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. */
124 /* Nodes that contain def-stmts of this node statements operands. */
125 vec
<slp_tree
> children
;
127 /* A group of scalar stmts to be vectorized together. */
128 vec
<stmt_vec_info
> stmts
;
129 /* A group of scalar operands to be vectorized together. */
131 /* The representative that should be used for analysis and
133 stmt_vec_info representative
;
135 /* Load permutation relative to the stores, NULL if there is no
137 vec
<unsigned> load_permutation
;
140 /* Vectorized stmt/s. */
141 vec
<stmt_vec_info
> vec_stmts
;
143 /* Number of vector stmts that are created to replace the group of scalar
144 stmts. It is calculated during the transformation phase as the number of
145 scalar elements in one scalar iteration (GROUP_SIZE) multiplied by VF
146 divided by vector size. */
147 unsigned int vec_stmts_size
;
149 /* Reference count in the SLP graph. */
151 /* The maximum number of vector elements for the subtree rooted
153 poly_uint64 max_nunits
;
154 /* Whether the scalar computations use two different operators. */
156 /* The DEF type of this node. */
157 enum vect_def_type def_type
;
161 /* SLP instance is a sequence of stmts in a loop that can be packed into
163 typedef class _slp_instance
{
165 /* The root of SLP tree. */
168 /* For vector constructors, the constructor stmt that the SLP tree is built
169 from, NULL otherwise. */
170 stmt_vec_info root_stmt
;
172 /* The unrolling factor required to vectorized this SLP instance. */
173 poly_uint64 unrolling_factor
;
175 /* The group of nodes that contain loads of this SLP instance. */
178 /* The SLP node containing the reduction PHIs. */
183 /* Access Functions. */
184 #define SLP_INSTANCE_TREE(S) (S)->root
185 #define SLP_INSTANCE_UNROLLING_FACTOR(S) (S)->unrolling_factor
186 #define SLP_INSTANCE_LOADS(S) (S)->loads
187 #define SLP_INSTANCE_ROOT_STMT(S) (S)->root_stmt
189 #define SLP_TREE_CHILDREN(S) (S)->children
190 #define SLP_TREE_SCALAR_STMTS(S) (S)->stmts
191 #define SLP_TREE_SCALAR_OPS(S) (S)->ops
192 #define SLP_TREE_VEC_STMTS(S) (S)->vec_stmts
193 #define SLP_TREE_VEC_DEFS(S) (S)->vec_defs
194 #define SLP_TREE_NUMBER_OF_VEC_STMTS(S) (S)->vec_stmts_size
195 #define SLP_TREE_LOAD_PERMUTATION(S) (S)->load_permutation
196 #define SLP_TREE_TWO_OPERATORS(S) (S)->two_operators
197 #define SLP_TREE_DEF_TYPE(S) (S)->def_type
198 #define SLP_TREE_VECTYPE(S) (S)->vectype
199 #define SLP_TREE_REPRESENTATIVE(S) (S)->representative
201 /* Key for map that records association between
202 scalar conditions and corresponding loop mask, and
203 is populated by vect_record_loop_mask. */
205 struct scalar_cond_masked_key
207 scalar_cond_masked_key (tree t
, unsigned ncopies_
)
210 get_cond_ops_from_tree (t
);
213 void get_cond_ops_from_tree (tree
);
222 struct default_hash_traits
<scalar_cond_masked_key
>
224 typedef scalar_cond_masked_key compare_type
;
225 typedef scalar_cond_masked_key value_type
;
227 static inline hashval_t
232 inchash::add_expr (v
.op0
, h
, 0);
233 inchash::add_expr (v
.op1
, h
, 0);
234 h
.add_int (v
.ncopies
);
239 equal (value_type existing
, value_type candidate
)
241 return (existing
.ncopies
== candidate
.ncopies
242 && existing
.code
== candidate
.code
243 && operand_equal_p (existing
.op0
, candidate
.op0
, 0)
244 && operand_equal_p (existing
.op1
, candidate
.op1
, 0));
247 static const bool empty_zero_p
= true;
250 mark_empty (value_type
&v
)
256 is_empty (value_type v
)
258 return v
.ncopies
== 0;
261 static inline void mark_deleted (value_type
&) {}
263 static inline bool is_deleted (const value_type
&)
268 static inline void remove (value_type
&) {}
271 typedef hash_set
<scalar_cond_masked_key
> scalar_cond_masked_set_type
;
273 /* Describes two objects whose addresses must be unequal for the vectorized
275 typedef std::pair
<tree
, tree
> vec_object_pair
;
277 /* Records that vectorization is only possible if abs (EXPR) >= MIN_VALUE.
278 UNSIGNED_P is true if we can assume that abs (EXPR) == EXPR. */
279 class vec_lower_bound
{
281 vec_lower_bound () {}
282 vec_lower_bound (tree e
, bool u
, poly_uint64 m
)
283 : expr (e
), unsigned_p (u
), min_value (m
) {}
287 poly_uint64 min_value
;
290 /* Vectorizer state shared between different analyses like vector sizes
291 of the same CFG region. */
292 class vec_info_shared
{
297 void save_datarefs();
298 void check_datarefs();
300 /* All data references. Freed by free_data_refs, so not an auto_vec. */
301 vec
<data_reference_p
> datarefs
;
302 vec
<data_reference
> datarefs_copy
;
304 /* The loop nest in which the data dependences are computed. */
305 auto_vec
<loop_p
> loop_nest
;
307 /* All data dependences. Freed by free_dependence_relations, so not
312 /* Vectorizer state common between loop and basic-block vectorization. */
315 typedef hash_set
<int_hash
<machine_mode
, E_VOIDmode
, E_BLKmode
> > mode_set
;
316 enum vec_kind
{ bb
, loop
};
318 vec_info (vec_kind
, void *, vec_info_shared
*);
321 stmt_vec_info
add_stmt (gimple
*);
322 stmt_vec_info
lookup_stmt (gimple
*);
323 stmt_vec_info
lookup_def (tree
);
324 stmt_vec_info
lookup_single_use (tree
);
325 class dr_vec_info
*lookup_dr (data_reference
*);
326 void move_dr (stmt_vec_info
, stmt_vec_info
);
327 void remove_stmt (stmt_vec_info
);
328 void replace_stmt (gimple_stmt_iterator
*, stmt_vec_info
, gimple
*);
330 /* The type of vectorization. */
333 /* Shared vectorizer state. */
334 vec_info_shared
*shared
;
336 /* The mapping of GIMPLE UID to stmt_vec_info. */
337 vec
<stmt_vec_info
> stmt_vec_infos
;
340 auto_vec
<slp_instance
> slp_instances
;
341 auto_vec
<slp_tree
> slp_loads
;
343 /* Maps base addresses to an innermost_loop_behavior that gives the maximum
344 known alignment for that base. */
345 vec_base_alignments base_alignments
;
347 /* All interleaving chains of stores, represented by the first
348 stmt in the chain. */
349 auto_vec
<stmt_vec_info
> grouped_stores
;
351 /* Cost data used by the target cost model. */
352 void *target_cost_data
;
354 /* The set of vector modes used in the vectorized region. */
355 mode_set used_vector_modes
;
357 /* The argument we should pass to related_vector_mode when looking up
358 the vector mode for a scalar mode, or VOIDmode if we haven't yet
359 made any decisions about which vector modes to use. */
360 machine_mode vector_mode
;
363 stmt_vec_info
new_stmt_vec_info (gimple
*stmt
);
364 void set_vinfo_for_stmt (gimple
*, stmt_vec_info
);
365 void free_stmt_vec_infos ();
366 void free_stmt_vec_info (stmt_vec_info
);
369 class _loop_vec_info
;
375 is_a_helper
<_loop_vec_info
*>::test (vec_info
*i
)
377 return i
->kind
== vec_info::loop
;
383 is_a_helper
<_bb_vec_info
*>::test (vec_info
*i
)
385 return i
->kind
== vec_info::bb
;
389 /* In general, we can divide the vector statements in a vectorized loop
390 into related groups ("rgroups") and say that for each rgroup there is
391 some nS such that the rgroup operates on nS values from one scalar
392 iteration followed by nS values from the next. That is, if VF is the
393 vectorization factor of the loop, the rgroup operates on a sequence:
395 (1,1) (1,2) ... (1,nS) (2,1) ... (2,nS) ... (VF,1) ... (VF,nS)
397 where (i,j) represents a scalar value with index j in a scalar
398 iteration with index i.
400 [ We use the term "rgroup" to emphasise that this grouping isn't
401 necessarily the same as the grouping of statements used elsewhere.
402 For example, if we implement a group of scalar loads using gather
403 loads, we'll use a separate gather load for each scalar load, and
404 thus each gather load will belong to its own rgroup. ]
406 In general this sequence will occupy nV vectors concatenated
407 together. If these vectors have nL lanes each, the total number
408 of scalar values N is given by:
410 N = nS * VF = nV * nL
412 None of nS, VF, nV and nL are required to be a power of 2. nS and nV
413 are compile-time constants but VF and nL can be variable (if the target
414 supports variable-length vectors).
416 In classical vectorization, each iteration of the vector loop would
417 handle exactly VF iterations of the original scalar loop. However,
418 in a fully-masked loop, a particular iteration of the vector loop
419 might handle fewer than VF iterations of the scalar loop. The vector
420 lanes that correspond to iterations of the scalar loop are said to be
421 "active" and the other lanes are said to be "inactive".
