2 Copyright (C) 2003-2020 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
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
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"
31 /* Used for naming of new temporaries. */
39 /* Defines type of operation. */
46 /* Define type of available alignment support. */
47 enum dr_alignment_support
{
48 dr_unaligned_unsupported
,
49 dr_unaligned_supported
,
51 dr_explicit_realign_optimized
,
55 /* Define type of def-use cross-iteration cycle. */
57 vect_uninitialized_def
= 0,
58 vect_constant_def
= 1,
63 vect_double_reduction_def
,
68 /* Define type of reduction. */
69 enum vect_reduction_type
{
72 INTEGER_INDUC_COND_REDUCTION
,
75 /* Retain a scalar phi and use a FOLD_EXTRACT_LAST within the loop
78 for (int i = 0; i < VF; ++i)
79 res = cond[i] ? val[i] : res; */
80 EXTRACT_LAST_REDUCTION
,
82 /* Use a folding reduction within the loop to implement:
84 for (int i = 0; i < VF; ++i)
87 (with no reassocation). */
91 #define VECTORIZABLE_CYCLE_DEF(D) (((D) == vect_reduction_def) \
92 || ((D) == vect_double_reduction_def) \
93 || ((D) == vect_nested_cycle))
95 /* Structure to encapsulate information about a group of like
96 instructions to be presented to the target cost model. */
97 struct stmt_info_for_cost
{
99 enum vect_cost_for_stmt kind
;
100 enum vect_cost_model_location where
;
101 stmt_vec_info stmt_info
;
106 typedef vec
<stmt_info_for_cost
> stmt_vector_for_cost
;
108 /* Maps base addresses to an innermost_loop_behavior that gives the maximum
109 known alignment for that base. */
110 typedef hash_map
<tree_operand_hash
,
111 innermost_loop_behavior
*> vec_base_alignments
;
113 /************************************************************************
115 ************************************************************************/
116 typedef struct _slp_tree
*slp_tree
;
118 /* A computation tree of an SLP instance. Each node corresponds to a group of
119 stmts to be packed in a SIMD stmt. */
121 /* Nodes that contain def-stmts of this node statements operands. */
122 vec
<slp_tree
> children
;
124 /* A group of scalar stmts to be vectorized together. */
125 vec
<stmt_vec_info
> stmts
;
126 /* A group of scalar operands to be vectorized together. */
129 /* Load permutation relative to the stores, NULL if there is no
131 vec
<unsigned> load_permutation
;
134 /* Vectorized stmt/s. */
135 vec
<stmt_vec_info
> vec_stmts
;
136 /* Number of vector stmts that are created to replace the group of scalar
137 stmts. It is calculated during the transformation phase as the number of
138 scalar elements in one scalar iteration (GROUP_SIZE) multiplied by VF
139 divided by vector size. */
140 unsigned int vec_stmts_size
;
142 /* Reference count in the SLP graph. */
144 /* The maximum number of vector elements for the subtree rooted
146 poly_uint64 max_nunits
;
147 /* Whether the scalar computations use two different operators. */
149 /* The DEF type of this node. */
150 enum vect_def_type def_type
;
154 /* SLP instance is a sequence of stmts in a loop that can be packed into
156 typedef class _slp_instance
{
158 /* The root of SLP tree. */
161 /* For vector constructors, the constructor stmt that the SLP tree is built
162 from, NULL otherwise. */
163 stmt_vec_info root_stmt
;
165 /* The unrolling factor required to vectorized this SLP instance. */
166 poly_uint64 unrolling_factor
;
168 /* The group of nodes that contain loads of this SLP instance. */
171 /* The SLP node containing the reduction PHIs. */
176 /* Access Functions. */
177 #define SLP_INSTANCE_TREE(S) (S)->root
178 #define SLP_INSTANCE_UNROLLING_FACTOR(S) (S)->unrolling_factor
179 #define SLP_INSTANCE_LOADS(S) (S)->loads
180 #define SLP_INSTANCE_ROOT_STMT(S) (S)->root_stmt
182 #define SLP_TREE_CHILDREN(S) (S)->children
183 #define SLP_TREE_SCALAR_STMTS(S) (S)->stmts
184 #define SLP_TREE_SCALAR_OPS(S) (S)->ops
185 #define SLP_TREE_VEC_STMTS(S) (S)->vec_stmts
186 #define SLP_TREE_NUMBER_OF_VEC_STMTS(S) (S)->vec_stmts_size
187 #define SLP_TREE_LOAD_PERMUTATION(S) (S)->load_permutation
188 #define SLP_TREE_TWO_OPERATORS(S) (S)->two_operators
189 #define SLP_TREE_DEF_TYPE(S) (S)->def_type
190 #define SLP_TREE_VECTYPE(S) (S)->vectype
192 /* Key for map that records association between
193 scalar conditions and corresponding loop mask, and
194 is populated by vect_record_loop_mask. */
196 struct scalar_cond_masked_key
198 scalar_cond_masked_key (tree t
, unsigned ncopies_
)
201 get_cond_ops_from_tree (t
);
204 void get_cond_ops_from_tree (tree
);
213 struct default_hash_traits
<scalar_cond_masked_key
>
215 typedef scalar_cond_masked_key compare_type
;
216 typedef scalar_cond_masked_key value_type
;
218 static inline hashval_t
223 inchash::add_expr (v
.op0
, h
, 0);
224 inchash::add_expr (v
.op1
, h
, 0);
225 h
.add_int (v
.ncopies
);
230 equal (value_type existing
, value_type candidate
)
232 return (existing
.ncopies
== candidate
.ncopies
233 && existing
.code
== candidate
.code
234 && operand_equal_p (existing
.op0
, candidate
.op0
, 0)
235 && operand_equal_p (existing
.op1
, candidate
.op1
, 0));
238 static const bool empty_zero_p
= true;
241 mark_empty (value_type
&v
)
247 is_empty (value_type v
)
249 return v
.ncopies
== 0;
252 static inline void mark_deleted (value_type
&) {}
254 static inline bool is_deleted (const value_type
&)
259 static inline void remove (value_type
&) {}
262 typedef hash_set
<scalar_cond_masked_key
> scalar_cond_masked_set_type
;
264 /* Describes two objects whose addresses must be unequal for the vectorized
266 typedef std::pair
<tree
, tree
> vec_object_pair
;
268 /* Records that vectorization is only possible if abs (EXPR) >= MIN_VALUE.
269 UNSIGNED_P is true if we can assume that abs (EXPR) == EXPR. */
270 class vec_lower_bound
{
272 vec_lower_bound () {}
273 vec_lower_bound (tree e
, bool u
, poly_uint64 m
)
274 : expr (e
), unsigned_p (u
), min_value (m
) {}
278 poly_uint64 min_value
;
281 /* Vectorizer state shared between different analyses like vector sizes
282 of the same CFG region. */
283 class vec_info_shared
{
288 void save_datarefs();
289 void check_datarefs();
291 /* All data references. Freed by free_data_refs, so not an auto_vec. */
292 vec
<data_reference_p
> datarefs
;
293 vec
<data_reference
> datarefs_copy
;
295 /* The loop nest in which the data dependences are computed. */
296 auto_vec
<loop_p
> loop_nest
;
298 /* All data dependences. Freed by free_dependence_relations, so not
303 /* Vectorizer state common between loop and basic-block vectorization. */
306 typedef hash_set
<int_hash
<machine_mode
, E_VOIDmode
, E_BLKmode
> > mode_set
;
307 enum vec_kind
{ bb
, loop
};
309 vec_info (vec_kind
, void *, vec_info_shared
*);
312 stmt_vec_info
add_stmt (gimple
*);
313 stmt_vec_info
lookup_stmt (gimple
*);
314 stmt_vec_info
lookup_def (tree
);
315 stmt_vec_info
lookup_single_use (tree
);
316 class dr_vec_info
*lookup_dr (data_reference
*);
317 void move_dr (stmt_vec_info
, stmt_vec_info
);
318 void remove_stmt (stmt_vec_info
);
319 void replace_stmt (gimple_stmt_iterator
*, stmt_vec_info
, gimple
*);
321 /* The type of vectorization. */
324 /* Shared vectorizer state. */
325 vec_info_shared
*shared
;
327 /* The mapping of GIMPLE UID to stmt_vec_info. */
328 vec
<stmt_vec_info
> stmt_vec_infos
;
331 auto_vec
<slp_instance
> slp_instances
;
332 auto_vec
<slp_tree
> slp_loads
;
334 /* Maps base addresses to an innermost_loop_behavior that gives the maximum
335 known alignment for that base. */
336 vec_base_alignments base_alignments
;
338 /* All interleaving chains of stores, represented by the first
339 stmt in the chain. */
340 auto_vec
<stmt_vec_info
> grouped_stores
;
342 /* Cost data used by the target cost model. */
343 void *target_cost_data
;
345 /* The set of vector modes used in the vectorized region. */
346 mode_set used_vector_modes
;
348 /* The argument we should pass to related_vector_mode when looking up
349 the vector mode for a scalar mode, or VOIDmode if we haven't yet
350 made any decisions about which vector modes to use. */
351 machine_mode vector_mode
;
354 stmt_vec_info
new_stmt_vec_info (gimple
*stmt
);
355 void set_vinfo_for_stmt (gimple
*, stmt_vec_info
);
356 void free_stmt_vec_infos ();
357 void free_stmt_vec_info (stmt_vec_info
);
360 class _loop_vec_info
;
366 is_a_helper
<_loop_vec_info
*>::test (vec_info
*i
)
368 return i
->kind
== vec_info::loop
;
374 is_a_helper
<_bb_vec_info
*>::test (vec_info
*i
)
376 return i
->kind
== vec_info::bb
;
380 /* In general, we can divide the vector statements in a vectorized loop
381 into related groups ("rgroups") and say that for each rgroup there is
382 some nS such that the rgroup operates on nS values from one scalar
383 iteration followed by nS values from the next. That is, if VF is the
384 vectorization factor of the loop, the rgroup operates on a sequence:
386 (1,1) (1,2) ... (1,nS) (2,1) ... (2,nS) ... (VF,1) ... (VF,nS)
388 where (i,j) represents a scalar value with index j in a scalar
389 iteration with index i.
