2 Copyright (C) 2003-2019 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
;
105 typedef vec
<stmt_info_for_cost
> stmt_vector_for_cost
;
107 /* Maps base addresses to an innermost_loop_behavior that gives the maximum
108 known alignment for that base. */
109 typedef hash_map
<tree_operand_hash
,
110 innermost_loop_behavior
*> vec_base_alignments
;
112 /************************************************************************
114 ************************************************************************/
115 typedef struct _slp_tree
*slp_tree
;
117 /* A computation tree of an SLP instance. Each node corresponds to a group of
118 stmts to be packed in a SIMD stmt. */
120 /* Nodes that contain def-stmts of this node statements operands. */
121 vec
<slp_tree
> children
;
122 /* A group of scalar stmts to be vectorized together. */
123 vec
<stmt_vec_info
> stmts
;
124 /* A group of scalar operands to be vectorized together. */
126 /* Load permutation relative to the stores, NULL if there is no
128 vec
<unsigned> load_permutation
;
129 /* Vectorized stmt/s. */
130 vec
<stmt_vec_info
> vec_stmts
;
131 /* Number of vector stmts that are created to replace the group of scalar
132 stmts. It is calculated during the transformation phase as the number of
133 scalar elements in one scalar iteration (GROUP_SIZE) multiplied by VF
134 divided by vector size. */
135 unsigned int vec_stmts_size
;
136 /* Reference count in the SLP graph. */
138 /* The maximum number of vector elements for the subtree rooted
140 poly_uint64 max_nunits
;
141 /* Whether the scalar computations use two different operators. */
143 /* The DEF type of this node. */
144 enum vect_def_type def_type
;
148 /* SLP instance is a sequence of stmts in a loop that can be packed into
150 typedef class _slp_instance
{
152 /* The root of SLP tree. */
155 /* For vector constructors, the constructor stmt that the SLP tree is built
156 from, NULL otherwise. */
157 stmt_vec_info root_stmt
;
159 /* Size of groups of scalar stmts that will be replaced by SIMD stmt/s. */
160 unsigned int group_size
;
162 /* The unrolling factor required to vectorized this SLP instance. */
163 poly_uint64 unrolling_factor
;
165 /* The group of nodes that contain loads of this SLP instance. */
168 /* The SLP node containing the reduction PHIs. */
173 /* Access Functions. */
174 #define SLP_INSTANCE_TREE(S) (S)->root
175 #define SLP_INSTANCE_GROUP_SIZE(S) (S)->group_size
176 #define SLP_INSTANCE_UNROLLING_FACTOR(S) (S)->unrolling_factor
177 #define SLP_INSTANCE_LOADS(S) (S)->loads
178 #define SLP_INSTANCE_ROOT_STMT(S) (S)->root_stmt
180 #define SLP_TREE_CHILDREN(S) (S)->children
181 #define SLP_TREE_SCALAR_STMTS(S) (S)->stmts
182 #define SLP_TREE_SCALAR_OPS(S) (S)->ops
183 #define SLP_TREE_VEC_STMTS(S) (S)->vec_stmts
184 #define SLP_TREE_NUMBER_OF_VEC_STMTS(S) (S)->vec_stmts_size
185 #define SLP_TREE_LOAD_PERMUTATION(S) (S)->load_permutation
186 #define SLP_TREE_TWO_OPERATORS(S) (S)->two_operators
187 #define SLP_TREE_DEF_TYPE(S) (S)->def_type
189 /* Key for map that records association between
190 scalar conditions and corresponding loop mask, and
191 is populated by vect_record_loop_mask. */
193 struct scalar_cond_masked_key
195 scalar_cond_masked_key (tree t
, unsigned ncopies_
)
198 get_cond_ops_from_tree (t
);
201 void get_cond_ops_from_tree (tree
);
210 struct default_hash_traits
<scalar_cond_masked_key
>
212 typedef scalar_cond_masked_key compare_type
;
213 typedef scalar_cond_masked_key value_type
;
215 static inline hashval_t
220 inchash::add_expr (v
.op0
, h
, 0);
221 inchash::add_expr (v
.op1
, h
, 0);
222 h
.add_int (v
.ncopies
);
227 equal (value_type existing
, value_type candidate
)
229 return (existing
.ncopies
== candidate
.ncopies
230 && existing
.code
== candidate
.code
231 && operand_equal_p (existing
.op0
, candidate
.op0
, 0)
232 && operand_equal_p (existing
.op1
, candidate
.op1
, 0));
236 mark_empty (value_type
&v
)
242 is_empty (value_type v
)
244 return v
.ncopies
== 0;
247 static inline void mark_deleted (value_type
&) {}
249 static inline bool is_deleted (const value_type
&)
254 static inline void remove (value_type
&) {}
257 typedef hash_set
<scalar_cond_masked_key
> scalar_cond_masked_set_type
;
259 /* Describes two objects whose addresses must be unequal for the vectorized
261 typedef std::pair
<tree
, tree
> vec_object_pair
;
263 /* Records that vectorization is only possible if abs (EXPR) >= MIN_VALUE.
264 UNSIGNED_P is true if we can assume that abs (EXPR) == EXPR. */
265 class vec_lower_bound
{
267 vec_lower_bound () {}
268 vec_lower_bound (tree e
, bool u
, poly_uint64 m
)
269 : expr (e
), unsigned_p (u
), min_value (m
) {}
273 poly_uint64 min_value
;
276 /* Vectorizer state shared between different analyses like vector sizes
277 of the same CFG region. */
278 class vec_info_shared
{
283 void save_datarefs();
284 void check_datarefs();
286 /* All data references. Freed by free_data_refs, so not an auto_vec. */
287 vec
<data_reference_p
> datarefs
;
288 vec
<data_reference
> datarefs_copy
;
290 /* The loop nest in which the data dependences are computed. */
291 auto_vec
<loop_p
> loop_nest
;
293 /* All data dependences. Freed by free_dependence_relations, so not
298 /* Vectorizer state common between loop and basic-block vectorization. */
301 enum vec_kind
{ bb
, loop
};
303 vec_info (vec_kind
, void *, vec_info_shared
*);
306 stmt_vec_info
add_stmt (gimple
*);
307 stmt_vec_info
lookup_stmt (gimple
*);
308 stmt_vec_info
lookup_def (tree
);
309 stmt_vec_info
lookup_single_use (tree
);
310 class dr_vec_info
*lookup_dr (data_reference
*);
311 void move_dr (stmt_vec_info
, stmt_vec_info
);
312 void remove_stmt (stmt_vec_info
);
313 void replace_stmt (gimple_stmt_iterator
*, stmt_vec_info
, gimple
*);
315 /* The type of vectorization. */
318 /* Shared vectorizer state. */
319 vec_info_shared
*shared
;
321 /* The mapping of GIMPLE UID to stmt_vec_info. */
322 vec
<stmt_vec_info
> stmt_vec_infos
;
324 /* All SLP instances. */
325 auto_vec
<slp_instance
> slp_instances
;
327 /* Maps base addresses to an innermost_loop_behavior that gives the maximum
328 known alignment for that base. */
329 vec_base_alignments base_alignments
;
331 /* All interleaving chains of stores, represented by the first
332 stmt in the chain. */
333 auto_vec
<stmt_vec_info
> grouped_stores
;
335 /* Cost data used by the target cost model. */
336 void *target_cost_data
;
338 /* The vector size for this loop in bytes, or 0 if we haven't picked
340 poly_uint64 vector_size
;
343 stmt_vec_info
new_stmt_vec_info (gimple
*stmt
);
344 void set_vinfo_for_stmt (gimple
*, stmt_vec_info
);
345 void free_stmt_vec_infos ();
346 void free_stmt_vec_info (stmt_vec_info
);
349 class _loop_vec_info
;
355 is_a_helper
<_loop_vec_info
*>::test (vec_info
*i
)
357 return i
->kind
== vec_info::loop
;
363 is_a_helper
<_bb_vec_info
*>::test (vec_info
*i
)
365 return i
->kind
== vec_info::bb
;
369 /* In general, we can divide the vector statements in a vectorized loop
370 into related groups ("rgroups") and say that for each rgroup there is
371 some nS such that the rgroup operates on nS values from one scalar
372 iteration followed by nS values from the next. That is, if VF is the
373 vectorization factor of the loop, the rgroup operates on a sequence:
375 (1,1) (1,2) ... (1,nS) (2,1) ... (2,nS) ... (VF,1) ... (VF,nS)
377 where (i,j) represents a scalar value with index j in a scalar
378 iteration with index i.
