1/* Vectorizer
2 Copyright (C) 2003-2017 Free Software Foundation, Inc.
3 Contributed by Dorit Naishlos <dorit@il.ibm.com>
4
5This file is part of GCC.
6
7GCC is free software; you can redistribute it and/or modify it under
8the terms of the GNU General Public License as published by the Free
9Software Foundation; either version 3, or (at your option) any later
10version.
11
12GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13WARRANTY; without even the implied warranty of MERCHANTABILITY or
14FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15for more details.
16
17You should have received a copy of the GNU General Public License
18along with GCC; see the file COPYING3. If not see
19<http://www.gnu.org/licenses/>. */
20
21#ifndef GCC_TREE_VECTORIZER_H
22#define GCC_TREE_VECTORIZER_H
23
24#include "tree-data-ref.h"
25#include "tree-hash-traits.h"
26#include "target.h"
27
28/* Used for naming of new temporaries. */
29enum vect_var_kind {
30 vect_simple_var,
31 vect_pointer_var,
32 vect_scalar_var,
33 vect_mask_var
34};
35
36/* Defines type of operation. */
37enum operation_type {
38 unary_op = 1,
39 binary_op,
40 ternary_op
41};
42
43/* Define type of available alignment support. */
44enum dr_alignment_support {
45 dr_unaligned_unsupported,
46 dr_unaligned_supported,
47 dr_explicit_realign,
48 dr_explicit_realign_optimized,
49 dr_aligned
50};
51
52/* Define type of def-use cross-iteration cycle. */
53enum vect_def_type {
54 vect_uninitialized_def = 0,
55 vect_constant_def = 1,
56 vect_external_def,
57 vect_internal_def,
58 vect_induction_def,
59 vect_reduction_def,
60 vect_double_reduction_def,
61 vect_nested_cycle,
62 vect_unknown_def_type
63};
64
65/* Define type of reduction. */
66enum vect_reduction_type {
67 TREE_CODE_REDUCTION,
68 COND_REDUCTION,
69 INTEGER_INDUC_COND_REDUCTION,
70 CONST_COND_REDUCTION
71};
72
73#define VECTORIZABLE_CYCLE_DEF(D) (((D) == vect_reduction_def) \
74 || ((D) == vect_double_reduction_def) \
75 || ((D) == vect_nested_cycle))
76
77/* Structure to encapsulate information about a group of like
78 instructions to be presented to the target cost model. */
79struct stmt_info_for_cost {
80 int count;
81 enum vect_cost_for_stmt kind;
82 gimple *stmt;
83 int misalign;
84};
85
86typedef vec<stmt_info_for_cost> stmt_vector_for_cost;
87
88/* Maps base addresses to an innermost_loop_behavior that gives the maximum
89 known alignment for that base. */
90typedef hash_map<tree_operand_hash,
91 innermost_loop_behavior *> vec_base_alignments;
92
93/************************************************************************
94 SLP
95 ************************************************************************/
96typedef struct _slp_tree *slp_tree;
97
98/* A computation tree of an SLP instance. Each node corresponds to a group of
99 stmts to be packed in a SIMD stmt. */
100struct _slp_tree {
101 /* Nodes that contain def-stmts of this node statements operands. */
102 vec<slp_tree> children;
103 /* A group of scalar stmts to be vectorized together. */
104 vec<gimple *> stmts;
105 /* Load permutation relative to the stores, NULL if there is no
106 permutation. */
107 vec<unsigned> load_permutation;
108 /* Vectorized stmt/s. */
109 vec<gimple *> vec_stmts;
110 /* Number of vector stmts that are created to replace the group of scalar
111 stmts. It is calculated during the transformation phase as the number of
112 scalar elements in one scalar iteration (GROUP_SIZE) multiplied by VF
113 divided by vector size. */
114 unsigned int vec_stmts_size;
115 /* Whether the scalar computations use two different operators. */
116 bool two_operators;
117 /* The DEF type of this node. */
118 enum vect_def_type def_type;
119};
120
121
122/* SLP instance is a sequence of stmts in a loop that can be packed into
123 SIMD stmts. */
124typedef struct _slp_instance {
125 /* The root of SLP tree. */
126 slp_tree root;
127
128 /* Size of groups of scalar stmts that will be replaced by SIMD stmt/s. */
129 unsigned int group_size;
130
131 /* The unrolling factor required to vectorized this SLP instance. */
132 unsigned int unrolling_factor;
133
134 /* The group of nodes that contain loads of this SLP instance. */
135 vec<slp_tree> loads;
136
137 /* The SLP node containing the reduction PHIs. */
138 slp_tree reduc_phis;
139} *slp_instance;
140
141
142/* Access Functions. */
143#define SLP_INSTANCE_TREE(S) (S)->root
144#define SLP_INSTANCE_GROUP_SIZE(S) (S)->group_size
145#define SLP_INSTANCE_UNROLLING_FACTOR(S) (S)->unrolling_factor
146#define SLP_INSTANCE_LOADS(S) (S)->loads
147
148#define SLP_TREE_CHILDREN(S) (S)->children
149#define SLP_TREE_SCALAR_STMTS(S) (S)->stmts
150#define SLP_TREE_VEC_STMTS(S) (S)->vec_stmts
151#define SLP_TREE_NUMBER_OF_VEC_STMTS(S) (S)->vec_stmts_size
152#define SLP_TREE_LOAD_PERMUTATION(S) (S)->load_permutation
153#define SLP_TREE_TWO_OPERATORS(S) (S)->two_operators
154#define SLP_TREE_DEF_TYPE(S) (S)->def_type
155
156
157
158/* Describes two objects whose addresses must be unequal for the vectorized
159 loop to be valid. */
160typedef std::pair<tree, tree> vec_object_pair;
161
162/* Vectorizer state common between loop and basic-block vectorization. */
163struct vec_info {
164 enum vec_kind { bb, loop };
165
166 vec_info (vec_kind, void *);
167 ~vec_info ();
168
169 /* The type of vectorization. */
170 vec_kind kind;
171
172 /* All SLP instances. */
173 auto_vec<slp_instance> slp_instances;
174
175 /* All data references. Freed by free_data_refs, so not an auto_vec. */
176 vec<data_reference_p> datarefs;
177
178 /* Maps base addresses to an innermost_loop_behavior that gives the maximum
179 known alignment for that base. */
180 vec_base_alignments base_alignments;
181
182 /* All data dependences. Freed by free_dependence_relations, so not
183 an auto_vec. */
184 vec<ddr_p> ddrs;
185
186 /* All interleaving chains of stores, represented by the first
187 stmt in the chain. */
188 auto_vec<gimple *> grouped_stores;
189
190 /* Cost data used by the target cost model. */
191 void *target_cost_data;
192};
193
194struct _loop_vec_info;
195struct _bb_vec_info;
196
197template<>
198template<>
199inline bool
200is_a_helper <_loop_vec_info *>::test (vec_info *i)
201{
202 return i->kind == vec_info::loop;
203}
204
205template<>
206template<>
207inline bool
208is_a_helper <_bb_vec_info *>::test (vec_info *i)
209{
210 return i->kind == vec_info::bb;
211}
212
213
214/*-----------------------------------------------------------------*/
215/* Info on vectorized loops. */
216/*-----------------------------------------------------------------*/
217typedef struct _loop_vec_info : public vec_info {
218 _loop_vec_info (struct loop *);
219 ~_loop_vec_info ();
220
221 /* The loop to which this info struct refers to. */
222 struct loop *loop;
223
224 /* The loop basic blocks. */
225 basic_block *bbs;
226
227 /* Number of latch executions. */
228 tree num_itersm1;
229 /* Number of iterations. */
230 tree num_iters;
231 /* Number of iterations of the original loop. */
232 tree num_iters_unchanged;
233 /* Condition under which this loop is analyzed and versioned. */
234 tree num_iters_assumptions;
235
236 /* Threshold of number of iterations below which vectorzation will not be
237 performed. It is calculated from MIN_PROFITABLE_ITERS and
238 PARAM_MIN_VECT_LOOP_BOUND. */
239 unsigned int th;
240
241 /* Unrolling factor */
242 int vectorization_factor;
243
244 /* Maximum runtime vectorization factor, or MAX_VECTORIZATION_FACTOR
245 if there is no particular limit. */
246 unsigned HOST_WIDE_INT max_vectorization_factor;
247
248 /* Unknown DRs according to which loop was peeled. */
249 struct data_reference *unaligned_dr;
250
251 /* peeling_for_alignment indicates whether peeling for alignment will take
252 place, and what the peeling factor should be:
253 peeling_for_alignment = X means:
254 If X=0: Peeling for alignment will not be applied.
255 If X>0: Peel first X iterations.
256 If X=-1: Generate a runtime test to calculate the number of iterations
257 to be peeled, using the dataref recorded in the field
258 unaligned_dr. */
259 int peeling_for_alignment;
260
261 /* The mask used to check the alignment of pointers or arrays. */
262 int ptr_mask;
263
264 /* The loop nest in which the data dependences are computed. */
265 auto_vec<loop_p> loop_nest;
266
267 /* Data Dependence Relations defining address ranges that are candidates
268 for a run-time aliasing check. */
269 auto_vec<ddr_p> may_alias_ddrs;
270
271 /* Data Dependence Relations defining address ranges together with segment
272 lengths from which the run-time aliasing check is built. */
273 auto_vec<dr_with_seg_len_pair_t> comp_alias_ddrs;
274
275 /* Check that the addresses of each pair of objects is unequal. */
276 auto_vec<vec_object_pair> check_unequal_addrs;
277
278 /* Statements in the loop that have data references that are candidates for a
279 runtime (loop versioning) misalignment check. */
280 auto_vec<gimple *> may_misalign_stmts;
281
282 /* Reduction cycles detected in the loop. Used in loop-aware SLP. */
283 auto_vec<gimple *> reductions;
284
285 /* All reduction chains in the loop, represented by the first
286 stmt in the chain. */
287 auto_vec<gimple *> reduction_chains;
288
289 /* Cost vector for a single scalar iteration. */
290 auto_vec<stmt_info_for_cost> scalar_cost_vec;
291
292 /* The unrolling factor needed to SLP the loop. In case of that pure SLP is
293 applied to the loop, i.e., no unrolling is needed, this is 1. */
294 unsigned slp_unrolling_factor;
295
296 /* Cost of a single scalar iteration. */
297 int single_scalar_iteration_cost;
298
299 /* Is the loop vectorizable? */
300 bool vectorizable;
301
302 /* When we have grouped data accesses with gaps, we may introduce invalid
303 memory accesses. We peel the last iteration of the loop to prevent
304 this. */
305 bool peeling_for_gaps;
306
307 /* When the number of iterations is not a multiple of the vector size
308 we need to peel off iterations at the end to form an epilogue loop. */
309 bool peeling_for_niter;
310
311 /* Reductions are canonicalized so that the last operand is the reduction
312 operand. If this places a constant into RHS1, this decanonicalizes
313 GIMPLE for other phases, so we must track when this has occurred and
314 fix it up. */
315 bool operands_swapped;
316
317 /* True if there are no loop carried data dependencies in the loop.
318 If loop->safelen <= 1, then this is always true, either the loop
319 didn't have any loop carried data dependencies, or the loop is being
320 vectorized guarded with some runtime alias checks, or couldn't
321 be vectorized at all, but then this field shouldn't be used.
322 For loop->safelen >= 2, the user has asserted that there are no
323 backward dependencies, but there still could be loop carried forward
324 dependencies in such loops. This flag will be false if normal
325 vectorizer data dependency analysis would fail or require versioning
326 for alias, but because of loop->safelen >= 2 it has been vectorized
327 even without versioning for alias. E.g. in:
328 #pragma omp simd
329 for (int i = 0; i < m; i++)
330 a[i] = a[i + k] * c;
331 (or #pragma simd or #pragma ivdep) we can vectorize this and it will
332 DTRT even for k > 0 && k < m, but without safelen we would not
333 vectorize this, so this field would be false. */
334 bool no_data_dependencies;
335
336 /* Mark loops having masked stores. */
337 bool has_mask_store;
338
339 /* If if-conversion versioned this loop before conversion, this is the
340 loop version without if-conversion. */
341 struct loop *scalar_loop;
342
343 /* For loops being epilogues of already vectorized loops
344 this points to the original vectorized loop. Otherwise NULL. */
345 _loop_vec_info *orig_loop_info;
346
347} *loop_vec_info;
348
349/* Access Functions. */
350#define LOOP_VINFO_LOOP(L) (L)->loop
351#define LOOP_VINFO_BBS(L) (L)->bbs
352#define LOOP_VINFO_NITERSM1(L) (L)->num_itersm1
353#define LOOP_VINFO_NITERS(L) (L)->num_iters
354/* Since LOOP_VINFO_NITERS and LOOP_VINFO_NITERSM1 can change after
355 prologue peeling retain total unchanged scalar loop iterations for
356 cost model. */
357#define LOOP_VINFO_NITERS_UNCHANGED(L) (L)->num_iters_unchanged
358#define LOOP_VINFO_NITERS_ASSUMPTIONS(L) (L)->num_iters_assumptions
359#define LOOP_VINFO_COST_MODEL_THRESHOLD(L) (L)->th
360#define LOOP_VINFO_VECTORIZABLE_P(L) (L)->vectorizable
361#define LOOP_VINFO_VECT_FACTOR(L) (L)->vectorization_factor
362#define LOOP_VINFO_MAX_VECT_FACTOR(L) (L)->max_vectorization_factor
363#define LOOP_VINFO_PTR_MASK(L) (L)->ptr_mask
364#define LOOP_VINFO_LOOP_NEST(L) (L)->loop_nest
365#define LOOP_VINFO_DATAREFS(L) (L)->datarefs
366#define LOOP_VINFO_DDRS(L) (L)->ddrs
367#define LOOP_VINFO_INT_NITERS(L) (TREE_INT_CST_LOW ((L)->num_iters))
368#define LOOP_VINFO_PEELING_FOR_ALIGNMENT(L) (L)->peeling_for_alignment
369#define LOOP_VINFO_UNALIGNED_DR(L) (L)->unaligned_dr
370#define LOOP_VINFO_MAY_MISALIGN_STMTS(L) (L)->may_misalign_stmts
371#define LOOP_VINFO_MAY_ALIAS_DDRS(L) (L)->may_alias_ddrs
372#define LOOP_VINFO_COMP_ALIAS_DDRS(L) (L)->comp_alias_ddrs
