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
;
133 /* Vectorized stmt/s. */
134 vec
<stmt_vec_info
> vec_stmts
;
135 /* Number of vector stmts that are created to replace the group of scalar
136 stmts. It is calculated during the transformation phase as the number of
137 scalar elements in one scalar iteration (GROUP_SIZE) multiplied by VF
138 divided by vector size. */
139 unsigned int vec_stmts_size
;
141 /* Reference count in the SLP graph. */
143 /* The maximum number of vector elements for the subtree rooted
145 poly_uint64 max_nunits
;
146 /* Whether the scalar computations use two different operators. */
148 /* The DEF type of this node. */
149 enum vect_def_type def_type
;
153 /* SLP instance is a sequence of stmts in a loop that can be packed into
155 typedef class _slp_instance
{
157 /* The root of SLP tree. */
160 /* For vector constructors, the constructor stmt that the SLP tree is built
161 from, NULL otherwise. */
162 stmt_vec_info root_stmt
;
164 /* The unrolling factor required to vectorized this SLP instance. */
165 poly_uint64 unrolling_factor
;
167 /* The group of nodes that contain loads of this SLP instance. */
170 /* The SLP node containing the reduction PHIs. */
175 /* Access Functions. */
176 #define SLP_INSTANCE_TREE(S) (S)->root
177 #define SLP_INSTANCE_UNROLLING_FACTOR(S) (S)->unrolling_factor
178 #define SLP_INSTANCE_LOADS(S) (S)->loads
179 #define SLP_INSTANCE_ROOT_STMT(S) (S)->root_stmt
181 #define SLP_TREE_CHILDREN(S) (S)->children
182 #define SLP_TREE_SCALAR_STMTS(S) (S)->stmts
183 #define SLP_TREE_SCALAR_OPS(S) (S)->ops
184 #define SLP_TREE_VEC_STMTS(S) (S)->vec_stmts
185 #define SLP_TREE_NUMBER_OF_VEC_STMTS(S) (S)->vec_stmts_size
186 #define SLP_TREE_LOAD_PERMUTATION(S) (S)->load_permutation
187 #define SLP_TREE_TWO_OPERATORS(S) (S)->two_operators
188 #define SLP_TREE_DEF_TYPE(S) (S)->def_type
190 /* Key for map that records association between
191 scalar conditions and corresponding loop mask, and
192 is populated by vect_record_loop_mask. */
194 struct scalar_cond_masked_key
196 scalar_cond_masked_key (tree t
, unsigned ncopies_
)
199 get_cond_ops_from_tree (t
);
202 void get_cond_ops_from_tree (tree
);
211 struct default_hash_traits
<scalar_cond_masked_key
>
213 typedef scalar_cond_masked_key compare_type
;
214 typedef scalar_cond_masked_key value_type
;
216 static inline hashval_t
221 inchash::add_expr (v
.op0
, h
, 0);
222 inchash::add_expr (v
.op1
, h
, 0);
223 h
.add_int (v
.ncopies
);
228 equal (value_type existing
, value_type candidate
)
230 return (existing
.ncopies
== candidate
.ncopies
231 && existing
.code
== candidate
.code
232 && operand_equal_p (existing
.op0
, candidate
.op0
, 0)
233 && operand_equal_p (existing
.op1
, candidate
.op1
, 0));
236 static const bool empty_zero_p
= true;
239 mark_empty (value_type
&v
)
245 is_empty (value_type v
)
247 return v
.ncopies
== 0;
250 static inline void mark_deleted (value_type
&) {}
252 static inline bool is_deleted (const value_type
&)
257 static inline void remove (value_type
&) {}
260 typedef hash_set
<scalar_cond_masked_key
> scalar_cond_masked_set_type
;
262 /* Describes two objects whose addresses must be unequal for the vectorized
264 typedef std::pair
<tree
, tree
> vec_object_pair
;
266 /* Records that vectorization is only possible if abs (EXPR) >= MIN_VALUE.
267 UNSIGNED_P is true if we can assume that abs (EXPR) == EXPR. */
268 class vec_lower_bound
{
270 vec_lower_bound () {}
271 vec_lower_bound (tree e
, bool u
, poly_uint64 m
)
272 : expr (e
), unsigned_p (u
), min_value (m
) {}
276 poly_uint64 min_value
;
279 /* Vectorizer state shared between different analyses like vector sizes
280 of the same CFG region. */
281 class vec_info_shared
{
286 void save_datarefs();
287 void check_datarefs();
289 /* All data references. Freed by free_data_refs, so not an auto_vec. */
290 vec
<data_reference_p
> datarefs
;
291 vec
<data_reference
> datarefs_copy
;
293 /* The loop nest in which the data dependences are computed. */
294 auto_vec
<loop_p
> loop_nest
;
296 /* All data dependences. Freed by free_dependence_relations, so not
301 /* Vectorizer state common between loop and basic-block vectorization. */
304 typedef hash_set
<int_hash
<machine_mode
, E_VOIDmode
, E_BLKmode
> > mode_set
;
305 enum vec_kind
{ bb
, loop
};
307 vec_info (vec_kind
, void *, vec_info_shared
*);
310 stmt_vec_info
add_stmt (gimple
*);
311 stmt_vec_info
lookup_stmt (gimple
*);
312 stmt_vec_info
lookup_def (tree
);
313 stmt_vec_info
lookup_single_use (tree
);
314 class dr_vec_info
*lookup_dr (data_reference
*);
315 void move_dr (stmt_vec_info
, stmt_vec_info
);
316 void remove_stmt (stmt_vec_info
);
317 void replace_stmt (gimple_stmt_iterator
*, stmt_vec_info
, gimple
*);
319 /* The type of vectorization. */
322 /* Shared vectorizer state. */
323 vec_info_shared
*shared
;
325 /* The mapping of GIMPLE UID to stmt_vec_info. */
326 vec
<stmt_vec_info
> stmt_vec_infos
;
329 auto_vec
<slp_instance
> slp_instances
;
330 auto_vec
<slp_tree
> slp_loads
;
332 /* Maps base addresses to an innermost_loop_behavior that gives the maximum
333 known alignment for that base. */
334 vec_base_alignments base_alignments
;
336 /* All interleaving chains of stores, represented by the first
337 stmt in the chain. */
338 auto_vec
<stmt_vec_info
> grouped_stores
;
340 /* Cost data used by the target cost model. */
341 void *target_cost_data
;
343 /* The set of vector modes used in the vectorized region. */
344 mode_set used_vector_modes
;
346 /* The argument we should pass to related_vector_mode when looking up
347 the vector mode for a scalar mode, or VOIDmode if we haven't yet
348 made any decisions about which vector modes to use. */
349 machine_mode vector_mode
;
352 stmt_vec_info
new_stmt_vec_info (gimple
*stmt
);
353 void set_vinfo_for_stmt (gimple
*, stmt_vec_info
);
354 void free_stmt_vec_infos ();
355 void free_stmt_vec_info (stmt_vec_info
);
358 class _loop_vec_info
;
364 is_a_helper
<_loop_vec_info
*>::test (vec_info
*i
)
366 return i
->kind
== vec_info::loop
;
372 is_a_helper
<_bb_vec_info
*>::test (vec_info
*i
)
374 return i
->kind
== vec_info::bb
;
378 /* In general, we can divide the vector statements in a vectorized loop
379 into related groups ("rgroups") and say that for each rgroup there is
380 some nS such that the rgroup operates on nS values from one scalar
381 iteration followed by nS values from the next. That is, if VF is the
382 vectorization factor of the loop, the rgroup operates on a sequence:
384 (1,1) (1,2) ... (1,nS) (2,1) ... (2,nS) ... (VF,1) ... (VF,nS)
386 where (i,j) represents a scalar value with index j in a scalar
387 iteration with index i.
389 [ We use the term "rgroup" to emphasise that this grouping isn't
390 necessarily the same as the grouping of statements used elsewhere.
391 For example, if we implement a group of scalar loads using gather
392 loads, we'll use a separate gather load for each scalar load, and
393 thus each gather load will belong to its own rgroup. ]
395 In general this sequence will occupy nV vectors concatenated
396 together. If these vectors have nL lanes each, the total number
397 of scalar values N is given by:
399 N = nS * VF = nV * nL
401 None of nS, VF, nV and nL are required to be a power of 2. nS and nV
402 are compile-time constants but VF and nL can be variable (if the target
403 supports variable-length vectors).
405 In classical vectorization, each iteration of the vector loop would
406 handle exactly VF iterations of the original scalar loop. However,
407 in a fully-masked loop, a particular iteration of the vector loop
408 might handle fewer than VF iterations of the scalar loop. The vector
409 lanes that correspond to iterations of the scalar loop are said to be
410 "active" and the other lanes are said to be "inactive".
412 In a fully-masked loop, many rgroups need to be masked to ensure that
413 they have no effect for the inactive lanes. Each such rgroup needs a
414 sequence of booleans in the same order as above, but with each (i,j)
415 replaced by a boolean that indicates whether iteration i is active.
416 This sequence occupies nV vector masks that again have nL lanes each.
417 Thus the mask sequence as a whole consists of VF independent booleans
418 that are each repeated nS times.