380 [ We use the term "rgroup" to emphasise that this grouping isn't
381 necessarily the same as the grouping of statements used elsewhere.
382 For example, if we implement a group of scalar loads using gather
383 loads, we'll use a separate gather load for each scalar load, and
384 thus each gather load will belong to its own rgroup. ]
386 In general this sequence will occupy nV vectors concatenated
387 together. If these vectors have nL lanes each, the total number
388 of scalar values N is given by:
390 N = nS * VF = nV * nL
392 None of nS, VF, nV and nL are required to be a power of 2. nS and nV
393 are compile-time constants but VF and nL can be variable (if the target
394 supports variable-length vectors).
396 In classical vectorization, each iteration of the vector loop would
397 handle exactly VF iterations of the original scalar loop. However,
398 in a fully-masked loop, a particular iteration of the vector loop
399 might handle fewer than VF iterations of the scalar loop. The vector
400 lanes that correspond to iterations of the scalar loop are said to be
401 "active" and the other lanes are said to be "inactive".
403 In a fully-masked loop, many rgroups need to be masked to ensure that
404 they have no effect for the inactive lanes. Each such rgroup needs a
405 sequence of booleans in the same order as above, but with each (i,j)
406 replaced by a boolean that indicates whether iteration i is active.
407 This sequence occupies nV vector masks that again have nL lanes each.
408 Thus the mask sequence as a whole consists of VF independent booleans
409 that are each repeated nS times.
411 We make the simplifying assumption that if a sequence of nV masks is
412 suitable for one (nS,nL) pair, we can reuse it for (nS/2,nL/2) by
413 VIEW_CONVERTing it. This holds for all current targets that support
414 fully-masked loops. For example, suppose the scalar loop is:
418 for (int i = 0; i < n; ++i)
420 f[i * 2 + 0] += 1.0f;
421 f[i * 2 + 1] += 2.0f;
425 and suppose that vectors have 256 bits. The vectorized f accesses
426 will belong to one rgroup and the vectorized d access to another:
428 f rgroup: nS = 2, nV = 1, nL = 8
429 d rgroup: nS = 1, nV = 1, nL = 4
432 [ In this simple example the rgroups do correspond to the normal
433 SLP grouping scheme. ]
435 If only the first three lanes are active, the masks we need are:
437 f rgroup: 1 1 | 1 1 | 1 1 | 0 0
438 d rgroup: 1 | 1 | 1 | 0
440 Here we can use a mask calculated for f's rgroup for d's, but not
443 Thus for each value of nV, it is enough to provide nV masks, with the
444 mask being calculated based on the highest nL (or, equivalently, based
445 on the highest nS) required by any rgroup with that nV. We therefore
446 represent the entire collection of masks as a two-level table, with the
447 first level being indexed by nV - 1 (since nV == 0 doesn't exist) and
448 the second being indexed by the mask index 0 <= i < nV. */
450 /* The masks needed by rgroups with nV vectors, according to the
451 description above. */
452 struct rgroup_masks
{
453 /* The largest nS for all rgroups that use these masks. */
454 unsigned int max_nscalars_per_iter
;
456 /* The type of mask to use, based on the highest nS recorded above. */
459 /* A vector of nV masks, in iteration order. */
463 typedef auto_vec
<rgroup_masks
> vec_loop_masks
;
465 typedef auto_vec
<std::pair
<data_reference
*, tree
> > drs_init_vec
;
467 /*-----------------------------------------------------------------*/
468 /* Info on vectorized loops. */
469 /*-----------------------------------------------------------------*/
470 typedef class _loop_vec_info
: public vec_info
{
472 _loop_vec_info (class loop
*, vec_info_shared
*);
475 /* The loop to which this info struct refers to. */
478 /* The loop basic blocks. */
481 /* Number of latch executions. */
483 /* Number of iterations. */
485 /* Number of iterations of the original loop. */
486 tree num_iters_unchanged
;
487 /* Condition under which this loop is analyzed and versioned. */
488 tree num_iters_assumptions
;
490 /* Threshold of number of iterations below which vectorization will not be
491 performed. It is calculated from MIN_PROFITABLE_ITERS and
492 param_min_vect_loop_bound. */
495 /* When applying loop versioning, the vector form should only be used
496 if the number of scalar iterations is >= this value, on top of all
497 the other requirements. Ignored when loop versioning is not being
499 poly_uint64 versioning_threshold
;
501 /* Unrolling factor */
502 poly_uint64 vectorization_factor
;
504 /* Maximum runtime vectorization factor, or MAX_VECTORIZATION_FACTOR
505 if there is no particular limit. */
506 unsigned HOST_WIDE_INT max_vectorization_factor
;
508 /* The masks that a fully-masked loop should use to avoid operating
509 on inactive scalars. */
510 vec_loop_masks masks
;
512 /* Set of scalar conditions that have loop mask applied. */
513 scalar_cond_masked_set_type scalar_cond_masked_set
;
515 /* If we are using a loop mask to align memory addresses, this variable
516 contains the number of vector elements that we should skip in the
517 first iteration of the vector loop (i.e. the number of leading
518 elements that should be false in the first mask). */
519 tree mask_skip_niters
;
521 /* Type of the variables to use in the WHILE_ULT call for fully-masked
523 tree mask_compare_type
;
525 /* For #pragma omp simd if (x) loops the x expression. If constant 0,
526 the loop should not be vectorized, if constant non-zero, simd_if_cond
527 shouldn't be set and loop vectorized normally, if SSA_NAME, the loop
528 should be versioned on that condition, using scalar loop if the condition
529 is false and vectorized loop otherwise. */
532 /* Type of the IV to use in the WHILE_ULT call for fully-masked
536 /* Unknown DRs according to which loop was peeled. */
537 class dr_vec_info
*unaligned_dr
;
539 /* peeling_for_alignment indicates whether peeling for alignment will take
540 place, and what the peeling factor should be:
541 peeling_for_alignment = X means:
542 If X=0: Peeling for alignment will not be applied.
543 If X>0: Peel first X iterations.
544 If X=-1: Generate a runtime test to calculate the number of iterations
545 to be peeled, using the dataref recorded in the field
547 int peeling_for_alignment
;
549 /* The mask used to check the alignment of pointers or arrays. */
552 /* Data Dependence Relations defining address ranges that are candidates
553 for a run-time aliasing check. */
554 auto_vec
<ddr_p
> may_alias_ddrs
;
556 /* Data Dependence Relations defining address ranges together with segment
557 lengths from which the run-time aliasing check is built. */
558 auto_vec
<dr_with_seg_len_pair_t
> comp_alias_ddrs
;
560 /* Check that the addresses of each pair of objects is unequal. */
561 auto_vec
<vec_object_pair
> check_unequal_addrs
;
563 /* List of values that are required to be nonzero. This is used to check
564 whether things like "x[i * n] += 1;" are safe and eventually gets added
565 to the checks for lower bounds below. */
566 auto_vec
<tree
> check_nonzero
;
568 /* List of values that need to be checked for a minimum value. */
569 auto_vec
<vec_lower_bound
> lower_bounds
;
571 /* Statements in the loop that have data references that are candidates for a
572 runtime (loop versioning) misalignment check. */
573 auto_vec
<stmt_vec_info
> may_misalign_stmts
;
575 /* Reduction cycles detected in the loop. Used in loop-aware SLP. */
576 auto_vec
<stmt_vec_info
> reductions
;
578 /* All reduction chains in the loop, represented by the first
579 stmt in the chain. */
580 auto_vec
<stmt_vec_info
> reduction_chains
;
582 /* Cost vector for a single scalar iteration. */
583 auto_vec
<stmt_info_for_cost
> scalar_cost_vec
;
585 /* Map of IV base/step expressions to inserted name in the preheader. */
586 hash_map
<tree_operand_hash
, tree
> *ivexpr_map
;
588 /* Map of OpenMP "omp simd array" scan variables to corresponding
589 rhs of the store of the initializer. */
590 hash_map
<tree
, tree
> *scan_map
;
592 /* The unrolling factor needed to SLP the loop. In case of that pure SLP is
593 applied to the loop, i.e., no unrolling is needed, this is 1. */
594 poly_uint64 slp_unrolling_factor
;
596 /* Cost of a single scalar iteration. */
597 int single_scalar_iteration_cost
;
599 /* Is the loop vectorizable? */
602 /* Records whether we still have the option of using a fully-masked loop. */
603 bool can_fully_mask_p
;
605 /* True if have decided to use a fully-masked loop. */
608 /* When we have grouped data accesses with gaps, we may introduce invalid
609 memory accesses. We peel the last iteration of the loop to prevent
611 bool peeling_for_gaps
;
613 /* When the number of iterations is not a multiple of the vector size
614 we need to peel off iterations at the end to form an epilogue loop. */
615 bool peeling_for_niter
;
617 /* True if there are no loop carried data dependencies in the loop.