373#define LOOP_VINFO_CHECK_UNEQUAL_ADDRS(L) (L)->check_unequal_addrs
374#define LOOP_VINFO_GROUPED_STORES(L) (L)->grouped_stores
375#define LOOP_VINFO_SLP_INSTANCES(L) (L)->slp_instances
376#define LOOP_VINFO_SLP_UNROLLING_FACTOR(L) (L)->slp_unrolling_factor
377#define LOOP_VINFO_REDUCTIONS(L) (L)->reductions
378#define LOOP_VINFO_REDUCTION_CHAINS(L) (L)->reduction_chains
379#define LOOP_VINFO_TARGET_COST_DATA(L) (L)->target_cost_data
380#define LOOP_VINFO_PEELING_FOR_GAPS(L) (L)->peeling_for_gaps
381#define LOOP_VINFO_OPERANDS_SWAPPED(L) (L)->operands_swapped
382#define LOOP_VINFO_PEELING_FOR_NITER(L) (L)->peeling_for_niter
383#define LOOP_VINFO_NO_DATA_DEPENDENCIES(L) (L)->no_data_dependencies
384#define LOOP_VINFO_SCALAR_LOOP(L) (L)->scalar_loop
385#define LOOP_VINFO_HAS_MASK_STORE(L) (L)->has_mask_store
386#define LOOP_VINFO_SCALAR_ITERATION_COST(L) (L)->scalar_cost_vec
387#define LOOP_VINFO_SINGLE_SCALAR_ITERATION_COST(L) (L)->single_scalar_iteration_cost
388#define LOOP_VINFO_ORIG_LOOP_INFO(L) (L)->orig_loop_info
389
390#define LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT(L) \
391 ((L)->may_misalign_stmts.length () > 0)
392#define LOOP_REQUIRES_VERSIONING_FOR_ALIAS(L) \
393 ((L)->comp_alias_ddrs.length () > 0 \
394 || (L)->check_unequal_addrs.length () > 0)
395#define LOOP_REQUIRES_VERSIONING_FOR_NITERS(L) \
396 (LOOP_VINFO_NITERS_ASSUMPTIONS (L))
397#define LOOP_REQUIRES_VERSIONING(L) \
398 (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (L) \
399 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (L) \
400 || LOOP_REQUIRES_VERSIONING_FOR_NITERS (L))
401
402#define LOOP_VINFO_NITERS_KNOWN_P(L) \
403 (tree_fits_shwi_p ((L)->num_iters) && tree_to_shwi ((L)->num_iters) > 0)
404
405#define LOOP_VINFO_EPILOGUE_P(L) \
406 (LOOP_VINFO_ORIG_LOOP_INFO (L) != NULL)
407
408#define LOOP_VINFO_ORIG_MAX_VECT_FACTOR(L) \
409 (LOOP_VINFO_MAX_VECT_FACTOR (LOOP_VINFO_ORIG_LOOP_INFO (L)))
410
411static inline loop_vec_info
412loop_vec_info_for_loop (struct loop *loop)
413{
414 return (loop_vec_info) loop->aux;
415}
416
417static inline bool
418nested_in_vect_loop_p (struct loop *loop, gimple *stmt)
419{
420 return (loop->inner
421 && (loop->inner == (gimple_bb (stmt))->loop_father));
422}
423
424typedef struct _bb_vec_info : public vec_info
425{
426 _bb_vec_info (gimple_stmt_iterator, gimple_stmt_iterator);
427 ~_bb_vec_info ();
428
429 basic_block bb;
430 gimple_stmt_iterator region_begin;
431 gimple_stmt_iterator region_end;
432} *bb_vec_info;
433
434#define BB_VINFO_BB(B) (B)->bb
435#define BB_VINFO_GROUPED_STORES(B) (B)->grouped_stores
436#define BB_VINFO_SLP_INSTANCES(B) (B)->slp_instances
437#define BB_VINFO_DATAREFS(B) (B)->datarefs
438#define BB_VINFO_DDRS(B) (B)->ddrs
439#define BB_VINFO_TARGET_COST_DATA(B) (B)->target_cost_data
440
441static inline bb_vec_info
442vec_info_for_bb (basic_block bb)
443{
444 return (bb_vec_info) bb->aux;
445}
446
447/*-----------------------------------------------------------------*/
448/* Info on vectorized defs. */
449/*-----------------------------------------------------------------*/
450enum stmt_vec_info_type {
451 undef_vec_info_type = 0,
452 load_vec_info_type,
453 store_vec_info_type,
454 shift_vec_info_type,
455 op_vec_info_type,
456 call_vec_info_type,
457 call_simd_clone_vec_info_type,
458 assignment_vec_info_type,
459 condition_vec_info_type,
460 comparison_vec_info_type,
461 reduc_vec_info_type,
462 induc_vec_info_type,
463 type_promotion_vec_info_type,
464 type_demotion_vec_info_type,
465 type_conversion_vec_info_type,
466 loop_exit_ctrl_vec_info_type
467};
468
469/* Indicates whether/how a variable is used in the scope of loop/basic
470 block. */
471enum vect_relevant {
472 vect_unused_in_scope = 0,
473
474 /* The def is only used outside the loop. */
475 vect_used_only_live,
476 /* The def is in the inner loop, and the use is in the outer loop, and the
477 use is a reduction stmt. */
478 vect_used_in_outer_by_reduction,
479 /* The def is in the inner loop, and the use is in the outer loop (and is
480 not part of reduction). */
481 vect_used_in_outer,
482
483 /* defs that feed computations that end up (only) in a reduction. These
484 defs may be used by non-reduction stmts, but eventually, any
485 computations/values that are affected by these defs are used to compute
486 a reduction (i.e. don't get stored to memory, for example). We use this
487 to identify computations that we can change the order in which they are
488 computed. */
489 vect_used_by_reduction,
490
491 vect_used_in_scope
492};
493
494/* The type of vectorization that can be applied to the stmt: regular loop-based
495 vectorization; pure SLP - the stmt is a part of SLP instances and does not
496 have uses outside SLP instances; or hybrid SLP and loop-based - the stmt is
497 a part of SLP instance and also must be loop-based vectorized, since it has
498 uses outside SLP sequences.
499
500 In the loop context the meanings of pure and hybrid SLP are slightly
501 different. By saying that pure SLP is applied to the loop, we mean that we
502 exploit only intra-iteration parallelism in the loop; i.e., the loop can be
503 vectorized without doing any conceptual unrolling, cause we don't pack
504 together stmts from different iterations, only within a single iteration.