420 We make the simplifying assumption that if a sequence of nV masks is
421 suitable for one (nS,nL) pair, we can reuse it for (nS/2,nL/2) by
422 VIEW_CONVERTing it. This holds for all current targets that support
423 fully-masked loops. For example, suppose the scalar loop is:
427 for (int i = 0; i < n; ++i)
429 f[i * 2 + 0] += 1.0f;
430 f[i * 2 + 1] += 2.0f;
434 and suppose that vectors have 256 bits. The vectorized f accesses
435 will belong to one rgroup and the vectorized d access to another:
437 f rgroup: nS = 2, nV = 1, nL = 8
438 d rgroup: nS = 1, nV = 1, nL = 4
441 [ In this simple example the rgroups do correspond to the normal
442 SLP grouping scheme. ]
444 If only the first three lanes are active, the masks we need are:
446 f rgroup: 1 1 | 1 1 | 1 1 | 0 0
447 d rgroup: 1 | 1 | 1 | 0
449 Here we can use a mask calculated for f's rgroup for d's, but not
452 Thus for each value of nV, it is enough to provide nV masks, with the
453 mask being calculated based on the highest nL (or, equivalently, based
454 on the highest nS) required by any rgroup with that nV. We therefore
455 represent the entire collection of masks as a two-level table, with the
456 first level being indexed by nV - 1 (since nV == 0 doesn't exist) and
457 the second being indexed by the mask index 0 <= i < nV. */
459 /* The masks needed by rgroups with nV vectors, according to the
460 description above. */
461 struct rgroup_masks
{
462 /* The largest nS for all rgroups that use these masks. */
463 unsigned int max_nscalars_per_iter
;
465 /* The type of mask to use, based on the highest nS recorded above. */
468 /* A vector of nV masks, in iteration order. */
472 typedef auto_vec
<rgroup_masks
> vec_loop_masks
;
474 typedef auto_vec
<std::pair
<data_reference
*, tree
> > drs_init_vec
;
476 /*-----------------------------------------------------------------*/
477 /* Info on vectorized loops. */
478 /*-----------------------------------------------------------------*/
479 typedef class _loop_vec_info
: public vec_info
{
481 _loop_vec_info (class loop
*, vec_info_shared
*);
484 /* The loop to which this info struct refers to. */
487 /* The loop basic blocks. */
490 /* Number of latch executions. */
492 /* Number of iterations. */
494 /* Number of iterations of the original loop. */
495 tree num_iters_unchanged
;
496 /* Condition under which this loop is analyzed and versioned. */
497 tree num_iters_assumptions
;
499 /* Threshold of number of iterations below which vectorization will not be
500 performed. It is calculated from MIN_PROFITABLE_ITERS and
501 param_min_vect_loop_bound. */
504 /* When applying loop versioning, the vector form should only be used
505 if the number of scalar iterations is >= this value, on top of all
506 the other requirements. Ignored when loop versioning is not being
508 poly_uint64 versioning_threshold
;
510 /* Unrolling factor */
511 poly_uint64 vectorization_factor
;
513 /* Maximum runtime vectorization factor, or MAX_VECTORIZATION_FACTOR
514 if there is no particular limit. */
515 unsigned HOST_WIDE_INT max_vectorization_factor
;
517 /* The masks that a fully-masked loop should use to avoid operating
518 on inactive scalars. */
519 vec_loop_masks masks
;
521 /* Set of scalar conditions that have loop mask applied. */
522 scalar_cond_masked_set_type scalar_cond_masked_set
;
524 /* If we are using a loop mask to align memory addresses, this variable
525 contains the number of vector elements that we should skip in the
526 first iteration of the vector loop (i.e. the number of leading
527 elements that should be false in the first mask). */
528 tree mask_skip_niters
;
530 /* Type of the variables to use in the WHILE_ULT call for fully-masked
532 tree mask_compare_type
;
534 /* For #pragma omp simd if (x) loops the x expression. If constant 0,
535 the loop should not be vectorized, if constant non-zero, simd_if_cond
536 shouldn't be set and loop vectorized normally, if SSA_NAME, the loop
537 should be versioned on that condition, using scalar loop if the condition
538 is false and vectorized loop otherwise. */
541 /* Type of the IV to use in the WHILE_ULT call for fully-masked
545 /* Unknown DRs according to which loop was peeled. */
546 class dr_vec_info
*unaligned_dr
;
548 /* peeling_for_alignment indicates whether peeling for alignment will take
549 place, and what the peeling factor should be:
550 peeling_for_alignment = X means:
551 If X=0: Peeling for alignment will not be applied.
552 If X>0: Peel first X iterations.
553 If X=-1: Generate a runtime test to calculate the number of iterations
554 to be peeled, using the dataref recorded in the field
556 int peeling_for_alignment
;
558 /* The mask used to check the alignment of pointers or arrays. */
561 /* Data Dependence Relations defining address ranges that are candidates
562 for a run-time aliasing check. */
563 auto_vec
<ddr_p
> may_alias_ddrs
;
565 /* Data Dependence Relations defining address ranges together with segment
566 lengths from which the run-time aliasing check is built. */
567 auto_vec
<dr_with_seg_len_pair_t
> comp_alias_ddrs
;
569 /* Check that the addresses of each pair of objects is unequal. */
570 auto_vec
<vec_object_pair
> check_unequal_addrs
;
572 /* List of values that are required to be nonzero. This is used to check
573 whether things like "x[i * n] += 1;" are safe and eventually gets added
574 to the checks for lower bounds below. */
575 auto_vec
<tree
> check_nonzero
;
577 /* List of values that need to be checked for a minimum value. */
578 auto_vec
<vec_lower_bound
> lower_bounds
;
580 /* Statements in the loop that have data references that are candidates for a
581 runtime (loop versioning) misalignment check. */
582 auto_vec
<stmt_vec_info
> may_misalign_stmts
;
584 /* Reduction cycles detected in the loop. Used in loop-aware SLP. */
585 auto_vec
<stmt_vec_info
> reductions
;
587 /* All reduction chains in the loop, represented by the first
588 stmt in the chain. */
589 auto_vec
<stmt_vec_info
> reduction_chains
;
591 /* Cost vector for a single scalar iteration. */
592 auto_vec
<stmt_info_for_cost
> scalar_cost_vec
;
594 /* Map of IV base/step expressions to inserted name in the preheader. */
595 hash_map
<tree_operand_hash
, tree
> *ivexpr_map
;
597 /* Map of OpenMP "omp simd array" scan variables to corresponding
598 rhs of the store of the initializer. */
599 hash_map
<tree
, tree
> *scan_map
;
601 /* The unrolling factor needed to SLP the loop. In case of that pure SLP is
602 applied to the loop, i.e., no unrolling is needed, this is 1. */
603 poly_uint64 slp_unrolling_factor
;
605 /* Cost of a single scalar iteration. */
606 int single_scalar_iteration_cost
;
608 /* The cost of the vector prologue and epilogue, including peeled
609 iterations and set-up code. */
610 int vec_outside_cost
;
612 /* The cost of the vector loop body. */
615 /* Is the loop vectorizable? */
618 /* Records whether we still have the option of using a fully-masked loop. */
619 bool can_fully_mask_p
;
621 /* True if have decided to use a fully-masked loop. */
624 /* When we have grouped data accesses with gaps, we may introduce invalid
625 memory accesses. We peel the last iteration of the loop to prevent
627 bool peeling_for_gaps
;
629 /* When the number of iterations is not a multiple of the vector size
630 we need to peel off iterations at the end to form an epilogue loop. */
631 bool peeling_for_niter
;
633 /* True if there are no loop carried data dependencies in the loop.
634 If loop->safelen <= 1, then this is always true, either the loop
635 didn't have any loop carried data dependencies, or the loop is being
636 vectorized guarded with some runtime alias checks, or couldn't
637 be vectorized at all, but then this field shouldn't be used.