618 If loop->safelen <= 1, then this is always true, either the loop
619 didn't have any loop carried data dependencies, or the loop is being
620 vectorized guarded with some runtime alias checks, or couldn't
621 be vectorized at all, but then this field shouldn't be used.
622 For loop->safelen >= 2, the user has asserted that there are no
623 backward dependencies, but there still could be loop carried forward
624 dependencies in such loops. This flag will be false if normal
625 vectorizer data dependency analysis would fail or require versioning
626 for alias, but because of loop->safelen >= 2 it has been vectorized
627 even without versioning for alias. E.g. in:
629 for (int i = 0; i < m; i++)
631 (or #pragma simd or #pragma ivdep) we can vectorize this and it will
632 DTRT even for k > 0 && k < m, but without safelen we would not
633 vectorize this, so this field would be false. */
634 bool no_data_dependencies
;
636 /* Mark loops having masked stores. */
639 /* Queued scaling factor for the scalar loop. */
640 profile_probability scalar_loop_scaling
;
642 /* If if-conversion versioned this loop before conversion, this is the
643 loop version without if-conversion. */
644 class loop
*scalar_loop
;
646 /* For loops being epilogues of already vectorized loops
647 this points to the original vectorized loop. Otherwise NULL. */
648 _loop_vec_info
*orig_loop_info
;
650 /* Used to store loop_vec_infos of epilogues of this loop during
652 vec
<_loop_vec_info
*> epilogue_vinfos
;
656 /* Access Functions. */
657 #define LOOP_VINFO_LOOP(L) (L)->loop
658 #define LOOP_VINFO_BBS(L) (L)->bbs
659 #define LOOP_VINFO_NITERSM1(L) (L)->num_itersm1
660 #define LOOP_VINFO_NITERS(L) (L)->num_iters
661 /* Since LOOP_VINFO_NITERS and LOOP_VINFO_NITERSM1 can change after
662 prologue peeling retain total unchanged scalar loop iterations for
664 #define LOOP_VINFO_NITERS_UNCHANGED(L) (L)->num_iters_unchanged
665 #define LOOP_VINFO_NITERS_ASSUMPTIONS(L) (L)->num_iters_assumptions
666 #define LOOP_VINFO_COST_MODEL_THRESHOLD(L) (L)->th
667 #define LOOP_VINFO_VERSIONING_THRESHOLD(L) (L)->versioning_threshold
668 #define LOOP_VINFO_VECTORIZABLE_P(L) (L)->vectorizable
669 #define LOOP_VINFO_CAN_FULLY_MASK_P(L) (L)->can_fully_mask_p
670 #define LOOP_VINFO_FULLY_MASKED_P(L) (L)->fully_masked_p
671 #define LOOP_VINFO_VECT_FACTOR(L) (L)->vectorization_factor
672 #define LOOP_VINFO_MAX_VECT_FACTOR(L) (L)->max_vectorization_factor
673 #define LOOP_VINFO_MASKS(L) (L)->masks
674 #define LOOP_VINFO_MASK_SKIP_NITERS(L) (L)->mask_skip_niters
675 #define LOOP_VINFO_MASK_COMPARE_TYPE(L) (L)->mask_compare_type
676 #define LOOP_VINFO_MASK_IV_TYPE(L) (L)->iv_type
677 #define LOOP_VINFO_PTR_MASK(L) (L)->ptr_mask
678 #define LOOP_VINFO_LOOP_NEST(L) (L)->shared->loop_nest
679 #define LOOP_VINFO_DATAREFS(L) (L)->shared->datarefs
680 #define LOOP_VINFO_DDRS(L) (L)->shared->ddrs
681 #define LOOP_VINFO_INT_NITERS(L) (TREE_INT_CST_LOW ((L)->num_iters))
682 #define LOOP_VINFO_PEELING_FOR_ALIGNMENT(L) (L)->peeling_for_alignment
683 #define LOOP_VINFO_UNALIGNED_DR(L) (L)->unaligned_dr
684 #define LOOP_VINFO_MAY_MISALIGN_STMTS(L) (L)->may_misalign_stmts
685 #define LOOP_VINFO_MAY_ALIAS_DDRS(L) (L)->may_alias_ddrs
686 #define LOOP_VINFO_COMP_ALIAS_DDRS(L) (L)->comp_alias_ddrs
687 #define LOOP_VINFO_CHECK_UNEQUAL_ADDRS(L) (L)->check_unequal_addrs
688 #define LOOP_VINFO_CHECK_NONZERO(L) (L)->check_nonzero
689 #define LOOP_VINFO_LOWER_BOUNDS(L) (L)->lower_bounds
690 #define LOOP_VINFO_GROUPED_STORES(L) (L)->grouped_stores
691 #define LOOP_VINFO_SLP_INSTANCES(L) (L)->slp_instances
692 #define LOOP_VINFO_SLP_UNROLLING_FACTOR(L) (L)->slp_unrolling_factor
693 #define LOOP_VINFO_REDUCTIONS(L) (L)->reductions
694 #define LOOP_VINFO_REDUCTION_CHAINS(L) (L)->reduction_chains
695 #define LOOP_VINFO_TARGET_COST_DATA(L) (L)->target_cost_data
696 #define LOOP_VINFO_PEELING_FOR_GAPS(L) (L)->peeling_for_gaps
697 #define LOOP_VINFO_PEELING_FOR_NITER(L) (L)->peeling_for_niter
698 #define LOOP_VINFO_NO_DATA_DEPENDENCIES(L) (L)->no_data_dependencies
699 #define LOOP_VINFO_SCALAR_LOOP(L) (L)->scalar_loop
700 #define LOOP_VINFO_SCALAR_LOOP_SCALING(L) (L)->scalar_loop_scaling
701 #define LOOP_VINFO_HAS_MASK_STORE(L) (L)->has_mask_store
702 #define LOOP_VINFO_SCALAR_ITERATION_COST(L) (L)->scalar_cost_vec
703 #define LOOP_VINFO_SINGLE_SCALAR_ITERATION_COST(L) (L)->single_scalar_iteration_cost
704 #define LOOP_VINFO_ORIG_LOOP_INFO(L) (L)->orig_loop_info
705 #define LOOP_VINFO_SIMD_IF_COND(L) (L)->simd_if_cond
707 #define LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT(L) \
708 ((L)->may_misalign_stmts.length () > 0)
709 #define LOOP_REQUIRES_VERSIONING_FOR_ALIAS(L) \
710 ((L)->comp_alias_ddrs.length () > 0 \
711 || (L)->check_unequal_addrs.length () > 0 \
712 || (L)->lower_bounds.length () > 0)
713 #define LOOP_REQUIRES_VERSIONING_FOR_NITERS(L) \
714 (LOOP_VINFO_NITERS_ASSUMPTIONS (L))
715 #define LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND(L) \
716 (LOOP_VINFO_SIMD_IF_COND (L))
717 #define LOOP_REQUIRES_VERSIONING(L) \
718 (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (L) \
719 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (L) \
720 || LOOP_REQUIRES_VERSIONING_FOR_NITERS (L) \
721 || LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND (L))
723 #define LOOP_VINFO_NITERS_KNOWN_P(L) \
724 (tree_fits_shwi_p ((L)->num_iters) && tree_to_shwi ((L)->num_iters) > 0)
726 #define LOOP_VINFO_EPILOGUE_P(L) \
727 (LOOP_VINFO_ORIG_LOOP_INFO (L) != NULL)
729 #define LOOP_VINFO_ORIG_MAX_VECT_FACTOR(L) \
730 (LOOP_VINFO_MAX_VECT_FACTOR (LOOP_VINFO_ORIG_LOOP_INFO (L)))
732 /* Wrapper for loop_vec_info, for tracking success/failure, where a non-NULL
733 value signifies success, and a NULL value signifies failure, supporting
734 propagating an opt_problem * describing the failure back up the call
736 typedef opt_pointer_wrapper
<loop_vec_info
> opt_loop_vec_info
;
738 static inline loop_vec_info
739 loop_vec_info_for_loop (class loop
*loop
)
741 return (loop_vec_info
) loop
->aux
;
744 typedef class _bb_vec_info
: public vec_info
747 _bb_vec_info (gimple_stmt_iterator
, gimple_stmt_iterator
, vec_info_shared
*);
751 gimple_stmt_iterator region_begin
;
752 gimple_stmt_iterator region_end
;
755 #define BB_VINFO_BB(B) (B)->bb
756 #define BB_VINFO_GROUPED_STORES(B) (B)->grouped_stores
757 #define BB_VINFO_SLP_INSTANCES(B) (B)->slp_instances