423 In a fully-masked loop, many rgroups need to be masked to ensure that
424 they have no effect for the inactive lanes. Each such rgroup needs a
425 sequence of booleans in the same order as above, but with each (i,j)
426 replaced by a boolean that indicates whether iteration i is active.
427 This sequence occupies nV vector masks that again have nL lanes each.
428 Thus the mask sequence as a whole consists of VF independent booleans
429 that are each repeated nS times.
431 We make the simplifying assumption that if a sequence of nV masks is
432 suitable for one (nS,nL) pair, we can reuse it for (nS/2,nL/2) by
433 VIEW_CONVERTing it. This holds for all current targets that support
434 fully-masked loops. For example, suppose the scalar loop is:
438 for (int i = 0; i < n; ++i)
440 f[i * 2 + 0] += 1.0f;
441 f[i * 2 + 1] += 2.0f;
445 and suppose that vectors have 256 bits. The vectorized f accesses
446 will belong to one rgroup and the vectorized d access to another:
448 f rgroup: nS = 2, nV = 1, nL = 8
449 d rgroup: nS = 1, nV = 1, nL = 4
452 [ In this simple example the rgroups do correspond to the normal
453 SLP grouping scheme. ]
455 If only the first three lanes are active, the masks we need are:
457 f rgroup: 1 1 | 1 1 | 1 1 | 0 0
458 d rgroup: 1 | 1 | 1 | 0
460 Here we can use a mask calculated for f's rgroup for d's, but not
463 Thus for each value of nV, it is enough to provide nV masks, with the
464 mask being calculated based on the highest nL (or, equivalently, based
465 on the highest nS) required by any rgroup with that nV. We therefore
466 represent the entire collection of masks as a two-level table, with the
467 first level being indexed by nV - 1 (since nV == 0 doesn't exist) and
468 the second being indexed by the mask index 0 <= i < nV. */
470 /* The masks needed by rgroups with nV vectors, according to the
471 description above. */
472 struct rgroup_masks
{
473 /* The largest nS for all rgroups that use these masks. */
474 unsigned int max_nscalars_per_iter
;
476 /* The type of mask to use, based on the highest nS recorded above. */
479 /* A vector of nV masks, in iteration order. */
483 typedef auto_vec
<rgroup_masks
> vec_loop_masks
;
485 typedef auto_vec
<std::pair
<data_reference
*, tree
> > drs_init_vec
;
487 /*-----------------------------------------------------------------*/
488 /* Info on vectorized loops. */
489 /*-----------------------------------------------------------------*/
490 typedef class _loop_vec_info
: public vec_info
{
492 _loop_vec_info (class loop
*, vec_info_shared
*);
495 /* The loop to which this info struct refers to. */
498 /* The loop basic blocks. */
501 /* Number of latch executions. */
503 /* Number of iterations. */
505 /* Number of iterations of the original loop. */
506 tree num_iters_unchanged
;
507 /* Condition under which this loop is analyzed and versioned. */
508 tree num_iters_assumptions
;
510 /* Threshold of number of iterations below which vectorization will not be
511 performed. It is calculated from MIN_PROFITABLE_ITERS and
512 param_min_vect_loop_bound. */
515 /* When applying loop versioning, the vector form should only be used
516 if the number of scalar iterations is >= this value, on top of all
517 the other requirements. Ignored when loop versioning is not being
519 poly_uint64 versioning_threshold
;
521 /* Unrolling factor */
522 poly_uint64 vectorization_factor
;
524 /* Maximum runtime vectorization factor, or MAX_VECTORIZATION_FACTOR
525 if there is no particular limit. */
526 unsigned HOST_WIDE_INT max_vectorization_factor
;
528 /* The masks that a fully-masked loop should use to avoid operating
529 on inactive scalars. */
530 vec_loop_masks masks
;
532 /* Set of scalar conditions that have loop mask applied. */
533 scalar_cond_masked_set_type scalar_cond_masked_set
;
535 /* If we are using a loop mask to align memory addresses, this variable
536 contains the number of vector elements that we should skip in the
537 first iteration of the vector loop (i.e. the number of leading
538 elements that should be false in the first mask). */
539 tree mask_skip_niters
;
541 /* Type of the variables to use in the WHILE_ULT call for fully-masked
543 tree mask_compare_type
;
545 /* For #pragma omp simd if (x) loops the x expression. If constant 0,
546 the loop should not be vectorized, if constant non-zero, simd_if_cond
547 shouldn't be set and loop vectorized normally, if SSA_NAME, the loop
548 should be versioned on that condition, using scalar loop if the condition
549 is false and vectorized loop otherwise. */
552 /* Type of the IV to use in the WHILE_ULT call for fully-masked
556 /* Unknown DRs according to which loop was peeled. */
557 class dr_vec_info
*unaligned_dr
;
559 /* peeling_for_alignment indicates whether peeling for alignment will take
560 place, and what the peeling factor should be:
561 peeling_for_alignment = X means:
562 If X=0: Peeling for alignment will not be applied.
563 If X>0: Peel first X iterations.
564 If X=-1: Generate a runtime test to calculate the number of iterations
565 to be peeled, using the dataref recorded in the field
567 int peeling_for_alignment
;
569 /* The mask used to check the alignment of pointers or arrays. */
572 /* Data Dependence Relations defining address ranges that are candidates
573 for a run-time aliasing check. */
574 auto_vec
<ddr_p
> may_alias_ddrs
;
576 /* Data Dependence Relations defining address ranges together with segment
577 lengths from which the run-time aliasing check is built. */
578 auto_vec
<dr_with_seg_len_pair_t
> comp_alias_ddrs
;
580 /* Check that the addresses of each pair of objects is unequal. */
581 auto_vec
<vec_object_pair
> check_unequal_addrs
;
583 /* List of values that are required to be nonzero. This is used to check
584 whether things like "x[i * n] += 1;" are safe and eventually gets added
585 to the checks for lower bounds below. */
586 auto_vec
<tree
> check_nonzero
;
588 /* List of values that need to be checked for a minimum value. */
589 auto_vec
<vec_lower_bound
> lower_bounds
;
591 /* Statements in the loop that have data references that are candidates for a
592 runtime (loop versioning) misalignment check. */
593 auto_vec
<stmt_vec_info
> may_misalign_stmts
;
595 /* Reduction cycles detected in the loop. Used in loop-aware SLP. */
596 auto_vec
<stmt_vec_info
> reductions
;
598 /* All reduction chains in the loop, represented by the first
599 stmt in the chain. */
600 auto_vec
<stmt_vec_info
> reduction_chains
;
602 /* Cost vector for a single scalar iteration. */
603 auto_vec
<stmt_info_for_cost
> scalar_cost_vec
;
605 /* Map of IV base/step expressions to inserted name in the preheader. */
606 hash_map
<tree_operand_hash
, tree
> *ivexpr_map
;
608 /* Map of OpenMP "omp simd array" scan variables to corresponding
609 rhs of the store of the initializer. */
610 hash_map
<tree
, tree
> *scan_map
;
612 /* The unrolling factor needed to SLP the loop. In case of that pure SLP is
613 applied to the loop, i.e., no unrolling is needed, this is 1. */
614 poly_uint64 slp_unrolling_factor
;
616 /* Cost of a single scalar iteration. */
617 int single_scalar_iteration_cost
;
619 /* The cost of the vector prologue and epilogue, including peeled
620 iterations and set-up code. */
621 int vec_outside_cost
;
623 /* The cost of the vector loop body. */
626 /* Is the loop vectorizable? */
629 /* Records whether we still have the option of using a fully-masked loop. */
630 bool can_fully_mask_p
;
632 /* True if have decided to use a fully-masked loop. */
635 /* When we have grouped data accesses with gaps, we may introduce invalid
636 memory accesses. We peel the last iteration of the loop to prevent
638 bool peeling_for_gaps
;
640 /* When the number of iterations is not a multiple of the vector size
641 we need to peel off iterations at the end to form an epilogue loop. */
642 bool peeling_for_niter
;
644 /* True if there are no loop carried data dependencies in the loop.
645 If loop->safelen <= 1, then this is always true, either the loop
646 didn't have any loop carried data dependencies, or the loop is being
647 vectorized guarded with some runtime alias checks, or couldn't
648 be vectorized at all, but then this field shouldn't be used.