391 [ We use the term "rgroup" to emphasise that this grouping isn't
392 necessarily the same as the grouping of statements used elsewhere.
393 For example, if we implement a group of scalar loads using gather
394 loads, we'll use a separate gather load for each scalar load, and
395 thus each gather load will belong to its own rgroup. ]
397 In general this sequence will occupy nV vectors concatenated
398 together. If these vectors have nL lanes each, the total number
399 of scalar values N is given by:
401 N = nS * VF = nV * nL
403 None of nS, VF, nV and nL are required to be a power of 2. nS and nV
404 are compile-time constants but VF and nL can be variable (if the target
405 supports variable-length vectors).
407 In classical vectorization, each iteration of the vector loop would
408 handle exactly VF iterations of the original scalar loop. However,
409 in a fully-masked loop, a particular iteration of the vector loop
410 might handle fewer than VF iterations of the scalar loop. The vector
411 lanes that correspond to iterations of the scalar loop are said to be
412 "active" and the other lanes are said to be "inactive".
414 In a fully-masked loop, many rgroups need to be masked to ensure that
415 they have no effect for the inactive lanes. Each such rgroup needs a
416 sequence of booleans in the same order as above, but with each (i,j)
417 replaced by a boolean that indicates whether iteration i is active.
418 This sequence occupies nV vector masks that again have nL lanes each.
419 Thus the mask sequence as a whole consists of VF independent booleans
420 that are each repeated nS times.
422 We make the simplifying assumption that if a sequence of nV masks is
423 suitable for one (nS,nL) pair, we can reuse it for (nS/2,nL/2) by
424 VIEW_CONVERTing it. This holds for all current targets that support
425 fully-masked loops. For example, suppose the scalar loop is:
429 for (int i = 0; i < n; ++i)
431 f[i * 2 + 0] += 1.0f;
432 f[i * 2 + 1] += 2.0f;
436 and suppose that vectors have 256 bits. The vectorized f accesses
437 will belong to one rgroup and the vectorized d access to another:
439 f rgroup: nS = 2, nV = 1, nL = 8
440 d rgroup: nS = 1, nV = 1, nL = 4
443 [ In this simple example the rgroups do correspond to the normal
444 SLP grouping scheme. ]
446 If only the first three lanes are active, the masks we need are:
448 f rgroup: 1 1 | 1 1 | 1 1 | 0 0
449 d rgroup: 1 | 1 | 1 | 0
451 Here we can use a mask calculated for f's rgroup for d's, but not
454 Thus for each value of nV, it is enough to provide nV masks, with the
455 mask being calculated based on the highest nL (or, equivalently, based
456 on the highest nS) required by any rgroup with that nV. We therefore
457 represent the entire collection of masks as a two-level table, with the
458 first level being indexed by nV - 1 (since nV == 0 doesn't exist) and
459 the second being indexed by the mask index 0 <= i < nV. */
461 /* The masks needed by rgroups with nV vectors, according to the
462 description above. */
463 struct rgroup_masks
{
464 /* The largest nS for all rgroups that use these masks. */
465 unsigned int max_nscalars_per_iter
;
467 /* The type of mask to use, based on the highest nS recorded above. */
470 /* A vector of nV masks, in iteration order. */
474 typedef auto_vec
<rgroup_masks
> vec_loop_masks
;
476 typedef auto_vec
<std::pair
<data_reference
*, tree
> > drs_init_vec
;
478 /*-----------------------------------------------------------------*/
479 /* Info on vectorized loops. */
480 /*-----------------------------------------------------------------*/
481 typedef class _loop_vec_info
: public vec_info
{
483 _loop_vec_info (class loop
*, vec_info_shared
*);
486 /* The loop to which this info struct refers to. */
489 /* The loop basic blocks. */
492 /* Number of latch executions. */
494 /* Number of iterations. */
496 /* Number of iterations of the original loop. */
497 tree num_iters_unchanged
;
498 /* Condition under which this loop is analyzed and versioned. */
499 tree num_iters_assumptions
;
501 /* Threshold of number of iterations below which vectorization will not be
502 performed. It is calculated from MIN_PROFITABLE_ITERS and
503 param_min_vect_loop_bound. */
506 /* When applying loop versioning, the vector form should only be used
507 if the number of scalar iterations is >= this value, on top of all
508 the other requirements. Ignored when loop versioning is not being
510 poly_uint64 versioning_threshold
;
512 /* Unrolling factor */
513 poly_uint64 vectorization_factor
;
515 /* Maximum runtime vectorization factor, or MAX_VECTORIZATION_FACTOR
516 if there is no particular limit. */
517 unsigned HOST_WIDE_INT max_vectorization_factor
;
519 /* The masks that a fully-masked loop should use to avoid operating
520 on inactive scalars. */
521 vec_loop_masks masks
;
523 /* Set of scalar conditions that have loop mask applied. */
524 scalar_cond_masked_set_type scalar_cond_masked_set
;
526 /* If we are using a loop mask to align memory addresses, this variable
527 contains the number of vector elements that we should skip in the
528 first iteration of the vector loop (i.e. the number of leading
529 elements that should be false in the first mask). */
530 tree mask_skip_niters
;
532 /* Type of the variables to use in the WHILE_ULT call for fully-masked
534 tree mask_compare_type
;
536 /* For #pragma omp simd if (x) loops the x expression. If constant 0,
537 the loop should not be vectorized, if constant non-zero, simd_if_cond
538 shouldn't be set and loop vectorized normally, if SSA_NAME, the loop
539 should be versioned on that condition, using scalar loop if the condition
540 is false and vectorized loop otherwise. */
543 /* Type of the IV to use in the WHILE_ULT call for fully-masked
547 /* Unknown DRs according to which loop was peeled. */
548 class dr_vec_info
*unaligned_dr
;
550 /* peeling_for_alignment indicates whether peeling for alignment will take
551 place, and what the peeling factor should be:
552 peeling_for_alignment = X means:
553 If X=0: Peeling for alignment will not be applied.
554 If X>0: Peel first X iterations.
555 If X=-1: Generate a runtime test to calculate the number of iterations
556 to be peeled, using the dataref recorded in the field
558 int peeling_for_alignment
;
560 /* The mask used to check the alignment of pointers or arrays. */
563 /* Data Dependence Relations defining address ranges that are candidates
564 for a run-time aliasing check. */
565 auto_vec
<ddr_p
> may_alias_ddrs
;
567 /* Data Dependence Relations defining address ranges together with segment
568 lengths from which the run-time aliasing check is built. */
569 auto_vec
<dr_with_seg_len_pair_t
> comp_alias_ddrs
;
571 /* Check that the addresses of each pair of objects is unequal. */
572 auto_vec
<vec_object_pair
> check_unequal_addrs
;
574 /* List of values that are required to be nonzero. This is used to check
575 whether things like "x[i * n] += 1;" are safe and eventually gets added
576 to the checks for lower bounds below. */
577 auto_vec
<tree
> check_nonzero
;
579 /* List of values that need to be checked for a minimum value. */
580 auto_vec
<vec_lower_bound
> lower_bounds
;
582 /* Statements in the loop that have data references that are candidates for a
583 runtime (loop versioning) misalignment check. */
584 auto_vec
<stmt_vec_info
> may_misalign_stmts
;
586 /* Reduction cycles detected in the loop. Used in loop-aware SLP. */
587 auto_vec
<stmt_vec_info
> reductions
;
589 /* All reduction chains in the loop, represented by the first
590 stmt in the chain. */
591 auto_vec
<stmt_vec_info
> reduction_chains
;
593 /* Cost vector for a single scalar iteration. */
594 auto_vec
<stmt_info_for_cost
> scalar_cost_vec
;
596 /* Map of IV base/step expressions to inserted name in the preheader. */
597 hash_map
<tree_operand_hash
, tree
> *ivexpr_map
;
599 /* Map of OpenMP "omp simd array" scan variables to corresponding
600 rhs of the store of the initializer. */
601 hash_map
<tree
, tree
> *scan_map
;
603 /* The unrolling factor needed to SLP the loop. In case of that pure SLP is
604 applied to the loop, i.e., no unrolling is needed, this is 1. */
605 poly_uint64 slp_unrolling_factor
;
607 /* Cost of a single scalar iteration. */
608 int single_scalar_iteration_cost
;
610 /* The cost of the vector prologue and epilogue, including peeled
611 iterations and set-up code. */
612 int vec_outside_cost
;
614 /* The cost of the vector loop body. */
617 /* Is the loop vectorizable? */
620 /* Records whether we still have the option of using a fully-masked loop. */
621 bool can_fully_mask_p
;
623 /* True if have decided to use a fully-masked loop. */
626 /* When we have grouped data accesses with gaps, we may introduce invalid
627 memory accesses. We peel the last iteration of the loop to prevent
629 bool peeling_for_gaps
;
631 /* When the number of iterations is not a multiple of the vector size
632 we need to peel off iterations at the end to form an epilogue loop. */
633 bool peeling_for_niter
;
635 /* True if there are no loop carried data dependencies in the loop.