638 For loop->safelen >= 2, the user has asserted that there are no
639 backward dependencies, but there still could be loop carried forward
640 dependencies in such loops. This flag will be false if normal
641 vectorizer data dependency analysis would fail or require versioning
642 for alias, but because of loop->safelen >= 2 it has been vectorized
643 even without versioning for alias. E.g. in:
645 for (int i = 0; i < m; i++)
647 (or #pragma simd or #pragma ivdep) we can vectorize this and it will
648 DTRT even for k > 0 && k < m, but without safelen we would not
649 vectorize this, so this field would be false. */
650 bool no_data_dependencies
;
652 /* Mark loops having masked stores. */
655 /* Queued scaling factor for the scalar loop. */
656 profile_probability scalar_loop_scaling
;
658 /* If if-conversion versioned this loop before conversion, this is the
659 loop version without if-conversion. */
660 class loop
*scalar_loop
;
662 /* For loops being epilogues of already vectorized loops
663 this points to the original vectorized loop. Otherwise NULL. */
664 _loop_vec_info
*orig_loop_info
;
666 /* Used to store loop_vec_infos of epilogues of this loop during
668 vec
<_loop_vec_info
*> epilogue_vinfos
;
672 /* Access Functions. */
673 #define LOOP_VINFO_LOOP(L) (L)->loop
674 #define LOOP_VINFO_BBS(L) (L)->bbs
675 #define LOOP_VINFO_NITERSM1(L) (L)->num_itersm1
676 #define LOOP_VINFO_NITERS(L) (L)->num_iters
677 /* Since LOOP_VINFO_NITERS and LOOP_VINFO_NITERSM1 can change after
678 prologue peeling retain total unchanged scalar loop iterations for
680 #define LOOP_VINFO_NITERS_UNCHANGED(L) (L)->num_iters_unchanged
681 #define LOOP_VINFO_NITERS_ASSUMPTIONS(L) (L)->num_iters_assumptions
682 #define LOOP_VINFO_COST_MODEL_THRESHOLD(L) (L)->th
683 #define LOOP_VINFO_VERSIONING_THRESHOLD(L) (L)->versioning_threshold
684 #define LOOP_VINFO_VECTORIZABLE_P(L) (L)->vectorizable
685 #define LOOP_VINFO_CAN_FULLY_MASK_P(L) (L)->can_fully_mask_p
686 #define LOOP_VINFO_FULLY_MASKED_P(L) (L)->fully_masked_p
687 #define LOOP_VINFO_VECT_FACTOR(L) (L)->vectorization_factor
688 #define LOOP_VINFO_MAX_VECT_FACTOR(L) (L)->max_vectorization_factor
689 #define LOOP_VINFO_MASKS(L) (L)->masks
690 #define LOOP_VINFO_MASK_SKIP_NITERS(L) (L)->mask_skip_niters
691 #define LOOP_VINFO_MASK_COMPARE_TYPE(L) (L)->mask_compare_type
692 #define LOOP_VINFO_MASK_IV_TYPE(L) (L)->iv_type
693 #define LOOP_VINFO_PTR_MASK(L) (L)->ptr_mask
694 #define LOOP_VINFO_LOOP_NEST(L) (L)->shared->loop_nest
695 #define LOOP_VINFO_DATAREFS(L) (L)->shared->datarefs
696 #define LOOP_VINFO_DDRS(L) (L)->shared->ddrs
697 #define LOOP_VINFO_INT_NITERS(L) (TREE_INT_CST_LOW ((L)->num_iters))
698 #define LOOP_VINFO_PEELING_FOR_ALIGNMENT(L) (L)->peeling_for_alignment
699 #define LOOP_VINFO_UNALIGNED_DR(L) (L)->unaligned_dr
700 #define LOOP_VINFO_MAY_MISALIGN_STMTS(L) (L)->may_misalign_stmts
701 #define LOOP_VINFO_MAY_ALIAS_DDRS(L) (L)->may_alias_ddrs
702 #define LOOP_VINFO_COMP_ALIAS_DDRS(L) (L)->comp_alias_ddrs
703 #define LOOP_VINFO_CHECK_UNEQUAL_ADDRS(L) (L)->check_unequal_addrs
704 #define LOOP_VINFO_CHECK_NONZERO(L) (L)->check_nonzero
705 #define LOOP_VINFO_LOWER_BOUNDS(L) (L)->lower_bounds
706 #define LOOP_VINFO_GROUPED_STORES(L) (L)->grouped_stores
707 #define LOOP_VINFO_SLP_INSTANCES(L) (L)->slp_instances
708 #define LOOP_VINFO_SLP_UNROLLING_FACTOR(L) (L)->slp_unrolling_factor
709 #define LOOP_VINFO_REDUCTIONS(L) (L)->reductions
710 #define LOOP_VINFO_REDUCTION_CHAINS(L) (L)->reduction_chains
711 #define LOOP_VINFO_TARGET_COST_DATA(L) (L)->target_cost_data
712 #define LOOP_VINFO_PEELING_FOR_GAPS(L) (L)->peeling_for_gaps
713 #define LOOP_VINFO_PEELING_FOR_NITER(L) (L)->peeling_for_niter
714 #define LOOP_VINFO_NO_DATA_DEPENDENCIES(L) (L)->no_data_dependencies
715 #define LOOP_VINFO_SCALAR_LOOP(L) (L)->scalar_loop
716 #define LOOP_VINFO_SCALAR_LOOP_SCALING(L) (L)->scalar_loop_scaling
717 #define LOOP_VINFO_HAS_MASK_STORE(L) (L)->has_mask_store
718 #define LOOP_VINFO_SCALAR_ITERATION_COST(L) (L)->scalar_cost_vec
719 #define LOOP_VINFO_SINGLE_SCALAR_ITERATION_COST(L) (L)->single_scalar_iteration_cost
720 #define LOOP_VINFO_ORIG_LOOP_INFO(L) (L)->orig_loop_info
721 #define LOOP_VINFO_SIMD_IF_COND(L) (L)->simd_if_cond
723 #define LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT(L) \
724 ((L)->may_misalign_stmts.length () > 0)
725 #define LOOP_REQUIRES_VERSIONING_FOR_ALIAS(L) \
726 ((L)->comp_alias_ddrs.length () > 0 \
727 || (L)->check_unequal_addrs.length () > 0 \
728 || (L)->lower_bounds.length () > 0)
729 #define LOOP_REQUIRES_VERSIONING_FOR_NITERS(L) \
730 (LOOP_VINFO_NITERS_ASSUMPTIONS (L))
731 #define LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND(L) \
732 (LOOP_VINFO_SIMD_IF_COND (L))
733 #define LOOP_REQUIRES_VERSIONING(L) \
734 (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (L) \
735 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (L) \
736 || LOOP_REQUIRES_VERSIONING_FOR_NITERS (L) \
737 || LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND (L))
739 #define LOOP_VINFO_NITERS_KNOWN_P(L) \
740 (tree_fits_shwi_p ((L)->num_iters) && tree_to_shwi ((L)->num_iters) > 0)
742 #define LOOP_VINFO_EPILOGUE_P(L) \
743 (LOOP_VINFO_ORIG_LOOP_INFO (L) != NULL)
745 #define LOOP_VINFO_ORIG_MAX_VECT_FACTOR(L) \
746 (LOOP_VINFO_MAX_VECT_FACTOR (LOOP_VINFO_ORIG_LOOP_INFO (L)))
748 /* Wrapper for loop_vec_info, for tracking success/failure, where a non-NULL
749 value signifies success, and a NULL value signifies failure, supporting
750 propagating an opt_problem * describing the failure back up the call
752 typedef opt_pointer_wrapper
<loop_vec_info
> opt_loop_vec_info
;
754 static inline loop_vec_info
755 loop_vec_info_for_loop (class loop
*loop
)
757 return (loop_vec_info
) loop
->aux
;
760 typedef class _bb_vec_info
: public vec_info
763 _bb_vec_info (gimple_stmt_iterator
, gimple_stmt_iterator
, vec_info_shared
*);
767 gimple_stmt_iterator region_begin
;
768 gimple_stmt_iterator region_end
;
771 #define BB_VINFO_BB(B) (B)->bb
772 #define BB_VINFO_GROUPED_STORES(B) (B)->grouped_stores
773 #define BB_VINFO_SLP_INSTANCES(B) (B)->slp_instances
774 #define BB_VINFO_DATAREFS(B) (B)->shared->datarefs
775 #define BB_VINFO_DDRS(B) (B)->shared->ddrs
776 #define BB_VINFO_TARGET_COST_DATA(B) (B)->target_cost_data
778 static inline bb_vec_info
779 vec_info_for_bb (basic_block bb
)
781 return (bb_vec_info
) bb
->aux
;
784 /*-----------------------------------------------------------------*/
785 /* Info on vectorized defs. */
786 /*-----------------------------------------------------------------*/
787 enum stmt_vec_info_type
{
788 undef_vec_info_type
= 0,
794 call_simd_clone_vec_info_type
,
795 assignment_vec_info_type
,
796 condition_vec_info_type
,
797 comparison_vec_info_type
,
800 type_promotion_vec_info_type
,
801 type_demotion_vec_info_type
,
802 type_conversion_vec_info_type
,
805 loop_exit_ctrl_vec_info_type
808 /* Indicates whether/how a variable is used in the scope of loop/basic
811 vect_unused_in_scope
= 0,
813 /* The def is only used outside the loop. */
815 /* The def is in the inner loop, and the use is in the outer loop, and the
816 use is a reduction stmt. */
817 vect_used_in_outer_by_reduction
,
818 /* The def is in the inner loop, and the use is in the outer loop (and is
819 not part of reduction). */
822 /* defs that feed computations that end up (only) in a reduction. These
823 defs may be used by non-reduction stmts, but eventually, any
824 computations/values that are affected by these defs are used to compute
825 a reduction (i.e. don't get stored to memory, for example). We use this
826 to identify computations that we can change the order in which they are
828 vect_used_by_reduction
,
833 /* The type of vectorization that can be applied to the stmt: regular loop-based
834 vectorization; pure SLP - the stmt is a part of SLP instances and does not
835 have uses outside SLP instances; or hybrid SLP and loop-based - the stmt is
836 a part of SLP instance and also must be loop-based vectorized, since it has
837 uses outside SLP sequences.
839 In the loop context the meanings of pure and hybrid SLP are slightly
840 different. By saying that pure SLP is applied to the loop, we mean that we
841 exploit only intra-iteration parallelism in the loop; i.e., the loop can be
842 vectorized without doing any conceptual unrolling, cause we don't pack
843 together stmts from different iterations, only within a single iteration.