649 For loop->safelen >= 2, the user has asserted that there are no
650 backward dependencies, but there still could be loop carried forward
651 dependencies in such loops. This flag will be false if normal
652 vectorizer data dependency analysis would fail or require versioning
653 for alias, but because of loop->safelen >= 2 it has been vectorized
654 even without versioning for alias. E.g. in:
656 for (int i = 0; i < m; i++)
658 (or #pragma simd or #pragma ivdep) we can vectorize this and it will
659 DTRT even for k > 0 && k < m, but without safelen we would not
660 vectorize this, so this field would be false. */
661 bool no_data_dependencies
;
663 /* Mark loops having masked stores. */
666 /* Queued scaling factor for the scalar loop. */
667 profile_probability scalar_loop_scaling
;
669 /* If if-conversion versioned this loop before conversion, this is the
670 loop version without if-conversion. */
671 class loop
*scalar_loop
;
673 /* For loops being epilogues of already vectorized loops
674 this points to the original vectorized loop. Otherwise NULL. */
675 _loop_vec_info
*orig_loop_info
;
677 /* Used to store loop_vec_infos of epilogues of this loop during
679 vec
<_loop_vec_info
*> epilogue_vinfos
;
683 /* Access Functions. */
684 #define LOOP_VINFO_LOOP(L) (L)->loop
685 #define LOOP_VINFO_BBS(L) (L)->bbs
686 #define LOOP_VINFO_NITERSM1(L) (L)->num_itersm1
687 #define LOOP_VINFO_NITERS(L) (L)->num_iters
688 /* Since LOOP_VINFO_NITERS and LOOP_VINFO_NITERSM1 can change after
689 prologue peeling retain total unchanged scalar loop iterations for
691 #define LOOP_VINFO_NITERS_UNCHANGED(L) (L)->num_iters_unchanged
692 #define LOOP_VINFO_NITERS_ASSUMPTIONS(L) (L)->num_iters_assumptions
693 #define LOOP_VINFO_COST_MODEL_THRESHOLD(L) (L)->th
694 #define LOOP_VINFO_VERSIONING_THRESHOLD(L) (L)->versioning_threshold
695 #define LOOP_VINFO_VECTORIZABLE_P(L) (L)->vectorizable
696 #define LOOP_VINFO_CAN_FULLY_MASK_P(L) (L)->can_fully_mask_p
697 #define LOOP_VINFO_FULLY_MASKED_P(L) (L)->fully_masked_p
698 #define LOOP_VINFO_VECT_FACTOR(L) (L)->vectorization_factor
699 #define LOOP_VINFO_MAX_VECT_FACTOR(L) (L)->max_vectorization_factor
700 #define LOOP_VINFO_MASKS(L) (L)->masks
701 #define LOOP_VINFO_MASK_SKIP_NITERS(L) (L)->mask_skip_niters
702 #define LOOP_VINFO_MASK_COMPARE_TYPE(L) (L)->mask_compare_type
703 #define LOOP_VINFO_MASK_IV_TYPE(L) (L)->iv_type
704 #define LOOP_VINFO_PTR_MASK(L) (L)->ptr_mask
705 #define LOOP_VINFO_LOOP_NEST(L) (L)->shared->loop_nest
706 #define LOOP_VINFO_DATAREFS(L) (L)->shared->datarefs
707 #define LOOP_VINFO_DDRS(L) (L)->shared->ddrs
708 #define LOOP_VINFO_INT_NITERS(L) (TREE_INT_CST_LOW ((L)->num_iters))
709 #define LOOP_VINFO_PEELING_FOR_ALIGNMENT(L) (L)->peeling_for_alignment
710 #define LOOP_VINFO_UNALIGNED_DR(L) (L)->unaligned_dr
711 #define LOOP_VINFO_MAY_MISALIGN_STMTS(L) (L)->may_misalign_stmts
712 #define LOOP_VINFO_MAY_ALIAS_DDRS(L) (L)->may_alias_ddrs
713 #define LOOP_VINFO_COMP_ALIAS_DDRS(L) (L)->comp_alias_ddrs
714 #define LOOP_VINFO_CHECK_UNEQUAL_ADDRS(L) (L)->check_unequal_addrs
715 #define LOOP_VINFO_CHECK_NONZERO(L) (L)->check_nonzero
716 #define LOOP_VINFO_LOWER_BOUNDS(L) (L)->lower_bounds
717 #define LOOP_VINFO_GROUPED_STORES(L) (L)->grouped_stores
718 #define LOOP_VINFO_SLP_INSTANCES(L) (L)->slp_instances
719 #define LOOP_VINFO_SLP_UNROLLING_FACTOR(L) (L)->slp_unrolling_factor
720 #define LOOP_VINFO_REDUCTIONS(L) (L)->reductions
721 #define LOOP_VINFO_REDUCTION_CHAINS(L) (L)->reduction_chains
722 #define LOOP_VINFO_TARGET_COST_DATA(L) (L)->target_cost_data
723 #define LOOP_VINFO_PEELING_FOR_GAPS(L) (L)->peeling_for_gaps
724 #define LOOP_VINFO_PEELING_FOR_NITER(L) (L)->peeling_for_niter
725 #define LOOP_VINFO_NO_DATA_DEPENDENCIES(L) (L)->no_data_dependencies
726 #define LOOP_VINFO_SCALAR_LOOP(L) (L)->scalar_loop
727 #define LOOP_VINFO_SCALAR_LOOP_SCALING(L) (L)->scalar_loop_scaling
728 #define LOOP_VINFO_HAS_MASK_STORE(L) (L)->has_mask_store
729 #define LOOP_VINFO_SCALAR_ITERATION_COST(L) (L)->scalar_cost_vec
730 #define LOOP_VINFO_SINGLE_SCALAR_ITERATION_COST(L) (L)->single_scalar_iteration_cost
731 #define LOOP_VINFO_ORIG_LOOP_INFO(L) (L)->orig_loop_info
732 #define LOOP_VINFO_SIMD_IF_COND(L) (L)->simd_if_cond
734 #define LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT(L) \
735 ((L)->may_misalign_stmts.length () > 0)
736 #define LOOP_REQUIRES_VERSIONING_FOR_ALIAS(L) \
737 ((L)->comp_alias_ddrs.length () > 0 \
738 || (L)->check_unequal_addrs.length () > 0 \
739 || (L)->lower_bounds.length () > 0)
740 #define LOOP_REQUIRES_VERSIONING_FOR_NITERS(L) \
741 (LOOP_VINFO_NITERS_ASSUMPTIONS (L))
742 #define LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND(L) \
743 (LOOP_VINFO_SIMD_IF_COND (L))
744 #define LOOP_REQUIRES_VERSIONING(L) \
745 (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (L) \
746 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (L) \
747 || LOOP_REQUIRES_VERSIONING_FOR_NITERS (L) \
748 || LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND (L))
750 #define LOOP_VINFO_NITERS_KNOWN_P(L) \
751 (tree_fits_shwi_p ((L)->num_iters) && tree_to_shwi ((L)->num_iters) > 0)
753 #define LOOP_VINFO_EPILOGUE_P(L) \
754 (LOOP_VINFO_ORIG_LOOP_INFO (L) != NULL)
756 #define LOOP_VINFO_ORIG_MAX_VECT_FACTOR(L) \
757 (LOOP_VINFO_MAX_VECT_FACTOR (LOOP_VINFO_ORIG_LOOP_INFO (L)))
759 /* Wrapper for loop_vec_info, for tracking success/failure, where a non-NULL
760 value signifies success, and a NULL value signifies failure, supporting
761 propagating an opt_problem * describing the failure back up the call
763 typedef opt_pointer_wrapper
<loop_vec_info
> opt_loop_vec_info
;
765 static inline loop_vec_info
766 loop_vec_info_for_loop (class loop
*loop
)
768 return (loop_vec_info
) loop
->aux
;
771 typedef class _bb_vec_info
: public vec_info
774 _bb_vec_info (gimple_stmt_iterator
, gimple_stmt_iterator
, vec_info_shared
*);
778 gimple_stmt_iterator region_begin
;
779 gimple_stmt_iterator region_end
;
782 #define BB_VINFO_BB(B) (B)->bb
783 #define BB_VINFO_GROUPED_STORES(B) (B)->grouped_stores
784 #define BB_VINFO_SLP_INSTANCES(B) (B)->slp_instances
785 #define BB_VINFO_DATAREFS(B) (B)->shared->datarefs
786 #define BB_VINFO_DDRS(B) (B)->shared->ddrs
787 #define BB_VINFO_TARGET_COST_DATA(B) (B)->target_cost_data
789 static inline bb_vec_info
790 vec_info_for_bb (basic_block bb
)
792 return (bb_vec_info
) bb
->aux
;
795 /*-----------------------------------------------------------------*/
796 /* Info on vectorized defs. */
797 /*-----------------------------------------------------------------*/
798 enum stmt_vec_info_type
{
799 undef_vec_info_type
= 0,
805 call_simd_clone_vec_info_type
,
806 assignment_vec_info_type
,
807 condition_vec_info_type
,
808 comparison_vec_info_type
,
811 type_promotion_vec_info_type
,
812 type_demotion_vec_info_type
,
813 type_conversion_vec_info_type
,
816 loop_exit_ctrl_vec_info_type
819 /* Indicates whether/how a variable is used in the scope of loop/basic
822 vect_unused_in_scope
= 0,
824 /* The def is only used outside the loop. */
826 /* The def is in the inner loop, and the use is in the outer loop, and the
827 use is a reduction stmt. */
828 vect_used_in_outer_by_reduction
,
829 /* The def is in the inner loop, and the use is in the outer loop (and is
830 not part of reduction). */
833 /* defs that feed computations that end up (only) in a reduction. These
834 defs may be used by non-reduction stmts, but eventually, any
835 computations/values that are affected by these defs are used to compute
836 a reduction (i.e. don't get stored to memory, for example). We use this
837 to identify computations that we can change the order in which they are
839 vect_used_by_reduction
,
844 /* The type of vectorization that can be applied to the stmt: regular loop-based
845 vectorization; pure SLP - the stmt is a part of SLP instances and does not
846 have uses outside SLP instances; or hybrid SLP and loop-based - the stmt is
847 a part of SLP instance and also must be loop-based vectorized, since it has
848 uses outside SLP sequences.
850 In the loop context the meanings of pure and hybrid SLP are slightly
851 different. By saying that pure SLP is applied to the loop, we mean that we
852 exploit only intra-iteration parallelism in the loop; i.e., the loop can be
853 vectorized without doing any conceptual unrolling, cause we don't pack
854 together stmts from different iterations, only within a single iteration.
