
175 bool *failed;
177 /* Allocate from slp_tree_pool. */
178 static void *operator new (size_t);
180 /* Return memory to slp_tree_pool. */
181 static void operator delete (void *, size_t);
183 /* Linked list of nodes to release when we free the slp_tree_pool. */
184 slp_tree next_node;
185 slp_tree prev_node;
188 /* The enum describes the type of operations that an SLP instance
189 can perform. */
191 enum slp_instance_kind {
192 slp_inst_kind_store,
193 slp_inst_kind_reduc_group,
194 slp_inst_kind_reduc_chain,
195 slp_inst_kind_bb_reduc,
196 slp_inst_kind_ctor
199 /* SLP instance is a sequence of stmts in a loop that can be packed into
200 SIMD stmts. */
201 typedef class _slp_instance {
202 public:
203 /* The root of SLP tree. */
204 slp_tree root;
206 /* For vector constructors, the constructor stmt that the SLP tree is built
207 from, NULL otherwise. */
208 vec<stmt_vec_info> root_stmts;
210 /* The unrolling factor required to vectorized this SLP instance. */
211 poly_uint64 unrolling_factor;
213 /* The group of nodes that contain loads of this SLP instance. */
214 vec<slp_tree> loads;
216 /* The SLP node containing the reduction PHIs. */
217 slp_tree reduc_phis;
219 /* Vector cost of this entry to the SLP graph. */
220 stmt_vector_for_cost cost_vec;
222 /* If this instance is the main entry of a subgraph the set of
223 entries into the same subgraph, including itself. */
224 vec<_slp_instance *> subgraph_entries;
226 /* The type of operation the SLP instance is performing. */
227 slp_instance_kind kind;
229 dump_user_location_t location () const;
230 } *slp_instance;
233 /* Access Functions. */
234 #define SLP_INSTANCE_TREE(S) (S)->root
235 #define SLP_INSTANCE_UNROLLING_FACTOR(S) (S)->unrolling_factor
236 #define SLP_INSTANCE_LOADS(S) (S)->loads
237 #define SLP_INSTANCE_ROOT_STMTS(S) (S)->root_stmts
238 #define SLP_INSTANCE_KIND(S) (S)->kind
240 #define SLP_TREE_CHILDREN(S) (S)->children
241 #define SLP_TREE_SCALAR_STMTS(S) (S)->stmts
242 #define SLP_TREE_SCALAR_OPS(S) (S)->ops
243 #define SLP_TREE_REF_COUNT(S) (S)->refcnt
244 #define SLP_TREE_VEC_STMTS(S) (S)->vec_stmts
245 #define SLP_TREE_VEC_DEFS(S) (S)->vec_defs
246 #define SLP_TREE_NUMBER_OF_VEC_STMTS(S) (S)->vec_stmts_size
247 #define SLP_TREE_LOAD_PERMUTATION(S) (S)->load_permutation
248 #define SLP_TREE_LANE_PERMUTATION(S) (S)->lane_permutation
249 #define SLP_TREE_DEF_TYPE(S) (S)->def_type
250 #define SLP_TREE_VECTYPE(S) (S)->vectype
251 #define SLP_TREE_REPRESENTATIVE(S) (S)->representative
252 #define SLP_TREE_LANES(S) (S)->lanes
253 #define SLP_TREE_CODE(S) (S)->code
255 /* Key for map that records association between
256 scalar conditions and corresponding loop mask, and
257 is populated by vect_record_loop_mask. */
259 struct scalar_cond_masked_key
261 scalar_cond_masked_key (tree t, unsigned ncopies_)
262 : ncopies (ncopies_)
264 get_cond_ops_from_tree (t);
267 void get_cond_ops_from_tree (tree);
269 unsigned ncopies;
270 bool inverted_p;
271 tree_code code;
272 tree op0;
273 tree op1;
276 template<>
277 struct default_hash_traits<scalar_cond_masked_key>
279 typedef scalar_cond_masked_key compare_type;
280 typedef scalar_cond_masked_key value_type;
282 static inline hashval_t
283 hash (value_type v)
285 inchash::hash h;
286 h.add_int (v.code);
287 inchash::add_expr (v.op0, h, 0);
288 inchash::add_expr (v.op1, h, 0);
289 h.add_int (v.ncopies);
290 h.add_flag (v.inverted_p);
291 return h.end ();
294 static inline bool
295 equal (value_type existing, value_type candidate)
297 return (existing.ncopies == candidate.ncopies
298 && existing.code == candidate.code
299 && existing.inverted_p == candidate.inverted_p
300 && operand_equal_p (existing.op0, candidate.op0, 0)
301 && operand_equal_p (existing.op1, candidate.op1, 0));
304 static const bool empty_zero_p = true;
306 static inline void
307 mark_empty (value_type &v)
309 v.ncopies = 0;
310 v.inverted_p = false;
313 static inline bool
314 is_empty (value_type v)
316 return v.ncopies == 0;
319 static inline void mark_deleted (value_type &) {}
321 static inline bool is_deleted (const value_type &)
323 return false;
326 static inline void remove (value_type &) {}
329 typedef hash_set<scalar_cond_masked_key> scalar_cond_masked_set_type;
331 /* Key and map that records association between vector conditions and
332 corresponding loop mask, and is populated by prepare_vec_mask. */
334 typedef pair_hash<tree_operand_hash, tree_operand_hash> tree_cond_mask_hash;
335 typedef hash_set<tree_cond_mask_hash> vec_cond_masked_set_type;
337 /* Describes two objects whose addresses must be unequal for the vectorized
338 loop to be valid. */
339 typedef std::pair<tree, tree> vec_object_pair;
341 /* Records that vectorization is only possible if abs (EXPR) >= MIN_VALUE.
342 UNSIGNED_P is true if we can assume that abs (EXPR) == EXPR. */
343 class vec_lower_bound {
344 public:
345 vec_lower_bound () {}
346 vec_lower_bound (tree e, bool u, poly_uint64 m)
347 : expr (e), unsigned_p (u), min_value (m) {}
349 tree expr;
350 bool unsigned_p;
351 poly_uint64 min_value;
354 /* Vectorizer state shared between different analyses like vector sizes
355 of the same CFG region. */
356 class vec_info_shared {
357 public:
358 vec_info_shared();
359 ~vec_info_shared();
361 void save_datarefs();
362 void check_datarefs();
364 /* The number of scalar stmts. */
365 unsigned n_stmts;
367 /* All data references. Freed by free_data_refs, so not an auto_vec. */
368 vec<data_reference_p> datarefs;
369 vec<data_reference> datarefs_copy;
371 /* The loop nest in which the data dependences are computed. */
372 auto_vec<loop_p> loop_nest;
374 /* All data dependences. Freed by free_dependence_relations, so not
375 an auto_vec. */
376 vec<ddr_p> ddrs;
379 /* Vectorizer state common between loop and basic-block vectorization. */
380 class vec_info {
381 public:
382 typedef hash_set<int_hash<machine_mode, E_VOIDmode, E_BLKmode> > mode_set;
383 enum vec_kind { bb, loop };
385 vec_info (vec_kind, vec_info_shared *);
386 ~vec_info ();
388 stmt_vec_info add_stmt (gimple *);
389 stmt_vec_info add_pattern_stmt (gimple *, stmt_vec_info);
390 stmt_vec_info lookup_stmt (gimple *);
391 stmt_vec_info lookup_def (tree);
392 stmt_vec_info lookup_single_use (tree);
393 class dr_vec_info *lookup_dr (data_reference *);
394 void move_dr (stmt_vec_info, stmt_vec_info);
395 void remove_stmt (stmt_vec_info);
396 void replace_stmt (gimple_stmt_iterator *, stmt_vec_info, gimple *);
397 void insert_on_entry (stmt_vec_info, gimple *);
398 void insert_seq_on_entry (stmt_vec_info, gimple_seq);
400 /* The type of vectorization. */
401 vec_kind kind;
403 /* Shared vectorizer state. */
404 vec_info_shared *shared;
406 /* The mapping of GIMPLE UID to stmt_vec_info. */
407 vec<stmt_vec_info> stmt_vec_infos;
408 /* Whether the above mapping is complete. */
409 bool stmt_vec_info_ro;
411 /* The SLP graph. */
412 auto_vec<slp_instance> slp_instances;
414 /* Maps base addresses to an innermost_loop_behavior that gives the maximum
415 known alignment for that base. */
416 vec_base_alignments base_alignments;
418 /* All interleaving chains of stores, represented by the first
419 stmt in the chain. */
420 auto_vec<stmt_vec_info> grouped_stores;
422 /* The set of vector modes used in the vectorized region. */
423 mode_set used_vector_modes;
425 /* The argument we should pass to related_vector_mode when looking up
426 the vector mode for a scalar mode, or VOIDmode if we haven't yet
427 made any decisions about which vector modes to use. */
428 machine_mode vector_mode;
430 private:
431 stmt_vec_info new_stmt_vec_info (gimple *stmt);
432 void set_vinfo_for_stmt (gimple *, stmt_vec_info, bool = true);
433 void free_stmt_vec_infos ();
434 void free_stmt_vec_info (stmt_vec_info);
437 class _loop_vec_info;
438 class _bb_vec_info;
440 template<>
441 template<>
442 inline bool
443 is_a_helper <_loop_vec_info *>::test (vec_info *i)
445 return i->kind == vec_info::loop;
448 template<>
449 template<>
450 inline bool
451 is_a_helper <_bb_vec_info *>::test (vec_info *i)
453 return i->kind == vec_info::bb;
456 /* In general, we can divide the vector statements in a vectorized loop
457 into related groups ("rgroups") and say that for each rgroup there is
458 some nS such that the rgroup operates on nS values from one scalar
459 iteration followed by nS values from the next. That is, if VF is the
460 vectorization factor of the loop, the rgroup operates on a sequence:
462 (1,1) (1,2) ... (1,nS) (2,1) ... (2,nS) ... (VF,1) ... (VF,nS)
464 where (i,j) represents a scalar value with index j in a scalar
465 iteration with index i.
467 [ We use the term "rgroup" to emphasise that this grouping isn't
468 necessarily the same as the grouping of statements used elsewhere.
469 For example, if we implement a group of scalar loads using gather
470 loads, we'll use a separate gather load for each scalar load, and
471 thus each gather load will belong to its own rgroup. ]
473 In general this sequence will occupy nV vectors concatenated
474 together. If these vectors have nL lanes each, the total number
475 of scalar values N is given by:
477 N = nS * VF = nV * nL
479 None of nS, VF, nV and nL are required to be a power of 2. nS and nV
480 are compile-time constants but VF and nL can be variable (if the target
481 supports variable-length vectors).
483 In classical vectorization, each iteration of the vector loop would
484 handle exactly VF iterations of the original scalar loop. However,
485 in vector loops that are able to operate on partial vectors, a
486 particular iteration of the vector loop might handle fewer than VF
487 iterations of the scalar loop. The vector lanes that correspond to
488 iterations of the scalar loop are said to be "active" and the other
489 lanes are said to be "inactive".
491 In such vector loops, many rgroups need to be controlled to ensure
492 that they have no effect for the inactive lanes. Conceptually, each
493 such rgroup needs a sequence of booleans in the same order as above,
494 but with each (i,j) replaced by a boolean that indicates whether
495 iteration i is active. This sequence occupies nV vector controls
496 that again have nL lanes each. Thus the control sequence as a whole
497 consists of VF independent booleans that are each repeated nS times.
499 Taking mask-based approach as a partially-populated vectors example.
500 We make the simplifying assumption that if a sequence of nV masks is
501 suitable for one (nS,nL) pair, we can reuse it for (nS/2,nL/2) by
502 VIEW_CONVERTing it. This holds for all current targets that support
503 fully-masked loops. For example, suppose the scalar loop is:
505 float *f;
506 double *d;
507 for (int i = 0; i < n; ++i)
509 f[i * 2 + 0] += 1.0f;
510 f[i * 2 + 1] += 2.0f;
511 d[i] += 3.0;
514 and suppose that vectors have 256 bits. The vectorized f accesses
515 will belong to one rgroup and the vectorized d access to another:
517 f rgroup: nS = 2, nV = 1, nL = 8
518 d rgroup: nS = 1, nV = 1, nL = 4
519 VF = 4
521 [ In this simple example the rgroups do correspond to the normal
522 SLP grouping scheme. ]
524 If only the first three lanes are active, the masks we need are:
526 f rgroup: 1 1 | 1 1 | 1 1 | 0 0
527 d rgroup: 1 | 1 | 1 | 0
529 Here we can use a mask calculated for f's rgroup for d's, but not
530 vice versa.