505 Loop hybrid SLP means that we exploit both intra-iteration and
506 inter-iteration parallelism (e.g., number of elements in the vector is 4
507 and the slp-group-size is 2, in which case we don't have enough parallelism
508 within an iteration, so we obtain the rest of the parallelism from subsequent
509 iterations by unrolling the loop by 2). */
510enum slp_vect_type {
511 loop_vect = 0,
512 pure_slp,
513 hybrid
514};
515
516/* Describes how we're going to vectorize an individual load or store,
517 or a group of loads or stores. */
518enum vect_memory_access_type {
519 /* An access to an invariant address. This is used only for loads. */
520 VMAT_INVARIANT,
521
522 /* A simple contiguous access. */
523 VMAT_CONTIGUOUS,
524
525 /* A contiguous access that goes down in memory rather than up,
526 with no additional permutation. This is used only for stores
527 of invariants. */
528 VMAT_CONTIGUOUS_DOWN,
529
530 /* A simple contiguous access in which the elements need to be permuted
531 after loading or before storing. Only used for loop vectorization;
532 SLP uses separate permutes. */
533 VMAT_CONTIGUOUS_PERMUTE,
534
535 /* A simple contiguous access in which the elements need to be reversed
536 after loading or before storing. */
537 VMAT_CONTIGUOUS_REVERSE,
538
539 /* An access that uses IFN_LOAD_LANES or IFN_STORE_LANES. */
540 VMAT_LOAD_STORE_LANES,
541
542 /* An access in which each scalar element is loaded or stored
543 individually. */
544 VMAT_ELEMENTWISE,
545
546 /* A hybrid of VMAT_CONTIGUOUS and VMAT_ELEMENTWISE, used for grouped
547 SLP accesses. Each unrolled iteration uses a contiguous load
548 or store for the whole group, but the groups from separate iterations
549 are combined in the same way as for VMAT_ELEMENTWISE. */
550 VMAT_STRIDED_SLP,
551
552 /* The access uses gather loads or scatter stores. */
553 VMAT_GATHER_SCATTER
554};
555
556typedef struct data_reference *dr_p;
557
558typedef struct _stmt_vec_info {
559
560 enum stmt_vec_info_type type;
561
562 /* Indicates whether this stmts is part of a computation whose result is
563 used outside the loop. */
564 bool live;
565
566 /* Stmt is part of some pattern (computation idiom) */
567 bool in_pattern_p;
568
569 /* Is this statement vectorizable or should it be skipped in (partial)
570 vectorization. */
571 bool vectorizable;
572
573 /* The stmt to which this info struct refers to. */
574 gimple *stmt;
575
576 /* The vec_info with respect to which STMT is vectorized. */
577 vec_info *vinfo;
578
579 /* The vector type to be used for the LHS of this statement. */
580 tree vectype;
581
582 /* The vectorized version of the stmt. */
583 gimple *vectorized_stmt;
584
585
586 /* The following is relevant only for stmts that contain a non-scalar
587 data-ref (array/pointer/struct access). A GIMPLE stmt is expected to have
588 at most one such data-ref. */
589
590 /* Information about the data-ref (access function, etc),
591 relative to the inner-most containing loop. */
592 struct data_reference *data_ref_info;
593
594 /* Information about the data-ref relative to this loop
595 nest (the loop that is being considered for vectorization). */
596 innermost_loop_behavior dr_wrt_vec_loop;
597
598 /* For loop PHI nodes, the base and evolution part of it. This makes sure
599 this information is still available in vect_update_ivs_after_vectorizer
600 where we may not be able to re-analyze the PHI nodes evolution as
601 peeling for the prologue loop can make it unanalyzable. The evolution
602 part is still correct after peeling, but the base may have changed from
603 the version here. */
604 tree loop_phi_evolution_base_unchanged;
605 tree loop_phi_evolution_part;
606
607 /* Used for various bookkeeping purposes, generally holding a pointer to
608 some other stmt S that is in some way "related" to this stmt.
609 Current use of this field is:
610 If this stmt is part of a pattern (i.e. the field 'in_pattern_p' is
611 true): S is the "pattern stmt" that represents (and replaces) the
612 sequence of stmts that constitutes the pattern. Similarly, the
613 related_stmt of the "pattern stmt" points back to this stmt (which is
614 the last stmt in the original sequence of stmts that constitutes the
615 pattern). */
616 gimple *related_stmt;
617
618 /* Used to keep a sequence of def stmts of a pattern stmt if such exists. */
619 gimple_seq pattern_def_seq;
620
621 /* List of datarefs that are known to have the same alignment as the dataref
622 of this stmt. */
623 vec<dr_p> same_align_refs;
624
625 /* Selected SIMD clone's function info. First vector element
626 is SIMD clone's function decl, followed by a pair of trees (base + step)
627 for linear arguments (pair of NULLs for other arguments). */
628 vec<tree> simd_clone_info;
629
630 /* Classify the def of this stmt. */
631 enum vect_def_type def_type;
632
633 /* Whether the stmt is SLPed, loop-based vectorized, or both. */
634 enum slp_vect_type slp_type;
635
636 /* Interleaving and reduction chains info. */
637 /* First element in the group. */
638 gimple *first_element;
639 /* Pointer to the next element in the group. */
640 gimple *next_element;
641 /* For data-refs, in case that two or more stmts share data-ref, this is the
642 pointer to the previously detected stmt with the same dr. */
643 gimple *same_dr_stmt;
644 /* The size of the group. */
645 unsigned int size;
646 /* For stores, number of stores from this group seen. We vectorize the last
647 one. */
648 unsigned int store_count;
649 /* For loads only, the gap from the previous load. For consecutive loads, GAP
650 is 1. */
651 unsigned int gap;
652
653 /* The minimum negative dependence distance this stmt participates in
654 or zero if none. */
655 unsigned int min_neg_dist;
656
657 /* Not all stmts in the loop need to be vectorized. e.g, the increment
658 of the loop induction variable and computation of array indexes. relevant
659 indicates whether the stmt needs to be vectorized. */
660 enum vect_relevant relevant;
661
662 /* For loads if this is a gather, for stores if this is a scatter. */
663 bool gather_scatter_p;
664
665 /* True if this is an access with loop-invariant stride. */
666 bool strided_p;
667
668 /* For both loads and stores. */
669 bool simd_lane_access_p;
670
671 /* Classifies how the load or store is going to be implemented
672 for loop vectorization. */
673 vect_memory_access_type memory_access_type;
674
675 /* For reduction loops, this is the type of reduction. */
676 enum vect_reduction_type v_reduc_type;
677
678 /* For CONST_COND_REDUCTION, record the reduc code. */
679 enum tree_code const_cond_reduc_code;
680
681 /* On a reduction PHI the reduction type as detected by
682 vect_force_simple_reduction. */
683 enum vect_reduction_type reduc_type;
684
685 /* On a reduction PHI the def returned by vect_force_simple_reduction.