855 Loop hybrid SLP means that we exploit both intra-iteration and
856 inter-iteration parallelism (e.g., number of elements in the vector is 4
857 and the slp-group-size is 2, in which case we don't have enough parallelism
858 within an iteration, so we obtain the rest of the parallelism from subsequent
859 iterations by unrolling the loop by 2). */
866 /* Says whether a statement is a load, a store of a vectorized statement
867 result, or a store of an invariant value. */
868 enum vec_load_store_type
{
874 /* Describes how we're going to vectorize an individual load or store,
875 or a group of loads or stores. */
876 enum vect_memory_access_type
{
877 /* An access to an invariant address. This is used only for loads. */
880 /* A simple contiguous access. */
883 /* A contiguous access that goes down in memory rather than up,
884 with no additional permutation. This is used only for stores
886 VMAT_CONTIGUOUS_DOWN
,
888 /* A simple contiguous access in which the elements need to be permuted
889 after loading or before storing. Only used for loop vectorization;
890 SLP uses separate permutes. */
891 VMAT_CONTIGUOUS_PERMUTE
,
893 /* A simple contiguous access in which the elements need to be reversed
894 after loading or before storing. */
895 VMAT_CONTIGUOUS_REVERSE
,
897 /* An access that uses IFN_LOAD_LANES or IFN_STORE_LANES. */
898 VMAT_LOAD_STORE_LANES
,
900 /* An access in which each scalar element is loaded or stored
904 /* A hybrid of VMAT_CONTIGUOUS and VMAT_ELEMENTWISE, used for grouped
905 SLP accesses. Each unrolled iteration uses a contiguous load
906 or store for the whole group, but the groups from separate iterations
907 are combined in the same way as for VMAT_ELEMENTWISE. */
910 /* The access uses gather loads or scatter stores. */
916 /* The data reference itself. */
918 /* The statement that contains the data reference. */
920 /* The misalignment in bytes of the reference, or -1 if not known. */
922 /* The byte alignment that we'd ideally like the reference to have,
923 and the value that misalignment is measured against. */
924 poly_uint64 target_alignment
;
925 /* If true the alignment of base_decl needs to be increased. */
926 bool base_misaligned
;
929 /* Stores current vectorized loop's offset. To be added to the DR's
930 offset to calculate current offset of data reference. */
934 typedef struct data_reference
*dr_p
;
936 class _stmt_vec_info
{
939 enum stmt_vec_info_type type
;
941 /* Indicates whether this stmts is part of a computation whose result is
942 used outside the loop. */
945 /* Stmt is part of some pattern (computation idiom) */
948 /* True if the statement was created during pattern recognition as
949 part of the replacement for RELATED_STMT. This implies that the
950 statement isn't part of any basic block, although for convenience
951 its gimple_bb is the same as for RELATED_STMT. */
954 /* Is this statement vectorizable or should it be skipped in (partial)
958 /* The stmt to which this info struct refers to. */
961 /* The vector type to be used for the LHS of this statement. */
964 /* The vectorized version of the stmt. */
965 stmt_vec_info vectorized_stmt
;
968 /* The following is relevant only for stmts that contain a non-scalar
969 data-ref (array/pointer/struct access). A GIMPLE stmt is expected to have
970 at most one such data-ref. */
974 /* Information about the data-ref relative to this loop
975 nest (the loop that is being considered for vectorization). */
976 innermost_loop_behavior dr_wrt_vec_loop
;
978 /* For loop PHI nodes, the base and evolution part of it. This makes sure
979 this information is still available in vect_update_ivs_after_vectorizer
980 where we may not be able to re-analyze the PHI nodes evolution as
981 peeling for the prologue loop can make it unanalyzable. The evolution
982 part is still correct after peeling, but the base may have changed from
984 tree loop_phi_evolution_base_unchanged
;
985 tree loop_phi_evolution_part
;
987 /* Used for various bookkeeping purposes, generally holding a pointer to
988 some other stmt S that is in some way "related" to this stmt.
989 Current use of this field is:
990 If this stmt is part of a pattern (i.e. the field 'in_pattern_p' is
991 true): S is the "pattern stmt" that represents (and replaces) the
992 sequence of stmts that constitutes the pattern. Similarly, the
993 related_stmt of the "pattern stmt" points back to this stmt (which is
994 the last stmt in the original sequence of stmts that constitutes the
996 stmt_vec_info related_stmt
;
998 /* Used to keep a sequence of def stmts of a pattern stmt if such exists.
999 The sequence is attached to the original statement rather than the
1000 pattern statement. */
1001 gimple_seq pattern_def_seq
;
1003 /* List of datarefs that are known to have the same alignment as the dataref
1005 vec
<dr_p
> same_align_refs
;
1007 /* Selected SIMD clone's function info. First vector element
1008 is SIMD clone's function decl, followed by a pair of trees (base + step)
1009 for linear arguments (pair of NULLs for other arguments). */
1010 vec
<tree
> simd_clone_info
;
1012 /* Classify the def of this stmt. */
1013 enum vect_def_type def_type
;
1015 /* Whether the stmt is SLPed, loop-based vectorized, or both. */
1016 enum slp_vect_type slp_type
;
1018 /* Interleaving and reduction chains info. */
1019 /* First element in the group. */
1020 stmt_vec_info first_element
;
1021 /* Pointer to the next element in the group. */
1022 stmt_vec_info next_element
;
1023 /* The size of the group. */
1025 /* For stores, number of stores from this group seen. We vectorize the last
1027 unsigned int store_count
;
1028 /* For loads only, the gap from the previous load. For consecutive loads, GAP
1032 /* The minimum negative dependence distance this stmt participates in
1034 unsigned int min_neg_dist
;
1036 /* Not all stmts in the loop need to be vectorized. e.g, the increment
1037 of the loop induction variable and computation of array indexes. relevant
1038 indicates whether the stmt needs to be vectorized. */
1039 enum vect_relevant relevant
;
1041 /* For loads if this is a gather, for stores if this is a scatter. */
1042 bool gather_scatter_p
;
1044 /* True if this is an access with loop-invariant stride. */
1047 /* For both loads and stores. */
1048 unsigned simd_lane_access_p
: 3;
1050 /* Classifies how the load or store is going to be implemented
1051 for loop vectorization. */
1052 vect_memory_access_type memory_access_type
;
1054 /* For INTEGER_INDUC_COND_REDUCTION, the initial value to be used. */
1055 tree induc_cond_initial_val
;
1057 /* If not NULL the value to be added to compute final reduction value. */
1058 tree reduc_epilogue_adjustment
;
1060 /* On a reduction PHI the reduction type as detected by
1061 vect_is_simple_reduction and vectorizable_reduction. */
1062 enum vect_reduction_type reduc_type
;
1064 /* The original reduction code, to be used in the epilogue. */
1065 enum tree_code reduc_code
;
1066 /* An internal function we should use in the epilogue. */
1067 internal_fn reduc_fn
;
1069 /* On a stmt participating in the reduction the index of the operand
1070 on the reduction SSA cycle. */
1073 /* On a reduction PHI the def returned by vect_force_simple_reduction.
1074 On the def returned by vect_force_simple_reduction the
1075 corresponding PHI. */
1076 stmt_vec_info reduc_def
;
1078 /* The vector input type relevant for reduction vectorization. */
1079 tree reduc_vectype_in
;
1081 /* The vector type for performing the actual reduction. */
1084 /* Whether we force a single cycle PHI during reduction vectorization. */
1085 bool force_single_cycle
;
1087 /* Whether on this stmt reduction meta is recorded. */
1090 /* The number of scalar stmt references from active SLP instances. */
1091 unsigned int num_slp_uses
;
1093 /* If nonzero, the lhs of the statement could be truncated to this
1094 many bits without affecting any users of the result. */
1095 unsigned int min_output_precision
;
1097 /* If nonzero, all non-boolean input operands have the same precision,
1098 and they could each be truncated to this many bits without changing
1100 unsigned int min_input_precision
;
1102 /* If OPERATION_BITS is nonzero, the statement could be performed on
1103 an integer with the sign and number of bits given by OPERATION_SIGN
1104 and OPERATION_BITS without changing the result. */
1105 unsigned int operation_precision
;
1106 signop operation_sign
;
1108 /* If the statement produces a boolean result, this value describes
1109 how we should choose the associated vector type. The possible
1112 - an integer precision N if we should use the vector mask type
1113 associated with N-bit integers. This is only used if all relevant
1114 input booleans also want the vector mask type for N-bit integers,
1115 or if we can convert them into that form by pattern-matching.
1117 - ~0U if we considered choosing a vector mask type but decided
1118 to treat the boolean as a normal integer type instead.