532 Thus for each value of nV, it is enough to provide nV masks, with the
533 mask being calculated based on the highest nL (or, equivalently, based
534 on the highest nS) required by any rgroup with that nV. We therefore
535 represent the entire collection of masks as a two-level table, with the
536 first level being indexed by nV - 1 (since nV == 0 doesn't exist) and
537 the second being indexed by the mask index 0 <= i < nV. */
539 /* The controls (like masks or lengths) needed by rgroups with nV vectors,
540 according to the description above. */
541 struct rgroup_controls {
542 /* The largest nS for all rgroups that use these controls. */
543 unsigned int max_nscalars_per_iter;
545 /* For the largest nS recorded above, the loop controls divide each scalar
546 into FACTOR equal-sized pieces. This is useful if we need to split
547 element-based accesses into byte-based accesses. */
548 unsigned int factor;
550 /* This is a vector type with MAX_NSCALARS_PER_ITER * VF / nV elements.
551 For mask-based controls, it is the type of the masks in CONTROLS.
552 For length-based controls, it can be any vector type that has the
553 specified number of elements; the type of the elements doesn't matter. */
554 tree type;
556 /* A vector of nV controls, in iteration order. */
557 vec<tree> controls;
560 typedef auto_vec<rgroup_controls> vec_loop_masks;
562 typedef auto_vec<rgroup_controls> vec_loop_lens;
564 typedef auto_vec<std::pair<data_reference*, tree> > drs_init_vec;
566 /* Information about a reduction accumulator from the main loop that could
567 conceivably be reused as the input to a reduction in an epilogue loop. */
568 struct vect_reusable_accumulator {
569 /* The final value of the accumulator, which forms the input to the
570 reduction operation. */
571 tree reduc_input;
573 /* The stmt_vec_info that describes the reduction (i.e. the one for
574 which is_reduc_info is true). */
575 stmt_vec_info reduc_info;
578 /*-----------------------------------------------------------------*/
579 /* Info on vectorized loops. */
580 /*-----------------------------------------------------------------*/
581 typedef class _loop_vec_info : public vec_info {
582 public:
583 _loop_vec_info (class loop *, vec_info_shared *);
584 ~_loop_vec_info ();
586 /* The loop to which this info struct refers to. */
587 class loop *loop;
589 /* The loop basic blocks. */
590 basic_block *bbs;
592 /* Number of latch executions. */
593 tree num_itersm1;
594 /* Number of iterations. */
595 tree num_iters;
596 /* Number of iterations of the original loop. */
597 tree num_iters_unchanged;
598 /* Condition under which this loop is analyzed and versioned. */
599 tree num_iters_assumptions;
601 /* The cost of the vector code. */
602 class vector_costs *vector_costs;
604 /* The cost of the scalar code. */
605 class vector_costs *scalar_costs;
607 /* Threshold of number of iterations below which vectorization will not be
608 performed. It is calculated from MIN_PROFITABLE_ITERS and
609 param_min_vect_loop_bound. */
610 unsigned int th;
612 /* When applying loop versioning, the vector form should only be used
613 if the number of scalar iterations is >= this value, on top of all
614 the other requirements. Ignored when loop versioning is not being
615 used. */
616 poly_uint64 versioning_threshold;
618 /* Unrolling factor */
619 poly_uint64 vectorization_factor;
621 /* If this loop is an epilogue loop whose main loop can be skipped,
622 MAIN_LOOP_EDGE is the edge from the main loop to this loop's
623 preheader. SKIP_MAIN_LOOP_EDGE is then the edge that skips the
624 main loop and goes straight to this loop's preheader.
626 Both fields are null otherwise. */
627 edge main_loop_edge;
628 edge skip_main_loop_edge;
630 /* If this loop is an epilogue loop that might be skipped after executing
631 the main loop, this edge is the one that skips the epilogue. */
632 edge skip_this_loop_edge;
634 /* The vectorized form of a standard reduction replaces the original
635 scalar code's final result (a loop-closed SSA PHI) with the result
636 of a vector-to-scalar reduction operation. After vectorization,
637 this variable maps these vector-to-scalar results to information
638 about the reductions that generated them. */
639 hash_map<tree, vect_reusable_accumulator> reusable_accumulators;
641 /* Maximum runtime vectorization factor, or MAX_VECTORIZATION_FACTOR
642 if there is no particular limit. */
643 unsigned HOST_WIDE_INT max_vectorization_factor;
645 /* The masks that a fully-masked loop should use to avoid operating
646 on inactive scalars. */
647 vec_loop_masks masks;
649 /* The lengths that a loop with length should use to avoid operating
650 on inactive scalars. */
651 vec_loop_lens lens;
653 /* Set of scalar conditions that have loop mask applied. */
654 scalar_cond_masked_set_type scalar_cond_masked_set;
656 /* Set of vector conditions that have loop mask applied. */
657 vec_cond_masked_set_type vec_cond_masked_set;
659 /* If we are using a loop mask to align memory addresses, this variable
660 contains the number of vector elements that we should skip in the
661 first iteration of the vector loop (i.e. the number of leading
662 elements that should be false in the first mask). */
663 tree mask_skip_niters;
665 /* The type that the loop control IV should be converted to before
666 testing which of the VF scalars are active and inactive.
667 Only meaningful if LOOP_VINFO_USING_PARTIAL_VECTORS_P. */
668 tree rgroup_compare_type;
670 /* For #pragma omp simd if (x) loops the x expression. If constant 0,
671 the loop should not be vectorized, if constant non-zero, simd_if_cond
672 shouldn't be set and loop vectorized normally, if SSA_NAME, the loop
673 should be versioned on that condition, using scalar loop if the condition
674 is false and vectorized loop otherwise. */
675 tree simd_if_cond;
677 /* The type that the vector loop control IV should have when
678 LOOP_VINFO_USING_PARTIAL_VECTORS_P is true. */
679 tree rgroup_iv_type;
681 /* Unknown DRs according to which loop was peeled. */
682 class dr_vec_info *unaligned_dr;
684 /* peeling_for_alignment indicates whether peeling for alignment will take
685 place, and what the peeling factor should be:
686 peeling_for_alignment = X means:
687 If X=0: Peeling for alignment will not be applied.
688 If X>0: Peel first X iterations.
689 If X=-1: Generate a runtime test to calculate the number of iterations
690 to be peeled, using the dataref recorded in the field
691 unaligned_dr. */
692 int peeling_for_alignment;
694 /* The mask used to check the alignment of pointers or arrays. */
695 int ptr_mask;
697 /* Data Dependence Relations defining address ranges that are candidates
698 for a run-time aliasing check. */
699 auto_vec<ddr_p> may_alias_ddrs;
701 /* Data Dependence Relations defining address ranges together with segment
702 lengths from which the run-time aliasing check is built. */
703 auto_vec<dr_with_seg_len_pair_t> comp_alias_ddrs;
705 /* Check that the addresses of each pair of objects is unequal. */
706 auto_vec<vec_object_pair> check_unequal_addrs;
708 /* List of values that are required to be nonzero. This is used to check
709 whether things like "x[i * n] += 1;" are safe and eventually gets added
710 to the checks for lower bounds below. */
711 auto_vec<tree> check_nonzero;
713 /* List of values that need to be checked for a minimum value. */
714 auto_vec<vec_lower_bound> lower_bounds;
716 /* Statements in the loop that have data references that are candidates for a
717 runtime (loop versioning) misalignment check. */
718 auto_vec<stmt_vec_info> may_misalign_stmts;
720 /* Reduction cycles detected in the loop. Used in loop-aware SLP. */
721 auto_vec<stmt_vec_info> reductions;
723 /* All reduction chains in the loop, represented by the first
724 stmt in the chain. */
725 auto_vec<stmt_vec_info> reduction_chains;
727 /* Cost vector for a single scalar iteration. */
728 auto_vec<stmt_info_for_cost> scalar_cost_vec;
730 /* Map of IV base/step expressions to inserted name in the preheader. */
731 hash_map<tree_operand_hash, tree> *ivexpr_map;
733 /* Map of OpenMP "omp simd array" scan variables to corresponding
734 rhs of the store of the initializer. */
735 hash_map<tree, tree> *scan_map;
737 /* The unrolling factor needed to SLP the loop. In case of that pure SLP is
738 applied to the loop, i.e., no unrolling is needed, this is 1. */
739 poly_uint64 slp_unrolling_factor;
741 /* The factor used to over weight those statements in an inner loop
742 relative to the loop being vectorized. */
743 unsigned int inner_loop_cost_factor;
745 /* Is the loop vectorizable? */
746 bool vectorizable;
748 /* Records whether we still have the option of vectorizing this loop
749 using partially-populated vectors; in other words, whether it is
750 still possible for one iteration of the vector loop to handle
751 fewer than VF scalars. */
752 bool can_use_partial_vectors_p;
754 /* True if we've decided to use partially-populated vectors, so that
755 the vector loop can handle fewer than VF scalars. */
756 bool using_partial_vectors_p;
758 /* True if we've decided to use partially-populated vectors for the
759 epilogue of loop. */
760 bool epil_using_partial_vectors_p;
762 /* When we have grouped data accesses with gaps, we may introduce invalid
763 memory accesses. We peel the last iteration of the loop to prevent
764 this. */
765 bool peeling_for_gaps;
767 /* When the number of iterations is not a multiple of the vector size
768 we need to peel off iterations at the end to form an epilogue loop. */
769 bool peeling_for_niter;
771 /* True if there are no loop carried data dependencies in the loop.
772 If loop->safelen <= 1, then this is always true, either the loop
773 didn't have any loop carried data dependencies, or the loop is being
774 vectorized guarded with some runtime alias checks, or couldn't
775 be vectorized at all, but then this field shouldn't be used.