636 If loop->safelen <= 1, then this is always true, either the loop
637 didn't have any loop carried data dependencies, or the loop is being
638 vectorized guarded with some runtime alias checks, or couldn't
639 be vectorized at all, but then this field shouldn't be used.
640 For loop->safelen >= 2, the user has asserted that there are no
641 backward dependencies, but there still could be loop carried forward
642 dependencies in such loops. This flag will be false if normal
643 vectorizer data dependency analysis would fail or require versioning
644 for alias, but because of loop->safelen >= 2 it has been vectorized
645 even without versioning for alias. E.g. in:
647 for (int i = 0; i < m; i++)
649 (or #pragma simd or #pragma ivdep) we can vectorize this and it will
650 DTRT even for k > 0 && k < m, but without safelen we would not
651 vectorize this, so this field would be false. */
652 bool no_data_dependencies
;
654 /* Mark loops having masked stores. */
657 /* Queued scaling factor for the scalar loop. */
658 profile_probability scalar_loop_scaling
;
660 /* If if-conversion versioned this loop before conversion, this is the
661 loop version without if-conversion. */
662 class loop
*scalar_loop
;
664 /* For loops being epilogues of already vectorized loops
665 this points to the original vectorized loop. Otherwise NULL. */
666 _loop_vec_info
*orig_loop_info
;
668 /* Used to store loop_vec_infos of epilogues of this loop during
670 vec
<_loop_vec_info
*> epilogue_vinfos
;
674 /* Access Functions. */
675 #define LOOP_VINFO_LOOP(L) (L)->loop
676 #define LOOP_VINFO_BBS(L) (L)->bbs
677 #define LOOP_VINFO_NITERSM1(L) (L)->num_itersm1
678 #define LOOP_VINFO_NITERS(L) (L)->num_iters
679 /* Since LOOP_VINFO_NITERS and LOOP_VINFO_NITERSM1 can change after
680 prologue peeling retain total unchanged scalar loop iterations for
682 #define LOOP_VINFO_NITERS_UNCHANGED(L) (L)->num_iters_unchanged
683 #define LOOP_VINFO_NITERS_ASSUMPTIONS(L) (L)->num_iters_assumptions
684 #define LOOP_VINFO_COST_MODEL_THRESHOLD(L) (L)->th
685 #define LOOP_VINFO_VERSIONING_THRESHOLD(L) (L)->versioning_threshold
686 #define LOOP_VINFO_VECTORIZABLE_P(L) (L)->vectorizable
687 #define LOOP_VINFO_CAN_FULLY_MASK_P(L) (L)->can_fully_mask_p
688 #define LOOP_VINFO_FULLY_MASKED_P(L) (L)->fully_masked_p
689 #define LOOP_VINFO_VECT_FACTOR(L) (L)->vectorization_factor
690 #define LOOP_VINFO_MAX_VECT_FACTOR(L) (L)->max_vectorization_factor
691 #define LOOP_VINFO_MASKS(L) (L)->masks
692 #define LOOP_VINFO_MASK_SKIP_NITERS(L) (L)->mask_skip_niters
693 #define LOOP_VINFO_MASK_COMPARE_TYPE(L) (L)->mask_compare_type
694 #define LOOP_VINFO_MASK_IV_TYPE(L) (L)->iv_type
695 #define LOOP_VINFO_PTR_MASK(L) (L)->ptr_mask
696 #define LOOP_VINFO_LOOP_NEST(L) (L)->shared->loop_nest
697 #define LOOP_VINFO_DATAREFS(L) (L)->shared->datarefs
698 #define LOOP_VINFO_DDRS(L) (L)->shared->ddrs
699 #define LOOP_VINFO_INT_NITERS(L) (TREE_INT_CST_LOW ((L)->num_iters))
700 #define LOOP_VINFO_PEELING_FOR_ALIGNMENT(L) (L)->peeling_for_alignment
701 #define LOOP_VINFO_UNALIGNED_DR(L) (L)->unaligned_dr
702 #define LOOP_VINFO_MAY_MISALIGN_STMTS(L) (L)->may_misalign_stmts
703 #define LOOP_VINFO_MAY_ALIAS_DDRS(L) (L)->may_alias_ddrs
704 #define LOOP_VINFO_COMP_ALIAS_DDRS(L) (L)->comp_alias_ddrs
705 #define LOOP_VINFO_CHECK_UNEQUAL_ADDRS(L) (L)->check_unequal_addrs
706 #define LOOP_VINFO_CHECK_NONZERO(L) (L)->check_nonzero
707 #define LOOP_VINFO_LOWER_BOUNDS(L) (L)->lower_bounds
708 #define LOOP_VINFO_GROUPED_STORES(L) (L)->grouped_stores
709 #define LOOP_VINFO_SLP_INSTANCES(L) (L)->slp_instances
710 #define LOOP_VINFO_SLP_UNROLLING_FACTOR(L) (L)->slp_unrolling_factor
711 #define LOOP_VINFO_REDUCTIONS(L) (L)->reductions
712 #define LOOP_VINFO_REDUCTION_CHAINS(L) (L)->reduction_chains
713 #define LOOP_VINFO_TARGET_COST_DATA(L) (L)->target_cost_data
714 #define LOOP_VINFO_PEELING_FOR_GAPS(L) (L)->peeling_for_gaps
715 #define LOOP_VINFO_PEELING_FOR_NITER(L) (L)->peeling_for_niter
716 #define LOOP_VINFO_NO_DATA_DEPENDENCIES(L) (L)->no_data_dependencies
717 #define LOOP_VINFO_SCALAR_LOOP(L) (L)->scalar_loop
718 #define LOOP_VINFO_SCALAR_LOOP_SCALING(L) (L)->scalar_loop_scaling
719 #define LOOP_VINFO_HAS_MASK_STORE(L) (L)->has_mask_store
720 #define LOOP_VINFO_SCALAR_ITERATION_COST(L) (L)->scalar_cost_vec
721 #define LOOP_VINFO_SINGLE_SCALAR_ITERATION_COST(L) (L)->single_scalar_iteration_cost
722 #define LOOP_VINFO_ORIG_LOOP_INFO(L) (L)->orig_loop_info
723 #define LOOP_VINFO_SIMD_IF_COND(L) (L)->simd_if_cond
725 #define LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT(L) \
726 ((L)->may_misalign_stmts.length () > 0)
727 #define LOOP_REQUIRES_VERSIONING_FOR_ALIAS(L) \
728 ((L)->comp_alias_ddrs.length () > 0 \
729 || (L)->check_unequal_addrs.length () > 0 \
730 || (L)->lower_bounds.length () > 0)
731 #define LOOP_REQUIRES_VERSIONING_FOR_NITERS(L) \
732 (LOOP_VINFO_NITERS_ASSUMPTIONS (L))
733 #define LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND(L) \
734 (LOOP_VINFO_SIMD_IF_COND (L))
735 #define LOOP_REQUIRES_VERSIONING(L) \
736 (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (L) \
737 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (L) \
738 || LOOP_REQUIRES_VERSIONING_FOR_NITERS (L) \
739 || LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND (L))
741 #define LOOP_VINFO_NITERS_KNOWN_P(L) \
742 (tree_fits_shwi_p ((L)->num_iters) && tree_to_shwi ((L)->num_iters) > 0)
744 #define LOOP_VINFO_EPILOGUE_P(L) \
745 (LOOP_VINFO_ORIG_LOOP_INFO (L) != NULL)
747 #define LOOP_VINFO_ORIG_MAX_VECT_FACTOR(L) \
748 (LOOP_VINFO_MAX_VECT_FACTOR (LOOP_VINFO_ORIG_LOOP_INFO (L)))
750 /* Wrapper for loop_vec_info, for tracking success/failure, where a non-NULL
751 value signifies success, and a NULL value signifies failure, supporting
752 propagating an opt_problem * describing the failure back up the call
754 typedef opt_pointer_wrapper
<loop_vec_info
> opt_loop_vec_info
;
756 static inline loop_vec_info
757 loop_vec_info_for_loop (class loop
*loop
)
759 return (loop_vec_info
) loop
->aux
;
762 typedef class _bb_vec_info
: public vec_info
765 _bb_vec_info (gimple_stmt_iterator
, gimple_stmt_iterator
, vec_info_shared
*);
769 gimple_stmt_iterator region_begin
;
770 gimple_stmt_iterator region_end
;
773 #define BB_VINFO_BB(B) (B)->bb
774 #define BB_VINFO_GROUPED_STORES(B) (B)->grouped_stores
775 #define BB_VINFO_SLP_INSTANCES(B) (B)->slp_instances
776 #define BB_VINFO_DATAREFS(B) (B)->shared->datarefs
777 #define BB_VINFO_DDRS(B) (B)->shared->ddrs
778 #define BB_VINFO_TARGET_COST_DATA(B) (B)->target_cost_data
780 static inline bb_vec_info
781 vec_info_for_bb (basic_block bb
)
783 return (bb_vec_info
) bb
->aux
;
786 /*-----------------------------------------------------------------*/
787 /* Info on vectorized defs. */
788 /*-----------------------------------------------------------------*/
789 enum stmt_vec_info_type
{
790 undef_vec_info_type
= 0,
796 call_simd_clone_vec_info_type
,
797 assignment_vec_info_type
,
798 condition_vec_info_type
,
799 comparison_vec_info_type
,
802 type_promotion_vec_info_type
,
803 type_demotion_vec_info_type
,
804 type_conversion_vec_info_type
,
807 loop_exit_ctrl_vec_info_type
810 /* Indicates whether/how a variable is used in the scope of loop/basic
813 vect_unused_in_scope
= 0,
815 /* The def is only used outside the loop. */
817 /* The def is in the inner loop, and the use is in the outer loop, and the
818 use is a reduction stmt. */
819 vect_used_in_outer_by_reduction
,
820 /* The def is in the inner loop, and the use is in the outer loop (and is
821 not part of reduction). */
824 /* defs that feed computations that end up (only) in a reduction. These
825 defs may be used by non-reduction stmts, but eventually, any
826 computations/values that are affected by these defs are used to compute
827 a reduction (i.e. don't get stored to memory, for example). We use this
828 to identify computations that we can change the order in which they are
830 vect_used_by_reduction
,
835 /* The type of vectorization that can be applied to the stmt: regular loop-based
836 vectorization; pure SLP - the stmt is a part of SLP instances and does not
837 have uses outside SLP instances; or hybrid SLP and loop-based - the stmt is
838 a part of SLP instance and also must be loop-based vectorized, since it has
839 uses outside SLP sequences.