1120 - 0 otherwise. This means either that the operation isn't one that
1121 could have a vector mask type (and so should have a normal vector
1122 type instead) or that we simply haven't made a choice either way. */
1123 unsigned int mask_precision
;
1125 /* True if this is only suitable for SLP vectorization. */
1126 bool slp_vect_only_p
;
1129 /* Information about a gather/scatter call. */
1130 struct gather_scatter_info
{
1131 /* The internal function to use for the gather/scatter operation,
1132 or IFN_LAST if a built-in function should be used instead. */
1135 /* The FUNCTION_DECL for the built-in gather/scatter function,
1136 or null if an internal function should be used instead. */
1139 /* The loop-invariant base value. */
1142 /* The original scalar offset, which is a non-loop-invariant SSA_NAME. */
1145 /* Each offset element should be multiplied by this amount before
1146 being added to the base. */
1149 /* The definition type for the vectorized offset. */
1150 enum vect_def_type offset_dt
;
1152 /* The type of the vectorized offset. */
1153 tree offset_vectype
;
1155 /* The type of the scalar elements after loading or before storing. */
1158 /* The type of the scalar elements being loaded or stored. */
1162 /* Access Functions. */
1163 #define STMT_VINFO_TYPE(S) (S)->type
1164 #define STMT_VINFO_STMT(S) (S)->stmt
1165 #define STMT_VINFO_RELEVANT(S) (S)->relevant
1166 #define STMT_VINFO_LIVE_P(S) (S)->live
1167 #define STMT_VINFO_VECTYPE(S) (S)->vectype
1168 #define STMT_VINFO_VEC_STMT(S) (S)->vectorized_stmt
1169 #define STMT_VINFO_VECTORIZABLE(S) (S)->vectorizable
1170 #define STMT_VINFO_DATA_REF(S) ((S)->dr_aux.dr + 0)
1171 #define STMT_VINFO_GATHER_SCATTER_P(S) (S)->gather_scatter_p
1172 #define STMT_VINFO_STRIDED_P(S) (S)->strided_p
1173 #define STMT_VINFO_MEMORY_ACCESS_TYPE(S) (S)->memory_access_type
1174 #define STMT_VINFO_SIMD_LANE_ACCESS_P(S) (S)->simd_lane_access_p
1175 #define STMT_VINFO_VEC_INDUC_COND_INITIAL_VAL(S) (S)->induc_cond_initial_val
1176 #define STMT_VINFO_REDUC_EPILOGUE_ADJUSTMENT(S) (S)->reduc_epilogue_adjustment
1177 #define STMT_VINFO_REDUC_IDX(S) (S)->reduc_idx
1178 #define STMT_VINFO_FORCE_SINGLE_CYCLE(S) (S)->force_single_cycle
1180 #define STMT_VINFO_DR_WRT_VEC_LOOP(S) (S)->dr_wrt_vec_loop
1181 #define STMT_VINFO_DR_BASE_ADDRESS(S) (S)->dr_wrt_vec_loop.base_address
1182 #define STMT_VINFO_DR_INIT(S) (S)->dr_wrt_vec_loop.init
1183 #define STMT_VINFO_DR_OFFSET(S) (S)->dr_wrt_vec_loop.offset
1184 #define STMT_VINFO_DR_STEP(S) (S)->dr_wrt_vec_loop.step
1185 #define STMT_VINFO_DR_BASE_ALIGNMENT(S) (S)->dr_wrt_vec_loop.base_alignment
1186 #define STMT_VINFO_DR_BASE_MISALIGNMENT(S) \
1187 (S)->dr_wrt_vec_loop.base_misalignment
1188 #define STMT_VINFO_DR_OFFSET_ALIGNMENT(S) \
1189 (S)->dr_wrt_vec_loop.offset_alignment
1190 #define STMT_VINFO_DR_STEP_ALIGNMENT(S) \
1191 (S)->dr_wrt_vec_loop.step_alignment
1193 #define STMT_VINFO_DR_INFO(S) \
1194 (gcc_checking_assert ((S)->dr_aux.stmt == (S)), &(S)->dr_aux)
1196 #define STMT_VINFO_IN_PATTERN_P(S) (S)->in_pattern_p
1197 #define STMT_VINFO_RELATED_STMT(S) (S)->related_stmt
1198 #define STMT_VINFO_PATTERN_DEF_SEQ(S) (S)->pattern_def_seq
1199 #define STMT_VINFO_SAME_ALIGN_REFS(S) (S)->same_align_refs
1200 #define STMT_VINFO_SIMD_CLONE_INFO(S) (S)->simd_clone_info
1201 #define STMT_VINFO_DEF_TYPE(S) (S)->def_type
1202 #define STMT_VINFO_GROUPED_ACCESS(S) \
1203 ((S)->dr_aux.dr && DR_GROUP_FIRST_ELEMENT(S))
1204 #define STMT_VINFO_LOOP_PHI_EVOLUTION_BASE_UNCHANGED(S) (S)->loop_phi_evolution_base_unchanged
1205 #define STMT_VINFO_LOOP_PHI_EVOLUTION_PART(S) (S)->loop_phi_evolution_part
1206 #define STMT_VINFO_MIN_NEG_DIST(S) (S)->min_neg_dist
1207 #define STMT_VINFO_NUM_SLP_USES(S) (S)->num_slp_uses
1208 #define STMT_VINFO_REDUC_TYPE(S) (S)->reduc_type
1209 #define STMT_VINFO_REDUC_CODE(S) (S)->reduc_code
1210 #define STMT_VINFO_REDUC_FN(S) (S)->reduc_fn
1211 #define STMT_VINFO_REDUC_DEF(S) (S)->reduc_def
1212 #define STMT_VINFO_REDUC_VECTYPE(S) (S)->reduc_vectype
1213 #define STMT_VINFO_REDUC_VECTYPE_IN(S) (S)->reduc_vectype_in
1214 #define STMT_VINFO_SLP_VECT_ONLY(S) (S)->slp_vect_only_p
1216 #define DR_GROUP_FIRST_ELEMENT(S) \
1217 (gcc_checking_assert ((S)->dr_aux.dr), (S)->first_element)
1218 #define DR_GROUP_NEXT_ELEMENT(S) \
1219 (gcc_checking_assert ((S)->dr_aux.dr), (S)->next_element)
1220 #define DR_GROUP_SIZE(S) \
1221 (gcc_checking_assert ((S)->dr_aux.dr), (S)->size)
1222 #define DR_GROUP_STORE_COUNT(S) \
1223 (gcc_checking_assert ((S)->dr_aux.dr), (S)->store_count)
1224 #define DR_GROUP_GAP(S) \
1225 (gcc_checking_assert ((S)->dr_aux.dr), (S)->gap)
1227 #define REDUC_GROUP_FIRST_ELEMENT(S) \
1228 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->first_element)
1229 #define REDUC_GROUP_NEXT_ELEMENT(S) \
1230 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->next_element)
1231 #define REDUC_GROUP_SIZE(S) \
1232 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->size)
1234 #define STMT_VINFO_RELEVANT_P(S) ((S)->relevant != vect_unused_in_scope)
1236 #define HYBRID_SLP_STMT(S) ((S)->slp_type == hybrid)
1237 #define PURE_SLP_STMT(S) ((S)->slp_type == pure_slp)
1238 #define STMT_SLP_TYPE(S) (S)->slp_type
1240 #define VECT_MAX_COST 1000
1242 /* The maximum number of intermediate steps required in multi-step type
1244 #define MAX_INTERM_CVT_STEPS 3
1246 #define MAX_VECTORIZATION_FACTOR INT_MAX
1248 /* Nonzero if TYPE represents a (scalar) boolean type or type
1249 in the middle-end compatible with it (unsigned precision 1 integral
1250 types). Used to determine which types should be vectorized as
1251 VECTOR_BOOLEAN_TYPE_P. */
1253 #define VECT_SCALAR_BOOLEAN_TYPE_P(TYPE) \
1254 (TREE_CODE (TYPE) == BOOLEAN_TYPE \
1255 || ((TREE_CODE (TYPE) == INTEGER_TYPE \
1256 || TREE_CODE (TYPE) == ENUMERAL_TYPE) \
1257 && TYPE_PRECISION (TYPE) == 1 \
1258 && TYPE_UNSIGNED (TYPE)))
1261 nested_in_vect_loop_p (class loop
*loop
, stmt_vec_info stmt_info
)
1264 && (loop
->inner
== (gimple_bb (stmt_info
->stmt
))->loop_father
));
1267 /* Return true if STMT_INFO should produce a vector mask type rather than
1268 a normal nonmask type. */
1271 vect_use_mask_type_p (stmt_vec_info stmt_info
)
1273 return stmt_info
->mask_precision
&& stmt_info
->mask_precision
!= ~0U;
1276 /* Return TRUE if a statement represented by STMT_INFO is a part of a
1280 is_pattern_stmt_p (stmt_vec_info stmt_info
)
1282 return stmt_info
->pattern_stmt_p
;
1285 /* If STMT_INFO is a pattern statement, return the statement that it
1286 replaces, otherwise return STMT_INFO itself. */
1288 inline stmt_vec_info
1289 vect_orig_stmt (stmt_vec_info stmt_info
)
1291 if (is_pattern_stmt_p (stmt_info
))
1292 return STMT_VINFO_RELATED_STMT (stmt_info
);
1296 /* Return the later statement between STMT1_INFO and STMT2_INFO. */
1298 static inline stmt_vec_info
1299 get_later_stmt (stmt_vec_info stmt1_info
, stmt_vec_info stmt2_info
)
1301 if (gimple_uid (vect_orig_stmt (stmt1_info
)->stmt
)
1302 > gimple_uid (vect_orig_stmt (stmt2_info
)->stmt
))
1308 /* If STMT_INFO has been replaced by a pattern statement, return the
1309 replacement statement, otherwise return STMT_INFO itself. */
1311 inline stmt_vec_info
1312 vect_stmt_to_vectorize (stmt_vec_info stmt_info
)
1314 if (STMT_VINFO_IN_PATTERN_P (stmt_info
))
1315 return STMT_VINFO_RELATED_STMT (stmt_info
);
1319 /* Return true if BB is a loop header. */
1322 is_loop_header_bb_p (basic_block bb
)
1324 if (bb
== (bb
->loop_father
)->header
)
1326 gcc_checking_assert (EDGE_COUNT (bb
->preds
) == 1);
1330 /* Return pow2 (X). */
1337 for (i
= 0; i