844 Loop hybrid SLP means that we exploit both intra-iteration and
845 inter-iteration parallelism (e.g., number of elements in the vector is 4
846 and the slp-group-size is 2, in which case we don't have enough parallelism
847 within an iteration, so we obtain the rest of the parallelism from subsequent
848 iterations by unrolling the loop by 2). */
855 /* Says whether a statement is a load, a store of a vectorized statement
856 result, or a store of an invariant value. */
857 enum vec_load_store_type
{
863 /* Describes how we're going to vectorize an individual load or store,
864 or a group of loads or stores. */
865 enum vect_memory_access_type
{
866 /* An access to an invariant address. This is used only for loads. */
869 /* A simple contiguous access. */
872 /* A contiguous access that goes down in memory rather than up,
873 with no additional permutation. This is used only for stores
875 VMAT_CONTIGUOUS_DOWN
,
877 /* A simple contiguous access in which the elements need to be permuted
878 after loading or before storing. Only used for loop vectorization;
879 SLP uses separate permutes. */
880 VMAT_CONTIGUOUS_PERMUTE
,
882 /* A simple contiguous access in which the elements need to be reversed
883 after loading or before storing. */
884 VMAT_CONTIGUOUS_REVERSE
,
886 /* An access that uses IFN_LOAD_LANES or IFN_STORE_LANES. */
887 VMAT_LOAD_STORE_LANES
,
889 /* An access in which each scalar element is loaded or stored
893 /* A hybrid of VMAT_CONTIGUOUS and VMAT_ELEMENTWISE, used for grouped
894 SLP accesses. Each unrolled iteration uses a contiguous load
895 or store for the whole group, but the groups from separate iterations
896 are combined in the same way as for VMAT_ELEMENTWISE. */
899 /* The access uses gather loads or scatter stores. */
905 /* The data reference itself. */
907 /* The statement that contains the data reference. */
909 /* The misalignment in bytes of the reference, or -1 if not known. */
911 /* The byte alignment that we'd ideally like the reference to have,
912 and the value that misalignment is measured against. */
913 poly_uint64 target_alignment
;
914 /* If true the alignment of base_decl needs to be increased. */
915 bool base_misaligned
;
918 /* Stores current vectorized loop's offset. To be added to the DR's
919 offset to calculate current offset of data reference. */
923 typedef struct data_reference
*dr_p
;
925 class _stmt_vec_info
{
928 enum stmt_vec_info_type type
;
930 /* Indicates whether this stmts is part of a computation whose result is
931 used outside the loop. */
934 /* Stmt is part of some pattern (computation idiom) */
937 /* True if the statement was created during pattern recognition as
938 part of the replacement for RELATED_STMT. This implies that the
939 statement isn't part of any basic block, although for convenience
940 its gimple_bb is the same as for RELATED_STMT. */
943 /* Is this statement vectorizable or should it be skipped in (partial)
947 /* The stmt to which this info struct refers to. */
950 /* The vector type to be used for the LHS of this statement. */
953 /* The vectorized version of the stmt. */
954 stmt_vec_info vectorized_stmt
;
957 /* The following is relevant only for stmts that contain a non-scalar
958 data-ref (array/pointer/struct access). A GIMPLE stmt is expected to have
959 at most one such data-ref. */
963 /* Information about the data-ref relative to this loop
964 nest (the loop that is being considered for vectorization). */
965 innermost_loop_behavior dr_wrt_vec_loop
;
967 /* For loop PHI nodes, the base and evolution part of it. This makes sure
968 this information is still available in vect_update_ivs_after_vectorizer
969 where we may not be able to re-analyze the PHI nodes evolution as
970 peeling for the prologue loop can make it unanalyzable. The evolution
971 part is still correct after peeling, but the base may have changed from
973 tree loop_phi_evolution_base_unchanged
;
974 tree loop_phi_evolution_part
;
976 /* Used for various bookkeeping purposes, generally holding a pointer to
977 some other stmt S that is in some way "related" to this stmt.
978 Current use of this field is:
979 If this stmt is part of a pattern (i.e. the field 'in_pattern_p' is
980 true): S is the "pattern stmt" that represents (and replaces) the
981 sequence of stmts that constitutes the pattern. Similarly, the
982 related_stmt of the "pattern stmt" points back to this stmt (which is
983 the last stmt in the original sequence of stmts that constitutes the
985 stmt_vec_info related_stmt
;
987 /* Used to keep a sequence of def stmts of a pattern stmt if such exists.
988 The sequence is attached to the original statement rather than the
989 pattern statement. */
990 gimple_seq pattern_def_seq
;
992 /* List of datarefs that are known to have the same alignment as the dataref
994 vec
<dr_p
> same_align_refs
;
996 /* Selected SIMD clone's function info. First vector element
997 is SIMD clone's function decl, followed by a pair of trees (base + step)
998 for linear arguments (pair of NULLs for other arguments). */
999 vec
<tree
> simd_clone_info
;
1001 /* Classify the def of this stmt. */
1002 enum vect_def_type def_type
;
1004 /* Whether the stmt is SLPed, loop-based vectorized, or both. */
1005 enum slp_vect_type slp_type
;
1007 /* Interleaving and reduction chains info. */
1008 /* First element in the group. */
1009 stmt_vec_info first_element
;
1010 /* Pointer to the next element in the group. */
1011 stmt_vec_info next_element
;
1012 /* The size of the group. */
1014 /* For stores, number of stores from this group seen. We vectorize the last
1016 unsigned int store_count
;
1017 /* For loads only, the gap from the previous load. For consecutive loads, GAP
1021 /* The minimum negative dependence distance this stmt participates in
1023 unsigned int min_neg_dist
;
1025 /* Not all stmts in the loop need to be vectorized. e.g, the increment
1026 of the loop induction variable and computation of array indexes. relevant
1027 indicates whether the stmt needs to be vectorized. */
1028 enum vect_relevant relevant
;
1030 /* For loads if this is a gather, for stores if this is a scatter. */
1031 bool gather_scatter_p
;
1033 /* True if this is an access with loop-invariant stride. */
1036 /* For both loads and stores. */
1037 unsigned simd_lane_access_p
: 3;
1039 /* Classifies how the load or store is going to be implemented
1040 for loop vectorization. */
1041 vect_memory_access_type memory_access_type
;
1043 /* For INTEGER_INDUC_COND_REDUCTION, the initial value to be used. */
1044 tree induc_cond_initial_val
;
1046 /* If not NULL the value to be added to compute final reduction value. */
1047 tree reduc_epilogue_adjustment
;
1049 /* On a reduction PHI the reduction type as detected by
1050 vect_is_simple_reduction and vectorizable_reduction. */
1051 enum vect_reduction_type reduc_type
;
1053 /* The original reduction code, to be used in the epilogue. */
1054 enum tree_code reduc_code
;
1055 /* An internal function we should use in the epilogue. */
1056 internal_fn reduc_fn
;
1058 /* On a stmt participating in the reduction the index of the operand
1059 on the reduction SSA cycle. */
1062 /* On a reduction PHI the def returned by vect_force_simple_reduction.
1063 On the def returned by vect_force_simple_reduction the
1064 corresponding PHI. */
1065 stmt_vec_info reduc_def
;
1067 /* The vector input type relevant for reduction vectorization. */
1068 tree reduc_vectype_in
;
1070 /* The vector type for performing the actual reduction. */
1073 /* Whether we force a single cycle PHI during reduction vectorization. */
1074 bool force_single_cycle
;
1076 /* Whether on this stmt reduction meta is recorded. */
1079 /* The number of scalar stmt references from active SLP instances. */
1080 unsigned int num_slp_uses
;
1082 /* If nonzero, the lhs of the statement could be truncated to this
1083 many bits without affecting any users of the result. */
1084 unsigned int min_output_precision
;
1086 /* If nonzero, all non-boolean input operands have the same precision,
1087 and they could each be truncated to this many bits without changing
1089 unsigned int min_input_precision
;
1091 /* If OPERATION_BITS is nonzero, the statement could be performed on
1092 an integer with the sign and number of bits given by OPERATION_SIGN
1093 and OPERATION_BITS without changing the result. */
1094 unsigned int operation_precision
;
1095 signop operation_sign
;
1097 /* If the statement produces a boolean result, this value describes
1098 how we should choose the associated vector type. The possible
1101 - an integer precision N if we should use the vector mask type
1102 associated with N-bit integers. This is only used if all relevant
1103 input booleans also want the vector mask type for N-bit integers,
1104 or if we can convert them into that form by pattern-matching.
1106 - ~0U if we considered choosing a vector mask type but decided
1107 to treat the boolean as a normal integer type instead.