841 In the loop context the meanings of pure and hybrid SLP are slightly
842 different. By saying that pure SLP is applied to the loop, we mean that we
843 exploit only intra-iteration parallelism in the loop; i.e., the loop can be
844 vectorized without doing any conceptual unrolling, cause we don't pack
845 together stmts from different iterations, only within a single iteration.
846 Loop hybrid SLP means that we exploit both intra-iteration and
847 inter-iteration parallelism (e.g., number of elements in the vector is 4
848 and the slp-group-size is 2, in which case we don't have enough parallelism
849 within an iteration, so we obtain the rest of the parallelism from subsequent
850 iterations by unrolling the loop by 2). */
857 /* Says whether a statement is a load, a store of a vectorized statement
858 result, or a store of an invariant value. */
859 enum vec_load_store_type
{
865 /* Describes how we're going to vectorize an individual load or store,
866 or a group of loads or stores. */
867 enum vect_memory_access_type
{
868 /* An access to an invariant address. This is used only for loads. */
871 /* A simple contiguous access. */
874 /* A contiguous access that goes down in memory rather than up,
875 with no additional permutation. This is used only for stores
877 VMAT_CONTIGUOUS_DOWN
,
879 /* A simple contiguous access in which the elements need to be permuted
880 after loading or before storing. Only used for loop vectorization;
881 SLP uses separate permutes. */
882 VMAT_CONTIGUOUS_PERMUTE
,
884 /* A simple contiguous access in which the elements need to be reversed
885 after loading or before storing. */
886 VMAT_CONTIGUOUS_REVERSE
,
888 /* An access that uses IFN_LOAD_LANES or IFN_STORE_LANES. */
889 VMAT_LOAD_STORE_LANES
,
891 /* An access in which each scalar element is loaded or stored
895 /* A hybrid of VMAT_CONTIGUOUS and VMAT_ELEMENTWISE, used for grouped
896 SLP accesses. Each unrolled iteration uses a contiguous load
897 or store for the whole group, but the groups from separate iterations
898 are combined in the same way as for VMAT_ELEMENTWISE. */
901 /* The access uses gather loads or scatter stores. */
907 /* The data reference itself. */
909 /* The statement that contains the data reference. */
911 /* The misalignment in bytes of the reference, or -1 if not known. */
913 /* The byte alignment that we'd ideally like the reference to have,
914 and the value that misalignment is measured against. */
915 poly_uint64 target_alignment
;
916 /* If true the alignment of base_decl needs to be increased. */
917 bool base_misaligned
;
920 /* Stores current vectorized loop's offset. To be added to the DR's
921 offset to calculate current offset of data reference. */
925 typedef struct data_reference
*dr_p
;
927 class _stmt_vec_info
{
930 enum stmt_vec_info_type type
;
932 /* Indicates whether this stmts is part of a computation whose result is
933 used outside the loop. */
936 /* Stmt is part of some pattern (computation idiom) */
939 /* True if the statement was created during pattern recognition as
940 part of the replacement for RELATED_STMT. This implies that the
941 statement isn't part of any basic block, although for convenience
942 its gimple_bb is the same as for RELATED_STMT. */
945 /* Is this statement vectorizable or should it be skipped in (partial)
949 /* The stmt to which this info struct refers to. */
952 /* The vector type to be used for the LHS of this statement. */
955 /* The vectorized version of the stmt. */
956 stmt_vec_info vectorized_stmt
;
959 /* The following is relevant only for stmts that contain a non-scalar
960 data-ref (array/pointer/struct access). A GIMPLE stmt is expected to have
961 at most one such data-ref. */
965 /* Information about the data-ref relative to this loop
966 nest (the loop that is being considered for vectorization). */
967 innermost_loop_behavior dr_wrt_vec_loop
;
969 /* For loop PHI nodes, the base and evolution part of it. This makes sure
970 this information is still available in vect_update_ivs_after_vectorizer
971 where we may not be able to re-analyze the PHI nodes evolution as
972 peeling for the prologue loop can make it unanalyzable. The evolution
973 part is still correct after peeling, but the base may have changed from
975 tree loop_phi_evolution_base_unchanged
;
976 tree loop_phi_evolution_part
;
978 /* Used for various bookkeeping purposes, generally holding a pointer to
979 some other stmt S that is in some way "related" to this stmt.
980 Current use of this field is:
981 If this stmt is part of a pattern (i.e. the field 'in_pattern_p' is
982 true): S is the "pattern stmt" that represents (and replaces) the
983 sequence of stmts that constitutes the pattern. Similarly, the
984 related_stmt of the "pattern stmt" points back to this stmt (which is
985 the last stmt in the original sequence of stmts that constitutes the
987 stmt_vec_info related_stmt
;
989 /* Used to keep a sequence of def stmts of a pattern stmt if such exists.
990 The sequence is attached to the original statement rather than the
991 pattern statement. */
992 gimple_seq pattern_def_seq
;
994 /* List of datarefs that are known to have the same alignment as the dataref
996 vec
<dr_p
> same_align_refs
;
998 /* Selected SIMD clone's function info. First vector element
999 is SIMD clone's function decl, followed by a pair of trees (base + step)
1000 for linear arguments (pair of NULLs for other arguments). */
1001 vec
<tree
> simd_clone_info
;
1003 /* Classify the def of this stmt. */
1004 enum vect_def_type def_type
;
1006 /* Whether the stmt is SLPed, loop-based vectorized, or both. */
1007 enum slp_vect_type slp_type
;
1009 /* Interleaving and reduction chains info. */
1010 /* First element in the group. */
1011 stmt_vec_info first_element
;
1012 /* Pointer to the next element in the group. */
1013 stmt_vec_info next_element
;
1014 /* The size of the group. */
1016 /* For stores, number of stores from this group seen. We vectorize the last
1018 unsigned int store_count
;
1019 /* For loads only, the gap from the previous load. For consecutive loads, GAP
1023 /* The minimum negative dependence distance this stmt participates in
1025 unsigned int min_neg_dist
;
1027 /* Not all stmts in the loop need to be vectorized. e.g, the increment
1028 of the loop induction variable and computation of array indexes. relevant
1029 indicates whether the stmt needs to be vectorized. */
1030 enum vect_relevant relevant
;
1032 /* For loads if this is a gather, for stores if this is a scatter. */
1033 bool gather_scatter_p
;
1035 /* True if this is an access with loop-invariant stride. */
1038 /* For both loads and stores. */
1039 unsigned simd_lane_access_p
: 3;
1041 /* Classifies how the load or store is going to be implemented
1042 for loop vectorization. */
1043 vect_memory_access_type memory_access_type
;
1045 /* For INTEGER_INDUC_COND_REDUCTION, the initial value to be used. */
1046 tree induc_cond_initial_val
;
1048 /* If not NULL the value to be added to compute final reduction value. */
1049 tree reduc_epilogue_adjustment
;
1051 /* On a reduction PHI the reduction type as detected by
1052 vect_is_simple_reduction and vectorizable_reduction. */
1053 enum vect_reduction_type reduc_type
;
1055 /* The original reduction code, to be used in the epilogue. */
1056 enum tree_code reduc_code
;
1057 /* An internal function we should use in the epilogue. */
1058 internal_fn reduc_fn
;
1060 /* On a stmt participating in the reduction the index of the operand
1061 on the reduction SSA cycle. */
1064 /* On a reduction PHI the def returned by vect_force_simple_reduction.
1065 On the def returned by vect_force_simple_reduction the
1066 corresponding PHI. */
1067 stmt_vec_info reduc_def
;
1069 /* The vector input type relevant for reduction vectorization. */
1070 tree reduc_vectype_in
;
1072 /* The vector type for performing the actual reduction. */
1075 /* Whether we force a single cycle PHI during reduction vectorization. */
1076 bool force_single_cycle
;
1078 /* Whether on this stmt reduction meta is recorded. */
1081 /* The number of scalar stmt references from active SLP instances. */
1082 unsigned int num_slp_uses
;
1084 /* If nonzero, the lhs of the statement could be truncated to this
1085 many bits without affecting any users of the result. */
1086 unsigned int min_output_precision
;
1088 /* If nonzero, all non-boolean input operands have the same precision,
1089 and they could each be truncated to this many bits without changing
1091 unsigned int min_input_precision
;
1093 /* If OPERATION_BITS is nonzero, the statement could be performed on
1094 an integer with the sign and number of bits given by OPERATION_SIGN
1095 and OPERATION_BITS without changing the result. */
1096 unsigned int operation_precision
;
1097 signop operation_sign
;
1099 /* If the statement produces a boolean result, this value describes
1100 how we should choose the associated vector type. The possible
1103 - an integer precision N if we should use the vector mask type
1104 associated with N-bit integers. This is only used if all relevant
1105 input booleans also want the vector mask type for N-bit integers,
1106 or if we can convert them into that form by pattern-matching.