< x
; i
++)
1343 /* Alias targetm.vectorize.builtin_vectorization_cost. */
1346 builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost
,
1347 tree vectype
, int misalign
)
1349 return targetm
.vectorize
.builtin_vectorization_cost (type_of_cost
,
1353 /* Get cost by calling cost target builtin. */
1356 int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost
)
1358 return builtin_vectorization_cost (type_of_cost
, NULL
, 0);
1361 /* Alias targetm.vectorize.init_cost. */
1363 static inline void *
1364 init_cost (class loop
*loop_info
)
1366 return targetm
.vectorize
.init_cost (loop_info
);
1369 extern void dump_stmt_cost (FILE *, void *, int, enum vect_cost_for_stmt
,
1370 stmt_vec_info
, tree
, int, unsigned,
1371 enum vect_cost_model_location
);
1373 /* Alias targetm.vectorize.add_stmt_cost. */
1375 static inline unsigned
1376 add_stmt_cost (vec_info
*vinfo
, void *data
, int count
,
1377 enum vect_cost_for_stmt kind
,
1378 stmt_vec_info stmt_info
, tree vectype
, int misalign
,
1379 enum vect_cost_model_location where
)
1381 unsigned cost
= targetm
.vectorize
.add_stmt_cost (vinfo
, data
, count
, kind
,
1384 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1385 dump_stmt_cost (dump_file
, data
, count
, kind
, stmt_info
, vectype
, misalign
,
1390 /* Alias targetm.vectorize.finish_cost. */
1393 finish_cost (void *data
, unsigned *prologue_cost
,
1394 unsigned *body_cost
, unsigned *epilogue_cost
)
1396 targetm
.vectorize
.finish_cost (data
, prologue_cost
, body_cost
, epilogue_cost
);
1399 /* Alias targetm.vectorize.destroy_cost_data. */
1402 destroy_cost_data (void *data
)
1404 targetm
.vectorize
.destroy_cost_data (data
);
1408 add_stmt_costs (vec_info
*vinfo
, void *data
, stmt_vector_for_cost
*cost_vec
)
1410 stmt_info_for_cost
*cost
;
1412 FOR_EACH_VEC_ELT (*cost_vec
, i
, cost
)
1413 add_stmt_cost (vinfo
, data
, cost
->count
, cost
->kind
, cost
->stmt_info
,
1414 cost
->vectype
, cost
->misalign
, cost
->where
);
1417 /*-----------------------------------------------------------------*/
1418 /* Info on data references alignment. */
1419 /*-----------------------------------------------------------------*/
1420 #define DR_MISALIGNMENT_UNKNOWN (-1)
1421 #define DR_MISALIGNMENT_UNINITIALIZED (-2)
1424 set_dr_misalignment (dr_vec_info
*dr_info
, int val
)
1426 dr_info
->misalignment
= val
;
1430 dr_misalignment (dr_vec_info
*dr_info
)
1432 int misalign
= dr_info
->misalignment
;
1433 gcc_assert (misalign
!= DR_MISALIGNMENT_UNINITIALIZED
);
1437 /* Reflects actual alignment of first access in the vectorized loop,
1438 taking into account peeling/versioning if applied. */
1439 #define DR_MISALIGNMENT(DR) dr_misalignment (DR)
1440 #define SET_DR_MISALIGNMENT(DR, VAL) set_dr_misalignment (DR, VAL)
1442 /* Only defined once DR_MISALIGNMENT is defined. */
1443 #define DR_TARGET_ALIGNMENT(DR) ((DR)->target_alignment)
1445 /* Return true if data access DR_INFO is aligned to its target alignment
1446 (which may be less than a full vector). */
1449 aligned_access_p (dr_vec_info
*dr_info
)
1451 return (DR_MISALIGNMENT (dr_info
) == 0);
1454 /* Return TRUE if the alignment of the data access is known, and FALSE
1458 known_alignment_for_access_p (dr_vec_info
*dr_info
)
1460 return (DR_MISALIGNMENT (dr_info
) != DR_MISALIGNMENT_UNKNOWN
);
1463 /* Return the minimum alignment in bytes that the vectorized version
1464 of DR_INFO is guaranteed to have. */
1466 static inline unsigned int
1467 vect_known_alignment_in_bytes (dr_vec_info
*dr_info
)
1469 if (DR_MISALIGNMENT (dr_info
) == DR_MISALIGNMENT_UNKNOWN
)
1470 return TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr_info
->dr
)));
1471 if (DR_MISALIGNMENT (dr_info
) == 0)
1472 return known_alignment (DR_TARGET_ALIGNMENT (dr_info
));
1473 return DR_MISALIGNMENT (dr_info
) & -DR_MISALIGNMENT (dr_info
);
1476 /* Return the behavior of DR_INFO with respect to the vectorization context
1477 (which for outer loop vectorization might not be the behavior recorded
1478 in DR_INFO itself). */
1480 static inline innermost_loop_behavior
*
1481 vect_dr_behavior (vec_info
*vinfo
, dr_vec_info
*dr_info
)
1483 stmt_vec_info stmt_info
= dr_info
->stmt
;
1484 loop_vec_info loop_vinfo
= dyn_cast
<loop_vec_info
> (vinfo
);
1485 if (loop_vinfo
== NULL
1486 || !nested_in_vect_loop_p (LOOP_VINFO_LOOP (loop_vinfo
), stmt_info
))
1487 return &DR_INNERMOST (dr_info
->dr
);
1489 return &STMT_VINFO_DR_WRT_VEC_LOOP (stmt_info
);
1492 /* Return the offset calculated by adding the offset of this DR_INFO to the
1493 corresponding data_reference's offset. If CHECK_OUTER then use
1494 vect_dr_behavior to select the appropriate data_reference to use. */
1497 get_dr_vinfo_offset (vec_info
*vinfo
,
1498 dr_vec_info
*dr_info
, bool check_outer
= false)
1500 innermost_loop_behavior
*base
;
1502 base
= vect_dr_behavior (vinfo
, dr_info
);
1504 base
= &dr_info
->dr
->innermost
;
1506 tree offset
= base
->offset
;
1508 if (!dr_info
->offset
)
1511 offset
= fold_convert (sizetype
, offset
);
1512 return fold_build2 (PLUS_EXPR
, TREE_TYPE (dr_info
->offset
), offset
,
1517 /* Return true if the vect cost model is unlimited. */
1519 unlimited_cost_model (loop_p loop
)
1521 if (loop
!= NULL
&& loop
->force_vectorize
1522 && flag_simd_cost_model
!= VECT_COST_MODEL_DEFAULT
)
1523 return flag_simd_cost_model
== VECT_COST_MODEL_UNLIMITED
;
1524 return (flag_vect_cost_model
== VECT_COST_MODEL_UNLIMITED
);
1527 /* Return true if the loop described by LOOP_VINFO is fully-masked and
1528 if the first iteration should use a partial mask in order to achieve
1532 vect_use_loop_mask_for_alignment_p (loop_vec_info loop_vinfo
)
1534 return (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo
)
1535 && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo
));
1538 /* Return the number of vectors of type VECTYPE that are needed to get
1539 NUNITS elements. NUNITS should be based on the vectorization factor,
1540 so it is always a known multiple of the number of elements in VECTYPE. */
1542 static inline unsigned int
1543 vect_get_num_vectors (poly_uint64 nunits
, tree vectype
)
1545 return exact_div (nunits
, TYPE_VECTOR_SUBPARTS (vectype
)).to_constant ();
1548 /* Return the number of copies needed for loop vectorization when
1549 a statement operates on vectors of type VECTYPE. This is the
1550 vectorization factor divided by the number of elements in
1551 VECTYPE and is always known at compile time. */
1553 static inline unsigned int
1554 vect_get_num_copies (loop_vec_info loop_vinfo
, tree vectype
)
1556 return vect_get_num_vectors (LOOP_VINFO_VECT_FACTOR (loop_vinfo
), vectype
);
1559 /* Update maximum unit count *MAX_NUNITS so that it accounts for
1560 NUNITS. *MAX_NUNITS can be 1 if we haven't yet recorded anything. */
1563 vect_update_max_nunits (poly_uint64
*max_nunits
, poly_uint64 nunits
)
1565 /* All unit counts have the form vec_info::vector_size * X for some
1566 rational X, so two unit sizes must have a common multiple.
1567 Everything is a multiple of the initial value of 1. */
1568 *max_nunits
= force_common_multiple (*max_nunits
, nunits
);
1571 /* Update maximum unit count *MAX_NUNITS so that it accounts for
1572 the number of units in vector type VECTYPE. *MAX_NUNITS can be 1
1573 if we haven't yet recorded any vector types. */
1576 vect_update_max_nunits (poly_uint64
*max_nunits
, tree vectype
)
1578 vect_update_max_nunits (max_nunits
, TYPE_VECTOR_SUBPARTS (vectype
));
1581 /* Return the vectorization factor that should be used for costing
1582 purposes while vectorizing the loop described by LOOP_VINFO.