758 #define BB_VINFO_DATAREFS(B) (B)->shared->datarefs
759 #define BB_VINFO_DDRS(B) (B)->shared->ddrs
760 #define BB_VINFO_TARGET_COST_DATA(B) (B)->target_cost_data
762 static inline bb_vec_info
763 vec_info_for_bb (basic_block bb
)
765 return (bb_vec_info
) bb
->aux
;
768 /*-----------------------------------------------------------------*/
769 /* Info on vectorized defs. */
770 /*-----------------------------------------------------------------*/
771 enum stmt_vec_info_type
{
772 undef_vec_info_type
= 0,
778 call_simd_clone_vec_info_type
,
779 assignment_vec_info_type
,
780 condition_vec_info_type
,
781 comparison_vec_info_type
,
784 type_promotion_vec_info_type
,
785 type_demotion_vec_info_type
,
786 type_conversion_vec_info_type
,
789 loop_exit_ctrl_vec_info_type
792 /* Indicates whether/how a variable is used in the scope of loop/basic
795 vect_unused_in_scope
= 0,
797 /* The def is only used outside the loop. */
799 /* The def is in the inner loop, and the use is in the outer loop, and the
800 use is a reduction stmt. */
801 vect_used_in_outer_by_reduction
,
802 /* The def is in the inner loop, and the use is in the outer loop (and is
803 not part of reduction). */
806 /* defs that feed computations that end up (only) in a reduction. These
807 defs may be used by non-reduction stmts, but eventually, any
808 computations/values that are affected by these defs are used to compute
809 a reduction (i.e. don't get stored to memory, for example). We use this
810 to identify computations that we can change the order in which they are
812 vect_used_by_reduction
,
817 /* The type of vectorization that can be applied to the stmt: regular loop-based
818 vectorization; pure SLP - the stmt is a part of SLP instances and does not
819 have uses outside SLP instances; or hybrid SLP and loop-based - the stmt is
820 a part of SLP instance and also must be loop-based vectorized, since it has
821 uses outside SLP sequences.
823 In the loop context the meanings of pure and hybrid SLP are slightly
824 different. By saying that pure SLP is applied to the loop, we mean that we
825 exploit only intra-iteration parallelism in the loop; i.e., the loop can be
826 vectorized without doing any conceptual unrolling, cause we don't pack
827 together stmts from different iterations, only within a single iteration.
828 Loop hybrid SLP means that we exploit both intra-iteration and
829 inter-iteration parallelism (e.g., number of elements in the vector is 4
830 and the slp-group-size is 2, in which case we don't have enough parallelism
831 within an iteration, so we obtain the rest of the parallelism from subsequent
832 iterations by unrolling the loop by 2). */
839 /* Says whether a statement is a load, a store of a vectorized statement
840 result, or a store of an invariant value. */
841 enum vec_load_store_type
{
847 /* Describes how we're going to vectorize an individual load or store,
848 or a group of loads or stores. */
849 enum vect_memory_access_type
{
850 /* An access to an invariant address. This is used only for loads. */
853 /* A simple contiguous access. */
856 /* A contiguous access that goes down in memory rather than up,
857 with no additional permutation. This is used only for stores
859 VMAT_CONTIGUOUS_DOWN
,
861 /* A simple contiguous access in which the elements need to be permuted
862 after loading or before storing. Only used for loop vectorization;
863 SLP uses separate permutes. */
864 VMAT_CONTIGUOUS_PERMUTE
,
866 /* A simple contiguous access in which the elements need to be reversed
867 after loading or before storing. */
868 VMAT_CONTIGUOUS_REVERSE
,
870 /* An access that uses IFN_LOAD_LANES or IFN_STORE_LANES. */
871 VMAT_LOAD_STORE_LANES
,
873 /* An access in which each scalar element is loaded or stored
877 /* A hybrid of VMAT_CONTIGUOUS and VMAT_ELEMENTWISE, used for grouped
878 SLP accesses. Each unrolled iteration uses a contiguous load
879 or store for the whole group, but the groups from separate iterations
880 are combined in the same way as for VMAT_ELEMENTWISE. */
883 /* The access uses gather loads or scatter stores. */
889 /* The data reference itself. */
891 /* The statement that contains the data reference. */
893 /* The misalignment in bytes of the reference, or -1 if not known. */
895 /* The byte alignment that we'd ideally like the reference to have,
896 and the value that misalignment is measured against. */
897 poly_uint64 target_alignment
;
898 /* If true the alignment of base_decl needs to be increased. */
899 bool base_misaligned
;
903 typedef struct data_reference
*dr_p
;
905 class _stmt_vec_info
{
908 enum stmt_vec_info_type type
;
910 /* Indicates whether this stmts is part of a computation whose result is
911 used outside the loop. */
914 /* Stmt is part of some pattern (computation idiom) */
917 /* True if the statement was created during pattern recognition as
918 part of the replacement for RELATED_STMT. This implies that the
919 statement isn't part of any basic block, although for convenience
920 its gimple_bb is the same as for RELATED_STMT. */
923 /* Is this statement vectorizable or should it be skipped in (partial)
927 /* The stmt to which this info struct refers to. */
930 /* The vec_info with respect to which STMT is vectorized. */
933 /* The vector type to be used for the LHS of this statement. */
936 /* The vectorized version of the stmt. */
937 stmt_vec_info vectorized_stmt
;
940 /* The following is relevant only for stmts that contain a non-scalar
941 data-ref (array/pointer/struct access). A GIMPLE stmt is expected to have
942 at most one such data-ref. */
946 /* Information about the data-ref relative to this loop
947 nest (the loop that is being considered for vectorization). */
948 innermost_loop_behavior dr_wrt_vec_loop
;
950 /* For loop PHI nodes, the base and evolution part of it. This makes sure
951 this information is still available in vect_update_ivs_after_vectorizer
952 where we may not be able to re-analyze the PHI nodes evolution as
953 peeling for the prologue loop can make it unanalyzable. The evolution
954 part is still correct after peeling, but the base may have changed from
956 tree loop_phi_evolution_base_unchanged
;
957 tree loop_phi_evolution_part
;
959 /* Used for various bookkeeping purposes, generally holding a pointer to
960 some other stmt S that is in some way "related" to this stmt.