1108 - ~0U if we considered choosing a vector mask type but decided
1109 to treat the boolean as a normal integer type instead.
1111 - 0 otherwise. This means either that the operation isn't one that
1112 could have a vector mask type (and so should have a normal vector
1113 type instead) or that we simply haven't made a choice either way. */
1114 unsigned int mask_precision
;
1116 /* True if this is only suitable for SLP vectorization. */
1117 bool slp_vect_only_p
;
1120 /* Information about a gather/scatter call. */
1121 struct gather_scatter_info
{
1122 /* The internal function to use for the gather/scatter operation,
1123 or IFN_LAST if a built-in function should be used instead. */
1126 /* The FUNCTION_DECL for the built-in gather/scatter function,
1127 or null if an internal function should be used instead. */
1130 /* The loop-invariant base value. */
1133 /* The original scalar offset, which is a non-loop-invariant SSA_NAME. */
1136 /* Each offset element should be multiplied by this amount before
1137 being added to the base. */
1140 /* The definition type for the vectorized offset. */
1141 enum vect_def_type offset_dt
;
1143 /* The type of the vectorized offset. */
1144 tree offset_vectype
;
1146 /* The type of the scalar elements after loading or before storing. */
1149 /* The type of the scalar elements being loaded or stored. */
1153 /* Access Functions. */
1154 #define STMT_VINFO_TYPE(S) (S)->type
1155 #define STMT_VINFO_STMT(S) (S)->stmt
1156 #define STMT_VINFO_RELEVANT(S) (S)->relevant
1157 #define STMT_VINFO_LIVE_P(S) (S)->live
1158 #define STMT_VINFO_VECTYPE(S) (S)->vectype
1159 #define STMT_VINFO_VEC_STMT(S) (S)->vectorized_stmt
1160 #define STMT_VINFO_VECTORIZABLE(S) (S)->vectorizable
1161 #define STMT_VINFO_DATA_REF(S) ((S)->dr_aux.dr + 0)
1162 #define STMT_VINFO_GATHER_SCATTER_P(S) (S)->gather_scatter_p
1163 #define STMT_VINFO_STRIDED_P(S) (S)->strided_p
1164 #define STMT_VINFO_MEMORY_ACCESS_TYPE(S) (S)->memory_access_type
1165 #define STMT_VINFO_SIMD_LANE_ACCESS_P(S) (S)->simd_lane_access_p
1166 #define STMT_VINFO_VEC_INDUC_COND_INITIAL_VAL(S) (S)->induc_cond_initial_val
1167 #define STMT_VINFO_REDUC_EPILOGUE_ADJUSTMENT(S) (S)->reduc_epilogue_adjustment
1168 #define STMT_VINFO_REDUC_IDX(S) (S)->reduc_idx
1169 #define STMT_VINFO_FORCE_SINGLE_CYCLE(S) (S)->force_single_cycle
1171 #define STMT_VINFO_DR_WRT_VEC_LOOP(S) (S)->dr_wrt_vec_loop
1172 #define STMT_VINFO_DR_BASE_ADDRESS(S) (S)->dr_wrt_vec_loop.base_address
1173 #define STMT_VINFO_DR_INIT(S) (S)->dr_wrt_vec_loop.init
1174 #define STMT_VINFO_DR_OFFSET(S) (S)->dr_wrt_vec_loop.offset
1175 #define STMT_VINFO_DR_STEP(S) (S)->dr_wrt_vec_loop.step
1176 #define STMT_VINFO_DR_BASE_ALIGNMENT(S) (S)->dr_wrt_vec_loop.base_alignment
1177 #define STMT_VINFO_DR_BASE_MISALIGNMENT(S) \
1178 (S)->dr_wrt_vec_loop.base_misalignment
1179 #define STMT_VINFO_DR_OFFSET_ALIGNMENT(S) \
1180 (S)->dr_wrt_vec_loop.offset_alignment
1181 #define STMT_VINFO_DR_STEP_ALIGNMENT(S) \
1182 (S)->dr_wrt_vec_loop.step_alignment
1184 #define STMT_VINFO_DR_INFO(S) \
1185 (gcc_checking_assert ((S)->dr_aux.stmt == (S)), &(S)->dr_aux)
1187 #define STMT_VINFO_IN_PATTERN_P(S) (S)->in_pattern_p
1188 #define STMT_VINFO_RELATED_STMT(S) (S)->related_stmt
1189 #define STMT_VINFO_PATTERN_DEF_SEQ(S) (S)->pattern_def_seq
1190 #define STMT_VINFO_SAME_ALIGN_REFS(S) (S)->same_align_refs
1191 #define STMT_VINFO_SIMD_CLONE_INFO(S) (S)->simd_clone_info
1192 #define STMT_VINFO_DEF_TYPE(S) (S)->def_type
1193 #define STMT_VINFO_GROUPED_ACCESS(S) \
1194 ((S)->dr_aux.dr && DR_GROUP_FIRST_ELEMENT(S))
1195 #define STMT_VINFO_LOOP_PHI_EVOLUTION_BASE_UNCHANGED(S) (S)->loop_phi_evolution_base_unchanged
1196 #define STMT_VINFO_LOOP_PHI_EVOLUTION_PART(S) (S)->loop_phi_evolution_part
1197 #define STMT_VINFO_MIN_NEG_DIST(S) (S)->min_neg_dist
1198 #define STMT_VINFO_NUM_SLP_USES(S) (S)->num_slp_uses
1199 #define STMT_VINFO_REDUC_TYPE(S) (S)->reduc_type
1200 #define STMT_VINFO_REDUC_CODE(S) (S)->reduc_code
1201 #define STMT_VINFO_REDUC_FN(S) (S)->reduc_fn
1202 #define STMT_VINFO_REDUC_DEF(S) (S)->reduc_def
1203 #define STMT_VINFO_REDUC_VECTYPE(S) (S)->reduc_vectype
1204 #define STMT_VINFO_REDUC_VECTYPE_IN(S) (S)->reduc_vectype_in
1205 #define STMT_VINFO_SLP_VECT_ONLY(S) (S)->slp_vect_only_p
1207 #define DR_GROUP_FIRST_ELEMENT(S) \
1208 (gcc_checking_assert ((S)->dr_aux.dr), (S)->first_element)
1209 #define DR_GROUP_NEXT_ELEMENT(S) \
1210 (gcc_checking_assert ((S)->dr_aux.dr), (S)->next_element)
1211 #define DR_GROUP_SIZE(S) \
1212 (gcc_checking_assert ((S)->dr_aux.dr), (S)->size)
1213 #define DR_GROUP_STORE_COUNT(S) \
1214 (gcc_checking_assert ((S)->dr_aux.dr), (S)->store_count)
1215 #define DR_GROUP_GAP(S) \
1216 (gcc_checking_assert ((S)->dr_aux.dr), (S)->gap)
1218 #define REDUC_GROUP_FIRST_ELEMENT(S) \
1219 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->first_element)
1220 #define REDUC_GROUP_NEXT_ELEMENT(S) \
1221 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->next_element)
1222 #define REDUC_GROUP_SIZE(S) \
1223 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->size)
1225 #define STMT_VINFO_RELEVANT_P(S) ((S)->relevant != vect_unused_in_scope)
1227 #define HYBRID_SLP_STMT(S) ((S)->slp_type == hybrid)
1228 #define PURE_SLP_STMT(S) ((S)->slp_type == pure_slp)
1229 #define STMT_SLP_TYPE(S) (S)->slp_type
1231 #define VECT_MAX_COST 1000
1233 /* The maximum number of intermediate steps required in multi-step type
1235 #define MAX_INTERM_CVT_STEPS 3
1237 #define MAX_VECTORIZATION_FACTOR INT_MAX
1239 /* Nonzero if TYPE represents a (scalar) boolean type or type
1240 in the middle-end compatible with it (unsigned precision 1 integral
1241 types). Used to determine which types should be vectorized as
1242 VECTOR_BOOLEAN_TYPE_P. */
1244 #define VECT_SCALAR_BOOLEAN_TYPE_P(TYPE) \
1245 (TREE_CODE (TYPE) == BOOLEAN_TYPE \
1246 || ((TREE_CODE (TYPE) == INTEGER_TYPE \
1247 || TREE_CODE (TYPE) == ENUMERAL_TYPE) \
1248 && TYPE_PRECISION (TYPE) == 1 \
1249 && TYPE_UNSIGNED (TYPE)))
1252 nested_in_vect_loop_p (class loop
*loop
, stmt_vec_info stmt_info
)
1255 && (loop
->inner
== (gimple_bb (stmt_info
->stmt
))->loop_father
));
1258 /* Return true if STMT_INFO should produce a vector mask type rather than
1259 a normal nonmask type. */
1262 vect_use_mask_type_p (stmt_vec_info stmt_info
)
1264 return stmt_info
->mask_precision
&& stmt_info
->mask_precision
!= ~0U;
1267 /* Return TRUE if a statement represented by STMT_INFO is a part of a
1271 is_pattern_stmt_p (stmt_vec_info stmt_info
)
1273 return stmt_info
->pattern_stmt_p
;
1276 /* If STMT_INFO is a pattern statement, return the statement that it
1277 replaces, otherwise return STMT_INFO itself. */
1279 inline stmt_vec_info
1280 vect_orig_stmt (stmt_vec_info stmt_info
)
1282 if (is_pattern_stmt_p (stmt_info
))
1283 return STMT_VINFO_RELATED_STMT (stmt_info
);
1287 /* Return the later statement between STMT1_INFO and STMT2_INFO. */
1289 static inline stmt_vec_info
1290 get_later_stmt (stmt_vec_info stmt1_info
, stmt_vec_info stmt2_info
)
1292 if (gimple_uid (vect_orig_stmt (stmt1_info
)->stmt
)
1293 > gimple_uid (vect_orig_stmt (stmt2_info
)->stmt
))
1299 /* If STMT_INFO has been replaced by a pattern statement, return the
1300 replacement statement, otherwise return STMT_INFO itself. */
1302 inline stmt_vec_info
1303 vect_stmt_to_vectorize (stmt_vec_info stmt_info