776 For loop->safelen >= 2, the user has asserted that there are no
777 backward dependencies, but there still could be loop carried forward
778 dependencies in such loops. This flag will be false if normal
779 vectorizer data dependency analysis would fail or require versioning
780 for alias, but because of loop->safelen >= 2 it has been vectorized
781 even without versioning for alias. E.g. in:
782 #pragma omp simd
783 for (int i = 0; i < m; i++)
784 a[i] = a[i + k] * c;
785 (or #pragma simd or #pragma ivdep) we can vectorize this and it will
786 DTRT even for k > 0 && k < m, but without safelen we would not
787 vectorize this, so this field would be false. */
788 bool no_data_dependencies;
790 /* Mark loops having masked stores. */
791 bool has_mask_store;
793 /* Queued scaling factor for the scalar loop. */
794 profile_probability scalar_loop_scaling;
796 /* If if-conversion versioned this loop before conversion, this is the
797 loop version without if-conversion. */
798 class loop *scalar_loop;
800 /* For loops being epilogues of already vectorized loops
801 this points to the original vectorized loop. Otherwise NULL. */
802 _loop_vec_info *orig_loop_info;
804 /* Used to store loop_vec_infos of epilogues of this loop during
805 analysis. */
806 vec<_loop_vec_info *> epilogue_vinfos;
808 } *loop_vec_info;
810 /* Access Functions. */
811 #define LOOP_VINFO_LOOP(L) (L)->loop
812 #define LOOP_VINFO_BBS(L) (L)->bbs
813 #define LOOP_VINFO_NITERSM1(L) (L)->num_itersm1
814 #define LOOP_VINFO_NITERS(L) (L)->num_iters
815 /* Since LOOP_VINFO_NITERS and LOOP_VINFO_NITERSM1 can change after
816 prologue peeling retain total unchanged scalar loop iterations for
817 cost model. */
818 #define LOOP_VINFO_NITERS_UNCHANGED(L) (L)->num_iters_unchanged
819 #define LOOP_VINFO_NITERS_ASSUMPTIONS(L) (L)->num_iters_assumptions
820 #define LOOP_VINFO_COST_MODEL_THRESHOLD(L) (L)->th
821 #define LOOP_VINFO_VERSIONING_THRESHOLD(L) (L)->versioning_threshold
822 #define LOOP_VINFO_VECTORIZABLE_P(L) (L)->vectorizable
823 #define LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P(L) (L)->can_use_partial_vectors_p
824 #define LOOP_VINFO_USING_PARTIAL_VECTORS_P(L) (L)->using_partial_vectors_p
825 #define LOOP_VINFO_EPIL_USING_PARTIAL_VECTORS_P(L) \
826 (L)->epil_using_partial_vectors_p
827 #define LOOP_VINFO_VECT_FACTOR(L) (L)->vectorization_factor
828 #define LOOP_VINFO_MAX_VECT_FACTOR(L) (L)->max_vectorization_factor
829 #define LOOP_VINFO_MASKS(L) (L)->masks
830 #define LOOP_VINFO_LENS(L) (L)->lens
831 #define LOOP_VINFO_MASK_SKIP_NITERS(L) (L)->mask_skip_niters
832 #define LOOP_VINFO_RGROUP_COMPARE_TYPE(L) (L)->rgroup_compare_type
833 #define LOOP_VINFO_RGROUP_IV_TYPE(L) (L)->rgroup_iv_type
834 #define LOOP_VINFO_PTR_MASK(L) (L)->ptr_mask
835 #define LOOP_VINFO_N_STMTS(L) (L)->shared->n_stmts
836 #define LOOP_VINFO_LOOP_NEST(L) (L)->shared->loop_nest
837 #define LOOP_VINFO_DATAREFS(L) (L)->shared->datarefs
838 #define LOOP_VINFO_DDRS(L) (L)->shared->ddrs
839 #define LOOP_VINFO_INT_NITERS(L) (TREE_INT_CST_LOW ((L)->num_iters))
840 #define LOOP_VINFO_PEELING_FOR_ALIGNMENT(L) (L)->peeling_for_alignment
841 #define LOOP_VINFO_UNALIGNED_DR(L) (L)->unaligned_dr
842 #define LOOP_VINFO_MAY_MISALIGN_STMTS(L) (L)->may_misalign_stmts
843 #define LOOP_VINFO_MAY_ALIAS_DDRS(L) (L)->may_alias_ddrs
844 #define LOOP_VINFO_COMP_ALIAS_DDRS(L) (L)->comp_alias_ddrs
845 #define LOOP_VINFO_CHECK_UNEQUAL_ADDRS(L) (L)->check_unequal_addrs
846 #define LOOP_VINFO_CHECK_NONZERO(L) (L)->check_nonzero
847 #define LOOP_VINFO_LOWER_BOUNDS(L) (L)->lower_bounds
848 #define LOOP_VINFO_GROUPED_STORES(L) (L)->grouped_stores
849 #define LOOP_VINFO_SLP_INSTANCES(L) (L)->slp_instances
850 #define LOOP_VINFO_SLP_UNROLLING_FACTOR(L) (L)->slp_unrolling_factor
851 #define LOOP_VINFO_REDUCTIONS(L) (L)->reductions
852 #define LOOP_VINFO_REDUCTION_CHAINS(L) (L)->reduction_chains
853 #define LOOP_VINFO_PEELING_FOR_GAPS(L) (L)->peeling_for_gaps
854 #define LOOP_VINFO_PEELING_FOR_NITER(L) (L)->peeling_for_niter
855 #define LOOP_VINFO_NO_DATA_DEPENDENCIES(L) (L)->no_data_dependencies
856 #define LOOP_VINFO_SCALAR_LOOP(L) (L)->scalar_loop
857 #define LOOP_VINFO_SCALAR_LOOP_SCALING(L) (L)->scalar_loop_scaling
858 #define LOOP_VINFO_HAS_MASK_STORE(L) (L)->has_mask_store
859 #define LOOP_VINFO_SCALAR_ITERATION_COST(L) (L)->scalar_cost_vec
860 #define LOOP_VINFO_ORIG_LOOP_INFO(L) (L)->orig_loop_info
861 #define LOOP_VINFO_SIMD_IF_COND(L) (L)->simd_if_cond
862 #define LOOP_VINFO_INNER_LOOP_COST_FACTOR(L) (L)->inner_loop_cost_factor
864 #define LOOP_VINFO_FULLY_MASKED_P(L) \
865 (LOOP_VINFO_USING_PARTIAL_VECTORS_P (L) \
866 && !LOOP_VINFO_MASKS (L).is_empty ())
868 #define LOOP_VINFO_FULLY_WITH_LENGTH_P(L) \
869 (LOOP_VINFO_USING_PARTIAL_VECTORS_P (L) \
870 && !LOOP_VINFO_LENS (L).is_empty ())
872 #define LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT(L) \
873 ((L)->may_misalign_stmts.length () > 0)
874 #define LOOP_REQUIRES_VERSIONING_FOR_ALIAS(L) \
875 ((L)->comp_alias_ddrs.length () > 0 \
876 || (L)->check_unequal_addrs.length () > 0 \
877 || (L)->lower_bounds.length () > 0)
878 #define LOOP_REQUIRES_VERSIONING_FOR_NITERS(L) \
879 (LOOP_VINFO_NITERS_ASSUMPTIONS (L))
880 #define LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND(L) \
881 (LOOP_VINFO_SIMD_IF_COND (L))
882 #define LOOP_REQUIRES_VERSIONING(L) \
883 (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (L) \
884 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (L) \
885 || LOOP_REQUIRES_VERSIONING_FOR_NITERS (L) \
886 || LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND (L))
888 #define LOOP_VINFO_NITERS_KNOWN_P(L) \
889 (tree_fits_shwi_p ((L)->num_iters) && tree_to_shwi ((L)->num_iters) > 0)
891 #define LOOP_VINFO_EPILOGUE_P(L) \
892 (LOOP_VINFO_ORIG_LOOP_INFO (L) != NULL)
894 #define LOOP_VINFO_ORIG_MAX_VECT_FACTOR(L) \
895 (LOOP_VINFO_MAX_VECT_FACTOR (LOOP_VINFO_ORIG_LOOP_INFO (L)))
897 /* Wrapper for loop_vec_info, for tracking success/failure, where a non-NULL
898 value signifies success, and a NULL value signifies failure, supporting
899 propagating an opt_problem * describing the failure back up the call
900 stack. */
901 typedef opt_pointer_wrapper <loop_vec_info> opt_loop_vec_info;
903 static inline loop_vec_info
904 loop_vec_info_for_loop (class loop *loop)
906 return (loop_vec_info) loop->aux;
909 struct slp_root
911 slp_root (slp_instance_kind kind_, vec<stmt_vec_info> stmts_,
912 vec<stmt_vec_info> roots_)
913 : kind(kind_), stmts(stmts_), roots(roots_) {}
914 slp_instance_kind kind;
915 vec<stmt_vec_info> stmts;
916 vec<stmt_vec_info> roots;
919 typedef class _bb_vec_info : public vec_info
921 public:
922 _bb_vec_info (vec<basic_block> bbs, vec_info_shared *);
923 ~_bb_vec_info ();
925 /* The region we are operating on. bbs[0] is the entry, excluding
926 its PHI nodes. In the future we might want to track an explicit
927 entry edge to cover bbs[0] PHI nodes and have a region entry
928 insert location. */
929 vec<basic_block> bbs;
931 vec<slp_root> roots;
932 } *bb_vec_info;
934 #define BB_VINFO_BB(B) (B)->bb
935 #define BB_VINFO_GROUPED_STORES(B) (B)->grouped_stores
936 #define BB_VINFO_SLP_INSTANCES(B) (B)->slp_instances
937 #define BB_VINFO_DATAREFS(B) (B)->shared->datarefs
938 #define BB_VINFO_DDRS(B) (B)->shared->ddrs
940 /*-----------------------------------------------------------------*/
941 /* Info on vectorized defs. */
942 /*-----------------------------------------------------------------*/
943 enum stmt_vec_info_type {
944 undef_vec_info_type = 0,
945 load_vec_info_type,
946 store_vec_info_type,
947 shift_vec_info_type,
948 op_vec_info_type,
949 call_vec_info_type,
950 call_simd_clone_vec_info_type,
951 assignment_vec_info_type,
952 condition_vec_info_type,
953 comparison_vec_info_type,
954 reduc_vec_info_type,
955 induc_vec_info_type,
956 type_promotion_vec_info_type,
957 type_demotion_vec_info_type,
958 type_conversion_vec_info_type,
959 cycle_phi_info_type,
960 lc_phi_info_type,
961 phi_info_type,
962 loop_exit_ctrl_vec_info_type
965 /* Indicates whether/how a variable is used in the scope of loop/basic
966 block. */
967 enum vect_relevant {
968 vect_unused_in_scope = 0,
970 /* The def is only used outside the loop. */
971 vect_used_only_live,
972 /* The def is in the inner loop, and the use is in the outer loop, and the
973 use is a reduction stmt. */
974 vect_used_in_outer_by_reduction,
975 /* The def is in the inner loop, and the use is in the outer loop (and is
976 not part of reduction). */
977 vect_used_in_outer,
979 /* defs that feed computations that end up (only) in a reduction. These
980 defs may be used by non-reduction stmts, but eventually, any
981 computations/values that are affected by these defs are used to compute
982 a reduction (i.e. don't get stored to memory, for example). We use this
983 to identify computations that we can change the order in which they are
984 computed. */
985 vect_used_by_reduction,
987 vect_used_in_scope
990 /* The type of vectorization that can be applied to the stmt: regular loop-based
991 vectorization; pure SLP - the stmt is a part of SLP instances and does not
992 have uses outside SLP instances; or hybrid SLP and loop-based - the stmt is
993 a part of SLP instance and also must be loop-based vectorized, since it has
994 uses outside SLP sequences.
996 In the loop context the meanings of pure and hybrid SLP are slightly
997 different. By saying that pure SLP is applied to the loop, we mean that we
998 exploit only intra-iteration parallelism in the loop; i.e., the loop can be
999 vectorized without doing any conceptual unrolling, cause we don't pack
1000 together stmts from different iterations, only within a single iteration.
1001 Loop hybrid SLP means that we exploit both intra-iteration and
1002 inter-iteration parallelism (e.g., number of elements in the vector is 4
1003 and the slp-group-size is 2, in which case we don't have enough parallelism
1004 within an iteration, so we obtain the rest of the parallelism from subsequent
1005 iterations by unrolling the loop by 2). */
1006 enum slp_vect_type {
1007 loop_vect = 0,
1008 pure_slp,
1009 hybrid
1012 /* Says whether a statement is a load, a store of a vectorized statement
1013 result, or a store of an invariant value. */
1014 enum vec_load_store_type {
1015 VLS_LOAD,
1016 VLS_STORE,
1017 VLS_STORE_INVARIANT
1020 /* Describes how we're going to vectorize an individual load or store,
1021 or a group of loads or stores. */
1022 enum vect_memory_access_type {
1023 /* An access to an invariant address. This is used only for loads. */
1024 VMAT_INVARIANT,
1026 /* A simple contiguous access. */
1027 VMAT_CONTIGUOUS,
1029 /* A contiguous access that goes down in memory rather than up,
1030 with no additional permutation. This is used only for stores
1031 of invariants. */
1032 VMAT_CONTIGUOUS_DOWN,
1034 /* A simple contiguous access in which the elements need to be permuted
1035 after loading or before storing. Only used for loop vectorization;
1036 SLP uses separate permutes. */
1037 VMAT_CONTIGUOUS_PERMUTE,
1039 /* A simple contiguous access in which the elements need to be reversed
1040 after loading or before storing. */
1041 VMAT_CONTIGUOUS_REVERSE,
1043 /* An access that uses IFN_LOAD_LANES or IFN_STORE_LANES. */
1044 VMAT_LOAD_STORE_LANES,
1046 /* An access in which each scalar element is loaded or stored
1047 individually. */
1048 VMAT_ELEMENTWISE,
1050 /* A hybrid of VMAT_CONTIGUOUS and VMAT_ELEMENTWISE, used for grouped
1051 SLP accesses. Each unrolled iteration uses a contiguous load
1052 or store for the whole group, but the groups from separate iterations
1053 are combined in the same way as for VMAT_ELEMENTWISE. */
1054 VMAT_STRIDED_SLP,
1056 /* The access uses gather loads or scatter stores. */
1057 VMAT_GATHER_SCATTER
1060 class dr_vec_info {
1061 public:
1062 /* The data reference itself. */
1063 data_reference *dr;
1064 /* The statement that contains the data reference. */
1065 stmt_vec_info stmt;
1066 /* The analysis group this DR belongs to when doing BB vectorization.