961 Current use of this field is:
962 If this stmt is part of a pattern (i.e. the field 'in_pattern_p' is
963 true): S is the "pattern stmt" that represents (and replaces) the
964 sequence of stmts that constitutes the pattern. Similarly, the
965 related_stmt of the "pattern stmt" points back to this stmt (which is
966 the last stmt in the original sequence of stmts that constitutes the
968 stmt_vec_info related_stmt
;
970 /* Used to keep a sequence of def stmts of a pattern stmt if such exists.
971 The sequence is attached to the original statement rather than the
972 pattern statement. */
973 gimple_seq pattern_def_seq
;
975 /* List of datarefs that are known to have the same alignment as the dataref
977 vec
<dr_p
> same_align_refs
;
979 /* Selected SIMD clone's function info. First vector element
980 is SIMD clone's function decl, followed by a pair of trees (base + step)
981 for linear arguments (pair of NULLs for other arguments). */
982 vec
<tree
> simd_clone_info
;
984 /* Classify the def of this stmt. */
985 enum vect_def_type def_type
;
987 /* Whether the stmt is SLPed, loop-based vectorized, or both. */
988 enum slp_vect_type slp_type
;
990 /* Interleaving and reduction chains info. */
991 /* First element in the group. */
992 stmt_vec_info first_element
;
993 /* Pointer to the next element in the group. */
994 stmt_vec_info next_element
;
995 /* The size of the group. */
997 /* For stores, number of stores from this group seen. We vectorize the last
999 unsigned int store_count
;
1000 /* For loads only, the gap from the previous load. For consecutive loads, GAP
1004 /* The minimum negative dependence distance this stmt participates in
1006 unsigned int min_neg_dist
;
1008 /* Not all stmts in the loop need to be vectorized. e.g, the increment
1009 of the loop induction variable and computation of array indexes. relevant
1010 indicates whether the stmt needs to be vectorized. */
1011 enum vect_relevant relevant
;
1013 /* For loads if this is a gather, for stores if this is a scatter. */
1014 bool gather_scatter_p
;
1016 /* True if this is an access with loop-invariant stride. */
1019 /* For both loads and stores. */
1020 unsigned simd_lane_access_p
: 3;
1022 /* Classifies how the load or store is going to be implemented
1023 for loop vectorization. */
1024 vect_memory_access_type memory_access_type
;
1026 /* For INTEGER_INDUC_COND_REDUCTION, the initial value to be used. */
1027 tree induc_cond_initial_val
;
1029 /* If not NULL the value to be added to compute final reduction value. */
1030 tree reduc_epilogue_adjustment
;
1032 /* On a reduction PHI the reduction type as detected by
1033 vect_is_simple_reduction and vectorizable_reduction. */
1034 enum vect_reduction_type reduc_type
;
1036 /* The original reduction code, to be used in the epilogue. */
1037 enum tree_code reduc_code
;
1038 /* An internal function we should use in the epilogue. */
1039 internal_fn reduc_fn
;
1041 /* On a stmt participating in the reduction the index of the operand
1042 on the reduction SSA cycle. */
1045 /* On a reduction PHI the def returned by vect_force_simple_reduction.
1046 On the def returned by vect_force_simple_reduction the
1047 corresponding PHI. */
1048 stmt_vec_info reduc_def
;
1050 /* The vector input type relevant for reduction vectorization. */
1051 tree reduc_vectype_in
;
1053 /* The vector type for performing the actual reduction. */
1056 /* Whether we force a single cycle PHI during reduction vectorization. */
1057 bool force_single_cycle
;
1059 /* Whether on this stmt reduction meta is recorded. */
1062 /* The number of scalar stmt references from active SLP instances. */
1063 unsigned int num_slp_uses
;
1065 /* If nonzero, the lhs of the statement could be truncated to this
1066 many bits without affecting any users of the result. */
1067 unsigned int min_output_precision
;
1069 /* If nonzero, all non-boolean input operands have the same precision,
1070 and they could each be truncated to this many bits without changing
1072 unsigned int min_input_precision
;
1074 /* If OPERATION_BITS is nonzero, the statement could be performed on
1075 an integer with the sign and number of bits given by OPERATION_SIGN
1076 and OPERATION_BITS without changing the result. */
1077 unsigned int operation_precision
;
1078 signop operation_sign
;
1080 /* True if this is only suitable for SLP vectorization. */
1081 bool slp_vect_only_p
;
1084 /* Information about a gather/scatter call. */
1085 struct gather_scatter_info
{
1086 /* The internal function to use for the gather/scatter operation,
1087 or IFN_LAST if a built-in function should be used instead. */
1090 /* The FUNCTION_DECL for the built-in gather/scatter function,
1091 or null if an internal function should be used instead. */
1094 /* The loop-invariant base value. */
1097 /* The original scalar offset, which is a non-loop-invariant SSA_NAME. */
1100 /* Each offset element should be multiplied by this amount before
1101 being added to the base. */
1104 /* The definition type for the vectorized offset. */
1105 enum vect_def_type offset_dt
;
1107 /* The type of the vectorized offset. */
1108 tree offset_vectype
;
1110 /* The type of the scalar elements after loading or before storing. */
1113 /* The type of the scalar elements being loaded or stored. */
1117 /* Access Functions. */
1118 #define STMT_VINFO_TYPE(S) (S)->type
1119 #define STMT_VINFO_STMT(S) (S)->stmt
1120 inline loop_vec_info
1121 STMT_VINFO_LOOP_VINFO (stmt_vec_info stmt_vinfo
)
1123 if (loop_vec_info loop_vinfo
= dyn_cast
<loop_vec_info
> (stmt_vinfo
->vinfo
))
1128 STMT_VINFO_BB_VINFO (stmt_vec_info stmt_vinfo
)
1130 if (bb_vec_info bb_vinfo
= dyn_cast
<bb_vec_info
> (stmt_vinfo
->vinfo
))
1134 #define STMT_VINFO_RELEVANT(S) (S)->relevant
1135 #define STMT_VINFO_LIVE_P(S) (S)->live
1136 #define STMT_VINFO_VECTYPE(S) (S)->vectype
1137 #define STMT_VINFO_VEC_STMT(S) (S)->vectorized_stmt
1138 #define STMT_VINFO_VECTORIZABLE(S) (S)->vectorizable
1139 #define STMT_VINFO_DATA_REF(S) ((S)->dr_aux.dr + 0)
1140 #define STMT_VINFO_GATHER_SCATTER_P(S) (S)->gather_scatter_p
1141 #define STMT_VINFO_STRIDED_P(S) (S)->strided_p
1142 #define STMT_VINFO_MEMORY_ACCESS_TYPE(S) (S)->memory_access_type
1143 #define STMT_VINFO_SIMD_LANE_ACCESS_P(S) (S)->simd_lane_access_p