)
1305 if (STMT_VINFO_IN_PATTERN_P (stmt_info
))
1306 return STMT_VINFO_RELATED_STMT (stmt_info
);
1310 /* Return true if BB is a loop header. */
1313 is_loop_header_bb_p (basic_block bb
)
1315 if (bb
== (bb
->loop_father
)->header
)
1317 gcc_checking_assert (EDGE_COUNT (bb
->preds
) == 1);
1321 /* Return pow2 (X). */
1328 for (i
= 0; i
< x
; i
++)
1334 /* Alias targetm.vectorize.builtin_vectorization_cost. */
1337 builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost
,
1338 tree vectype
, int misalign
)
1340 return targetm
.vectorize
.builtin_vectorization_cost (type_of_cost
,
1344 /* Get cost by calling cost target builtin. */
1347 int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost
)
1349 return builtin_vectorization_cost (type_of_cost
, NULL
, 0);
1352 /* Alias targetm.vectorize.init_cost. */
1354 static inline void *
1355 init_cost (class loop
*loop_info
)
1357 return targetm
.vectorize
.init_cost (loop_info
);
1360 extern void dump_stmt_cost (FILE *, void *, int, enum vect_cost_for_stmt
,
1361 stmt_vec_info
, tree
, int, unsigned,
1362 enum vect_cost_model_location
);
1364 /* Alias targetm.vectorize.add_stmt_cost. */
1366 static inline unsigned
1367 add_stmt_cost (vec_info
*vinfo
, void *data
, int count
,
1368 enum vect_cost_for_stmt kind
,
1369 stmt_vec_info stmt_info
, tree vectype
, int misalign
,
1370 enum vect_cost_model_location where
)
1372 unsigned cost
= targetm
.vectorize
.add_stmt_cost (vinfo
, data
, count
, kind
,
1375 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1376 dump_stmt_cost (dump_file
, data
, count
, kind
, stmt_info
, vectype
, misalign
,
1381 /* Alias targetm.vectorize.finish_cost. */
1384 finish_cost (void *data
, unsigned *prologue_cost
,
1385 unsigned *body_cost
, unsigned *epilogue_cost
)
1387 targetm
.vectorize
.finish_cost (data
, prologue_cost
, body_cost
, epilogue_cost
);
1390 /* Alias targetm.vectorize.destroy_cost_data. */
1393 destroy_cost_data (void *data
)
1395 targetm
.vectorize
.destroy_cost_data (data
);
1399 add_stmt_costs (vec_info
*vinfo
, void *data
, stmt_vector_for_cost
*cost_vec
)
1401 stmt_info_for_cost
*cost
;
1403 FOR_EACH_VEC_ELT (*cost_vec
, i
, cost
)
1404 add_stmt_cost (vinfo
, data
, cost
->count
, cost
->kind
, cost
->stmt_info
,
1405 cost
->vectype
, cost
->misalign
, cost
->where
);
1408 /*-----------------------------------------------------------------*/
1409 /* Info on data references alignment. */
1410 /*-----------------------------------------------------------------*/
1411 #define DR_MISALIGNMENT_UNKNOWN (-1)
1412 #define DR_MISALIGNMENT_UNINITIALIZED (-2)
1415 set_dr_misalignment (dr_vec_info
*dr_info
, int val
)
1417 dr_info
->misalignment
= val
;
1421 dr_misalignment (dr_vec_info
*dr_info
)
1423 int misalign
= dr_info
->misalignment
;
1424 gcc_assert (misalign
!= DR_MISALIGNMENT_UNINITIALIZED
);
1428 /* Reflects actual alignment of first access in the vectorized loop,
1429 taking into account peeling/versioning if applied. */
1430 #define DR_MISALIGNMENT(DR) dr_misalignment (DR)
1431 #define SET_DR_MISALIGNMENT(DR, VAL) set_dr_misalignment (DR, VAL)
1433 /* Only defined once DR_MISALIGNMENT is defined. */
1434 #define DR_TARGET_ALIGNMENT(DR) ((DR)->target_alignment)
1436 /* Return true if data access DR_INFO is aligned to its target alignment
1437 (which may be less than a full vector). */
1440 aligned_access_p (dr_vec_info
*dr_info
)
1442 return (DR_MISALIGNMENT (dr_info
) == 0);
1445 /* Return TRUE if the alignment of the data access is known, and FALSE
1449 known_alignment_for_access_p (dr_vec_info
*dr_info
)
1451 return (DR_MISALIGNMENT (dr_info
) != DR_MISALIGNMENT_UNKNOWN
);
1454 /* Return the minimum alignment in bytes that the vectorized version
1455 of DR_INFO is guaranteed to have. */
1457 static inline unsigned int
1458 vect_known_alignment_in_bytes (dr_vec_info
*dr_info
)
1460 if (DR_MISALIGNMENT (dr_info
) == DR_MISALIGNMENT_UNKNOWN
)
1461 return TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr_info
->dr
)));
1462 if (DR_MISALIGNMENT (dr_info
) == 0)
1463 return known_alignment (DR_TARGET_ALIGNMENT (dr_info
));
1464 return DR_MISALIGNMENT (dr_info
) & -DR_MISALIGNMENT (dr_info
);
1467 /* Return the behavior of DR_INFO with respect to the vectorization context
1468 (which for outer loop vectorization might not be the behavior recorded
1469 in DR_INFO itself). */
1471 static inline innermost_loop_behavior
*
1472 vect_dr_behavior (vec_info
*vinfo
, dr_vec_info
*dr_info
)
1474 stmt_vec_info stmt_info
= dr_info
->stmt
;
1475 loop_vec_info loop_vinfo
= dyn_cast
<loop_vec_info
> (vinfo
);
1476 if (loop_vinfo
== NULL
1477 || !nested_in_vect_loop_p (LOOP_VINFO_LOOP (loop_vinfo
), stmt_info
))
1478 return &DR_INNERMOST (dr_info
->dr
);
1480 return &STMT_VINFO_DR_WRT_VEC_LOOP (stmt_info
);
1483 /* Return the offset calculated by adding the offset of this DR_INFO to the
1484 corresponding data_reference's offset. If CHECK_OUTER then use
1485 vect_dr_behavior to select the appropriate data_reference to use. */
1488 get_dr_vinfo_offset (vec_info
*vinfo
,
1489 dr_vec_info
*dr_info
, bool check_outer
= false)
1491 innermost_loop_behavior
*base
;
1493 base
= vect_dr_behavior (vinfo
, dr_info
);
1495 base
= &dr_info
->dr
->innermost
;
1497 tree offset
= base
->offset
;
1499 if (!dr_info
->offset
)
1502 offset
= fold_convert (sizetype
, offset
);
1503 return fold_build2 (PLUS_EXPR
, TREE_TYPE (dr_info
->offset
), offset
,
1508 /* Return true if the vect cost model is unlimited. */
1510 unlimited_cost_model (loop_p loop
)
1512 if (loop
!= NULL
&& loop
->force_vectorize
1513 && flag_simd_cost_model
!= VECT_COST_MODEL_DEFAULT
)
1514 return flag_simd_cost_model
== VECT_COST_MODEL_UNLIMITED
;
1515 return (flag_vect_cost_model
== VECT_COST_MODEL_UNLIMITED
);
1518 /* Return true if the loop described by LOOP_VINFO is fully-masked and
1519 if the first iteration should use a partial mask in order to achieve
1523 vect_use_loop_mask_for_alignment_p (loop_vec_info loop_vinfo
)
1525 return (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo
)
1526 && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo
));
1529 /* Return the number of vectors of type VECTYPE that are needed to get
1530 NUNITS elements. NUNITS should be based on the vectorization factor,
1531 so it is always a known multiple of the number of elements in VECTYPE. */
1533 static inline unsigned int
1534 vect_get_num_vectors (poly_uint64 nunits
, tree vectype
)
1536 return exact_div (nunits
, TYPE_VECTOR_SUBPARTS (vectype
)).to_constant ();
1539 /* Return the number of copies needed for loop vectorization when
1540 a statement operates on vectors of type VECTYPE. This is the
1541 vectorization factor divided by the number of elements in
1542 VECTYPE and is always known at compile time. */
1544 static inline unsigned int
1545 vect_get_num_copies (loop_vec_info loop_vinfo
, tree vectype
)
1547 return vect_get_num_vectors (LOOP_VINFO_VECT_FACTOR (loop_vinfo
), vectype
);
1550 /* Update maximum unit count *MAX_NUNITS so that it accounts for
1551 NUNITS. *MAX_NUNITS can be 1 if we haven't yet recorded anything. */
1554 vect_update_max_nunits (poly_uint64
*max_nunits
, poly_uint64 nunits
)
1556 /* All unit counts have the form vec_info::vector_size * X for some
1557 rational X, so two unit sizes must have a common multiple.