686 On the def returned by vect_force_simple_reduction the
687 corresponding PHI. */
688 gimple *reduc_def;
689
690 /* The number of scalar stmt references from active SLP instances. */
691 unsigned int num_slp_uses;
692} *stmt_vec_info;
693
694/* Information about a gather/scatter call. */
695struct gather_scatter_info {
696 /* The FUNCTION_DECL for the built-in gather/scatter function. */
697 tree decl;
698
699 /* The loop-invariant base value. */
700 tree base;
701
702 /* The original scalar offset, which is a non-loop-invariant SSA_NAME. */
703 tree offset;
704
705 /* Each offset element should be multiplied by this amount before
706 being added to the base. */
707 int scale;
708
709 /* The definition type for the vectorized offset. */
710 enum vect_def_type offset_dt;
711
712 /* The type of the vectorized offset. */
713 tree offset_vectype;
714};
715
716/* Access Functions. */
717#define STMT_VINFO_TYPE(S) (S)->type
718#define STMT_VINFO_STMT(S) (S)->stmt
719inline loop_vec_info
720STMT_VINFO_LOOP_VINFO (stmt_vec_info stmt_vinfo)
721{
722 if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (stmt_vinfo->vinfo))
723 return loop_vinfo;
724 return NULL;
725}
726inline bb_vec_info
727STMT_VINFO_BB_VINFO (stmt_vec_info stmt_vinfo)
728{
729 if (bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (stmt_vinfo->vinfo))
730 return bb_vinfo;
731 return NULL;
732}
733#define STMT_VINFO_RELEVANT(S) (S)->relevant
734#define STMT_VINFO_LIVE_P(S) (S)->live
735#define STMT_VINFO_VECTYPE(S) (S)->vectype
736#define STMT_VINFO_VEC_STMT(S) (S)->vectorized_stmt
737#define STMT_VINFO_VECTORIZABLE(S) (S)->vectorizable
738#define STMT_VINFO_DATA_REF(S) (S)->data_ref_info
739#define STMT_VINFO_GATHER_SCATTER_P(S) (S)->gather_scatter_p
740#define STMT_VINFO_STRIDED_P(S) (S)->strided_p
741#define STMT_VINFO_MEMORY_ACCESS_TYPE(S) (S)->memory_access_type
742#define STMT_VINFO_SIMD_LANE_ACCESS_P(S) (S)->simd_lane_access_p
743#define STMT_VINFO_VEC_REDUCTION_TYPE(S) (S)->v_reduc_type
744#define STMT_VINFO_VEC_CONST_COND_REDUC_CODE(S) (S)->const_cond_reduc_code
745
746#define STMT_VINFO_DR_WRT_VEC_LOOP(S) (S)->dr_wrt_vec_loop
747#define STMT_VINFO_DR_BASE_ADDRESS(S) (S)->dr_wrt_vec_loop.base_address
748#define STMT_VINFO_DR_INIT(S) (S)->dr_wrt_vec_loop.init
749#define STMT_VINFO_DR_OFFSET(S) (S)->dr_wrt_vec_loop.offset
750#define STMT_VINFO_DR_STEP(S) (S)->dr_wrt_vec_loop.step
751#define STMT_VINFO_DR_BASE_ALIGNMENT(S) (S)->dr_wrt_vec_loop.base_alignment
752#define STMT_VINFO_DR_BASE_MISALIGNMENT(S) \
753 (S)->dr_wrt_vec_loop.base_misalignment
754#define STMT_VINFO_DR_OFFSET_ALIGNMENT(S) \
755 (S)->dr_wrt_vec_loop.offset_alignment
756#define STMT_VINFO_DR_STEP_ALIGNMENT(S) \
757 (S)->dr_wrt_vec_loop.step_alignment
758
759#define STMT_VINFO_IN_PATTERN_P(S) (S)->in_pattern_p
760#define STMT_VINFO_RELATED_STMT(S) (S)->related_stmt
761#define STMT_VINFO_PATTERN_DEF_SEQ(S) (S)->pattern_def_seq
762#define STMT_VINFO_SAME_ALIGN_REFS(S) (S)->same_align_refs
763#define STMT_VINFO_SIMD_CLONE_INFO(S) (S)->simd_clone_info
764#define STMT_VINFO_DEF_TYPE(S) (S)->def_type
765#define STMT_VINFO_GROUP_FIRST_ELEMENT(S) (S)->first_element
766#define STMT_VINFO_GROUP_NEXT_ELEMENT(S) (S)->next_element
767#define STMT_VINFO_GROUP_SIZE(S) (S)->size
768#define STMT_VINFO_GROUP_STORE_COUNT(S) (S)->store_count
769#define STMT_VINFO_GROUP_GAP(S) (S)->gap
770#define STMT_VINFO_GROUP_SAME_DR_STMT(S) (S)->same_dr_stmt
771#define STMT_VINFO_GROUPED_ACCESS(S) ((S)->first_element != NULL && (S)->data_ref_info)
772#define STMT_VINFO_LOOP_PHI_EVOLUTION_BASE_UNCHANGED(S) (S)->loop_phi_evolution_base_unchanged
773#define STMT_VINFO_LOOP_PHI_EVOLUTION_PART(S) (S)->loop_phi_evolution_part
774#define STMT_VINFO_MIN_NEG_DIST(S) (S)->min_neg_dist
775#define STMT_VINFO_NUM_SLP_USES(S) (S)->num_slp_uses
776#define STMT_VINFO_REDUC_TYPE(S) (S)->reduc_type
777#define STMT_VINFO_REDUC_DEF(S) (S)->reduc_def
778
779#define GROUP_FIRST_ELEMENT(S) (S)->first_element
780#define GROUP_NEXT_ELEMENT(S) (S)->next_element
781#define GROUP_SIZE(S) (S)->size
782#define GROUP_STORE_COUNT(S) (S)->store_count
783#define GROUP_GAP(S) (S)->gap
784#define GROUP_SAME_DR_STMT(S) (S)->same_dr_stmt
785
786#define STMT_VINFO_RELEVANT_P(S) ((S)->relevant != vect_unused_in_scope)
787
788#define HYBRID_SLP_STMT(S) ((S)->slp_type == hybrid)
789#define PURE_SLP_STMT(S) ((S)->slp_type == pure_slp)
790#define STMT_SLP_TYPE(S) (S)->slp_type
791
792struct dataref_aux {
793 /* The misalignment in bytes of the reference, or -1 if not known. */
794 int misalignment;
795 /* The byte alignment that we'd ideally like the reference to have,
796 and the value that misalignment is measured against. */
797 int target_alignment;
798 /* If true the alignment of base_decl needs to be increased. */
799 bool base_misaligned;
800 tree base_decl;
801};
802
803#define DR_VECT_AUX(dr) ((dataref_aux *)(dr)->aux)
804
805#define VECT_MAX_COST 1000
806
807/* The maximum number of intermediate steps required in multi-step type
808 conversion. */
809#define MAX_INTERM_CVT_STEPS 3
810
811/* The maximum vectorization factor supported by any target (V64QI). */
812#define MAX_VECTORIZATION_FACTOR 64
813
814/* Nonzero if TYPE represents a (scalar) boolean type or type
815 in the middle-end compatible with it (unsigned precision 1 integral
816 types). Used to determine which types should be vectorized as
817 VECTOR_BOOLEAN_TYPE_P. */
818
819#define VECT_SCALAR_BOOLEAN_TYPE_P(TYPE) \
820 (TREE_CODE (TYPE) == BOOLEAN_TYPE \
821 || ((TREE_CODE (TYPE) == INTEGER_TYPE \
822 || TREE_CODE (TYPE) == ENUMERAL_TYPE) \
823 && TYPE_PRECISION (TYPE) == 1 \
824 && TYPE_UNSIGNED (TYPE)))
825
826extern vec<stmt_vec_info> stmt_vec_info_vec;
827
828void init_stmt_vec_info_vec (void);
829void free_stmt_vec_info_vec (void);
830
831/* Return a stmt_vec_info corresponding to STMT. */