1067 DRs of the same group belong to the same conditional execution context. */
1068 unsigned group;
1069 /* The misalignment in bytes of the reference, or -1 if not known. */
1070 int misalignment;
1071 /* The byte alignment that we'd ideally like the reference to have,
1072 and the value that misalignment is measured against. */
1073 poly_uint64 target_alignment;
1074 /* If true the alignment of base_decl needs to be increased. */
1075 bool base_misaligned;
1076 tree base_decl;
1078 /* Stores current vectorized loop's offset. To be added to the DR's
1079 offset to calculate current offset of data reference. */
1080 tree offset;
1083 typedef struct data_reference *dr_p;
1085 class _stmt_vec_info {
1086 public:
1088 enum stmt_vec_info_type type;
1090 /* Indicates whether this stmts is part of a computation whose result is
1091 used outside the loop. */
1092 bool live;
1094 /* Stmt is part of some pattern (computation idiom) */
1095 bool in_pattern_p;
1097 /* True if the statement was created during pattern recognition as
1098 part of the replacement for RELATED_STMT. This implies that the
1099 statement isn't part of any basic block, although for convenience
1100 its gimple_bb is the same as for RELATED_STMT. */
1101 bool pattern_stmt_p;
1103 /* Is this statement vectorizable or should it be skipped in (partial)
1104 vectorization. */
1105 bool vectorizable;
1107 /* The stmt to which this info struct refers to. */
1108 gimple *stmt;
1110 /* The vector type to be used for the LHS of this statement. */
1111 tree vectype;
1113 /* The vectorized stmts. */
1114 vec<gimple *> vec_stmts;
1116 /* The following is relevant only for stmts that contain a non-scalar
1117 data-ref (array/pointer/struct access). A GIMPLE stmt is expected to have
1118 at most one such data-ref. */
1120 dr_vec_info dr_aux;
1122 /* Information about the data-ref relative to this loop
1123 nest (the loop that is being considered for vectorization). */
1124 innermost_loop_behavior dr_wrt_vec_loop;
1126 /* For loop PHI nodes, the base and evolution part of it. This makes sure
1127 this information is still available in vect_update_ivs_after_vectorizer
1128 where we may not be able to re-analyze the PHI nodes evolution as
1129 peeling for the prologue loop can make it unanalyzable. The evolution
1130 part is still correct after peeling, but the base may have changed from
1131 the version here. */
1132 tree loop_phi_evolution_base_unchanged;
1133 tree loop_phi_evolution_part;
1135 /* Used for various bookkeeping purposes, generally holding a pointer to
1136 some other stmt S that is in some way "related" to this stmt.
1137 Current use of this field is:
1138 If this stmt is part of a pattern (i.e. the field 'in_pattern_p' is
1139 true): S is the "pattern stmt" that represents (and replaces) the
1140 sequence of stmts that constitutes the pattern. Similarly, the
1141 related_stmt of the "pattern stmt" points back to this stmt (which is
1142 the last stmt in the original sequence of stmts that constitutes the
1143 pattern). */
1144 stmt_vec_info related_stmt;
1146 /* Used to keep a sequence of def stmts of a pattern stmt if such exists.
1147 The sequence is attached to the original statement rather than the
1148 pattern statement. */
1149 gimple_seq pattern_def_seq;
1151 /* Selected SIMD clone's function info. First vector element
1152 is SIMD clone's function decl, followed by a pair of trees (base + step)
1153 for linear arguments (pair of NULLs for other arguments). */
1154 vec<tree> simd_clone_info;
1156 /* Classify the def of this stmt. */
1157 enum vect_def_type def_type;
1159 /* Whether the stmt is SLPed, loop-based vectorized, or both. */
1160 enum slp_vect_type slp_type;
1162 /* Interleaving and reduction chains info. */
1163 /* First element in the group. */
1164 stmt_vec_info first_element;
1165 /* Pointer to the next element in the group. */
1166 stmt_vec_info next_element;
1167 /* The size of the group. */
1168 unsigned int size;
1169 /* For stores, number of stores from this group seen. We vectorize the last
1170 one. */
1171 unsigned int store_count;
1172 /* For loads only, the gap from the previous load. For consecutive loads, GAP
1173 is 1. */
1174 unsigned int gap;
1176 /* The minimum negative dependence distance this stmt participates in
1177 or zero if none. */
1178 unsigned int min_neg_dist;
1180 /* Not all stmts in the loop need to be vectorized. e.g, the increment
1181 of the loop induction variable and computation of array indexes. relevant
1182 indicates whether the stmt needs to be vectorized. */
1183 enum vect_relevant relevant;
1185 /* For loads if this is a gather, for stores if this is a scatter. */
1186 bool gather_scatter_p;
1188 /* True if this is an access with loop-invariant stride. */
1189 bool strided_p;
1191 /* For both loads and stores. */
1192 unsigned simd_lane_access_p : 3;
1194 /* Classifies how the load or store is going to be implemented
1195 for loop vectorization. */
1196 vect_memory_access_type memory_access_type;
1198 /* For INTEGER_INDUC_COND_REDUCTION, the initial value to be used. */
1199 tree induc_cond_initial_val;
1201 /* If not NULL the value to be added to compute final reduction value. */
1202 tree reduc_epilogue_adjustment;
1204 /* On a reduction PHI the reduction type as detected by
1205 vect_is_simple_reduction and vectorizable_reduction. */
1206 enum vect_reduction_type reduc_type;
1208 /* The original reduction code, to be used in the epilogue. */
1209 code_helper reduc_code;
1210 /* An internal function we should use in the epilogue. */
1211 internal_fn reduc_fn;
1213 /* On a stmt participating in the reduction the index of the operand
1214 on the reduction SSA cycle. */
1215 int reduc_idx;
1217 /* On a reduction PHI the def returned by vect_force_simple_reduction.
1218 On the def returned by vect_force_simple_reduction the
1219 corresponding PHI. */
1220 stmt_vec_info reduc_def;
1222 /* The vector input type relevant for reduction vectorization. */
1223 tree reduc_vectype_in;
1225 /* The vector type for performing the actual reduction. */
1226 tree reduc_vectype;
1228 /* If IS_REDUC_INFO is true and if the vector code is performing
1229 N scalar reductions in parallel, this variable gives the initial
1230 scalar values of those N reductions. */
1231 vec<tree> reduc_initial_values;
1233 /* If IS_REDUC_INFO is true and if the vector code is performing
1234 N scalar reductions in parallel, this variable gives the vectorized code's
1235 final (scalar) result for each of those N reductions. In other words,
1236 REDUC_SCALAR_RESULTS[I] replaces the original scalar code's loop-closed
1237 SSA PHI for reduction number I. */
1238 vec<tree> reduc_scalar_results;
1240 /* Only meaningful if IS_REDUC_INFO. If non-null, the reduction is
1241 being performed by an epilogue loop and we have decided to reuse
1242 this accumulator from the main loop. */
1243 vect_reusable_accumulator *reused_accumulator;
1245 /* Whether we force a single cycle PHI during reduction vectorization. */
1246 bool force_single_cycle;
1248 /* Whether on this stmt reduction meta is recorded. */
1249 bool is_reduc_info;
1251 /* If nonzero, the lhs of the statement could be truncated to this
1252 many bits without affecting any users of the result. */
1253 unsigned int min_output_precision;
1255 /* If nonzero, all non-boolean input operands have the same precision,
1256 and they could each be truncated to this many bits without changing
1257 the result. */
1258 unsigned int min_input_precision;
1260 /* If OPERATION_BITS is nonzero, the statement could be performed on
1261 an integer with the sign and number of bits given by OPERATION_SIGN
1262 and OPERATION_BITS without changing the result. */
1263 unsigned int operation_precision;
1264 signop operation_sign;
1266 /* If the statement produces a boolean result, this value describes
1267 how we should choose the associated vector type. The possible
1268 values are:
1270 - an integer precision N if we should use the vector mask type
1271 associated with N-bit integers. This is only used if all relevant
1272 input booleans also want the vector mask type for N-bit integers,
1273 or if we can convert them into that form by pattern-matching.
1275 - ~0U if we considered choosing a vector mask type but decided
1276 to treat the boolean as a normal integer type instead.
1278 - 0 otherwise. This means either that the operation isn't one that
1279 could have a vector mask type (and so should have a normal vector
1280 type instead) or that we simply haven't made a choice either way. */
1281 unsigned int mask_precision;
1283 /* True if this is only suitable for SLP vectorization. */
1284 bool slp_vect_only_p;
1286 /* True if this is a pattern that can only be handled by SLP
1287 vectorization. */
1288 bool slp_vect_pattern_only_p;
1291 /* Information about a gather/scatter call. */
1292 struct gather_scatter_info {
1293 /* The internal function to use for the gather/scatter operation,
1294 or IFN_LAST if a built-in function should be used instead. */
1295 internal_fn ifn;
1297 /* The FUNCTION_DECL for the built-in gather/scatter function,
1298 or null if an internal function should be used instead. */
1299 tree decl;
1301 /* The loop-invariant base value. */
1302 tree base;
1304 /* The original scalar offset, which is a non-loop-invariant SSA_NAME. */
1305 tree offset;
1307 /* Each offset element should be multiplied by this amount before
1308 being added to the base. */
1309 int scale;
1311 /* The definition type for the vectorized offset. */
1312 enum vect_def_type offset_dt;
1314 /* The type of the vectorized offset. */
1315 tree offset_vectype;
1317 /* The type of the scalar elements after loading or before storing. */
1318 tree element_type;
1320 /* The type of the scalar elements being loaded or stored. */
1321 tree memory_type;
1324 /* Access Functions. */
1325 #define STMT_VINFO_TYPE(S) (S)->type
1326 #define STMT_VINFO_STMT(S) (S)->stmt
1327 #define STMT_VINFO_RELEVANT(S) (S)->relevant
1328 #define STMT_VINFO_LIVE_P(S) (S)->live
1329 #define STMT_VINFO_VECTYPE(S) (S)->vectype
1330 #define STMT_VINFO_VEC_STMTS(S) (S)->vec_stmts
1331 #define STMT_VINFO_VECTORIZABLE(S) (S)->vectorizable
1332 #define STMT_VINFO_DATA_REF(S) ((S)->dr_aux.dr + 0)
1333 #define STMT_VINFO_GATHER_SCATTER_P(S) (S)->gather_scatter_p
1334 #define STMT_VINFO_STRIDED_P(S) (S)->strided_p
1335 #define STMT_VINFO_MEMORY_ACCESS_TYPE(S) (S)->memory_access_type
1336 #define STMT_VINFO_SIMD_LANE_ACCESS_P(S) (S)->simd_lane_access_p
1337 #define STMT_VINFO_VEC_INDUC_COND_INITIAL_VAL(S) (S)->induc_cond_initial_val
1338 #define STMT_VINFO_REDUC_EPILOGUE_ADJUSTMENT(S) (S)->reduc_epilogue_adjustment
1339 #define STMT_VINFO_REDUC_IDX(S) (S)->reduc_idx
1340 #define STMT_VINFO_FORCE_SINGLE_CYCLE(S) (S)->force_single_cycle
1342 #define STMT_VINFO_DR_WRT_VEC_LOOP(S) (S)->dr_wrt_vec_loop
1343 #define STMT_VINFO_DR_BASE_ADDRESS(S) (S)->dr_wrt_vec_loop.base_address
1344 #define STMT_VINFO_DR_INIT(S) (S)->dr_wrt_vec_loop.init
1345 #define STMT_VINFO_DR_OFFSET(S) (S)->dr_wrt_vec_loop.offset
1346 #define STMT_VINFO_DR_STEP(S) (S)->dr_wrt_vec_loop.step
1347 #define STMT_VINFO_DR_BASE_ALIGNMENT(S) (S)->dr_wrt_vec_loop.base_alignment
1348 #define STMT_VINFO_DR_BASE_MISALIGNMENT(S) \
1349 (S)->dr_wrt_vec_loop.base_misalignment
1350 #define STMT_VINFO_DR_OFFSET_ALIGNMENT(S) \
1351 (S)->dr_wrt_vec_loop.offset_alignment
1352 #define STMT_VINFO_DR_STEP_ALIGNMENT(S) \
1353 (S)->dr_wrt_vec_loop.step_alignment
1355 #define STMT_VINFO_DR_INFO(S) \
1356 (gcc_checking_assert ((S)->dr_aux.stmt == (S)), &(S)->dr_aux)
1358 #define STMT_VINFO_IN_PATTERN_P(S) (S)->in_pattern_p
1359 #define STMT_VINFO_RELATED_STMT(S) (S)->related_stmt
1360 #define STMT_VINFO_PATTERN_DEF_SEQ(S) (S)->pattern_def_seq
1361 #define STMT_VINFO_SIMD_CLONE_INFO(S) (S)->simd_clone_info
1362 #define STMT_VINFO_DEF_TYPE(S) (S)->def_type
1363 #define STMT_VINFO_GROUPED_ACCESS(S) \
1364 ((S)->dr_aux.dr && DR_GROUP_FIRST_ELEMENT(S))
1365 #define STMT_VINFO_LOOP_PHI_EVOLUTION_BASE_UNCHANGED(S) (S)->loop_phi_evolution_base_unchanged
1366 #define STMT_VINFO_LOOP_PHI_EVOLUTION_PART(S) (S)->loop_phi_evolution_part
1367 #define STMT_VINFO_MIN_NEG_DIST(S) (S)->min_neg_dist
1368 #define STMT_VINFO_REDUC_TYPE(S) (S)->reduc_type
1369 #define STMT_VINFO_REDUC_CODE(S) (S)->reduc_code
1370 #define STMT_VINFO_REDUC_FN(S) (S)->reduc_fn
1371 #define STMT_VINFO_REDUC_DEF(S) (S)->reduc_def
1372 #define STMT_VINFO_REDUC_VECTYPE(S) (S)->reduc_vectype
1373 #define STMT_VINFO_REDUC_VECTYPE_IN(S) (S)->reduc_vectype_in
1374 #define STMT_VINFO_SLP_VECT_ONLY(S) (S)->slp_vect_only_p
1375 #define STMT_VINFO_SLP_VECT_ONLY_PATTERN(S) (S)->slp_vect_pattern_only_p
1377 #define DR_GROUP_FIRST_ELEMENT(S) \
1378 (gcc_checking_assert ((S)->dr_aux.dr), (S)->first_element)
1379 #define DR_GROUP_NEXT_ELEMENT(S) \
1380 (gcc_checking_assert ((S)->dr_aux.dr), (S)->next_element)
1381 #define DR_GROUP_SIZE(S) \
1382 (gcc_checking_assert ((S)->dr_aux.dr), (S)->size)
1383 #define DR_GROUP_STORE_COUNT(S) \
1384 (gcc_checking_assert ((S)->dr_aux.dr), (S)->store_count)
1385 #define DR_GROUP_GAP(S) \
1386 (gcc_checking_assert ((S)->dr_aux.dr), (S)->gap)
1388 #define REDUC_GROUP_FIRST_ELEMENT(S) \
1389 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->first_element)
1390 #define REDUC_GROUP_NEXT_ELEMENT(S) \
1391 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->next_element)
1392 #define REDUC_GROUP_SIZE(S) \
1393 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->size)
1395 #define STMT_VINFO_RELEVANT_P(S) ((S)->relevant != vect_unused_in_scope)
1397 #define HYBRID_SLP_STMT(S) ((S)->slp_type == hybrid)
1398 #define PURE_SLP_STMT(S) ((S)->slp_type == pure_slp)
1399 #define STMT_SLP_TYPE(S) (S)->slp_type
1401 /* Contains the scalar or vector costs for a vec_info. */
1402 class vector_costs
1404 public:
1405 vector_costs (vec_info *, bool);
1406 virtual ~vector_costs () {}
1408 /* Update the costs in response to adding COUNT copies of a statement.