1558 Everything is a multiple of the initial value of 1. */
1559 *max_nunits
= force_common_multiple (*max_nunits
, nunits
);
1562 /* Update maximum unit count *MAX_NUNITS so that it accounts for
1563 the number of units in vector type VECTYPE. *MAX_NUNITS can be 1
1564 if we haven't yet recorded any vector types. */
1567 vect_update_max_nunits (poly_uint64
*max_nunits
, tree vectype
)
1569 vect_update_max_nunits (max_nunits
, TYPE_VECTOR_SUBPARTS (vectype
));
1572 /* Return the vectorization factor that should be used for costing
1573 purposes while vectorizing the loop described by LOOP_VINFO.
1574 Pick a reasonable estimate if the vectorization factor isn't
1575 known at compile time. */
1577 static inline unsigned int
1578 vect_vf_for_cost (loop_vec_info loop_vinfo
)
1580 return estimated_poly_value (LOOP_VINFO_VECT_FACTOR (loop_vinfo
));
1583 /* Estimate the number of elements in VEC_TYPE for costing purposes.
1584 Pick a reasonable estimate if the exact number isn't known at
1587 static inline unsigned int
1588 vect_nunits_for_cost (tree vec_type
)
1590 return estimated_poly_value (TYPE_VECTOR_SUBPARTS (vec_type
));
1593 /* Return the maximum possible vectorization factor for LOOP_VINFO. */
1595 static inline unsigned HOST_WIDE_INT
1596 vect_max_vf (loop_vec_info loop_vinfo
)
1598 unsigned HOST_WIDE_INT vf
;
1599 if (LOOP_VINFO_VECT_FACTOR (loop_vinfo
).is_constant (&vf
))
1601 return MAX_VECTORIZATION_FACTOR
;
1604 /* Return the size of the value accessed by unvectorized data reference
1605 DR_INFO. This is only valid once STMT_VINFO_VECTYPE has been calculated
1606 for the associated gimple statement, since that guarantees that DR_INFO
1607 accesses either a scalar or a scalar equivalent. ("Scalar equivalent"
1608 here includes things like V1SI, which can be vectorized in the same way
1612 vect_get_scalar_dr_size (dr_vec_info
*dr_info
)
1614 return tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_info
->dr
))));
1617 /* Return true if LOOP_VINFO requires a runtime check for whether the
1618 vector loop is profitable. */
1621 vect_apply_runtime_profitability_check_p (loop_vec_info loop_vinfo
)
1623 unsigned int th
= LOOP_VINFO_COST_MODEL_THRESHOLD (loop_vinfo
);
1624 return (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
)
1625 && th
>= vect_vf_for_cost (loop_vinfo
));
1628 /* Source location + hotness information. */
1629 extern dump_user_location_t vect_location
;
1631 /* A macro for calling:
1632 dump_begin_scope (MSG, vect_location);
1633 via an RAII object, thus printing "=== MSG ===\n" to the dumpfile etc,
1636 once the object goes out of scope, thus capturing the nesting of
1639 These scopes affect dump messages within them: dump messages at the
1640 top level implicitly default to MSG_PRIORITY_USER_FACING, whereas those
1641 in a nested scope implicitly default to MSG_PRIORITY_INTERNALS. */
1643 #define DUMP_VECT_SCOPE(MSG) \
1644 AUTO_DUMP_SCOPE (MSG, vect_location)
1646 /* A sentinel class for ensuring that the "vect_location" global gets
1647 reset at the end of a scope.
1649 The "vect_location" global is used during dumping and contains a
1650 location_t, which could contain references to a tree block via the
1651 ad-hoc data. This data is used for tracking inlining information,
1652 but it's not a GC root; it's simply assumed that such locations never
1653 get accessed if the blocks are optimized away.
1655 Hence we need to ensure that such locations are purged at the end
1656 of any operations using them (e.g. via this class). */
1658 class auto_purge_vect_location
1661 ~auto_purge_vect_location ();
1664 /*-----------------------------------------------------------------*/
1665 /* Function prototypes. */
1666 /*-----------------------------------------------------------------*/
1668 /* Simple loop peeling and versioning utilities for vectorizer's purposes -
1669 in tree-vect-loop-manip.c. */
1670 extern void vect_set_loop_condition (class loop
*, loop_vec_info
,
1671 tree
, tree
, tree
, bool);
1672 extern bool slpeel_can_duplicate_loop_p (const class loop
*, const_edge
);
1673 class loop
*slpeel_tree_duplicate_loop_to_edge_cfg (class loop
*,
1674 class loop
*, edge
);
1675 class loop
*vect_loop_versioning (loop_vec_info
, gimple
*);
1676 extern class loop
*vect_do_peeling (loop_vec_info
, tree
, tree
,
1677 tree
*, tree
*, tree
*, int, bool, bool,
1679 extern void vect_prepare_for_masked_peels (loop_vec_info
);
1680 extern dump_user_location_t
find_loop_location (class loop
*);
1681 extern bool vect_can_advance_ivs_p (loop_vec_info
);
1682 extern void vect_update_inits_of_drs (loop_vec_info
, tree
, tree_code
);
1684 /* In tree-vect-stmts.c. */
1685 extern tree
get_related_vectype_for_scalar_type (machine_mode
, tree
,
1687 extern tree
get_vectype_for_scalar_type (vec_info
*, tree
, unsigned int = 0);
1688 extern tree
get_vectype_for_scalar_type (vec_info
*, tree
, slp_tree
);
1689 extern tree
get_mask_type_for_scalar_type (vec_info
*, tree
, unsigned int = 0);
1690 extern tree
get_same_sized_vectype (tree
, tree
);
1691 extern bool vect_chooses_same_modes_p (vec_info
*, machine_mode
);
1692 extern bool vect_get_loop_mask_type (loop_vec_info
);
1693 extern bool vect_is_simple_use (tree
, vec_info
*, enum vect_def_type
*,
1694 stmt_vec_info
* = NULL
, gimple
** = NULL
);
1695 extern bool vect_is_simple_use (tree
, vec_info
*, enum vect_def_type
*,
1696 tree
*, stmt_vec_info
* = NULL
,
1698 extern bool supportable_widening_operation (vec_info
*,
1699 enum tree_code
, stmt_vec_info
,
1700 tree
, tree
, enum tree_code
*,
1701 enum tree_code
*, int *,
1703 extern bool supportable_narrowing_operation (enum tree_code
, tree
, tree
,
1704 enum tree_code
*, int *,
1707 extern unsigned record_stmt_cost (stmt_vector_for_cost
*, int,
1708 enum vect_cost_for_stmt
, stmt_vec_info
,
1709 tree
, int, enum vect_cost_model_location
);
1711 /* Overload of record_stmt_cost with VECTYPE derived from STMT_INFO. */
1713 static inline unsigned
1714 record_stmt_cost (stmt_vector_for_cost
*body_cost_vec
, int count
,
1715 enum vect_cost_for_stmt kind
, stmt_vec_info stmt_info
,
1716 int misalign
, enum vect_cost_model_location where
)
1718 return record_stmt_cost (body_cost_vec
, count
, kind
, stmt_info
,
1719 STMT_VINFO_VECTYPE (stmt_info
), misalign
, where
);
1722 extern stmt_vec_info
vect_finish_replace_stmt (vec_info
*,
1723 stmt_vec_info
, gimple
*);
1724 extern stmt_vec_info
vect_finish_stmt_generation (vec_info
*,
1725 stmt_vec_info
, gimple
*,
1726 gimple_stmt_iterator
*);
1727 extern opt_result
vect_mark_stmts_to_be_vectorized (loop_vec_info
, bool *);
1728 extern tree
vect_get_store_rhs (stmt_vec_info
);
1729 extern tree
vect_get_vec_def_for_operand_1 (stmt_vec_info
, enum vect_def_type
);
1730 extern tree
vect_get_vec_def_for_operand (vec_info
*, tree
,
1731 stmt_vec_info
, tree
= NULL
);
1732 extern void vect_get_vec_defs (vec_info
*, tree
, tree
, stmt_vec_info
,
1733 vec
<tree
> *, vec
<tree
> *, slp_tree
);
1734 extern void vect_get_vec_defs_for_stmt_copy (vec_info
*,
1735 vec
<tree
> *, vec
<tree
> *);
1736 extern tree
vect_init_vector (vec_info
*, stmt_vec_info
, tree
, tree
,
1737 gimple_stmt_iterator
*);
1738 extern tree
vect_get_vec_def_for_stmt_copy (vec_info
*, tree
);
1739 extern bool vect_transform_stmt (vec_info
*, stmt_vec_info
,
1740 gimple_stmt_iterator
*,
1741 slp_tree
, slp_instance
);
1742 extern void vect_remove_stores (vec_info
*, stmt_vec_info
);
1743 extern bool vect_nop_conversion_p (stmt_vec_info
);
1744 extern opt_result
vect_analyze_stmt (vec_info
*, stmt_vec_info
, bool *,
1746 slp_instance
, stmt_vector_for_cost
*);
1747 extern void vect_get_load_cost (vec_info
*, stmt_vec_info
, int, bool,
1748 unsigned int *, unsigned int *,
1749 stmt_vector_for_cost
*,
1750 stmt_vector_for_cost
*, bool);