1109 - 0 otherwise. This means either that the operation isn't one that
1110 could have a vector mask type (and so should have a normal vector
1111 type instead) or that we simply haven't made a choice either way. */
1112 unsigned int mask_precision
;
1114 /* True if this is only suitable for SLP vectorization. */
1115 bool slp_vect_only_p
;
1118 /* Information about a gather/scatter call. */
1119 struct gather_scatter_info
{
1120 /* The internal function to use for the gather/scatter operation,
1121 or IFN_LAST if a built-in function should be used instead. */
1124 /* The FUNCTION_DECL for the built-in gather/scatter function,
1125 or null if an internal function should be used instead. */
1128 /* The loop-invariant base value. */
1131 /* The original scalar offset, which is a non-loop-invariant SSA_NAME. */
1134 /* Each offset element should be multiplied by this amount before
1135 being added to the base. */
1138 /* The definition type for the vectorized offset. */
1139 enum vect_def_type offset_dt
;
1141 /* The type of the vectorized offset. */
1142 tree offset_vectype
;
1144 /* The type of the scalar elements after loading or before storing. */
1147 /* The type of the scalar elements being loaded or stored. */
1151 /* Access Functions. */
1152 #define STMT_VINFO_TYPE(S) (S)->type
1153 #define STMT_VINFO_STMT(S) (S)->stmt
1154 #define STMT_VINFO_RELEVANT(S) (S)->relevant
1155 #define STMT_VINFO_LIVE_P(S) (S)->live
1156 #define STMT_VINFO_VECTYPE(S) (S)->vectype
1157 #define STMT_VINFO_VEC_STMT(S) (S)->vectorized_stmt
1158 #define STMT_VINFO_VECTORIZABLE(S) (S)->vectorizable
1159 #define STMT_VINFO_DATA_REF(S) ((S)->dr_aux.dr + 0)
1160 #define STMT_VINFO_GATHER_SCATTER_P(S) (S)->gather_scatter_p
1161 #define STMT_VINFO_STRIDED_P(S) (S)->strided_p
1162 #define STMT_VINFO_MEMORY_ACCESS_TYPE(S) (S)->memory_access_type
1163 #define STMT_VINFO_SIMD_LANE_ACCESS_P(S) (S)->simd_lane_access_p
1164 #define STMT_VINFO_VEC_INDUC_COND_INITIAL_VAL(S) (S)->induc_cond_initial_val
1165 #define STMT_VINFO_REDUC_EPILOGUE_ADJUSTMENT(S) (S)->reduc_epilogue_adjustment
1166 #define STMT_VINFO_REDUC_IDX(S) (S)->reduc_idx
1167 #define STMT_VINFO_FORCE_SINGLE_CYCLE(S) (S)->force_single_cycle
1169 #define STMT_VINFO_DR_WRT_VEC_LOOP(S) (S)->dr_wrt_vec_loop
1170 #define STMT_VINFO_DR_BASE_ADDRESS(S) (S)->dr_wrt_vec_loop.base_address
1171 #define STMT_VINFO_DR_INIT(S) (S)->dr_wrt_vec_loop.init
1172 #define STMT_VINFO_DR_OFFSET(S) (S)->dr_wrt_vec_loop.offset
1173 #define STMT_VINFO_DR_STEP(S) (S)->dr_wrt_vec_loop.step
1174 #define STMT_VINFO_DR_BASE_ALIGNMENT(S) (S)->dr_wrt_vec_loop.base_alignment
1175 #define STMT_VINFO_DR_BASE_MISALIGNMENT(S) \
1176 (S)->dr_wrt_vec_loop.base_misalignment
1177 #define STMT_VINFO_DR_OFFSET_ALIGNMENT(S) \
1178 (S)->dr_wrt_vec_loop.offset_alignment
1179 #define STMT_VINFO_DR_STEP_ALIGNMENT(S) \
1180 (S)->dr_wrt_vec_loop.step_alignment
1182 #define STMT_VINFO_DR_INFO(S) \
1183 (gcc_checking_assert ((S)->dr_aux.stmt == (S)), &(S)->dr_aux)
1185 #define STMT_VINFO_IN_PATTERN_P(S) (S)->in_pattern_p
1186 #define STMT_VINFO_RELATED_STMT(S) (S)->related_stmt
1187 #define STMT_VINFO_PATTERN_DEF_SEQ(S) (S)->pattern_def_seq
1188 #define STMT_VINFO_SAME_ALIGN_REFS(S) (S)->same_align_refs
1189 #define STMT_VINFO_SIMD_CLONE_INFO(S) (S)->simd_clone_info
1190 #define STMT_VINFO_DEF_TYPE(S) (S)->def_type
1191 #define STMT_VINFO_GROUPED_ACCESS(S) \
1192 ((S)->dr_aux.dr && DR_GROUP_FIRST_ELEMENT(S))
1193 #define STMT_VINFO_LOOP_PHI_EVOLUTION_BASE_UNCHANGED(S) (S)->loop_phi_evolution_base_unchanged
1194 #define STMT_VINFO_LOOP_PHI_EVOLUTION_PART(S) (S)->loop_phi_evolution_part
1195 #define STMT_VINFO_MIN_NEG_DIST(S) (S)->min_neg_dist
1196 #define STMT_VINFO_NUM_SLP_USES(S) (S)->num_slp_uses
1197 #define STMT_VINFO_REDUC_TYPE(S) (S)->reduc_type
1198 #define STMT_VINFO_REDUC_CODE(S) (S)->reduc_code
1199 #define STMT_VINFO_REDUC_FN(S) (S)->reduc_fn
1200 #define STMT_VINFO_REDUC_DEF(S) (S)->reduc_def
1201 #define STMT_VINFO_REDUC_VECTYPE(S) (S)->reduc_vectype
1202 #define STMT_VINFO_REDUC_VECTYPE_IN(S) (S)->reduc_vectype_in
1203 #define STMT_VINFO_SLP_VECT_ONLY(S) (S)->slp_vect_only_p
1205 #define DR_GROUP_FIRST_ELEMENT(S) \
1206 (gcc_checking_assert ((S)->dr_aux.dr), (S)->first_element)
1207 #define DR_GROUP_NEXT_ELEMENT(S) \
1208 (gcc_checking_assert ((S)->dr_aux.dr), (S)->next_element)
1209 #define DR_GROUP_SIZE(S) \
1210 (gcc_checking_assert ((S)->dr_aux.dr), (S)->size)
1211 #define DR_GROUP_STORE_COUNT(S) \
1212 (gcc_checking_assert ((S)->dr_aux.dr), (S)->store_count)
1213 #define DR_GROUP_GAP(S) \
1214 (gcc_checking_assert ((S)->dr_aux.dr), (S)->gap)
1216 #define REDUC_GROUP_FIRST_ELEMENT(S) \
1217 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->first_element)
1218 #define REDUC_GROUP_NEXT_ELEMENT(S) \
1219 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->next_element)
1220 #define REDUC_GROUP_SIZE(S) \
1221 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->size)
1223 #define STMT_VINFO_RELEVANT_P(S) ((S)->relevant != vect_unused_in_scope)
1225 #define HYBRID_SLP_STMT(S) ((S)->slp_type == hybrid)
1226 #define PURE_SLP_STMT(S) ((S)->slp_type == pure_slp)
1227 #define STMT_SLP_TYPE(S) (S)->slp_type
1229 #define VECT_MAX_COST 1000
1231 /* The maximum number of intermediate steps required in multi-step type
1233 #define MAX_INTERM_CVT_STEPS 3
1235 #define MAX_VECTORIZATION_FACTOR INT_MAX
1237 /* Nonzero if TYPE represents a (scalar) boolean type or type
1238 in the middle-end compatible with it (unsigned precision 1 integral
1239 types). Used to determine which types should be vectorized as
1240 VECTOR_BOOLEAN_TYPE_P. */
1242 #define VECT_SCALAR_BOOLEAN_TYPE_P(TYPE) \
1243 (TREE_CODE (TYPE) == BOOLEAN_TYPE \
1244 || ((TREE_CODE (TYPE) == INTEGER_TYPE \
1245 || TREE_CODE (TYPE) == ENUMERAL_TYPE) \
1246 && TYPE_PRECISION (TYPE) == 1 \
1247 && TYPE_UNSIGNED (TYPE)))
1250 nested_in_vect_loop_p (class loop
*loop
, stmt_vec_info stmt_info
)
1253 && (loop
->inner
== (gimple_bb (stmt_info
->stmt
))->loop_father
));
1256 /* Return true if STMT_INFO should produce a vector mask type rather than
1257 a normal nonmask type. */
1260 vect_use_mask_type_p (stmt_vec_info stmt_info
)
1262 return stmt_info
->mask_precision
&& stmt_info
->mask_precision
!= ~0U;
1265 /* Return TRUE if a statement represented by STMT_INFO is a part of a
1269 is_pattern_stmt_p (stmt_vec_info stmt_info
)
1271 return stmt_info
->pattern_stmt_p
;
1274 /* If STMT_INFO is a pattern statement, return the statement that it
1275 replaces, otherwise return STMT_INFO itself. */
1277 inline stmt_vec_info
1278 vect_orig_stmt (stmt_vec_info stmt_info
)
1280 if (is_pattern_stmt_p (stmt_info
))
1281 return STMT_VINFO_RELATED_STMT (stmt_info
);
1285 /* Return the later statement between STMT1_INFO and STMT2_INFO. */
1287 static inline stmt_vec_info
1288 get_later_stmt (stmt_vec_info stmt1_info
, stmt_vec_info stmt2_info
)
1290 if (gimple_uid (vect_orig_stmt (stmt1_info
)->stmt
)
1291 > gimple_uid (vect_orig_stmt (stmt2_info
)->stmt
))
1297 /* If STMT_INFO has been replaced by a pattern statement, return the
1298 replacement statement, otherwise return STMT_INFO itself. */
1300 inline stmt_vec_info
1301 vect_stmt_to_vectorize (stmt_vec_info stmt_info
)
1303 if (STMT_VINFO_IN_PATTERN_P (stmt_info
))
1304 return STMT_VINFO_RELATED_STMT (stmt_info
);
1308 /* Return true if BB is a loop header. */
1311 is_loop_header_bb_p (basic_block bb
)
1313 if (bb
== (bb
->loop_father
)->header
)
1315 gcc_checking_assert (EDGE_COUNT (bb
->preds
) == 1);
1319 /* Return pow2 (X). */
1326 for (i
= 0; i