1410 - WHERE specifies whether the cost occurs in the loop prologue,
1411 the loop body, or the loop epilogue.
1412 - KIND is the kind of statement, which is always meaningful.
1413 - STMT_INFO, if nonnull, describes the statement that will be
1414 vectorized.
1415 - VECTYPE, if nonnull, is the vector type that the vectorized
1416 statement will operate on. Note that this should be used in
1417 preference to STMT_VINFO_VECTYPE (STMT_INFO) since the latter
1418 is not correct for SLP.
1419 - for unaligned_load and unaligned_store statements, MISALIGN is
1420 the byte misalignment of the load or store relative to the target's
1421 preferred alignment for VECTYPE, or DR_MISALIGNMENT_UNKNOWN
1422 if the misalignment is not known.
1424 Return the calculated cost as well as recording it. The return
1425 value is used for dumping purposes. */
1426 virtual unsigned int add_stmt_cost (int count, vect_cost_for_stmt kind,
1427 stmt_vec_info stmt_info, tree vectype,
1428 int misalign,
1429 vect_cost_model_location where);
1431 /* Finish calculating the cost of the code. The results can be
1432 read back using the functions below.
1434 If the costs describe vector code, SCALAR_COSTS gives the costs
1435 of the corresponding scalar code, otherwise it is null. */
1436 virtual void finish_cost (const vector_costs *scalar_costs);
1438 /* The costs in THIS and OTHER both describe ways of vectorizing
1439 a main loop. Return true if the costs described by THIS are
1440 cheaper than the costs described by OTHER. Return false if any
1441 of the following are true:
1443 - THIS and OTHER are of equal cost
1444 - OTHER is better than THIS
1445 - we can't be sure about the relative costs of THIS and OTHER. */
1446 virtual bool better_main_loop_than_p (const vector_costs *other) const;
1448 /* Likewise, but the costs in THIS and OTHER both describe ways of
1449 vectorizing an epilogue loop of MAIN_LOOP. */
1450 virtual bool better_epilogue_loop_than_p (const vector_costs *other,
1451 loop_vec_info main_loop) const;
1453 unsigned int prologue_cost () const;
1454 unsigned int body_cost () const;
1455 unsigned int epilogue_cost () const;
1456 unsigned int outside_cost () const;
1457 unsigned int total_cost () const;
1459 protected:
1460 unsigned int record_stmt_cost (stmt_vec_info, vect_cost_model_location,
1461 unsigned int);
1462 unsigned int adjust_cost_for_freq (stmt_vec_info, vect_cost_model_location,
1463 unsigned int);
1464 int compare_inside_loop_cost (const vector_costs *) const;
1465 int compare_outside_loop_cost (const vector_costs *) const;
1467 /* The region of code that we're considering vectorizing. */
1468 vec_info *m_vinfo;
1470 /* True if we're costing the scalar code, false if we're costing
1471 the vector code. */
1472 bool m_costing_for_scalar;
1474 /* The costs of the three regions, indexed by vect_cost_model_location. */
1475 unsigned int m_costs[3];
1477 /* True if finish_cost has been called. */
1478 bool m_finished;
1481 /* Create costs for VINFO. COSTING_FOR_SCALAR is true if the costs
1482 are for scalar code, false if they are for vector code. */
1484 inline
1485 vector_costs::vector_costs (vec_info *vinfo, bool costing_for_scalar)
1486 : m_vinfo (vinfo),
1487 m_costing_for_scalar (costing_for_scalar),
1488 m_costs (),
1489 m_finished (false)
1493 /* Return the cost of the prologue code (in abstract units). */
1495 inline unsigned int
1496 vector_costs::prologue_cost () const
1498 gcc_checking_assert (m_finished);
1499 return m_costs[vect_prologue];
1502 /* Return the cost of the body code (in abstract units). */
1504 inline unsigned int
1505 vector_costs::body_cost () const
1507 gcc_checking_assert (m_finished);
1508 return m_costs[vect_body];
1511 /* Return the cost of the epilogue code (in abstract units). */
1513 inline unsigned int
1514 vector_costs::epilogue_cost () const
1516 gcc_checking_assert (m_finished);
1517 return m_costs[vect_epilogue];
1520 /* Return the cost of the prologue and epilogue code (in abstract units). */
1522 inline unsigned int
1523 vector_costs::outside_cost () const
1525 return prologue_cost () + epilogue_cost ();
1528 /* Return the cost of the prologue, body and epilogue code
1529 (in abstract units). */
1531 inline unsigned int
1532 vector_costs::total_cost () const
1534 return body_cost () + outside_cost ();
1537 #define VECT_MAX_COST 1000
1539 /* The maximum number of intermediate steps required in multi-step type
1540 conversion. */
1541 #define MAX_INTERM_CVT_STEPS 3
1543 #define MAX_VECTORIZATION_FACTOR INT_MAX
1545 /* Nonzero if TYPE represents a (scalar) boolean type or type
1546 in the middle-end compatible with it (unsigned precision 1 integral
1547 types). Used to determine which types should be vectorized as
1548 VECTOR_BOOLEAN_TYPE_P. */
1550 #define VECT_SCALAR_BOOLEAN_TYPE_P(TYPE) \
1551 (TREE_CODE (TYPE) == BOOLEAN_TYPE \
1552 || ((TREE_CODE (TYPE) == INTEGER_TYPE \
1553 || TREE_CODE (TYPE) == ENUMERAL_TYPE) \
1554 && TYPE_PRECISION (TYPE) == 1 \
1555 && TYPE_UNSIGNED (TYPE)))
1557 static inline bool
1558 nested_in_vect_loop_p (class loop *loop, stmt_vec_info stmt_info)
1560 return (loop->inner
1561 && (loop->inner == (gimple_bb (stmt_info->stmt))->loop_father));
1564 /* PHI is either a scalar reduction phi or a scalar induction phi.