832
833static inline stmt_vec_info
834vinfo_for_stmt (gimple *stmt)
835{
836 int uid = gimple_uid (stmt);
837 if (uid <= 0)
838 return NULL;
839
840 return stmt_vec_info_vec[uid - 1];
841}
842
843/* Set vectorizer information INFO for STMT. */
844
845static inline void
846set_vinfo_for_stmt (gimple *stmt, stmt_vec_info info)
847{
848 unsigned int uid = gimple_uid (stmt);
849 if (uid == 0)
850 {
851 gcc_checking_assert (info);
852 uid = stmt_vec_info_vec.length () + 1;
853 gimple_set_uid (stmt, uid);
854 stmt_vec_info_vec.safe_push (info);
855 }
856 else
857 {
858 gcc_checking_assert (info == NULL);
859 stmt_vec_info_vec[uid - 1] = info;
860 }
861}
862
863/* Return the earlier statement between STMT1 and STMT2. */
864
865static inline gimple *
866get_earlier_stmt (gimple *stmt1, gimple *stmt2)
867{
868 unsigned int uid1, uid2;
869
870 if (stmt1 == NULL)
871 return stmt2;
872
873 if (stmt2 == NULL)
874 return stmt1;
875
876 uid1 = gimple_uid (stmt1);
877 uid2 = gimple_uid (stmt2);
878
879 if (uid1 == 0 || uid2 == 0)
880 return NULL;
881
882 gcc_checking_assert (uid1 <= stmt_vec_info_vec.length ()
883 && uid2 <= stmt_vec_info_vec.length ());
884
885 if (uid1 < uid2)
886 return stmt1;
887 else
888 return stmt2;
889}
890
891/* Return the later statement between STMT1 and STMT2. */
892
893static inline gimple *
894get_later_stmt (gimple *stmt1, gimple *stmt2)
895{
896 unsigned int uid1, uid2;
897
898 if (stmt1 == NULL)
899 return stmt2;
900
901 if (stmt2 == NULL)
902 return stmt1;
903
904 uid1 = gimple_uid (stmt1);
905 uid2 = gimple_uid (stmt2);
906
907 if (uid1 == 0 || uid2 == 0)
908 return NULL;
909
910 gcc_assert (uid1 <= stmt_vec_info_vec.length ());
911 gcc_assert (uid2 <= stmt_vec_info_vec.length ());
912
913 if (uid1 > uid2)
914 return stmt1;
915 else
916 return stmt2;
917}
918
919/* Return TRUE if a statement represented by STMT_INFO is a part of a
920 pattern. */
921
922static inline bool
923is_pattern_stmt_p (stmt_vec_info stmt_info)
924{
925 gimple *related_stmt;
926 stmt_vec_info related_stmt_info;
927
928 related_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
929 if (related_stmt
930 && (related_stmt_info = vinfo_for_stmt (related_stmt))
931 && STMT_VINFO_IN_PATTERN_P (related_stmt_info))
932 return true;
933
934 return false;
935}
936
937/* Return true if BB is a loop header. */
938
939static inline bool
940is_loop_header_bb_p (basic_block bb)
941{
942 if (bb == (bb->loop_father)->header)
943 return true;
944 gcc_checking_assert (EDGE_COUNT (bb->preds) == 1);
945 return false;
946}
947
948/* Return pow2 (X). */
949
950static inline int
951vect_pow2 (int x)
952{
953 int i, res = 1;
954
955 for (i = 0; i < x; i++)
956 res *= 2;
957
958 return res;
959}
960
961/* Alias targetm.vectorize.builtin_vectorization_cost. */
962
963static inline int
964builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost,
965 tree vectype, int misalign)
966{
967 return targetm.vectorize.builtin_vectorization_cost (type_of_cost,
968 vectype, misalign);
969}
970
971/* Get cost by calling cost target builtin. */
972
973static inline
974int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost)
975{
976 return builtin_vectorization_cost (type_of_cost, NULL, 0);
977}
978
979/* Alias targetm.vectorize.init_cost. */
980
981static inline void *
982init_cost (struct loop *loop_info)
983{
984 return targetm.vectorize.init_cost (loop_info);
985}
986
987/* Alias targetm.vectorize.add_stmt_cost. */
988
989static inline unsigned
990add_stmt_cost (void *data, int count, enum vect_cost_for_stmt kind,
991 stmt_vec_info stmt_info, int misalign,
992 enum vect_cost_model_location where)
993{
994 return targetm.vectorize.add_stmt_cost (data, count, kind,
995 stmt_info, misalign, where);
996}
997
998/* Alias targetm.vectorize.finish_cost. */
999
1000static inline void
1001finish_cost (void *data, unsigned *prologue_cost,
1002 unsigned *body_cost, unsigned *epilogue_cost)
1003{
1004 targetm.vectorize.finish_cost (data, prologue_cost, body_cost, epilogue_cost);
1005}
1006
1007/* Alias targetm.vectorize.destroy_cost_data. */
1008
1009static inline void
1010destroy_cost_data (void *data)
1011{
1012 targetm.vectorize.destroy_cost_data (data);
1013}
1014
1015/*-----------------------------------------------------------------*/
1016/* Info on data references alignment. */
1017/*-----------------------------------------------------------------*/
1018inline void
1019set_dr_misalignment (struct data_reference *dr, int val)
1020{
1021 dataref_aux *data_aux = DR_VECT_AUX (dr);
1022
1023 if (!data_aux)
1024 {
1025 data_aux = XCNEW (dataref_aux);
1026 dr->aux = data_aux;
1027 }
1028
1029 data_aux->misalignment = val;
1030}
1031
1032inline int
1033dr_misalignment (struct data_reference *dr)
1034{
1035 return DR_VECT_AUX (dr)->misalignment;
1036}
1037
1038/* Reflects actual alignment of first access in the vectorized loop,
1039 taking into account peeling/versioning if applied. */
1040#define DR_MISALIGNMENT(DR) dr_misalignment (DR)
1041#define SET_DR_MISALIGNMENT(DR, VAL) set_dr_misalignment (DR, VAL)
1042#define DR_MISALIGNMENT_UNKNOWN (-1)
1043
1044/* Only defined once DR_MISALIGNMENT is defined. */
1045#define DR_TARGET_ALIGNMENT(DR) DR_VECT_AUX (DR)->target_alignment
1046
1047/* Return true if data access DR is aligned to its target alignment
1048 (which may be less than a full vector). */
1049
1050static inline bool
1051aligned_access_p (struct data_reference *data_ref_info)
1052{
1053 return (DR_MISALIGNMENT (data_ref_info) == 0);
1054}
1055
1056/* Return TRUE if the alignment of the data access is known, and FALSE
1057 otherwise. */
1058
1059static inline bool
1060known_alignment_for_access_p (struct data_reference *data_ref_info)
1061{
1062 return (DR_MISALIGNMENT (data_ref_info) != DR_MISALIGNMENT_UNKNOWN);
1063}
1064