< x
; i
++)
1332 /* Alias targetm.vectorize.builtin_vectorization_cost. */
1335 builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost
,
1336 tree vectype
, int misalign
)
1338 return targetm
.vectorize
.builtin_vectorization_cost (type_of_cost
,
1342 /* Get cost by calling cost target builtin. */
1345 int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost
)
1347 return builtin_vectorization_cost (type_of_cost
, NULL
, 0);
1350 /* Alias targetm.vectorize.init_cost. */
1352 static inline void *
1353 init_cost (class loop
*loop_info
)
1355 return targetm
.vectorize
.init_cost (loop_info
);
1358 extern void dump_stmt_cost (FILE *, void *, int, enum vect_cost_for_stmt
,
1359 stmt_vec_info
, tree
, int, unsigned,
1360 enum vect_cost_model_location
);
1362 /* Alias targetm.vectorize.add_stmt_cost. */
1364 static inline unsigned
1365 add_stmt_cost (vec_info
*vinfo
, void *data
, int count
,
1366 enum vect_cost_for_stmt kind
,
1367 stmt_vec_info stmt_info
, tree vectype
, int misalign
,
1368 enum vect_cost_model_location where
)
1370 unsigned cost
= targetm
.vectorize
.add_stmt_cost (vinfo
, data
, count
, kind
,
1373 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1374 dump_stmt_cost (dump_file
, data
, count
, kind
, stmt_info
, vectype
, misalign
,
1379 /* Alias targetm.vectorize.finish_cost. */
1382 finish_cost (void *data
, unsigned *prologue_cost
,
1383 unsigned *body_cost
, unsigned *epilogue_cost
)
1385 targetm
.vectorize
.finish_cost (data
, prologue_cost
, body_cost
, epilogue_cost
);
1388 /* Alias targetm.vectorize.destroy_cost_data. */
1391 destroy_cost_data (void *data
)
1393 targetm
.vectorize
.destroy_cost_data (data
);
1397 add_stmt_costs (vec_info
*vinfo
, void *data
, stmt_vector_for_cost
*cost_vec
)
1399 stmt_info_for_cost
*cost
;
1401 FOR_EACH_VEC_ELT (*cost_vec
, i
, cost
)
1402 add_stmt_cost (vinfo
, data
, cost
->count
, cost
->kind
, cost
->stmt_info
,
1403 cost
->vectype
, cost
->misalign
, cost
->where
);
1406 /*-----------------------------------------------------------------*/
1407 /* Info on data references alignment. */
1408 /*-----------------------------------------------------------------*/
1409 #define DR_MISALIGNMENT_UNKNOWN (-1)
1410 #define DR_MISALIGNMENT_UNINITIALIZED (-2)
1413 set_dr_misalignment (dr_vec_info
*dr_info
, int val
)
1415 dr_info
->misalignment
= val
;
1419 dr_misalignment (dr_vec_info
*dr_info
)
1421 int misalign
= dr_info
->misalignment
;
1422 gcc_assert (misalign
!= DR_MISALIGNMENT_UNINITIALIZED
);
1426 /* Reflects actual alignment of first access in the vectorized loop,
1427 taking into account peeling/versioning if applied. */
1428 #define DR_MISALIGNMENT(DR) dr_misalignment (DR)
1429 #define SET_DR_MISALIGNMENT(DR, VAL) set_dr_misalignment (DR, VAL)
1431 /* Only defined once DR_MISALIGNMENT is defined. */
1432 #define DR_TARGET_ALIGNMENT(DR) ((DR)->target_alignment)
1434 /* Return true if data access DR_INFO is aligned to its target alignment
1435 (which may be less than a full vector). */
1438 aligned_access_p (dr_vec_info
*dr_info
)
1440 return (DR_MISALIGNMENT (dr_info
) == 0);
1443 /* Return TRUE if the alignment of the data access is known, and FALSE
1447 known_alignment_for_access_p (dr_vec_info
*dr_info
)
1449 return (DR_MISALIGNMENT (dr_info
) != DR_MISALIGNMENT_UNKNOWN
);
1452 /* Return the minimum alignment in bytes that the vectorized version
1453 of DR_INFO is guaranteed to have. */
1455 static inline unsigned int
1456 vect_known_alignment_in_bytes (dr_vec_info
*dr_info
)
1458 if (DR_MISALIGNMENT (dr_info
) == DR_MISALIGNMENT_UNKNOWN
)
1459 return TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr_info
->dr
)));
1460 if (DR_MISALIGNMENT (dr_info
) == 0)
1461 return known_alignment (DR_TARGET_ALIGNMENT (dr_info
));
1462 return DR_MISALIGNMENT (dr_info
) & -DR_MISALIGNMENT (dr_info
);
1465 /* Return the behavior of DR_INFO with respect to the vectorization context
1466 (which for outer loop vectorization might not be the behavior recorded
1467 in DR_INFO itself). */
1469 static inline innermost_loop_behavior
*
1470 vect_dr_behavior (vec_info
*vinfo
, dr_vec_info
*dr_info
)
1472 stmt_vec_info stmt_info
= dr_info
->stmt
;
1473 loop_vec_info loop_vinfo
= dyn_cast
<loop_vec_info
> (vinfo
);
1474 if (loop_vinfo
== NULL
1475 || !nested_in_vect_loop_p (LOOP_VINFO_LOOP (loop_vinfo
), stmt_info
))
1476 return &DR_INNERMOST (dr_info
->dr
);
1478 return &STMT_VINFO_DR_WRT_VEC_LOOP (stmt_info
);
1481 /* Return the offset calculated by adding the offset of this DR_INFO to the
1482 corresponding data_reference's offset. If CHECK_OUTER then use
1483 vect_dr_behavior to select the appropriate data_reference to use. */
1486 get_dr_vinfo_offset (vec_info
*vinfo
,
1487 dr_vec_info
*dr_info
, bool check_outer
= false)
1489 innermost_loop_behavior
*base
;
1491 base
= vect_dr_behavior (vinfo
, dr_info
);
1493 base
= &dr_info
->dr
->innermost
;
1495 tree offset
= base
->offset
;
1497 if (!dr_info
->offset
)
1500 offset
= fold_convert (sizetype
, offset
);
1501 return fold_build2 (PLUS_EXPR
, TREE_TYPE (dr_info
->offset
), offset
,
1506 /* Return true if the vect cost model is unlimited. */
1508 unlimited_cost_model (loop_p loop
)
1510 if (loop
!= NULL
&& loop
->force_vectorize
1511 && flag_simd_cost_model
!= VECT_COST_MODEL_DEFAULT
)
1512 return flag_simd_cost_model
== VECT_COST_MODEL_UNLIMITED
;
1513 return (flag_vect_cost_model
== VECT_COST_MODEL_UNLIMITED
);
1516 /* Return true if the loop described by LOOP_VINFO is fully-masked and
1517 if the first iteration should use a partial mask in order to achieve
1521 vect_use_loop_mask_for_alignment_p (loop_vec_info loop_vinfo
)
1523 return (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo
)
1524 && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo
));
1527 /* Return the number of vectors of type VECTYPE that are needed to get
1528 NUNITS elements. NUNITS should be based on the vectorization factor,
1529 so it is always a known multiple of the number of elements in VECTYPE. */
1531 static inline unsigned int
1532 vect_get_num_vectors (poly_uint64 nunits
, tree vectype
)
1534 return exact_div (nunits
, TYPE_VECTOR_SUBPARTS (vectype
)).to_constant ();
1537 /* Return the number of copies needed for loop vectorization when
1538 a statement operates on vectors of type VECTYPE. This is the
1539 vectorization factor divided by the number of elements in
1540 VECTYPE and is always known at compile time. */
1542 static inline unsigned int
1543 vect_get_num_copies (loop_vec_info loop_vinfo
, tree vectype
)
1545 return vect_get_num_vectors (LOOP_VINFO_VECT_FACTOR (loop_vinfo
), vectype
);
1548 /* Update maximum unit count *MAX_NUNITS so that it accounts for
1549 NUNITS. *MAX_NUNITS can be 1 if we haven't yet recorded anything. */
1552 vect_update_max_nunits (poly_uint64
*max_nunits
, poly_uint64 nunits
)
1554 /* All unit counts have the form vec_info::vector_size * X for some
1555 rational X, so two unit sizes must have a common multiple.
1556 Everything is a multiple of the initial value of 1. */
1557 *max_nunits
= force_common_multiple (*max_nunits
, nunits
);
1560 /* Update maximum unit count *MAX_NUNITS so that it accounts for
1561 the number of units in vector type VECTYPE. *MAX_NUNITS can be 1
1562 if we haven't yet recorded any vector types. */
1565 vect_update_max_nunits (poly_uint64
*max_nunits
, tree vectype
)
1567 vect_update_max_nunits (max_nunits
, TYPE_VECTOR_SUBPARTS (vectype
));
1570 /* Return the vectorization factor that should be used for costing
1571 purposes while vectorizing the loop described by LOOP_VINFO.