1565 Return the initial value of the variable on entry to the containing
1566 loop. */
1568 static inline tree
1569 vect_phi_initial_value (gphi *phi)
1571 basic_block bb = gimple_bb (phi);
1572 edge pe = loop_preheader_edge (bb->loop_father);
1573 gcc_assert (pe->dest == bb);
1574 return PHI_ARG_DEF_FROM_EDGE (phi, pe);
1577 /* Return true if STMT_INFO should produce a vector mask type rather than
1578 a normal nonmask type. */
1580 static inline bool
1581 vect_use_mask_type_p (stmt_vec_info stmt_info)
1583 return stmt_info->mask_precision && stmt_info->mask_precision != ~0U;
1586 /* Return TRUE if a statement represented by STMT_INFO is a part of a
1587 pattern. */
1589 static inline bool
1590 is_pattern_stmt_p (stmt_vec_info stmt_info)
1592 return stmt_info->pattern_stmt_p;
1595 /* If STMT_INFO is a pattern statement, return the statement that it
1596 replaces, otherwise return STMT_INFO itself. */
1598 inline stmt_vec_info
1599 vect_orig_stmt (stmt_vec_info stmt_info)
1601 if (is_pattern_stmt_p (stmt_info))
1602 return STMT_VINFO_RELATED_STMT (stmt_info);
1603 return stmt_info;
1606 /* Return the later statement between STMT1_INFO and STMT2_INFO. */
1608 static inline stmt_vec_info
1609 get_later_stmt (stmt_vec_info stmt1_info, stmt_vec_info stmt2_info)
1611 if (gimple_uid (vect_orig_stmt (stmt1_info)->stmt)
1612 > gimple_uid (vect_orig_stmt (stmt2_info)->stmt))
1613 return stmt1_info;
1614 else
1615 return stmt2_info;
1618 /* If STMT_INFO has been replaced by a pattern statement, return the
1619 replacement statement, otherwise return STMT_INFO itself. */
1621 inline stmt_vec_info
1622 vect_stmt_to_vectorize (stmt_vec_info stmt_info)
1624 if (STMT_VINFO_IN_PATTERN_P (stmt_info))
1625 return STMT_VINFO_RELATED_STMT (stmt_info);
1626 return stmt_info;
1629 /* Return true if BB is a loop header. */
1631 static inline bool
1632 is_loop_header_bb_p (basic_block bb)
1634 if (bb == (bb->loop_father)->header)
1635 return true;
1636 gcc_checking_assert (EDGE_COUNT (bb->preds) == 1);
1637 return false;
1640 /* Return pow2 (X). */
1642 static inline int
1643 vect_pow2 (int x)
1645 int i, res = 1;
1647 for (i = 0; i < x; i++)
1648 res *= 2;
1650 return res;
1653 /* Alias targetm.vectorize.builtin_vectorization_cost. */
1655 static inline int
1656 builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost,
1657 tree vectype, int misalign)
1659 return targetm.vectorize.builtin_vectorization_cost (type_of_cost,
1660 vectype, misalign);
1663 /* Get cost by calling cost target builtin. */
1665 static inline
1666 int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost)
1668 return builtin_vectorization_cost (type_of_cost, NULL, 0);
1671 /* Alias targetm.vectorize.init_cost. */
1673 static inline vector_costs *
1674 init_cost (vec_info *vinfo, bool costing_for_scalar)
1676 return targetm.vectorize.create_costs (vinfo, costing_for_scalar);
1679 extern void dump_stmt_cost (FILE *, int, enum vect_cost_for_stmt,
1680 stmt_vec_info, tree, int, unsigned,
1681 enum vect_cost_model_location);
1683 /* Alias targetm.vectorize.add_stmt_cost. */
1685 static inline unsigned
1686 add_stmt_cost (vector_costs *costs, int count,
1687 enum vect_cost_for_stmt kind,
1688 stmt_vec_info stmt_info, tree vectype, int misalign,
1689 enum vect_cost_model_location where)
1691 unsigned cost = costs->add_stmt_cost (count, kind, stmt_info, vectype,
1692 misalign, where);
1693 if (dump_file && (dump_flags & TDF_DETAILS))
1694 dump_stmt_cost (dump_file, count, kind, stmt_info, vectype, misalign,
1695 cost, where);
1696 return cost;
1699 /* Alias targetm.vectorize.add_stmt_cost. */
1701 static inline unsigned
1702 add_stmt_cost (vector_costs *costs, stmt_info_for_cost *i)
1704 return add_stmt_cost (costs, i->count, i->kind, i->stmt_info,
1705 i->vectype, i->misalign, i->where);
1708 /* Alias targetm.vectorize.finish_cost. */
1710 static inline void
1711 finish_cost (vector_costs *costs, const vector_costs *scalar_costs,
1712 unsigned *prologue_cost, unsigned *body_cost,
1713 unsigned *epilogue_cost)
1715 costs->finish_cost (scalar_costs);
1716 *prologue_cost = costs->prologue_cost ();
1717 *body_cost = costs->body_cost ();
1718 *epilogue_cost = costs->epilogue_cost ();
1721 inline void
1722 add_stmt_costs (vector_costs *costs, stmt_vector_for_cost *cost_vec)
1724 stmt_info_for_cost *cost;
1725 unsigned i;
1726 FOR_EACH_VEC_ELT (*cost_vec, i, cost)
1727 add_stmt_cost (costs, cost->count, cost->kind, cost->stmt_info,
1728 cost->vectype, cost->misalign, cost->where);
1731 /*-----------------------------------------------------------------*/
1732 /* Info on data references alignment. */
1733 /*-----------------------------------------------------------------*/
1734 #define DR_MISALIGNMENT_UNKNOWN (-1)
1735 #define DR_MISALIGNMENT_UNINITIALIZED (-2)
1737 inline void
1738 set_dr_misalignment (dr_vec_info *dr_info, int val)
1740 dr_info->misalignment = val;
1743 extern int dr_misalignment (dr_vec_info *dr_info, tree vectype,
1744 poly_int64 offset = 0);
1746 #define SET_DR_MISALIGNMENT(DR, VAL) set_dr_misalignment (DR, VAL)
1748 /* Only defined once DR_MISALIGNMENT is defined. */
1749 static inline const poly_uint64
1750 dr_target_alignment (dr_vec_info *dr_info)
1752 if (STMT_VINFO_GROUPED_ACCESS (dr_info->stmt))
1753 dr_info = STMT_VINFO_DR_INFO (DR_GROUP_FIRST_ELEMENT (dr_info->stmt));
1754 return dr_info->target_alignment;
1756 #define DR_TARGET_ALIGNMENT(DR) dr_target_alignment (DR)
1758 static inline void
1759 set_dr_target_alignment (dr_vec_info *dr_info, poly_uint64 val)
1761 dr_info->target_alignment = val;
1763 #define SET_DR_TARGET_ALIGNMENT(DR, VAL) set_dr_target_alignment (DR, VAL)
1765 /* Return true if data access DR_INFO is aligned to the targets
1766 preferred alignment for VECTYPE (which may be less than a full vector). */
1768 static inline bool
1769 aligned_access_p (dr_vec_info *dr_info, tree vectype)
1771 return (dr_misalignment (dr_info, vectype) == 0);
1774 /* Return TRUE if the (mis-)alignment of the data access is known with
1775 respect to the targets preferred alignment for VECTYPE, and FALSE
1776 otherwise. */
1778 static inline bool
1779 known_alignment_for_access_p (dr_vec_info *dr_info, tree vectype)
1781 return (dr_misalignment (dr_info, vectype) != DR_MISALIGNMENT_UNKNOWN);
1784 /* Return the minimum alignment in bytes that the vectorized version
1785 of DR_INFO is guaranteed to have. */
1787 static inline unsigned int
1788 vect_known_alignment_in_bytes (dr_vec_info *dr_info, tree vectype)
1790 int misalignment = dr_misalignment (dr_info, vectype);
1791 if (misalignment == DR_MISALIGNMENT_UNKNOWN)
1792 return TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr_info->dr)));
1793 else if (misalignment == 0)
1794 return known_alignment (DR_TARGET_ALIGNMENT (dr_info));
1795 return misalignment & -misalignment;
1798 /* Return the behavior of DR_INFO with respect to the vectorization context
1799 (which for outer loop vectorization might not be the behavior recorded
1800 in DR_INFO itself). */
1802 static inline innermost_loop_behavior *
1803 vect_dr_behavior (vec_info *vinfo, dr_vec_info *dr_info)
1805 stmt_vec_info stmt_info = dr_info->stmt;
1806 loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo);
1807 if (loop_vinfo == NULL
1808 || !nested_in_vect_loop_p (LOOP_VINFO_LOOP (loop_vinfo), stmt_info))
1809 return &DR_INNERMOST (dr_info->dr);
1810 else
1811 return &STMT_VINFO_DR_WRT_VEC_LOOP (stmt_info);
1814 /* Return the offset calculated by adding the offset of this DR_INFO to the
1815 corresponding data_reference's offset. If CHECK_OUTER then use
1816 vect_dr_behavior to select the appropriate data_reference to use. */
1818 inline tree
1819 get_dr_vinfo_offset (vec_info *vinfo,
1820 dr_vec_info *dr_info, bool check_outer = false)
1822 innermost_loop_behavior *base;
1823 if (check_outer)
1824 base = vect_dr_behavior (vinfo, dr_info);
1825 else
1826 base = &dr_info->dr->innermost;
1828 tree offset = base->offset;
1830 if (!dr_info->offset)
1831 return offset;
1833 offset = fold_convert (sizetype, offset);
1834 return fold_build2 (PLUS_EXPR, TREE_TYPE (dr_info->offset), offset,
1835 dr_info->offset);
1839 /* Return the vect cost model for LOOP. */
1840 static inline enum vect_cost_model
1841 loop_cost_model (loop_p loop)
1843 if (loop != NULL
1844 && loop->force_vectorize
1845 && flag_simd_cost_model != VECT_COST_MODEL_DEFAULT)
1846 return flag_simd_cost_model;
1847 return flag_vect_cost_model;
1850 /* Return true if the vect cost model is unlimited. */
1851 static inline bool
1852 unlimited_cost_model (loop_p loop)
1854 return loop_cost_model (loop) == VECT_COST_MODEL_UNLIMITED;
1857 /* Return true if the loop described by LOOP_VINFO is fully-masked and
1858 if the first iteration should use a partial mask in order to achieve
1859 alignment. */
1861 static inline bool
1862 vect_use_loop_mask_for_alignment_p (loop_vec_info loop_vinfo)
1864 return (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo)
1865 && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo));
1868 /* Return the number of vectors of type VECTYPE that are needed to get
1869 NUNITS elements. NUNITS should be based on the vectorization factor,
1870 so it is always a known multiple of the number of elements in VECTYPE. */
1872 static inline unsigned int
1873 vect_get_num_vectors (poly_uint64 nunits, tree vectype)
1875 return exact_div (nunits, TYPE_VECTOR_SUBPARTS (vectype)).to_constant ();
1878 /* Return the number of copies needed for loop vectorization when
1879 a statement operates on vectors of type VECTYPE. This is the
1880 vectorization factor divided by the number of elements in
1881 VECTYPE and is always known at compile time. */
1883 static inline unsigned int
1884 vect_get_num_copies (loop_vec_info loop_vinfo, tree vectype)
1886 return vect_get_num_vectors (LOOP_VINFO_VECT_FACTOR (loop_vinfo), vectype);
1889 /* Update maximum unit count *MAX_NUNITS so that it accounts for
1890 NUNITS. *MAX_NUNITS can be 1 if we haven't yet recorded anything. */
1892 static inline void
1893 vect_update_max_nunits (poly_uint64 *max_nunits, poly_uint64 nunits)
1895 /* All unit counts have the form vec_info::vector_size * X for some
1896 rational X, so two unit sizes must have a common multiple.
1897 Everything is a multiple of the initial value of 1. */
1898 *max_nunits = force_common_multiple (*max_nunits, nunits);
1901 /* Update maximum unit count *MAX_NUNITS so that it accounts for
1902 the number of units in vector type VECTYPE. *MAX_NUNITS can be 1
1903 if we haven't yet recorded any vector types. */
1905 static inline void
1906 vect_update_max_nunits (poly_uint64 *max_nunits, tree vectype)
1908 vect_update_max_nunits (max_nunits, TYPE_VECTOR_SUBPARTS (vectype));
1911 /* Return the vectorization factor that should be used for costing
1912 purposes while vectorizing the loop described by LOOP_VINFO.
1913 Pick a reasonable estimate if the vectorization factor isn't
1914 known at compile time. */
1916 static inline unsigned int
1917 vect_vf_for_cost (loop_vec_info loop_vinfo)
1919 return estimated_poly_value (LOOP_VINFO_VECT_FACTOR (loop_vinfo));
1922 /* Estimate the number of elements in VEC_TYPE for costing purposes.
1923 Pick a reasonable estimate if the exact number isn't known at
1924 compile time. */
1926 static inline unsigned int
1927 vect_nunits_for_cost (tree vec_type)
1929 return estimated_poly_value (TYPE_VECTOR_SUBPARTS (vec_type));
1932 /* Return the maximum possible vectorization factor for LOOP_VINFO. */
1934 static inline unsigned HOST_WIDE_INT
1935 vect_max_vf (loop_vec_info loop_vinfo)
1937 unsigned HOST_WIDE_INT vf;
1938 if (LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (&vf))
1939 return vf;
1940 return MAX_VECTORIZATION_FACTOR;
1943 /* Return the size of the value accessed by unvectorized data reference
1944 DR_INFO. This is only valid once STMT_VINFO_VECTYPE has been calculated
1945 for the associated gimple statement, since that guarantees that DR_INFO
1946 accesses either a scalar or a scalar equivalent. ("Scalar equivalent"
1947 here includes things like V1SI, which can be vectorized in the same way
1948 as a plain SI.) */
1950 inline unsigned int
1951 vect_get_scalar_dr_size (dr_vec_info *dr_info)
1953 return tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_info->dr))));
1956 /* Return true if LOOP_VINFO requires a runtime check for whether the
1957 vector loop is profitable. */
1959 inline bool
1960 vect_apply_runtime_profitability_check_p (loop_vec_info loop_vinfo)
1962 unsigned int th = LOOP_VINFO_COST_MODEL_THRESHOLD (loop_vinfo);
1963 return (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)
1964 && th >= vect_vf_for_cost (loop_vinfo));
1967 /* Source location + hotness information. */
1968 extern dump_user_location_t vect_location;
1970 /* A macro for calling:
1971 dump_begin_scope (MSG, vect_location);
1972 via an RAII object, thus printing "=== MSG ===\n" to the dumpfile etc,
1973 and then calling
1974 dump_end_scope ();
1975 once the object goes out of scope, thus capturing the nesting of
1976 the scopes.
1978 These scopes affect dump messages within them: dump messages at the
1979 top level implicitly default to MSG_PRIORITY_USER_FACING, whereas those
1980 in a nested scope implicitly default to MSG_PRIORITY_INTERNALS. */
1982 #define DUMP_VECT_SCOPE(MSG) \
1983 AUTO_DUMP_SCOPE (MSG, vect_location)
1985 /* A sentinel class for ensuring that the "vect_location" global gets
1986 reset at the end of a scope.