1144 #define STMT_VINFO_VEC_INDUC_COND_INITIAL_VAL(S) (S)->induc_cond_initial_val
1145 #define STMT_VINFO_REDUC_EPILOGUE_ADJUSTMENT(S) (S)->reduc_epilogue_adjustment
1146 #define STMT_VINFO_REDUC_IDX(S) (S)->reduc_idx
1147 #define STMT_VINFO_FORCE_SINGLE_CYCLE(S) (S)->force_single_cycle
1149 #define STMT_VINFO_DR_WRT_VEC_LOOP(S) (S)->dr_wrt_vec_loop
1150 #define STMT_VINFO_DR_BASE_ADDRESS(S) (S)->dr_wrt_vec_loop.base_address
1151 #define STMT_VINFO_DR_INIT(S) (S)->dr_wrt_vec_loop.init
1152 #define STMT_VINFO_DR_OFFSET(S) (S)->dr_wrt_vec_loop.offset
1153 #define STMT_VINFO_DR_STEP(S) (S)->dr_wrt_vec_loop.step
1154 #define STMT_VINFO_DR_BASE_ALIGNMENT(S) (S)->dr_wrt_vec_loop.base_alignment
1155 #define STMT_VINFO_DR_BASE_MISALIGNMENT(S) \
1156 (S)->dr_wrt_vec_loop.base_misalignment
1157 #define STMT_VINFO_DR_OFFSET_ALIGNMENT(S) \
1158 (S)->dr_wrt_vec_loop.offset_alignment
1159 #define STMT_VINFO_DR_STEP_ALIGNMENT(S) \
1160 (S)->dr_wrt_vec_loop.step_alignment
1162 #define STMT_VINFO_DR_INFO(S) \
1163 (gcc_checking_assert ((S)->dr_aux.stmt == (S)), &(S)->dr_aux)
1165 #define STMT_VINFO_IN_PATTERN_P(S) (S)->in_pattern_p
1166 #define STMT_VINFO_RELATED_STMT(S) (S)->related_stmt
1167 #define STMT_VINFO_PATTERN_DEF_SEQ(S) (S)->pattern_def_seq
1168 #define STMT_VINFO_SAME_ALIGN_REFS(S) (S)->same_align_refs
1169 #define STMT_VINFO_SIMD_CLONE_INFO(S) (S)->simd_clone_info
1170 #define STMT_VINFO_DEF_TYPE(S) (S)->def_type
1171 #define STMT_VINFO_GROUPED_ACCESS(S) \
1172 ((S)->dr_aux.dr && DR_GROUP_FIRST_ELEMENT(S))
1173 #define STMT_VINFO_LOOP_PHI_EVOLUTION_BASE_UNCHANGED(S) (S)->loop_phi_evolution_base_unchanged
1174 #define STMT_VINFO_LOOP_PHI_EVOLUTION_PART(S) (S)->loop_phi_evolution_part
1175 #define STMT_VINFO_MIN_NEG_DIST(S) (S)->min_neg_dist
1176 #define STMT_VINFO_NUM_SLP_USES(S) (S)->num_slp_uses
1177 #define STMT_VINFO_REDUC_TYPE(S) (S)->reduc_type
1178 #define STMT_VINFO_REDUC_CODE(S) (S)->reduc_code
1179 #define STMT_VINFO_REDUC_FN(S) (S)->reduc_fn
1180 #define STMT_VINFO_REDUC_DEF(S) (S)->reduc_def
1181 #define STMT_VINFO_REDUC_VECTYPE(S) (S)->reduc_vectype
1182 #define STMT_VINFO_REDUC_VECTYPE_IN(S) (S)->reduc_vectype_in
1183 #define STMT_VINFO_SLP_VECT_ONLY(S) (S)->slp_vect_only_p
1185 #define DR_GROUP_FIRST_ELEMENT(S) \
1186 (gcc_checking_assert ((S)->dr_aux.dr), (S)->first_element)
1187 #define DR_GROUP_NEXT_ELEMENT(S) \
1188 (gcc_checking_assert ((S)->dr_aux.dr), (S)->next_element)
1189 #define DR_GROUP_SIZE(S) \
1190 (gcc_checking_assert ((S)->dr_aux.dr), (S)->size)
1191 #define DR_GROUP_STORE_COUNT(S) \
1192 (gcc_checking_assert ((S)->dr_aux.dr), (S)->store_count)
1193 #define DR_GROUP_GAP(S) \
1194 (gcc_checking_assert ((S)->dr_aux.dr), (S)->gap)
1196 #define REDUC_GROUP_FIRST_ELEMENT(S) \
1197 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->first_element)
1198 #define REDUC_GROUP_NEXT_ELEMENT(S) \
1199 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->next_element)
1200 #define REDUC_GROUP_SIZE(S) \
1201 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->size)
1203 #define STMT_VINFO_RELEVANT_P(S) ((S)->relevant != vect_unused_in_scope)
1205 #define HYBRID_SLP_STMT(S) ((S)->slp_type == hybrid)
1206 #define PURE_SLP_STMT(S) ((S)->slp_type == pure_slp)
1207 #define STMT_SLP_TYPE(S) (S)->slp_type
1209 #define VECT_MAX_COST 1000
1211 /* The maximum number of intermediate steps required in multi-step type
1213 #define MAX_INTERM_CVT_STEPS 3
1215 #define MAX_VECTORIZATION_FACTOR INT_MAX
1217 /* Nonzero if TYPE represents a (scalar) boolean type or type
1218 in the middle-end compatible with it (unsigned precision 1 integral
1219 types). Used to determine which types should be vectorized as
1220 VECTOR_BOOLEAN_TYPE_P. */
1222 #define VECT_SCALAR_BOOLEAN_TYPE_P(TYPE) \
1223 (TREE_CODE (TYPE) == BOOLEAN_TYPE \
1224 || ((TREE_CODE (TYPE) == INTEGER_TYPE \
1225 || TREE_CODE (TYPE) == ENUMERAL_TYPE) \
1226 && TYPE_PRECISION (TYPE) == 1 \
1227 && TYPE_UNSIGNED (TYPE)))
1230 nested_in_vect_loop_p (class loop
*loop
, stmt_vec_info stmt_info
)
1233 && (loop
->inner
== (gimple_bb (stmt_info
->stmt
))->loop_father
));
1236 /* Return TRUE if a statement represented by STMT_INFO is a part of a
1240 is_pattern_stmt_p (stmt_vec_info stmt_info
)
1242 return stmt_info
->pattern_stmt_p
;
1245 /* If STMT_INFO is a pattern statement, return the statement that it
1246 replaces, otherwise return STMT_INFO itself. */
1248 inline stmt_vec_info
1249 vect_orig_stmt (stmt_vec_info stmt_info
)
1251 if (is_pattern_stmt_p (stmt_info
))
1252 return STMT_VINFO_RELATED_STMT (stmt_info
);
1256 /* Return the later statement between STMT1_INFO and STMT2_INFO. */
1258 static inline stmt_vec_info
1259 get_later_stmt (stmt_vec_info stmt1_info
, stmt_vec_info stmt2_info
)
1261 if (gimple_uid (vect_orig_stmt (stmt1_info
)->stmt
)
1262 > gimple_uid (vect_orig_stmt (stmt2_info
)->stmt
))
1268 /* If STMT_INFO has been replaced by a pattern statement, return the
1269 replacement statement, otherwise return STMT_INFO itself. */
1271 inline stmt_vec_info
1272 vect_stmt_to_vectorize (stmt_vec_info stmt_info
)
1274 if (STMT_VINFO_IN_PATTERN_P (stmt_info
))
1275 return STMT_VINFO_RELATED_STMT (stmt_info
);
1279 /* Return true if BB is a loop header. */
1282 is_loop_header_bb_p (basic_block bb
)
1284 if (bb
== (bb
->loop_father
)->header
)
1286 gcc_checking_assert (EDGE_COUNT (bb
->preds
) == 1);
1290 /* Return pow2 (X). */
1297 for (i
= 0; i
< x
; i
++)
1303 /* Alias targetm.vectorize.builtin_vectorization_cost. */
1306 builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost
,
1307 tree vectype
, int misalign
)
1309 return targetm
.vectorize
.builtin_vectorization_cost (type_of_cost
,
1313 /* Get cost by calling cost target builtin. */
1316 int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost
)
1318 return builtin_vectorization_cost (type_of_cost
, NULL
, 0);
1321 /* Alias targetm.vectorize.init_cost. */
1323 static inline void *
1324 init_cost (class loop
*loop_info
)
1326 return targetm
.vectorize
.init_cost (loop_info
);
1329 extern void dump_stmt_cost (FILE *, void *, int, enum vect_cost_for_stmt
,
1330 stmt_vec_info
, int, unsigned,
1331 enum vect_cost_model_location
);
1333 /* Alias targetm.vectorize.add_stmt_cost. */
1335 static inline unsigned
1336 add_stmt_cost (void *data
, int count