1065/* Return the minimum alignment in bytes that the vectorized version
1066 of DR is guaranteed to have. */
1067
1068static inline unsigned int
1069vect_known_alignment_in_bytes (struct data_reference *dr)
1070{
1071 if (DR_MISALIGNMENT (dr) == DR_MISALIGNMENT_UNKNOWN)
1072 return TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr)));
1073 if (DR_MISALIGNMENT (dr) == 0)
1074 return DR_TARGET_ALIGNMENT (dr);
1075 return DR_MISALIGNMENT (dr) & -DR_MISALIGNMENT (dr);
1076}
1077
1078/* Return the behavior of DR with respect to the vectorization context
1079 (which for outer loop vectorization might not be the behavior recorded
1080 in DR itself). */
1081
1082static inline innermost_loop_behavior *
1083vect_dr_behavior (data_reference *dr)
1084{
1085 gimple *stmt = DR_STMT (dr);
1086 stmt_vec_info stmt_info = vinfo_for_stmt (stmt);
1087 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info);
1088 if (loop_vinfo == NULL
1089 || !nested_in_vect_loop_p (LOOP_VINFO_LOOP (loop_vinfo), stmt))
1090 return &DR_INNERMOST (dr);
1091 else
1092 return &STMT_VINFO_DR_WRT_VEC_LOOP (stmt_info);
1093}
1094
1095/* Return true if the vect cost model is unlimited. */
1096static inline bool
1097unlimited_cost_model (loop_p loop)
1098{
1099 if (loop != NULL && loop->force_vectorize
1100 && flag_simd_cost_model != VECT_COST_MODEL_DEFAULT)
1101 return flag_simd_cost_model == VECT_COST_MODEL_UNLIMITED;
1102 return (flag_vect_cost_model == VECT_COST_MODEL_UNLIMITED);
1103}
1104
1105/* Return the number of copies needed for loop vectorization when
1106 a statement operates on vectors of type VECTYPE. This is the
1107 vectorization factor divided by the number of elements in
1108 VECTYPE and is always known at compile time. */
1109
1110static inline unsigned int
1111vect_get_num_copies (loop_vec_info loop_vinfo, tree vectype)
1112{
1113 gcc_checking_assert (LOOP_VINFO_VECT_FACTOR (loop_vinfo)
1114 % TYPE_VECTOR_SUBPARTS (vectype) == 0);
1115 return (LOOP_VINFO_VECT_FACTOR (loop_vinfo)
1116 / TYPE_VECTOR_SUBPARTS (vectype));
1117}
1118
1119/* Return the size of the value accessed by unvectorized data reference DR.
1120 This is only valid once STMT_VINFO_VECTYPE has been calculated for the
1121 associated gimple statement, since that guarantees that DR accesses
1122 either a scalar or a scalar equivalent. ("Scalar equivalent" here
1123 includes things like V1SI, which can be vectorized in the same way
1124 as a plain SI.) */
1125
1126inline unsigned int
1127vect_get_scalar_dr_size (struct data_reference *dr)
1128{
1129 return tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr))));
1130}
1131
1132/* Source location */
1133extern source_location vect_location;
1134
1135/*-----------------------------------------------------------------*/
1136/* Function prototypes. */
1137/*-----------------------------------------------------------------*/
1138
1139/* Simple loop peeling and versioning utilities for vectorizer's purposes -
1140 in tree-vect-loop-manip.c. */
1141extern void slpeel_make_loop_iterate_ntimes (struct loop *, tree);
1142extern bool slpeel_can_duplicate_loop_p (const struct loop *, const_edge);
1143struct loop *slpeel_tree_duplicate_loop_to_edge_cfg (struct loop *,
1144 struct loop *, edge);
1145extern void vect_loop_versioning (loop_vec_info, unsigned int, bool);
1146extern struct loop *vect_do_peeling (loop_vec_info, tree, tree,
1147 tree *, int, bool, bool);
1148extern source_location find_loop_location (struct loop *);
1149extern bool vect_can_advance_ivs_p (loop_vec_info);
1150
1151/* In tree-vect-stmts.c. */
1152extern unsigned int current_vector_size;
1153extern tree get_vectype_for_scalar_type (tree);
1154extern tree get_mask_type_for_scalar_type (tree);
1155extern tree get_same_sized_vectype (tree, tree);
1156extern bool vect_is_simple_use (tree, vec_info *, gimple **,
1157 enum vect_def_type *);
1158extern bool vect_is_simple_use (tree, vec_info *, gimple **,
1159 enum vect_def_type *, tree *);
1160extern bool supportable_widening_operation (enum tree_code, gimple *, tree,
1161 tree, enum tree_code *,
1162 enum tree_code *, int *,
1163 vec<tree> *);
1164extern bool supportable_narrowing_operation (enum tree_code, tree, tree,
1165 enum tree_code *,
1166 int *, vec<tree> *);
1167extern stmt_vec_info new_stmt_vec_info (gimple *stmt, vec_info *);
1168extern void free_stmt_vec_info (gimple *stmt);
1169extern void vect_model_simple_cost (stmt_vec_info, int, enum vect_def_type *,
1170 int, stmt_vector_for_cost *,
1171 stmt_vector_for_cost *);
1172extern void vect_model_store_cost (stmt_vec_info, int, vect_memory_access_type,
1173 enum vect_def_type, slp_tree,
1174 stmt_vector_for_cost *,
1175 stmt_vector_for_cost *);
1176extern void vect_model_load_cost (stmt_vec_info, int, vect_memory_access_type,
1177 slp_tree, stmt_vector_for_cost *,
1178 stmt_vector_for_cost *);
1179extern unsigned record_stmt_cost (stmt_vector_for_cost *, int,
1180 enum vect_cost_for_stmt, stmt_vec_info,
1181 int, enum vect_cost_model_location);
1182extern void vect_finish_stmt_generation (gimple *, gimple *,
1183 gimple_stmt_iterator *);
1184extern bool vect_mark_stmts_to_be_vectorized (loop_vec_info);
1185extern tree vect_get_vec_def_for_operand_1 (gimple *, enum vect_def_type);
1186extern tree vect_get_vec_def_for_operand (tree, gimple *, tree = NULL);
1187extern void vect_get_vec_defs (tree, tree, gimple *, vec<tree> *,
1188 vec<tree> *, slp_tree);
1189extern void vect_get_vec_defs_for_stmt_copy (enum vect_def_type *,
1190 vec<tree> *, vec<tree> *);
1191extern tree vect_init_vector (gimple *, tree, tree,
1192 gimple_stmt_iterator *);
1193extern tree vect_get_vec_def_for_stmt_copy (enum vect_def_type, tree);
1194extern bool vect_transform_stmt (gimple *, gimple_stmt_iterator *,
1195 bool *, slp_tree, slp_instance);
1196extern void vect_remove_stores (gimple *);