1583 Pick a reasonable estimate if the vectorization factor isn't
1584 known at compile time. */
1586 static inline unsigned int
1587 vect_vf_for_cost (loop_vec_info loop_vinfo
)
1589 return estimated_poly_value (LOOP_VINFO_VECT_FACTOR (loop_vinfo
));
1592 /* Estimate the number of elements in VEC_TYPE for costing purposes.
1593 Pick a reasonable estimate if the exact number isn't known at
1596 static inline unsigned int
1597 vect_nunits_for_cost (tree vec_type
)
1599 return estimated_poly_value (TYPE_VECTOR_SUBPARTS (vec_type
));
1602 /* Return the maximum possible vectorization factor for LOOP_VINFO. */
1604 static inline unsigned HOST_WIDE_INT
1605 vect_max_vf (loop_vec_info loop_vinfo
)
1607 unsigned HOST_WIDE_INT vf
;
1608 if (LOOP_VINFO_VECT_FACTOR (loop_vinfo
).is_constant (&vf
))
1610 return MAX_VECTORIZATION_FACTOR
;
1613 /* Return the size of the value accessed by unvectorized data reference
1614 DR_INFO. This is only valid once STMT_VINFO_VECTYPE has been calculated
1615 for the associated gimple statement, since that guarantees that DR_INFO
1616 accesses either a scalar or a scalar equivalent. ("Scalar equivalent"
1617 here includes things like V1SI, which can be vectorized in the same way
1621 vect_get_scalar_dr_size (dr_vec_info
*dr_info
)
1623 return tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_info
->dr
))));
1626 /* Return true if LOOP_VINFO requires a runtime check for whether the
1627 vector loop is profitable. */
1630 vect_apply_runtime_profitability_check_p (loop_vec_info loop_vinfo
)
1632 unsigned int th
= LOOP_VINFO_COST_MODEL_THRESHOLD (loop_vinfo
);
1633 return (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo
)
1634 && th
>= vect_vf_for_cost (loop_vinfo
));
1637 /* Source location + hotness information. */
1638 extern dump_user_location_t vect_location
;
1640 /* A macro for calling:
1641 dump_begin_scope (MSG, vect_location);
1642 via an RAII object, thus printing "=== MSG ===\n" to the dumpfile etc,
1645 once the object goes out of scope, thus capturing the nesting of
1648 These scopes affect dump messages within them: dump messages at the
1649 top level implicitly default to MSG_PRIORITY_USER_FACING, whereas those
1650 in a nested scope implicitly default to MSG_PRIORITY_INTERNALS. */
1652 #define DUMP_VECT_SCOPE(MSG) \
1653 AUTO_DUMP_SCOPE (MSG, vect_location)
1655 /* A sentinel class for ensuring that the "vect_location" global gets
1656 reset at the end of a scope.
1658 The "vect_location" global is used during dumping and contains a
1659 location_t, which could contain references to a tree block via the
1660 ad-hoc data. This data is used for tracking inlining information,
1661 but it's not a GC root; it's simply assumed that such locations never
1662 get accessed if the blocks are optimized away.
1664 Hence we need to ensure that such locations are purged at the end
1665 of any operations using them (e.g. via this class). */
1667 class auto_purge_vect_location
1670 ~auto_purge_vect_location ();
1673 /*-----------------------------------------------------------------*/
1674 /* Function prototypes. */
1675 /*-----------------------------------------------------------------*/
1677 /* Simple loop peeling and versioning utilities for vectorizer's purposes -
1678 in tree-vect-loop-manip.c. */
1679 extern void vect_set_loop_condition (class loop
*, loop_vec_info
,
1680 tree
, tree
, tree
, bool);
1681 extern bool slpeel_can_duplicate_loop_p (const class loop
*, const_edge
);
1682 class loop
*slpeel_tree_duplicate_loop_to_edge_cfg (class loop
*,
1683 class loop
*, edge
);
1684 class loop
*vect_loop_versioning (loop_vec_info
, gimple
*);
1685 extern class loop
*vect_do_peeling (loop_vec_info
, tree
, tree
,
1686 tree
*, tree
*, tree
*, int, bool, bool,
1688 extern void vect_prepare_for_masked_peels (loop_vec_info
);
1689 extern dump_user_location_t
find_loop_location (class loop
*);
1690 extern bool vect_can_advance_ivs_p (loop_vec_info
);
1691 extern void vect_update_inits_of_drs (loop_vec_info
, tree
, tree_code
);
1693 /* In tree-vect-stmts.c. */
1694 extern tree
get_related_vectype_for_scalar_type (machine_mode
, tree
,
1696 extern tree
get_vectype_for_scalar_type (vec_info
*, tree
, unsigned int = 0);
1697 extern tree
get_vectype_for_scalar_type (vec_info
*, tree
, slp_tree
);
1698 extern tree
get_mask_type_for_scalar_type (vec_info
*, tree
, unsigned int = 0);
1699 extern tree
get_same_sized_vectype (tree
, tree
);
1700 extern bool vect_chooses_same_modes_p (vec_info
*, machine_mode
);
1701 extern bool vect_get_loop_mask_type (loop_vec_info
);
1702 extern bool vect_is_simple_use (tree
, vec_info
*, enum vect_def_type
*,
1703 stmt_vec_info
* = NULL
, gimple
** = NULL
);
1704 extern bool vect_is_simple_use (tree
, vec_info
*, enum vect_def_type
*,
1705 tree
*, stmt_vec_info
* = NULL
,
1707 extern bool vect_is_simple_use (vec_info
*, stmt_vec_info
, slp_tree
,
1708 unsigned, tree
*, slp_tree
*,
1709 enum vect_def_type
*,
1710 tree
*, stmt_vec_info
* = NULL
);
1711 extern bool vect_maybe_update_slp_op_vectype (slp_tree
, tree
);
1712 extern bool supportable_widening_operation (vec_info
*,
1713 enum tree_code
, stmt_vec_info
,
1714 tree
, tree
, enum tree_code
*,
1715 enum tree_code
*, int *,
1717 extern bool supportable_narrowing_operation (enum tree_code
, tree
, tree
,
1718 enum tree_code
*, int *,
1721 extern unsigned record_stmt_cost (stmt_vector_for_cost
*, int,
1722 enum vect_cost_for_stmt
, stmt_vec_info
,
1723 tree
, int, enum vect_cost_model_location
);
1725 /* Overload of record_stmt_cost with VECTYPE derived from STMT_INFO. */
1727 static inline unsigned
1728 record_stmt_cost (stmt_vector_for_cost
*body_cost_vec
, int count
,
1729 enum vect_cost_for_stmt kind
, stmt_vec_info stmt_info
,
1730 int misalign
, enum vect_cost_model_location where
)
1732 return record_stmt_cost (body_cost_vec
, count
, kind
, stmt_info
,
1733 STMT_VINFO_VECTYPE (stmt_info
), misalign
, where
);
1736 extern stmt_vec_info
vect_finish_replace_stmt (vec_info
*,
1737 stmt_vec_info
, gimple
*);
1738 extern stmt_vec_info
vect_finish_stmt_generation (vec_info
*,
1739 stmt_vec_info
, gimple
*,
1740 gimple_stmt_iterator
*);
1741 extern opt_result
vect_mark_stmts_to_be_vectorized (loop_vec_info
, bool *);
1742 extern tree
vect_get_store_rhs (stmt_vec_info
);
1743 extern tree
vect_get_vec_def_for_operand_1 (stmt_vec_info
, enum vect_def_type
);
1744 extern tree
vect_get_vec_def_for_operand (vec_info
*, tree
,
1745 stmt_vec_info
, tree
= NULL
);
1746 extern void vect_get_vec_defs (vec_info
*, tree
, tree
, stmt_vec_info
,
1747 vec
<tree
> *, vec
<tree
> *, slp_tree
);
1748 extern void vect_get_vec_defs_for_stmt_copy (vec_info
*,
1749 vec
<tree
> *, vec
<tree
> *);
1750 extern tree
vect_init_vector (vec_info
*, stmt_vec_info
, tree
, tree
,
1751 gimple_stmt_iterator
*);
1752 extern tree
vect_get_vec_def_for_stmt_copy (vec_info
*, tree
);
1753 extern bool vect_transform_stmt (vec_info
*, stmt_vec_info
,
1754 gimple_stmt_iterator
*,
1755 slp_tree
, slp_instance
);
1756 extern void vect_remove_stores (vec_info
*, stmt_vec_info
);
1757 extern bool vect_nop_conversion_p (stmt_vec_info
);
1758 extern opt_result
vect_analyze_stmt (vec_info
*, stmt_vec_info
, bool *,
1760 slp_instance
, stmt_vector_for_cost
*);
1761 extern void vect_get_load_cost (vec_info
*, stmt_vec_info
, int, bool,
1762 unsigned int *, unsigned int *,
1763 stmt_vector_for_cost
*,
1764 stmt_vector_for_cost
*, bool);
1765 extern void vect_get_store_cost (vec_info
*, stmt_vec_info
, int,
1766 unsigned int *, stmt_vector_for_cost
*);
1767 extern bool vect_supportable_shift (vec_info
*, enum tree_code