, enum vect_cost_for_stmt kind
,
1337 stmt_vec_info stmt_info
, int misalign
,
1338 enum vect_cost_model_location where
)
1340 unsigned cost
= targetm
.vectorize
.add_stmt_cost (data
, count
, kind
,
1341 stmt_info
, misalign
, where
);
1342 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1343 dump_stmt_cost (dump_file
, data
, count
, kind
, stmt_info
, misalign
,
1348 /* Alias targetm.vectorize.finish_cost. */
1351 finish_cost (void *data
, unsigned *prologue_cost
,
1352 unsigned *body_cost
, unsigned *epilogue_cost
)
1354 targetm
.vectorize
.finish_cost (data
, prologue_cost
, body_cost
, epilogue_cost
);
1357 /* Alias targetm.vectorize.destroy_cost_data. */
1360 destroy_cost_data (void *data
)
1362 targetm
.vectorize
.destroy_cost_data (data
);
1366 add_stmt_costs (void *data
, stmt_vector_for_cost
*cost_vec
)
1368 stmt_info_for_cost
*cost
;
1370 FOR_EACH_VEC_ELT (*cost_vec
, i
, cost
)
1371 add_stmt_cost (data
, cost
->count
, cost
->kind
, cost
->stmt_info
,
1372 cost
->misalign
, cost
->where
);
1375 /*-----------------------------------------------------------------*/
1376 /* Info on data references alignment. */
1377 /*-----------------------------------------------------------------*/
1378 #define DR_MISALIGNMENT_UNKNOWN (-1)
1379 #define DR_MISALIGNMENT_UNINITIALIZED (-2)
1382 set_dr_misalignment (dr_vec_info
*dr_info
, int val
)
1384 dr_info
->misalignment
= val
;
1388 dr_misalignment (dr_vec_info
*dr_info
)
1390 int misalign
= dr_info
->misalignment
;
1391 gcc_assert (misalign
!= DR_MISALIGNMENT_UNINITIALIZED
);
1395 /* Reflects actual alignment of first access in the vectorized loop,
1396 taking into account peeling/versioning if applied. */
1397 #define DR_MISALIGNMENT(DR) dr_misalignment (DR)
1398 #define SET_DR_MISALIGNMENT(DR, VAL) set_dr_misalignment (DR, VAL)
1400 /* Only defined once DR_MISALIGNMENT is defined. */
1401 #define DR_TARGET_ALIGNMENT(DR) ((DR)->target_alignment)
1403 /* Return true if data access DR_INFO is aligned to its target alignment
1404 (which may be less than a full vector). */
1407 aligned_access_p (dr_vec_info
*dr_info
)
1409 return (DR_MISALIGNMENT (dr_info
) == 0);
1412 /* Return TRUE if the alignment of the data access is known, and FALSE
1416 known_alignment_for_access_p (dr_vec_info
*dr_info
)
1418 return (DR_MISALIGNMENT (dr_info
) != DR_MISALIGNMENT_UNKNOWN
);
1421 /* Return the minimum alignment in bytes that the vectorized version
1422 of DR_INFO is guaranteed to have. */
1424 static inline unsigned int
1425 vect_known_alignment_in_bytes (dr_vec_info
*dr_info
)
1427 if (DR_MISALIGNMENT (dr_info
) == DR_MISALIGNMENT_UNKNOWN
)
1428 return TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr_info
->dr
)));
1429 if (DR_MISALIGNMENT (dr_info
) == 0)
1430 return known_alignment (DR_TARGET_ALIGNMENT (dr_info
));
1431 return DR_MISALIGNMENT (dr_info
) & -DR_MISALIGNMENT (dr_info
);
1434 /* Return the behavior of DR_INFO with respect to the vectorization context
1435 (which for outer loop vectorization might not be the behavior recorded
1436 in DR_INFO itself). */
1438 static inline innermost_loop_behavior
*
1439 vect_dr_behavior (dr_vec_info
*dr_info
)
1441 stmt_vec_info stmt_info
= dr_info
->stmt
;
1442 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
1443 if (loop_vinfo
== NULL
1444 || !nested_in_vect_loop_p (LOOP_VINFO_LOOP (loop_vinfo
), stmt_info
))
1445 return &DR_INNERMOST (dr_info
->dr
);
1447 return &STMT_VINFO_DR_WRT_VEC_LOOP (stmt_info
);
1450 /* Return true if the vect cost model is unlimited. */
1452 unlimited_cost_model (loop_p loop
)
1454 if (loop
!= NULL
&& loop
->force_vectorize
1455 && flag_simd_cost_model
!= VECT_COST_MODEL_DEFAULT
)
1456 return flag_simd_cost_model
== VECT_COST_MODEL_UNLIMITED
;
1457 return (flag_vect_cost_model
== VECT_COST_MODEL_UNLIMITED
);
1460 /* Return true if the loop described by LOOP_VINFO is fully-masked and
1461 if the first iteration should use a partial mask in order to achieve
1465 vect_use_loop_mask_for_alignment_p (loop_vec_info loop_vinfo
)
1467 return (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo
)
1468 && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo
));
1471 /* Return the number of vectors of type VECTYPE that are needed to get
1472 NUNITS elements. NUNITS should be based on the vectorization factor,
1473 so it is always a known multiple of the number of elements in VECTYPE. */
1475 static inline unsigned int
1476 vect_get_num_vectors (poly_uint64 nunits
, tree vectype
)
1478 return exact_div (nunits
, TYPE_VECTOR_SUBPARTS (vectype
)).to_constant ();
1481 /* Return the number of copies needed for loop vectorization when
1482 a statement operates on vectors of type VECTYPE. This is the
1483 vectorization factor divided by the number of elements in
1484 VECTYPE and is always known at compile time. */
1486 static inline unsigned int
1487 vect_get_num_copies (loop_vec_info loop_vinfo
, tree vectype
)
1489 return vect_get_num_vectors (LOOP_VINFO_VECT_FACTOR (loop_vinfo
), vectype
);
1492 /* Update maximum unit count *MAX_NUNITS so that it accounts for
1493 NUNITS. *MAX_NUNITS can be 1 if we haven't yet recorded anything. */
1496 vect_update_max_nunits (poly_uint64
*max_nunits
, poly_uint64 nunits
)
1498 /* All unit counts have the form vec_info::vector_size * X for some
1499 rational X, so two unit sizes must have a common multiple.
1500 Everything is a multiple of the initial value of 1. */
1501 *max_nunits
= force_common_multiple (*max_nunits
, nunits
);
1504 /* Update maximum unit count *MAX_NUNITS so that it accounts for
1505 the number of units in vector type VECTYPE. *MAX_NUNITS can be 1
1506 if we haven't yet recorded any vector types. */
1509 vect_update_max_nunits (poly_uint64
*max_nunits
, tree vectype
)
1511 vect_update_max_nunits (max_nunits
, TYPE_VECTOR_SUBPARTS (vectype
));
1514 /* Return the vectorization factor that should be used for costing
1515 purposes while vectorizing the loop described by LOOP_VINFO.
1516 Pick a reasonable estimate if the vectorization factor isn't
1517 known at compile time. */
1519 static inline unsigned int
1520 vect_vf_for_cost (loop_vec_info loop_vinfo
)
1522 return estimated_poly_value (LOOP_VINFO_VECT_FACTOR (loop_vinfo
));
1525 /* Estimate the number of elements in VEC_TYPE for costing purposes.