1572 Pick a reasonable estimate if the vectorization factor isn't
1573 known at compile time. */
1575 static inline unsigned int
1576 vect_vf_for_cost (loop_vec_info loop_vinfo
)
1578 return estimated_poly_value (LOOP_VINFO_VECT_FACTOR (loop_vinfo
));
1581 /* Estimate the number of elements in VEC_TYPE for costing purposes.
1582 Pick a reasonable estimate if the exact number isn't known at
1585 static inline unsigned int
1586 vect_nunits_for_cost (tree vec_type
)
1588 return estimated_poly_value (TYPE_VECTOR_SUBPARTS (vec_type
));
1591 /* Return the maximum possible vectorization factor for LOOP_VINFO. */
1593 static inline unsigned HOST_WIDE_INT
1594 vect_max_vf (loop_vec_info loop_vinfo
)
1596 unsigned HOST_WIDE_INT vf
;
1597 if (LOOP_VINFO_VECT_FACTOR (loop_vinfo
).is_constant (&vf
))
1599 return MAX_VECTORIZATION_FACTOR
;
1602 /* Return the size of the value accessed by unvectorized data reference
1603 DR_INFO. This is only valid once STMT_VINFO_VECTYPE has been calculated
1604 for the associated gimple statement, since that guarantees that DR_INFO
1605 accesses either a scalar or a scalar equivalent. ("Scalar equivalent"
1606 here includes things like V1SI, which can be vectorized in the same way
1610 vect_get_scalar_dr_size (dr_vec_info
*dr_info
)
1612 return tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_info
->dr
))));
1615 /* Return true if LOOP_VINFO requires a runtime check for whether the
1616 vector loop is profitable. */
1619 vect_apply_runtime_profitability_check_p (loop_vec_info loop_vinfo
)
1621 unsigned int th
= LOOP_VINFO_COST_MODEL_THRESHOLD (loop_vinfo
);
1622 return (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
)
1623 && th
>= vect_vf_for_cost (loop_vinfo
));
1626 /* Source location + hotness information. */
1627 extern dump_user_location_t vect_location
;
1629 /* A macro for calling:
1630 dump_begin_scope (MSG, vect_location);
1631 via an RAII object, thus printing "=== MSG ===\n" to the dumpfile etc,
1634 once the object goes out of scope, thus capturing the nesting of
1637 These scopes affect dump messages within them: dump messages at the
1638 top level implicitly default to MSG_PRIORITY_USER_FACING, whereas those
1639 in a nested scope implicitly default to MSG_PRIORITY_INTERNALS. */
1641 #define DUMP_VECT_SCOPE(MSG) \
1642 AUTO_DUMP_SCOPE (MSG, vect_location)
1644 /* A sentinel class for ensuring that the "vect_location" global gets
1645 reset at the end of a scope.
1647 The "vect_location" global is used during dumping and contains a
1648 location_t, which could contain references to a tree block via the
1649 ad-hoc data. This data is used for tracking inlining information,
1650 but it's not a GC root; it's simply assumed that such locations never
1651 get accessed if the blocks are optimized away.
1653 Hence we need to ensure that such locations are purged at the end
1654 of any operations using them (e.g. via this class). */
1656 class auto_purge_vect_location
1659 ~auto_purge_vect_location ();
1662 /*-----------------------------------------------------------------*/
1663 /* Function prototypes. */
1664 /*-----------------------------------------------------------------*/
1666 /* Simple loop peeling and versioning utilities for vectorizer's purposes -
1667 in tree-vect-loop-manip.c. */
1668 extern void vect_set_loop_condition (class loop
*, loop_vec_info
,
1669 tree
, tree
, tree
, bool);
1670 extern bool slpeel_can_duplicate_loop_p (const class loop
*, const_edge
);
1671 class loop
*slpeel_tree_duplicate_loop_to_edge_cfg (class loop
*,
1672 class loop
*, edge
);
1673 class loop
*vect_loop_versioning (loop_vec_info
, gimple
*);
1674 extern class loop
*vect_do_peeling (loop_vec_info
, tree
, tree
,
1675 tree
*, tree
*, tree
*, int, bool, bool,
1677 extern void vect_prepare_for_masked_peels (loop_vec_info
);
1678 extern dump_user_location_t
find_loop_location (class loop
*);
1679 extern bool vect_can_advance_ivs_p (loop_vec_info
);
1680 extern void vect_update_inits_of_drs (loop_vec_info
, tree
, tree_code
);
1682 /* In tree-vect-stmts.c. */
1683 extern tree
get_related_vectype_for_scalar_type (machine_mode
, tree
,
1685 extern tree
get_vectype_for_scalar_type (vec_info
*, tree
, unsigned int = 0);
1686 extern tree
get_vectype_for_scalar_type (vec_info
*, tree
, slp_tree
);
1687 extern tree
get_mask_type_for_scalar_type (vec_info
*, tree
, unsigned int = 0);
1688 extern tree
get_same_sized_vectype (tree
, tree
);
1689 extern bool vect_chooses_same_modes_p (vec_info
*, machine_mode
);
1690 extern bool vect_get_loop_mask_type (loop_vec_info
);
1691 extern bool vect_is_simple_use (tree
, vec_info
*, enum vect_def_type
*,
1692 stmt_vec_info
* = NULL
, gimple
** = NULL
);
1693 extern bool vect_is_simple_use (tree
, vec_info
*, enum vect_def_type
*,
1694 tree
*, stmt_vec_info
* = NULL
,
1696 extern bool supportable_widening_operation (vec_info
*,
1697 enum tree_code
, stmt_vec_info
,
1698 tree
, tree
, enum tree_code
*,
1699 enum tree_code
*, int *,
1701 extern bool supportable_narrowing_operation (enum tree_code
, tree
, tree
,
1702 enum tree_code
*, int *,
1705 extern unsigned record_stmt_cost (stmt_vector_for_cost
*, int,
1706 enum vect_cost_for_stmt
, stmt_vec_info
,
1707 tree
, int, enum vect_cost_model_location
);
1709 /* Overload of record_stmt_cost with VECTYPE derived from STMT_INFO. */
1711 static inline unsigned
1712 record_stmt_cost (stmt_vector_for_cost
*body_cost_vec
, int count
,
1713 enum vect_cost_for_stmt kind
, stmt_vec_info stmt_info
,
1714 int misalign
, enum vect_cost_model_location where
)
1716 return record_stmt_cost (body_cost_vec
, count
, kind
, stmt_info
,
1717 STMT_VINFO_VECTYPE (stmt_info
), misalign
, where
);
1720 extern stmt_vec_info
vect_finish_replace_stmt (vec_info
*,
1721 stmt_vec_info
, gimple
*);
1722 extern stmt_vec_info
vect_finish_stmt_generation (vec_info
*,
1723 stmt_vec_info
, gimple
*,
1724 gimple_stmt_iterator
*);
1725 extern opt_result
vect_mark_stmts_to_be_vectorized (loop_vec_info
, bool *);
1726 extern tree
vect_get_store_rhs (stmt_vec_info
);
1727 extern tree
vect_get_vec_def_for_operand_1 (stmt_vec_info
, enum vect_def_type
);
1728 extern tree
vect_get_vec_def_for_operand (vec_info
*, tree
,
1729 stmt_vec_info
, tree
= NULL
);
1730 extern void vect_get_vec_defs (vec_info
*, tree
, tree
, stmt_vec_info
,
1731 vec
<tree
> *, vec
<tree
> *, slp_tree
);
1732 extern void vect_get_vec_defs_for_stmt_copy (vec_info
*,
1733 vec
<tree
> *, vec
<tree
> *);
1734 extern tree
vect_init_vector (vec_info
*, stmt_vec_info
, tree
, tree
,
1735 gimple_stmt_iterator
*);
1736 extern tree
vect_get_vec_def_for_stmt_copy (vec_info
*, tree
);
1737 extern bool vect_transform_stmt (vec_info
*, stmt_vec_info
,
1738 gimple_stmt_iterator
*,
1739 slp_tree
, slp_instance
);
1740 extern void vect_remove_stores (vec_info
*, stmt_vec_info
);
1741 extern bool vect_nop_conversion_p (stmt_vec_info
);
1742 extern opt_result
vect_analyze_stmt (vec_info
*, stmt_vec_info
, bool *,
1744 slp_instance
, stmt_vector_for_cost
*);
1745 extern void vect_get_load_cost (vec_info
*, stmt_vec_info
, int, bool,
1746 unsigned int *, unsigned int *,
1747 stmt_vector_for_cost
*,
1748 stmt_vector_for_cost
*, bool);
1749 extern void vect_get_store_cost (vec_info
*, stmt_vec_info
, int,
1750 unsigned int *, stmt_vector_for_cost
*);
1751 extern bool vect_supportable_shift (vec_info
*, enum tree_code
, tree
);
1752 extern tree
vect_gen_perm_mask_any (tree
, const vec_perm_indices
&);
1753 extern tree
vect_gen_perm_mask_checked (tree