, tree
);
1768 extern tree
vect_gen_perm_mask_any (tree
, const vec_perm_indices
&);
1769 extern tree
vect_gen_perm_mask_checked (tree
, const vec_perm_indices
&);
1770 extern void optimize_mask_stores (class loop
*);
1771 extern gcall
*vect_gen_while (tree
, tree
, tree
);
1772 extern tree
vect_gen_while_not (gimple_seq
*, tree
, tree
, tree
);
1773 extern opt_result
vect_get_vector_types_for_stmt (vec_info
*,
1774 stmt_vec_info
, tree
*,
1775 tree
*, unsigned int = 0);
1776 extern opt_tree
vect_get_mask_type_for_stmt (stmt_vec_info
, unsigned int = 0);
1778 /* In tree-vect-data-refs.c. */
1779 extern bool vect_can_force_dr_alignment_p (const_tree
, poly_uint64
);
1780 extern enum dr_alignment_support vect_supportable_dr_alignment
1781 (vec_info
*, dr_vec_info
*, bool);
1782 extern tree
vect_get_smallest_scalar_type (stmt_vec_info
, HOST_WIDE_INT
*,
1784 extern opt_result
vect_analyze_data_ref_dependences (loop_vec_info
, unsigned int *);
1785 extern bool vect_slp_analyze_instance_dependence (vec_info
*, slp_instance
);
1786 extern opt_result
vect_enhance_data_refs_alignment (loop_vec_info
);
1787 extern opt_result
vect_analyze_data_refs_alignment (loop_vec_info
);
1788 extern opt_result
vect_verify_datarefs_alignment (loop_vec_info
);
1789 extern bool vect_slp_analyze_and_verify_instance_alignment (vec_info
*,
1791 extern opt_result
vect_analyze_data_ref_accesses (vec_info
*);
1792 extern opt_result
vect_prune_runtime_alias_test_list (loop_vec_info
);
1793 extern bool vect_gather_scatter_fn_p (vec_info
*, bool, bool, tree
, tree
,
1794 tree
, int, internal_fn
*, tree
*);
1795 extern bool vect_check_gather_scatter (stmt_vec_info
, loop_vec_info
,
1796 gather_scatter_info
*);
1797 extern opt_result
vect_find_stmt_data_reference (loop_p
, gimple
*,
1798 vec
<data_reference_p
> *);
1799 extern opt_result
vect_analyze_data_refs (vec_info
*, poly_uint64
*, bool *);
1800 extern void vect_record_base_alignments (vec_info
*);
1801 extern tree
vect_create_data_ref_ptr (vec_info
*,
1802 stmt_vec_info
, tree
, class loop
*, tree
,
1803 tree
*, gimple_stmt_iterator
*,
1805 tree
= NULL_TREE
, tree
= NULL_TREE
);
1806 extern tree
bump_vector_ptr (vec_info
*, tree
, gimple
*, gimple_stmt_iterator
*,
1807 stmt_vec_info
, tree
);
1808 extern void vect_copy_ref_info (tree
, tree
);
1809 extern tree
vect_create_destination_var (tree
, tree
);
1810 extern bool vect_grouped_store_supported (tree
, unsigned HOST_WIDE_INT
);
1811 extern bool vect_store_lanes_supported (tree
, unsigned HOST_WIDE_INT
, bool);
1812 extern bool vect_grouped_load_supported (tree
, bool, unsigned HOST_WIDE_INT
);
1813 extern bool vect_load_lanes_supported (tree
, unsigned HOST_WIDE_INT
, bool);
1814 extern void vect_permute_store_chain (vec_info
*,
1815 vec
<tree
> ,unsigned int, stmt_vec_info
,
1816 gimple_stmt_iterator
*, vec
<tree
> *);
1817 extern tree
vect_setup_realignment (vec_info
*,
1818 stmt_vec_info
, gimple_stmt_iterator
*,
1819 tree
*, enum dr_alignment_support
, tree
,
1821 extern void vect_transform_grouped_load (vec_info
*, stmt_vec_info
, vec
<tree
>,
1822 int, gimple_stmt_iterator
*);
1823 extern void vect_record_grouped_load_vectors (vec_info
*,
1824 stmt_vec_info
, vec
<tree
>);
1825 extern tree
vect_get_new_vect_var (tree
, enum vect_var_kind
, const char *);
1826 extern tree
vect_get_new_ssa_name (tree
, enum vect_var_kind
,
1827 const char * = NULL
);
1828 extern tree
vect_create_addr_base_for_vector_ref (vec_info
*,
1829 stmt_vec_info
, gimple_seq
*,
1830 tree
, tree
= NULL_TREE
);
1832 /* In tree-vect-loop.c. */
1833 extern widest_int
vect_iv_limit_for_full_masking (loop_vec_info loop_vinfo
);
1834 /* Used in tree-vect-loop-manip.c */
1835 extern void determine_peel_for_niter (loop_vec_info
);
1836 /* Used in gimple-loop-interchange.c and tree-parloops.c. */
1837 extern bool check_reduction_path (dump_user_location_t
, loop_p
, gphi
*, tree
,
1839 extern bool needs_fold_left_reduction_p (tree
, tree_code
);
1840 /* Drive for loop analysis stage. */
1841 extern opt_loop_vec_info
vect_analyze_loop (class loop
*, vec_info_shared
*);
1842 extern tree
vect_build_loop_niters (loop_vec_info
, bool * = NULL
);
1843 extern void vect_gen_vector_loop_niters (loop_vec_info
, tree
, tree
*,
1845 extern tree
vect_halve_mask_nunits (tree
, machine_mode
);
1846 extern tree
vect_double_mask_nunits (tree
, machine_mode
);
1847 extern void vect_record_loop_mask (loop_vec_info
, vec_loop_masks
*,
1848 unsigned int, tree
, tree
);
1849 extern tree
vect_get_loop_mask (gimple_stmt_iterator
*, vec_loop_masks
*,
1850 unsigned int, tree
, unsigned int);
1851 extern stmt_vec_info
info_for_reduction (vec_info
*, stmt_vec_info
);
1853 /* Drive for loop transformation stage. */
1854 extern class loop
*vect_transform_loop (loop_vec_info
, gimple
*);
1855 extern opt_loop_vec_info
vect_analyze_loop_form (class loop
*,
1857 extern bool vectorizable_live_operation (loop_vec_info
,
1858 stmt_vec_info
, gimple_stmt_iterator
*,
1859 slp_tree
, slp_instance
, int,
1860 bool, stmt_vector_for_cost
*);
1861 extern bool vectorizable_reduction (loop_vec_info
, stmt_vec_info
,
1862 slp_tree
, slp_instance
,
1863 stmt_vector_for_cost
*);
1864 extern bool vectorizable_induction (loop_vec_info
, stmt_vec_info
,
1865 gimple_stmt_iterator
*,
1866 stmt_vec_info
*, slp_tree
,
1867 stmt_vector_for_cost
*);
1868 extern bool vect_transform_reduction (loop_vec_info
, stmt_vec_info
,
1869 gimple_stmt_iterator
*,
1870 stmt_vec_info
*, slp_tree
);
1871 extern bool vect_transform_cycle_phi (loop_vec_info
, stmt_vec_info
,
1873 slp_tree
, slp_instance
);
1874 extern bool vectorizable_lc_phi (loop_vec_info
, stmt_vec_info
,
1875 stmt_vec_info
*, slp_tree
);
1876 extern bool vect_worthwhile_without_simd_p (vec_info
*, tree_code
);
1877 extern int vect_get_known_peeling_cost (loop_vec_info
, int, int *,
1878 stmt_vector_for_cost
*,
1879 stmt_vector_for_cost
*,
1880 stmt_vector_for_cost
*);
1881 extern tree
cse_and_gimplify_to_preheader (loop_vec_info
, tree
);
1883 /* In tree-vect-slp.c. */
1884 extern void vect_free_slp_instance (slp_instance
, bool);
1885 extern bool vect_transform_slp_perm_load (vec_info
*, slp_tree
, vec
<tree
>,
1886 gimple_stmt_iterator
*, poly_uint64
,
1888 extern bool vect_slp_analyze_operations (vec_info
*);
1889 extern void vect_schedule_slp (vec_info
*);
1890 extern opt_result
vect_analyze_slp (vec_info
*, unsigned);
1891 extern bool vect_make_slp_decision (loop_vec_info
);
1892 extern void vect_detect_hybrid_slp (loop_vec_info
);
1893 extern void vect_optimize_slp (vec_info
*);
1894 extern void vect_get_slp_defs (vec_info
*, slp_tree
, vec
<vec
<tree
> > *,
1896 extern bool vect_slp_bb (basic_block
);
1897 extern stmt_vec_info
vect_find_last_scalar_stmt_in_slp (slp_tree
);
1898 extern bool is_simple_and_all_uses_invariant (stmt_vec_info
, loop_vec_info
);
1899 extern bool can_duplicate_and_interleave_p (vec_info
*, unsigned int, tree
,
1900 unsigned int * = NULL
,
1901 tree
* = NULL
, tree
* = NULL
);
1902 extern void duplicate_and_interleave (vec_info
*, gimple_seq
*, tree
,
1903 vec
<tree
>, unsigned int, vec
<tree
> &);
1904 extern int vect_get_place_in_interleaving_chain (stmt_vec_info
, stmt_vec_info
);
1906 /* In tree-vect-patterns.c. */
1907 /* Pattern recognition functions.
1908 Additional pattern recognition functions can (and will) be added
1910 void vect_pattern_recog (vec_info
*);
1912 /* In tree-vectorizer.c. */
1913 unsigned vectorize_loops (void);
1914 void vect_free_loop_info_assumptions (class loop
*);
1915 gimple
*vect_loop_vectorized_call (class loop
*, gcond
**cond
= NULL
);
1918 #endif /* GCC_TREE_VECTORIZER_H */