1988 The "vect_location" global is used during dumping and contains a
1989 location_t, which could contain references to a tree block via the
1990 ad-hoc data. This data is used for tracking inlining information,
1991 but it's not a GC root; it's simply assumed that such locations never
1992 get accessed if the blocks are optimized away.
1994 Hence we need to ensure that such locations are purged at the end
1995 of any operations using them (e.g. via this class). */
1997 class auto_purge_vect_location
1999 public:
2000 ~auto_purge_vect_location ();
2003 /*-----------------------------------------------------------------*/
2004 /* Function prototypes. */
2005 /*-----------------------------------------------------------------*/
2007 /* Simple loop peeling and versioning utilities for vectorizer's purposes -
2008 in tree-vect-loop-manip.c. */
2009 extern void vect_set_loop_condition (class loop *, loop_vec_info,
2010 tree, tree, tree, bool);
2011 extern bool slpeel_can_duplicate_loop_p (const class loop *, const_edge);
2012 class loop *slpeel_tree_duplicate_loop_to_edge_cfg (class loop *,
2013 class loop *, edge);
2014 class loop *vect_loop_versioning (loop_vec_info, gimple *);
2015 extern class loop *vect_do_peeling (loop_vec_info, tree, tree,
2016 tree *, tree *, tree *, int, bool, bool,
2017 tree *);
2018 extern tree vect_get_main_loop_result (loop_vec_info, tree, tree);
2019 extern void vect_prepare_for_masked_peels (loop_vec_info);
2020 extern dump_user_location_t find_loop_location (class loop *);
2021 extern bool vect_can_advance_ivs_p (loop_vec_info);
2022 extern void vect_update_inits_of_drs (loop_vec_info, tree, tree_code);
2024 /* In tree-vect-stmts.c. */
2025 extern tree get_related_vectype_for_scalar_type (machine_mode, tree,
2026 poly_uint64 = 0);
2027 extern tree get_vectype_for_scalar_type (vec_info *, tree, unsigned int = 0);
2028 extern tree get_vectype_for_scalar_type (vec_info *, tree, slp_tree);
2029 extern tree get_mask_type_for_scalar_type (vec_info *, tree, unsigned int = 0);
2030 extern tree get_same_sized_vectype (tree, tree);
2031 extern bool vect_chooses_same_modes_p (vec_info *, machine_mode);
2032 extern bool vect_get_loop_mask_type (loop_vec_info);
2033 extern bool vect_is_simple_use (tree, vec_info *, enum vect_def_type *,
2034 stmt_vec_info * = NULL, gimple ** = NULL);
2035 extern bool vect_is_simple_use (tree, vec_info *, enum vect_def_type *,
2036 tree *, stmt_vec_info * = NULL,
2037 gimple ** = NULL);
2038 extern bool vect_is_simple_use (vec_info *, stmt_vec_info, slp_tree,
2039 unsigned, tree *, slp_tree *,
2040 enum vect_def_type *,
2041 tree *, stmt_vec_info * = NULL);
2042 extern bool vect_maybe_update_slp_op_vectype (slp_tree, tree);
2043 extern bool supportable_widening_operation (vec_info *,
2044 enum tree_code, stmt_vec_info,
2045 tree, tree, enum tree_code *,
2046 enum tree_code *, int *,
2047 vec<tree> *);
2048 extern bool supportable_narrowing_operation (enum tree_code, tree, tree,
2049 enum tree_code *, int *,
2050 vec<tree> *);
2052 extern unsigned record_stmt_cost (stmt_vector_for_cost *, int,
2053 enum vect_cost_for_stmt, stmt_vec_info,
2054 tree, int, enum vect_cost_model_location);
2056 /* Overload of record_stmt_cost with VECTYPE derived from STMT_INFO. */
2058 static inline unsigned
2059 record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count,
2060 enum vect_cost_for_stmt kind, stmt_vec_info stmt_info,
2061 int misalign, enum vect_cost_model_location where)
2063 return record_stmt_cost (body_cost_vec, count, kind, stmt_info,
2064 STMT_VINFO_VECTYPE (stmt_info), misalign, where);
2067 extern void vect_finish_replace_stmt (vec_info *, stmt_vec_info, gimple *);
2068 extern void vect_finish_stmt_generation (vec_info *, stmt_vec_info, gimple *,
2069 gimple_stmt_iterator *);
2070 extern opt_result vect_mark_stmts_to_be_vectorized (loop_vec_info, bool *);
2071 extern tree vect_get_store_rhs (stmt_vec_info);
2072 void vect_get_vec_defs_for_operand (vec_info *vinfo, stmt_vec_info, unsigned,
2073 tree op, vec<tree> *, tree = NULL);
2074 void vect_get_vec_defs (vec_info *, stmt_vec_info, slp_tree, unsigned,
2075 tree, vec<tree> *,
2076 tree = NULL, vec<tree> * = NULL,
2077 tree = NULL, vec<tree> * = NULL,
2078 tree = NULL, vec<tree> * = NULL);
2079 void vect_get_vec_defs (vec_info *, stmt_vec_info, slp_tree, unsigned,
2080 tree, vec<tree> *, tree,
2081 tree = NULL, vec<tree> * = NULL, tree = NULL,
2082 tree = NULL, vec<tree> * = NULL, tree = NULL,
2083 tree = NULL, vec<tree> * = NULL, tree = NULL);
2084 extern tree vect_init_vector (vec_info *, stmt_vec_info, tree, tree,
2085 gimple_stmt_iterator *);
2086 extern tree vect_get_slp_vect_def (slp_tree, unsigned);
2087 extern bool vect_transform_stmt (vec_info *, stmt_vec_info,
2088 gimple_stmt_iterator *,
2089 slp_tree, slp_instance);
2090 extern void vect_remove_stores (vec_info *, stmt_vec_info);
2091 extern bool vect_nop_conversion_p (stmt_vec_info);
2092 extern opt_result vect_analyze_stmt (vec_info *, stmt_vec_info, bool *,
2093 slp_tree,
2094 slp_instance, stmt_vector_for_cost *);
2095 extern void vect_get_load_cost (vec_info *, stmt_vec_info, int,
2096 dr_alignment_support, int, bool,
2097 unsigned int *, unsigned int *,
2098 stmt_vector_for_cost *,
2099 stmt_vector_for_cost *, bool);
2100 extern void vect_get_store_cost (vec_info *, stmt_vec_info, int,
2101 dr_alignment_support, int,
2102 unsigned int *, stmt_vector_for_cost *);
2103 extern bool vect_supportable_shift (vec_info *, enum tree_code, tree);
2104 extern tree vect_gen_perm_mask_any (tree, const vec_perm_indices &);
2105 extern tree vect_gen_perm_mask_checked (tree, const vec_perm_indices &);
2106 extern void optimize_mask_stores (class loop*);
2107 extern tree vect_gen_while (gimple_seq *, tree, tree, tree,
2108 const char * = nullptr);
2109 extern tree vect_gen_while_not (gimple_seq *, tree, tree, tree);
2110 extern opt_result vect_get_vector_types_for_stmt (vec_info *,
2111 stmt_vec_info, tree *,
2112 tree *, unsigned int = 0);
2113 extern opt_tree vect_get_mask_type_for_stmt (stmt_vec_info, unsigned int = 0);
2115 /* In tree-vect-data-refs.c. */
2116 extern bool vect_can_force_dr_alignment_p (const_tree, poly_uint64);
2117 extern enum dr_alignment_support vect_supportable_dr_alignment
2118 (vec_info *, dr_vec_info *, tree, int);
2119 extern tree vect_get_smallest_scalar_type (stmt_vec_info, tree);
2120 extern opt_result vect_analyze_data_ref_dependences (loop_vec_info, unsigned int *);
2121 extern bool vect_slp_analyze_instance_dependence (vec_info *, slp_instance);
2122 extern opt_result vect_enhance_data_refs_alignment (loop_vec_info);
2123 extern opt_result vect_analyze_data_refs_alignment (loop_vec_info);
2124 extern bool vect_slp_analyze_instance_alignment (vec_info *, slp_instance);
2125 extern opt_result vect_analyze_data_ref_accesses (vec_info *, vec<int> *);
2126 extern opt_result vect_prune_runtime_alias_test_list (loop_vec_info);
2127 extern bool vect_gather_scatter_fn_p (vec_info *, bool, bool, tree, tree,
2128 tree, int, internal_fn *, tree *);
2129 extern bool vect_check_gather_scatter (stmt_vec_info, loop_vec_info,
2130 gather_scatter_info *);
2131 extern opt_result vect_find_stmt_data_reference (loop_p, gimple *,
2132 vec<data_reference_p> *,
2133 vec<int> *, int);
2134 extern opt_result vect_analyze_data_refs (vec_info *, poly_uint64 *, bool *);
2135 extern void vect_record_base_alignments (vec_info *);
2136 extern tree vect_create_data_ref_ptr (vec_info *,
2137 stmt_vec_info, tree, class loop *, tree,
2138 tree *, gimple_stmt_iterator *,
2139 gimple **, bool,
2140 tree = NULL_TREE);
2141 extern tree bump_vector_ptr (vec_info *, tree, gimple *, gimple_stmt_iterator *,
2142 stmt_vec_info, tree);
2143 extern void vect_copy_ref_info (tree, tree);
2144 extern tree vect_create_destination_var (tree, tree);
2145 extern bool vect_grouped_store_supported (tree, unsigned HOST_WIDE_INT);
2146 extern bool vect_store_lanes_supported (tree, unsigned HOST_WIDE_INT, bool);
2147 extern bool vect_grouped_load_supported (tree, bool, unsigned HOST_WIDE_INT);
2148 extern bool vect_load_lanes_supported (tree, unsigned HOST_WIDE_INT, bool);
2149 extern void vect_permute_store_chain (vec_info *, vec<tree> &,
2150 unsigned int, stmt_vec_info,
2151 gimple_stmt_iterator *, vec<tree> *);
2152 extern tree vect_setup_realignment (vec_info *,
2153 stmt_vec_info, gimple_stmt_iterator *,
2154 tree *, enum dr_alignment_support, tree,
2155 class loop **);
2156 extern void vect_transform_grouped_load (vec_info *, stmt_vec_info, vec<tree>,
2157 int, gimple_stmt_iterator *);
2158 extern void vect_record_grouped_load_vectors (vec_info *,
2159 stmt_vec_info, vec<tree>);
2160 extern tree vect_get_new_vect_var (tree, enum vect_var_kind, const char *);
2161 extern tree vect_get_new_ssa_name (tree, enum vect_var_kind,
2162 const char * = NULL);
2163 extern tree vect_create_addr_base_for_vector_ref (vec_info *,
2164 stmt_vec_info, gimple_seq *,
2165 tree);
2167 /* In tree-vect-loop.c. */
2168 extern tree neutral_op_for_reduction (tree, code_helper, tree);
2169 extern widest_int vect_iv_limit_for_partial_vectors (loop_vec_info loop_vinfo);
2170 bool vect_rgroup_iv_might_wrap_p (loop_vec_info, rgroup_controls *);
2171 /* Used in tree-vect-loop-manip.c */
2172 extern opt_result vect_determine_partial_vectors_and_peeling (loop_vec_info,
2173 bool);
2174 /* Used in gimple-loop-interchange.c and tree-parloops.c. */
2175 extern bool check_reduction_path (dump_user_location_t, loop_p, gphi *, tree,
2176 enum tree_code);
2177 extern bool needs_fold_left_reduction_p (tree, code_helper);
2178 /* Drive for loop analysis stage. */
2179 extern opt_loop_vec_info vect_analyze_loop (class loop *, vec_info_shared *);
2180 extern tree vect_build_loop_niters (loop_vec_info, bool * = NULL);
2181 extern void vect_gen_vector_loop_niters (loop_vec_info, tree, tree *,
2182 tree *, bool);
2183 extern tree vect_halve_mask_nunits (tree, machine_mode);
2184 extern tree vect_double_mask_nunits (tree, machine_mode);
2185 extern void vect_record_loop_mask (loop_vec_info, vec_loop_masks *,
2186 unsigned int, tree, tree);
2187 extern tree vect_get_loop_mask (gimple_stmt_iterator *, vec_loop_masks *,
2188 unsigned int, tree, unsigned int);
2189 extern void vect_record_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
2190 tree, unsigned int);
2191 extern tree vect_get_loop_len (loop_vec_info, vec_loop_lens *, unsigned int,
2192 unsigned int);
2193 extern gimple_seq vect_gen_len (tree, tree, tree, tree);