, const vec_perm_indices
&);
1754 extern void optimize_mask_stores (class loop
*);
1755 extern gcall
*vect_gen_while (tree
, tree
, tree
);
1756 extern tree
vect_gen_while_not (gimple_seq
*, tree
, tree
, tree
);
1757 extern opt_result
vect_get_vector_types_for_stmt (vec_info
*,
1758 stmt_vec_info
, tree
*,
1759 tree
*, unsigned int = 0);
1760 extern opt_tree
vect_get_mask_type_for_stmt (stmt_vec_info
, unsigned int = 0);
1762 /* In tree-vect-data-refs.c. */
1763 extern bool vect_can_force_dr_alignment_p (const_tree
, poly_uint64
);
1764 extern enum dr_alignment_support vect_supportable_dr_alignment
1765 (vec_info
*, dr_vec_info
*, bool);
1766 extern tree
vect_get_smallest_scalar_type (stmt_vec_info
, HOST_WIDE_INT
*,
1768 extern opt_result
vect_analyze_data_ref_dependences (loop_vec_info
, unsigned int *);
1769 extern bool vect_slp_analyze_instance_dependence (vec_info
*, slp_instance
);
1770 extern opt_result
vect_enhance_data_refs_alignment (loop_vec_info
);
1771 extern opt_result
vect_analyze_data_refs_alignment (loop_vec_info
);
1772 extern opt_result
vect_verify_datarefs_alignment (loop_vec_info
);
1773 extern bool vect_slp_analyze_and_verify_instance_alignment (vec_info
*,
1775 extern opt_result
vect_analyze_data_ref_accesses (vec_info
*);
1776 extern opt_result
vect_prune_runtime_alias_test_list (loop_vec_info
);
1777 extern bool vect_gather_scatter_fn_p (vec_info
*, bool, bool, tree
, tree
,
1778 tree
, int, internal_fn
*, tree
*);
1779 extern bool vect_check_gather_scatter (stmt_vec_info
, loop_vec_info
,
1780 gather_scatter_info
*);
1781 extern opt_result
vect_find_stmt_data_reference (loop_p
, gimple
*,
1782 vec
<data_reference_p
> *);
1783 extern opt_result
vect_analyze_data_refs (vec_info
*, poly_uint64
*, bool *);
1784 extern void vect_record_base_alignments (vec_info
*);
1785 extern tree
vect_create_data_ref_ptr (vec_info
*,
1786 stmt_vec_info
, tree
, class loop
*, tree
,
1787 tree
*, gimple_stmt_iterator
*,
1789 tree
= NULL_TREE
, tree
= NULL_TREE
);
1790 extern tree
bump_vector_ptr (vec_info
*, tree
, gimple
*, gimple_stmt_iterator
*,
1791 stmt_vec_info
, tree
);
1792 extern void vect_copy_ref_info (tree
, tree
);
1793 extern tree
vect_create_destination_var (tree
, tree
);
1794 extern bool vect_grouped_store_supported (tree
, unsigned HOST_WIDE_INT
);
1795 extern bool vect_store_lanes_supported (tree
, unsigned HOST_WIDE_INT
, bool);
1796 extern bool vect_grouped_load_supported (tree
, bool, unsigned HOST_WIDE_INT
);
1797 extern bool vect_load_lanes_supported (tree
, unsigned HOST_WIDE_INT
, bool);
1798 extern void vect_permute_store_chain (vec_info
*,
1799 vec
<tree
> ,unsigned int, stmt_vec_info
,
1800 gimple_stmt_iterator
*, vec
<tree
> *);
1801 extern tree
vect_setup_realignment (vec_info
*,
1802 stmt_vec_info
, gimple_stmt_iterator
*,
1803 tree
*, enum dr_alignment_support
, tree
,
1805 extern void vect_transform_grouped_load (vec_info
*, stmt_vec_info
, vec
<tree
>,
1806 int, gimple_stmt_iterator
*);
1807 extern void vect_record_grouped_load_vectors (vec_info
*,
1808 stmt_vec_info
, vec
<tree
>);
1809 extern tree
vect_get_new_vect_var (tree
, enum vect_var_kind
, const char *);
1810 extern tree
vect_get_new_ssa_name (tree
, enum vect_var_kind
,
1811 const char * = NULL
);
1812 extern tree
vect_create_addr_base_for_vector_ref (vec_info
*,
1813 stmt_vec_info
, gimple_seq
*,
1814 tree
, tree
= NULL_TREE
);
1816 /* In tree-vect-loop.c. */
1817 extern widest_int
vect_iv_limit_for_full_masking (loop_vec_info loop_vinfo
);
1818 /* Used in tree-vect-loop-manip.c */
1819 extern void determine_peel_for_niter (loop_vec_info
);
1820 /* Used in gimple-loop-interchange.c and tree-parloops.c. */
1821 extern bool check_reduction_path (dump_user_location_t
, loop_p
, gphi
*, tree
,
1823 extern bool needs_fold_left_reduction_p (tree
, tree_code
);
1824 /* Drive for loop analysis stage. */
1825 extern opt_loop_vec_info
vect_analyze_loop (class loop
*, vec_info_shared
*);
1826 extern tree
vect_build_loop_niters (loop_vec_info
, bool * = NULL
);
1827 extern void vect_gen_vector_loop_niters (loop_vec_info
, tree
, tree
*,
1829 extern tree
vect_halve_mask_nunits (tree
, machine_mode
);
1830 extern tree
vect_double_mask_nunits (tree
, machine_mode
);
1831 extern void vect_record_loop_mask (loop_vec_info
, vec_loop_masks
*,
1832 unsigned int, tree
, tree
);
1833 extern tree
vect_get_loop_mask (gimple_stmt_iterator
*, vec_loop_masks
*,
1834 unsigned int, tree
, unsigned int);
1835 extern stmt_vec_info
info_for_reduction (vec_info
*, stmt_vec_info
);
1837 /* Drive for loop transformation stage. */
1838 extern class loop
*vect_transform_loop (loop_vec_info
, gimple
*);
1839 extern opt_loop_vec_info
vect_analyze_loop_form (class loop
*,
1841 extern bool vectorizable_live_operation (loop_vec_info
,
1842 stmt_vec_info
, gimple_stmt_iterator
*,
1843 slp_tree
, slp_instance
, int,
1844 bool, stmt_vector_for_cost
*);
1845 extern bool vectorizable_reduction (loop_vec_info
, stmt_vec_info
,
1846 slp_tree
, slp_instance
,
1847 stmt_vector_for_cost
*);
1848 extern bool vectorizable_induction (loop_vec_info
, stmt_vec_info
,
1849 gimple_stmt_iterator
*,
1850 stmt_vec_info
*, slp_tree
,
1851 stmt_vector_for_cost
*);
1852 extern bool vect_transform_reduction (loop_vec_info
, stmt_vec_info
,
1853 gimple_stmt_iterator
*,
1854 stmt_vec_info
*, slp_tree
);
1855 extern bool vect_transform_cycle_phi (loop_vec_info
, stmt_vec_info
,
1857 slp_tree
, slp_instance
);
1858 extern bool vectorizable_lc_phi (loop_vec_info
, stmt_vec_info
,
1859 stmt_vec_info
*, slp_tree
);
1860 extern bool vect_worthwhile_without_simd_p (vec_info
*, tree_code
);
1861 extern int vect_get_known_peeling_cost (loop_vec_info
, int, int *,
1862 stmt_vector_for_cost
*,
1863 stmt_vector_for_cost
*,
1864 stmt_vector_for_cost
*);
1865 extern tree
cse_and_gimplify_to_preheader (loop_vec_info
, tree
);
1867 /* In tree-vect-slp.c. */
1868 extern void vect_free_slp_instance (slp_instance
, bool);
1869 extern bool vect_transform_slp_perm_load (vec_info
*, slp_tree
, vec
<tree
>,
1870 gimple_stmt_iterator
*, poly_uint64
,
1872 extern bool vect_slp_analyze_operations (vec_info
*);
1873 extern void vect_schedule_slp (vec_info
*);
1874 extern opt_result
vect_analyze_slp (vec_info
*, unsigned);
1875 extern bool vect_make_slp_decision (loop_vec_info
);
1876 extern void vect_detect_hybrid_slp (loop_vec_info
);
1877 extern void vect_optimize_slp (vec_info
*);
1878 extern void vect_get_slp_defs (vec_info
*, slp_tree
, vec
<vec
<tree
> > *,
1880 extern bool vect_slp_bb (basic_block
);
1881 extern stmt_vec_info
vect_find_last_scalar_stmt_in_slp (slp_tree
);
1882 extern bool is_simple_and_all_uses_invariant (stmt_vec_info
, loop_vec_info
);
1883 extern bool can_duplicate_and_interleave_p (vec_info
*, unsigned int, tree
,
1884 unsigned int * = NULL
,
1885 tree
* = NULL
, tree
* = NULL
);
1886 extern void duplicate_and_interleave (vec_info
*, gimple_seq
*, tree
,
1887 vec
<tree
>, unsigned int, vec
<tree
> &);
1888 extern int vect_get_place_in_interleaving_chain (stmt_vec_info
, stmt_vec_info
);
1890 /* In tree-vect-patterns.c. */
1891 /* Pattern recognition functions.
1892 Additional pattern recognition functions can (and will) be added
1894 void vect_pattern_recog (vec_info
*);
1896 /* In tree-vectorizer.c. */
1897 unsigned vectorize_loops (void);
1898 void vect_free_loop_info_assumptions (class loop
*);
1899 gimple
*vect_loop_vectorized_call (class loop
*, gcond
**cond
= NULL
);
1902 #endif /* GCC_TREE_VECTORIZER_H */