1197extern bool vect_analyze_stmt (gimple *, bool *, slp_tree, slp_instance);
1198extern bool vectorizable_condition (gimple *, gimple_stmt_iterator *,
1199 gimple **, tree, int, slp_tree);
1200extern void vect_get_load_cost (struct data_reference *, int, bool,
1201 unsigned int *, unsigned int *,
1202 stmt_vector_for_cost *,
1203 stmt_vector_for_cost *, bool);
1204extern void vect_get_store_cost (struct data_reference *, int,
1205 unsigned int *, stmt_vector_for_cost *);
1206extern bool vect_supportable_shift (enum tree_code, tree);
1207extern tree vect_gen_perm_mask_any (tree, vec_perm_indices);
1208extern tree vect_gen_perm_mask_checked (tree, vec_perm_indices);
1209extern void optimize_mask_stores (struct loop*);
1210
1211/* In tree-vect-data-refs.c. */
1212extern bool vect_can_force_dr_alignment_p (const_tree, unsigned int);
1213extern enum dr_alignment_support vect_supportable_dr_alignment
1214 (struct data_reference *, bool);
1215extern tree vect_get_smallest_scalar_type (gimple *, HOST_WIDE_INT *,
1216 HOST_WIDE_INT *);
1217extern bool vect_analyze_data_ref_dependences (loop_vec_info, int *);
1218extern bool vect_slp_analyze_instance_dependence (slp_instance);
1219extern bool vect_enhance_data_refs_alignment (loop_vec_info);
1220extern bool vect_analyze_data_refs_alignment (loop_vec_info);
1221extern bool vect_verify_datarefs_alignment (loop_vec_info);
1222extern bool vect_slp_analyze_and_verify_instance_alignment (slp_instance);
1223extern bool vect_analyze_data_ref_accesses (vec_info *);
1224extern bool vect_prune_runtime_alias_test_list (loop_vec_info);
1225extern bool vect_check_gather_scatter (gimple *, loop_vec_info,
1226 gather_scatter_info *);
1227extern bool vect_analyze_data_refs (vec_info *, int *);
1228extern void vect_record_base_alignments (vec_info *);
1229extern tree vect_create_data_ref_ptr (gimple *, tree, struct loop *, tree,
1230 tree *, gimple_stmt_iterator *,
1231 gimple **, bool, bool *,
1232 tree = NULL_TREE);
1233extern tree bump_vector_ptr (tree, gimple *, gimple_stmt_iterator *, gimple *,
1234 tree);
1235extern tree vect_create_destination_var (tree, tree);
1236extern bool vect_grouped_store_supported (tree, unsigned HOST_WIDE_INT);
1237extern bool vect_store_lanes_supported (tree, unsigned HOST_WIDE_INT);
1238extern bool vect_grouped_load_supported (tree, bool, unsigned HOST_WIDE_INT);
1239extern bool vect_load_lanes_supported (tree, unsigned HOST_WIDE_INT);
1240extern void vect_permute_store_chain (vec<tree> ,unsigned int, gimple *,
1241 gimple_stmt_iterator *, vec<tree> *);
1242extern tree vect_setup_realignment (gimple *, gimple_stmt_iterator *, tree *,
1243 enum dr_alignment_support, tree,
1244 struct loop **);
1245extern void vect_transform_grouped_load (gimple *, vec<tree> , int,
1246 gimple_stmt_iterator *);
1247extern void vect_record_grouped_load_vectors (gimple *, vec<tree> );
1248extern tree vect_get_new_vect_var (tree, enum vect_var_kind, const char *);
1249extern tree vect_get_new_ssa_name (tree, enum vect_var_kind,
1250 const char * = NULL);
1251extern tree vect_create_addr_base_for_vector_ref (gimple *, gimple_seq *,
1252 tree, tree = NULL_TREE);
1253
1254/* In tree-vect-loop.c. */
1255/* FORNOW: Used in tree-parloops.c. */
1256extern gimple *vect_force_simple_reduction (loop_vec_info, gimple *,
1257 bool *, bool);
1258/* Used in gimple-loop-interchange.c. */
1259extern bool check_reduction_path (location_t, loop_p, gphi *, tree,
1260 enum tree_code);
1261/* Drive for loop analysis stage. */
1262extern loop_vec_info vect_analyze_loop (struct loop *, loop_vec_info);
1263extern tree vect_build_loop_niters (loop_vec_info, bool * = NULL);
1264extern void vect_gen_vector_loop_niters (loop_vec_info, tree, tree *, bool);
1265/* Drive for loop transformation stage. */
1266extern struct loop *vect_transform_loop (loop_vec_info);
1267extern loop_vec_info vect_analyze_loop_form (struct loop *);
1268extern bool vectorizable_live_operation (gimple *, gimple_stmt_iterator *,
1269 slp_tree, int, gimple **);
1270extern bool vectorizable_reduction (gimple *, gimple_stmt_iterator *,
1271 gimple **, slp_tree, slp_instance);
1272extern bool vectorizable_induction (gimple *, gimple_stmt_iterator *,
1273 gimple **, slp_tree);
1274extern tree get_initial_def_for_reduction (gimple *, tree, tree *);
1275extern bool vect_worthwhile_without_simd_p (vec_info *, tree_code);
1276extern int vect_get_known_peeling_cost (loop_vec_info, int, int *,
1277 stmt_vector_for_cost *,
1278 stmt_vector_for_cost *,
1279 stmt_vector_for_cost *);
1280
1281/* In tree-vect-slp.c. */
1282extern void vect_free_slp_instance (slp_instance);
1283extern bool vect_transform_slp_perm_load (slp_tree, vec<tree> ,
1284 gimple_stmt_iterator *, int,
1285 slp_instance, bool, unsigned *);
1286extern bool vect_slp_analyze_operations (vec_info *);
1287extern bool vect_schedule_slp (vec_info *);
1288extern bool vect_analyze_slp (vec_info *, unsigned);
1289extern bool vect_make_slp_decision (loop_vec_info);
1290extern void vect_detect_hybrid_slp (loop_vec_info);
1291extern void vect_get_slp_defs (vec<tree> , slp_tree, vec<vec<tree> > *);
1292extern bool vect_slp_bb (basic_block);
1293extern gimple *vect_find_last_scalar_stmt_in_slp (slp_tree);
1294extern bool is_simple_and_all_uses_invariant (gimple *, loop_vec_info);
1295
1296/* In tree-vect-patterns.c. */
1297/* Pattern recognition functions.
1298 Additional pattern recognition functions can (and will) be added
1299 in the future. */
1300typedef gimple *(* vect_recog_func_ptr) (vec<gimple *> *, tree *, tree *);
1301#define NUM_PATTERNS 14
1302void vect_pattern_recog (vec_info *);
1303
1304/* In tree-vectorizer.c. */
1305unsigned vectorize_loops (void);
1306bool vect_stmt_in_region_p (vec_info *, gimple *);
1307void vect_free_loop_info_assumptions (struct loop *);
1308
1309#endif /* GCC_TREE_VECTORIZER_H */
1310