1751 extern void vect_get_store_cost (vec_info
*, stmt_vec_info
, int,
1752 unsigned int *, stmt_vector_for_cost
*);
1753 extern bool vect_supportable_shift (vec_info
*, enum tree_code
, tree
);
1754 extern tree
vect_gen_perm_mask_any (tree
, const vec_perm_indices
&);
1755 extern tree
vect_gen_perm_mask_checked (tree
, const vec_perm_indices
&);
1756 extern void optimize_mask_stores (class loop
*);
1757 extern gcall
*vect_gen_while (tree
, tree
, tree
);
1758 extern tree
vect_gen_while_not (gimple_seq
*, tree
, tree
, tree
);
1759 extern opt_result
vect_get_vector_types_for_stmt (vec_info
*,
1760 stmt_vec_info
, tree
*,
1761 tree
*, unsigned int = 0);
1762 extern opt_tree
vect_get_mask_type_for_stmt (stmt_vec_info
, unsigned int = 0);
1764 /* In tree-vect-data-refs.c. */
1765 extern bool vect_can_force_dr_alignment_p (const_tree
, poly_uint64
);
1766 extern enum dr_alignment_support vect_supportable_dr_alignment
1767 (vec_info
*, dr_vec_info
*, bool);
1768 extern tree
vect_get_smallest_scalar_type (stmt_vec_info
, HOST_WIDE_INT
*,
1770 extern opt_result
vect_analyze_data_ref_dependences (loop_vec_info
, unsigned int *);
1771 extern bool vect_slp_analyze_instance_dependence (vec_info
*, slp_instance
);
1772 extern opt_result
vect_enhance_data_refs_alignment (loop_vec_info
);
1773 extern opt_result
vect_analyze_data_refs_alignment (loop_vec_info
);
1774 extern opt_result
vect_verify_datarefs_alignment (loop_vec_info
);
1775 extern bool vect_slp_analyze_and_verify_instance_alignment (vec_info
*,
1777 extern opt_result
vect_analyze_data_ref_accesses (vec_info
*);
1778 extern opt_result
vect_prune_runtime_alias_test_list (loop_vec_info
);
1779 extern bool vect_gather_scatter_fn_p (vec_info
*, bool, bool, tree
, tree
,
1780 tree
, int, internal_fn
*, tree
*);
1781 extern bool vect_check_gather_scatter (stmt_vec_info
, loop_vec_info
,
1782 gather_scatter_info
*);
1783 extern opt_result
vect_find_stmt_data_reference (loop_p
, gimple
*,
1784 vec
<data_reference_p
> *);
1785 extern opt_result
vect_analyze_data_refs (vec_info
*, poly_uint64
*, bool *);
1786 extern void vect_record_base_alignments (vec_info
*);
1787 extern tree
vect_create_data_ref_ptr (vec_info
*,
1788 stmt_vec_info
, tree
, class loop
*, tree
,
1789 tree
*, gimple_stmt_iterator
*,
1791 tree
= NULL_TREE
, tree
= NULL_TREE
);
1792 extern tree
bump_vector_ptr (vec_info
*, tree
, gimple
*, gimple_stmt_iterator
*,
1793 stmt_vec_info
, tree
);
1794 extern void vect_copy_ref_info (tree
, tree
);
1795 extern tree
vect_create_destination_var (tree
, tree
);
1796 extern bool vect_grouped_store_supported (tree
, unsigned HOST_WIDE_INT
);
1797 extern bool vect_store_lanes_supported (tree
, unsigned HOST_WIDE_INT
, bool);
1798 extern bool vect_grouped_load_supported (tree
, bool, unsigned HOST_WIDE_INT
);
1799 extern bool vect_load_lanes_supported (tree
, unsigned HOST_WIDE_INT
, bool);
1800 extern void vect_permute_store_chain (vec_info
*,
1801 vec
<tree
> ,unsigned int, stmt_vec_info
,
1802 gimple_stmt_iterator
*, vec
<tree
> *);
1803 extern tree
vect_setup_realignment (vec_info
*,
1804 stmt_vec_info
, gimple_stmt_iterator
*,
1805 tree
*, enum dr_alignment_support
, tree
,
1807 extern void vect_transform_grouped_load (vec_info
*, stmt_vec_info
, vec
<tree
>,
1808 int, gimple_stmt_iterator
*);
1809 extern void vect_record_grouped_load_vectors (vec_info
*,
1810 stmt_vec_info
, vec
<tree
>);
1811 extern tree
vect_get_new_vect_var (tree
, enum vect_var_kind
, const char *);
1812 extern tree
vect_get_new_ssa_name (tree
, enum vect_var_kind
,
1813 const char * = NULL
);
1814 extern tree
vect_create_addr_base_for_vector_ref (vec_info
*,
1815 stmt_vec_info
, gimple_seq
*,
1816 tree
, tree
= NULL_TREE
);
1818 /* In tree-vect-loop.c. */
1819 extern widest_int
vect_iv_limit_for_full_masking (loop_vec_info loop_vinfo
);
1820 /* Used in tree-vect-loop-manip.c */
1821 extern void determine_peel_for_niter (loop_vec_info
);
1822 /* Used in gimple-loop-interchange.c and tree-parloops.c. */
1823 extern bool check_reduction_path (dump_user_location_t
, loop_p
, gphi
*, tree
,
1825 extern bool needs_fold_left_reduction_p (tree
, tree_code
);
1826 /* Drive for loop analysis stage. */
1827 extern opt_loop_vec_info
vect_analyze_loop (class loop
*, vec_info_shared
*);
1828 extern tree
vect_build_loop_niters (loop_vec_info
, bool * = NULL
);
1829 extern void vect_gen_vector_loop_niters (loop_vec_info
, tree
, tree
*,
1831 extern tree
vect_halve_mask_nunits (tree
, machine_mode
);
1832 extern tree
vect_double_mask_nunits (tree
, machine_mode
);
1833 extern void vect_record_loop_mask (loop_vec_info
, vec_loop_masks
*,
1834 unsigned int, tree
, tree
);
1835 extern tree
vect_get_loop_mask (gimple_stmt_iterator
*, vec_loop_masks
*,
1836 unsigned int, tree
, unsigned int);
1837 extern stmt_vec_info
info_for_reduction (vec_info
*, stmt_vec_info
);
1839 /* Drive for loop transformation stage. */
1840 extern class loop
*vect_transform_loop (loop_vec_info
, gimple
*);
1841 extern opt_loop_vec_info
vect_analyze_loop_form (class loop
*,
1843 extern bool vectorizable_live_operation (loop_vec_info
,
1844 stmt_vec_info
, gimple_stmt_iterator
*,
1845 slp_tree
, slp_instance
, int,
1846 bool, stmt_vector_for_cost
*);
1847 extern bool vectorizable_reduction (loop_vec_info
, stmt_vec_info
,
1848 slp_tree
, slp_instance
,
1849 stmt_vector_for_cost
*);
1850 extern bool vectorizable_induction (loop_vec_info
, stmt_vec_info
,
1851 gimple_stmt_iterator
*,
1852 stmt_vec_info
*, slp_tree
,
1853 stmt_vector_for_cost
*);
1854 extern bool vect_transform_reduction (loop_vec_info
, stmt_vec_info
,
1855 gimple_stmt_iterator
*,
1856 stmt_vec_info
*, slp_tree
);
1857 extern bool vect_transform_cycle_phi (loop_vec_info
, stmt_vec_info
,
1859 slp_tree
, slp_instance
);
1860 extern bool vectorizable_lc_phi (loop_vec_info
, stmt_vec_info
,
1861 stmt_vec_info
*, slp_tree
);
1862 extern bool vect_worthwhile_without_simd_p (vec_info
*, tree_code
);
1863 extern int vect_get_known_peeling_cost (loop_vec_info
, int, int *,
1864 stmt_vector_for_cost
*,
1865 stmt_vector_for_cost
*,
1866 stmt_vector_for_cost
*);
1867 extern tree
cse_and_gimplify_to_preheader (loop_vec_info
, tree
);
1869 /* In tree-vect-slp.c. */
1870 extern void vect_free_slp_instance (slp_instance
, bool);
1871 extern bool vect_transform_slp_perm_load (vec_info
*, slp_tree
, vec
<tree
>,
1872 gimple_stmt_iterator
*, poly_uint64
,
1874 extern bool vect_slp_analyze_operations (vec_info
*);
1875 extern void vect_schedule_slp (vec_info
*);
1876 extern opt_result
vect_analyze_slp (vec_info
*, unsigned);
1877 extern bool vect_make_slp_decision (loop_vec_info
);
1878 extern void vect_detect_hybrid_slp (loop_vec_info
);
1879 extern void vect_optimize_slp (vec_info
*);
1880 extern void vect_get_slp_defs (vec_info
*, slp_tree
, vec
<vec
<tree
> > *,
1882 extern bool vect_slp_bb (basic_block
);
1883 extern stmt_vec_info
vect_find_last_scalar_stmt_in_slp (slp_tree
);
1884 extern bool is_simple_and_all_uses_invariant (stmt_vec_info
, loop_vec_info
);
1885 extern bool can_duplicate_and_interleave_p (vec_info
*, unsigned int, tree
,
1886 unsigned int * = NULL
,
1887 tree
* = NULL
, tree
* = NULL
);
1888 extern void duplicate_and_interleave (vec_info
*, gimple_seq
*, tree
,
1889 vec
<tree
>, unsigned int, vec
<tree
> &);
1890 extern int vect_get_place_in_interleaving_chain (stmt_vec_info
, stmt_vec_info
);
1892 /* In tree-vect-patterns.c. */
1893 /* Pattern recognition functions.
1894 Additional pattern recognition functions can (and will) be added
1896 void vect_pattern_recog (vec_info
*);
1898 /* In tree-vectorizer.c. */
1899 unsigned vectorize_loops (void);
1900 void vect_free_loop_info_assumptions (class loop
*);
1901 gimple
*vect_loop_vectorized_call (class loop
*, gcond
**cond
= NULL
);
1904 #endif /* GCC_TREE_VECTORIZER_H */