2194 extern stmt_vec_info info_for_reduction (vec_info *, stmt_vec_info);
2195 extern bool reduction_fn_for_scalar_code (code_helper, internal_fn *);
2197 /* Drive for loop transformation stage. */
2198 extern class loop *vect_transform_loop (loop_vec_info, gimple *);
2199 struct vect_loop_form_info
2201 tree number_of_iterations;
2202 tree number_of_iterationsm1;
2203 tree assumptions;
2204 gcond *loop_cond;
2205 gcond *inner_loop_cond;
2207 extern opt_result vect_analyze_loop_form (class loop *, vect_loop_form_info *);
2208 extern loop_vec_info vect_create_loop_vinfo (class loop *, vec_info_shared *,
2209 const vect_loop_form_info *,
2210 loop_vec_info = nullptr);
2211 extern bool vectorizable_live_operation (vec_info *,
2212 stmt_vec_info, gimple_stmt_iterator *,
2213 slp_tree, slp_instance, int,
2214 bool, stmt_vector_for_cost *);
2215 extern bool vectorizable_reduction (loop_vec_info, stmt_vec_info,
2216 slp_tree, slp_instance,
2217 stmt_vector_for_cost *);
2218 extern bool vectorizable_induction (loop_vec_info, stmt_vec_info,
2219 gimple **, slp_tree,
2220 stmt_vector_for_cost *);
2221 extern bool vect_transform_reduction (loop_vec_info, stmt_vec_info,
2222 gimple_stmt_iterator *,
2223 gimple **, slp_tree);
2224 extern bool vect_transform_cycle_phi (loop_vec_info, stmt_vec_info,
2225 gimple **,
2226 slp_tree, slp_instance);
2227 extern bool vectorizable_lc_phi (loop_vec_info, stmt_vec_info,
2228 gimple **, slp_tree);
2229 extern bool vectorizable_phi (vec_info *, stmt_vec_info, gimple **, slp_tree,
2230 stmt_vector_for_cost *);
2231 extern bool vect_emulated_vector_p (tree);
2232 extern bool vect_can_vectorize_without_simd_p (tree_code);
2233 extern bool vect_can_vectorize_without_simd_p (code_helper);
2234 extern int vect_get_known_peeling_cost (loop_vec_info, int, int *,
2235 stmt_vector_for_cost *,
2236 stmt_vector_for_cost *,
2237 stmt_vector_for_cost *);
2238 extern tree cse_and_gimplify_to_preheader (loop_vec_info, tree);
2240 /* In tree-vect-slp.c. */
2241 extern void vect_slp_init (void);
2242 extern void vect_slp_fini (void);
2243 extern void vect_free_slp_instance (slp_instance);
2244 extern bool vect_transform_slp_perm_load (vec_info *, slp_tree, const vec<tree> &,
2245 gimple_stmt_iterator *, poly_uint64,
2246 bool, unsigned *,
2247 unsigned * = nullptr, bool = false);
2248 extern bool vect_slp_analyze_operations (vec_info *);
2249 extern void vect_schedule_slp (vec_info *, const vec<slp_instance> &);
2250 extern opt_result vect_analyze_slp (vec_info *, unsigned);
2251 extern bool vect_make_slp_decision (loop_vec_info);
2252 extern void vect_detect_hybrid_slp (loop_vec_info);
2253 extern void vect_optimize_slp (vec_info *);
2254 extern void vect_gather_slp_loads (vec_info *);
2255 extern void vect_get_slp_defs (slp_tree, vec<tree> *);
2256 extern void vect_get_slp_defs (vec_info *, slp_tree, vec<vec<tree> > *,
2257 unsigned n = -1U);
2258 extern bool vect_slp_if_converted_bb (basic_block bb, loop_p orig_loop);
2259 extern bool vect_slp_function (function *);
2260 extern stmt_vec_info vect_find_last_scalar_stmt_in_slp (slp_tree);
2261 extern stmt_vec_info vect_find_first_scalar_stmt_in_slp (slp_tree);
2262 extern bool is_simple_and_all_uses_invariant (stmt_vec_info, loop_vec_info);
2263 extern bool can_duplicate_and_interleave_p (vec_info *, unsigned int, tree,
2264 unsigned int * = NULL,
2265 tree * = NULL, tree * = NULL);
2266 extern void duplicate_and_interleave (vec_info *, gimple_seq *, tree,
2267 const vec<tree> &, unsigned int, vec<tree> &);
2268 extern int vect_get_place_in_interleaving_chain (stmt_vec_info, stmt_vec_info);
2269 extern slp_tree vect_create_new_slp_node (unsigned, tree_code);
2270 extern void vect_free_slp_tree (slp_tree);
2272 /* In tree-vect-patterns.c. */
2273 extern void
2274 vect_mark_pattern_stmts (vec_info *, stmt_vec_info, gimple *, tree);
2276 /* Pattern recognition functions.
2277 Additional pattern recognition functions can (and will) be added
2278 in the future. */
2279 void vect_pattern_recog (vec_info *);
2281 /* In tree-vectorizer.c. */
2282 unsigned vectorize_loops (void);
2283 void vect_free_loop_info_assumptions (class loop *);
2284 gimple *vect_loop_vectorized_call (class loop *, gcond **cond = NULL);
2285 bool vect_stmt_dominates_stmt_p (gimple *, gimple *);
2287 /* SLP Pattern matcher types, tree-vect-slp-patterns.c. */
2289 /* Forward declaration of possible two operands operation that can be matched
2290 by the complex numbers pattern matchers. */
2291 enum _complex_operation : unsigned;
2293 /* All possible load permute values that could result from the partial data-flow
2294 analysis. */
2295 typedef enum _complex_perm_kinds {
2296 PERM_UNKNOWN,
2297 PERM_EVENODD,
2298 PERM_ODDEVEN,
2299 PERM_ODDODD,
2300 PERM_EVENEVEN,
2301 /* Can be combined with any other PERM values. */
2302 PERM_TOP
2303 } complex_perm_kinds_t;
2305 /* Cache from nodes to the load permutation they represent. */
2306 typedef hash_map <slp_tree, complex_perm_kinds_t>
2307 slp_tree_to_load_perm_map_t;
2309 /* Vector pattern matcher base class. All SLP pattern matchers must inherit
2310 from this type. */
2312 class vect_pattern
2314 protected:
2315 /* The number of arguments that the IFN requires. */
2316 unsigned m_num_args;
2318 /* The internal function that will be used when a pattern is created. */
2319 internal_fn m_ifn;
2321 /* The current node being inspected. */
2322 slp_tree *m_node;
2324 /* The list of operands to be the children for the node produced when the
2325 internal function is created. */
2326 vec<slp_tree> m_ops;
2328 /* Default constructor where NODE is the root of the tree to inspect. */
2329 vect_pattern (slp_tree *node, vec<slp_tree> *m_ops, internal_fn ifn)
2331 this->m_ifn = ifn;
2332 this->m_node = node;
2333 this->m_ops.create (0);
2334 if (m_ops)
2335 this->m_ops.safe_splice (*m_ops);
2338 public:
2340 /* Create a new instance of the pattern matcher class of the given type. */
2341 static vect_pattern* recognize (slp_tree_to_load_perm_map_t *, slp_tree *);
2343 /* Build the pattern from the data collected so far. */
2344 virtual void build (vec_info *) = 0;
2346 /* Default destructor. */
2347 virtual ~vect_pattern ()
2349 this->m_ops.release ();
2353 /* Function pointer to create a new pattern matcher from a generic type. */
2354 typedef vect_pattern* (*vect_pattern_decl_t) (slp_tree_to_load_perm_map_t *,
2355 slp_tree *);
2357 /* List of supported pattern matchers. */
2358 extern vect_pattern_decl_t slp_patterns[];
2360 /* Number of supported pattern matchers. */
2361 extern size_t num__slp_patterns;
2363 /* ----------------------------------------------------------------------
2364 Target support routines
2365 -----------------------------------------------------------------------
2366 The following routines are provided to simplify costing decisions in
2367 target code. Please add more as needed. */
2369 /* Return true if an operaton of kind KIND for STMT_INFO represents
2370 the extraction of an element from a vector in preparation for
2371 storing the element to memory. */
2372 inline bool
2373 vect_is_store_elt_extraction (vect_cost_for_stmt kind, stmt_vec_info stmt_info)
2375 return (kind == vec_to_scalar
2376 && STMT_VINFO_DATA_REF (stmt_info)
2377 && DR_IS_WRITE (STMT_VINFO_DATA_REF (stmt_info)));
2380 /* Return true if STMT_INFO represents part of a reduction. */
2381 inline bool
2382 vect_is_reduction (stmt_vec_info stmt_info)
2384 return STMT_VINFO_REDUC_IDX (stmt_info) >= 0;
2387 /* If STMT_INFO describes a reduction, return the vect_reduction_type
2388 of the reduction it describes, otherwise return -1. */
2389 inline int
2390 vect_reduc_type (vec_info *vinfo, stmt_vec_info stmt_info)
2392 if (loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo))
2393 if (STMT_VINFO_REDUC_DEF (stmt_info))
2395 stmt_vec_info reduc_info = info_for_reduction (loop_vinfo, stmt_info);
2396 return int (STMT_VINFO_REDUC_TYPE (reduc_info));
2398 return -1;
2401 /* If STMT_INFO is a COND_EXPR that includes an embedded comparison, return the
2402 scalar type of the values being compared. Return null otherwise. */
2403 inline tree
2404 vect_embedded_comparison_type (stmt_vec_info stmt_info)
2406 if (auto *assign = dyn_cast<gassign *> (stmt_info->stmt))
2407 if (gimple_assign_rhs_code (assign) == COND_EXPR)
2409 tree cond = gimple_assign_rhs1 (assign);
2410 if (COMPARISON_CLASS_P (cond))
2411 return TREE_TYPE (TREE_OPERAND (cond, 0));
2413 return NULL_TREE;
2416 /* If STMT_INFO is a comparison or contains an embedded comparison, return the
2417 scalar type of the values being compared. Return null otherwise. */
2418 inline tree
2419 vect_comparison_type (stmt_vec_info stmt_info)
2421 if (auto *assign = dyn_cast<gassign *> (stmt_info->stmt))
2422 if (TREE_CODE_CLASS (gimple_assign_rhs_code (assign)) == tcc_comparison)
2423 return TREE_TYPE (gimple_assign_rhs1 (assign));
2424 return vect_embedded_comparison_type (stmt_info);
2427 /* Return true if STMT_INFO extends the result of a load. */
2428 inline bool
2429 vect_is_extending_load (class vec_info *vinfo, stmt_vec_info stmt_info)
2431 /* Although this is quite large for an inline function, this part
2432 at least should be inline. */
2433 gassign *assign = dyn_cast <gassign *> (stmt_info->stmt);
2434 if (!assign || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (assign)))
2435 return false;
2437 tree rhs = gimple_assign_rhs1 (stmt_info->stmt);
2438 tree lhs_type = TREE_TYPE (gimple_assign_lhs (assign));
2439 tree rhs_type = TREE_TYPE (rhs);
2440 if (!INTEGRAL_TYPE_P (lhs_type)
2441 || !INTEGRAL_TYPE_P (rhs_type)
2442 || TYPE_PRECISION (lhs_type) <= TYPE_PRECISION (rhs_type))
2443 return false;
2445 stmt_vec_info def_stmt_info = vinfo->lookup_def (rhs);
2446 return (def_stmt_info
2447 && STMT_VINFO_DATA_REF (def_stmt_info)
2448 && DR_IS_READ (STMT_VINFO_DATA_REF (def_stmt_info)));
2451 /* Return true if STMT_INFO is an integer truncation. */
2452 inline bool
2453 vect_is_integer_truncation (stmt_vec_info stmt_info)
2455 gassign *assign = dyn_cast <gassign *> (stmt_info->stmt);
2456 if (!assign || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (assign)))
2457 return false;
2459 tree lhs_type = TREE_TYPE (gimple_assign_lhs (assign));
2460 tree rhs_type = TREE_TYPE (gimple_assign_rhs1 (assign));
2461 return (INTEGRAL_TYPE_P (lhs_type)
2462 && INTEGRAL_TYPE_P (rhs_type)
2463 && TYPE_PRECISION (lhs_type) < TYPE_PRECISION (rhs_type));
2466 #endif /* GCC_TREE_VECTORIZER_H */