1526 Pick a reasonable estimate if the exact number isn't known at
1529 static inline unsigned int
1530 vect_nunits_for_cost (tree vec_type
)
1532 return estimated_poly_value (TYPE_VECTOR_SUBPARTS (vec_type
));
1535 /* Return the maximum possible vectorization factor for LOOP_VINFO. */
1537 static inline unsigned HOST_WIDE_INT
1538 vect_max_vf (loop_vec_info loop_vinfo
)
1540 unsigned HOST_WIDE_INT vf
;
1541 if (LOOP_VINFO_VECT_FACTOR (loop_vinfo
).is_constant (&vf
))
1543 return MAX_VECTORIZATION_FACTOR
;
1546 /* Return the size of the value accessed by unvectorized data reference
1547 DR_INFO. This is only valid once STMT_VINFO_VECTYPE has been calculated
1548 for the associated gimple statement, since that guarantees that DR_INFO
1549 accesses either a scalar or a scalar equivalent. ("Scalar equivalent"
1550 here includes things like V1SI, which can be vectorized in the same way
1554 vect_get_scalar_dr_size (dr_vec_info
*dr_info
)
1556 return tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_info
->dr
))));
1559 /* Source location + hotness information. */
1560 extern dump_user_location_t vect_location
;
1562 /* A macro for calling:
1563 dump_begin_scope (MSG, vect_location);
1564 via an RAII object, thus printing "=== MSG ===\n" to the dumpfile etc,
1567 once the object goes out of scope, thus capturing the nesting of
1570 These scopes affect dump messages within them: dump messages at the
1571 top level implicitly default to MSG_PRIORITY_USER_FACING, whereas those
1572 in a nested scope implicitly default to MSG_PRIORITY_INTERNALS. */
1574 #define DUMP_VECT_SCOPE(MSG) \
1575 AUTO_DUMP_SCOPE (MSG, vect_location)
1577 /* A sentinel class for ensuring that the "vect_location" global gets
1578 reset at the end of a scope.
1580 The "vect_location" global is used during dumping and contains a
1581 location_t, which could contain references to a tree block via the
1582 ad-hoc data. This data is used for tracking inlining information,
1583 but it's not a GC root; it's simply assumed that such locations never
1584 get accessed if the blocks are optimized away.
1586 Hence we need to ensure that such locations are purged at the end
1587 of any operations using them (e.g. via this class). */
1589 class auto_purge_vect_location
1592 ~auto_purge_vect_location ();
1595 /*-----------------------------------------------------------------*/
1596 /* Function prototypes. */
1597 /*-----------------------------------------------------------------*/
1599 /* Simple loop peeling and versioning utilities for vectorizer's purposes -
1600 in tree-vect-loop-manip.c. */
1601 extern void vect_set_loop_condition (class loop
*, loop_vec_info
,
1602 tree
, tree
, tree
, bool);
1603 extern bool slpeel_can_duplicate_loop_p (const class loop
*, const_edge
);
1604 class loop
*slpeel_tree_duplicate_loop_to_edge_cfg (class loop
*,
1605 class loop
*, edge
);
1606 class loop
*vect_loop_versioning (loop_vec_info
);
1607 extern class loop
*vect_do_peeling (loop_vec_info
, tree
, tree
,
1608 tree
*, tree
*, tree
*, int, bool, bool,
1609 tree
*, drs_init_vec
&);
1610 extern void vect_prepare_for_masked_peels (loop_vec_info
);
1611 extern dump_user_location_t
find_loop_location (class loop
*);
1612 extern bool vect_can_advance_ivs_p (loop_vec_info
);
1613 extern void vect_update_inits_of_drs (loop_vec_info
, tree
, tree_code
);
1615 /* In tree-vect-stmts.c. */
1616 extern tree
get_vectype_for_scalar_type (vec_info
*, tree
);
1617 extern tree
get_vectype_for_scalar_type_and_size (tree
, poly_uint64
);
1618 extern tree
get_mask_type_for_scalar_type (vec_info
*, tree
);
1619 extern tree
get_same_sized_vectype (tree
, tree
);
1620 extern bool vect_get_loop_mask_type (loop_vec_info
);
1621 extern bool vect_is_simple_use (tree
, vec_info
*, enum vect_def_type
*,
1622 stmt_vec_info
* = NULL
, gimple
** = NULL
);
1623 extern bool vect_is_simple_use (tree
, vec_info
*, enum vect_def_type
*,
1624 tree
*, stmt_vec_info
* = NULL
,
1626 extern bool supportable_widening_operation (enum tree_code
, stmt_vec_info
,
1627 tree
, tree
, enum tree_code
*,
1628 enum tree_code
*, int *,
1630 extern bool supportable_narrowing_operation (vec_info
*, enum tree_code
, tree
,
1631 tree
, enum tree_code
*,
1632 int *, vec
<tree
> *);
1633 extern unsigned record_stmt_cost (stmt_vector_for_cost
*, int,
1634 enum vect_cost_for_stmt
, stmt_vec_info
,
1635 int, enum vect_cost_model_location
);
1636 extern stmt_vec_info
vect_finish_replace_stmt (stmt_vec_info
, gimple
*);
1637 extern stmt_vec_info
vect_finish_stmt_generation (stmt_vec_info
, gimple
*,
1638 gimple_stmt_iterator
*);
1639 extern opt_result
vect_mark_stmts_to_be_vectorized (loop_vec_info
, bool *);
1640 extern tree
vect_get_store_rhs (stmt_vec_info
);
1641 extern tree
vect_get_vec_def_for_operand_1 (stmt_vec_info
, enum vect_def_type
);
1642 extern tree
vect_get_vec_def_for_operand (tree
, stmt_vec_info
, tree
= NULL
);
1643 extern void vect_get_vec_defs (tree
, tree
, stmt_vec_info
, vec
<tree
> *,
1644 vec
<tree
> *, slp_tree
);
1645 extern void vect_get_vec_defs_for_stmt_copy (vec_info
*,
1646 vec
<tree
> *, vec
<tree
> *);
1647 extern tree
vect_init_vector (stmt_vec_info
, tree
, tree
,
1648 gimple_stmt_iterator
*);
1649 extern tree
vect_get_vec_def_for_stmt_copy (vec_info
*, tree
);
1650 extern bool vect_transform_stmt (stmt_vec_info
, gimple_stmt_iterator
*,
1651 slp_tree
, slp_instance
);
1652 extern void vect_remove_stores (stmt_vec_info
);
1653 extern opt_result
vect_analyze_stmt (stmt_vec_info
, bool *, slp_tree
,
1654 slp_instance
, stmt_vector_for_cost
*);
1655 extern void vect_get_load_cost (stmt_vec_info
, int, bool,
1656 unsigned int *, unsigned int *,
1657 stmt_vector_for_cost
*,
1658 stmt_vector_for_cost
*, bool);
1659 extern void vect_get_store_cost (stmt_vec_info
, int,
1660 unsigned int *, stmt_vector_for_cost
*);
1661 extern bool vect_supportable_shift (vec_info
*, enum tree_code
, tree
);
1662 extern tree
vect_gen_perm_mask_any (tree
, const vec_perm_indices
&);
1663 extern tree
vect_gen_perm_mask_checked (tree
, const vec_perm_indices
&);
1664 extern void optimize_mask_stores (class loop
*);
1665 extern gcall
*vect_gen_while (tree
, tree
, tree
);
1666 extern tree
vect_gen_while_not (gimple_seq
*, tree
, tree
, tree
);
1667 extern opt_result
vect_get_vector_types_for_stmt (stmt_vec_info
, tree
*,
1669 extern opt_tree
vect_get_mask_type_for_stmt (stmt_vec_info
);
1671 /* In tree-vect-data-refs.c. */
1672 extern bool vect_can_force_dr_alignment_p (const_tree
, poly_uint64
);
1673 extern enum dr_alignment_support vect_supportable_dr_alignment
1674 (dr_vec_info
*, bool);
1675 extern tree
vect_get_smallest_scalar_type (stmt_vec_info
, HOST_WIDE_INT
*,
1677 extern opt_result
vect_analyze_data_ref_dependences (loop_vec_info
, unsigned int *);
1678 extern bool vect_slp_analyze_instance_dependence (slp_instance
);
1679 extern opt_result
vect_enhance_data_refs_alignment (loop_vec_info
);
1680 extern opt_result
vect_analyze_data_refs_alignment (loop_vec_info
);
1681 extern opt_result
vect_verify_datarefs_alignment (loop_vec_info
);
1682 extern bool vect_slp_analyze_and_verify_instance_alignment (slp_instance
);
1683 extern opt_result
vect_analyze_data_ref_accesses (vec_info
*);
1684 extern opt_result
vect_prune_runtime_alias_test_list (loop_vec_info
);
1685 extern bool vect_gather_scatter_fn_p (vec_info
*, bool, bool, tree
, tree
,
1686 tree
, int, internal_fn
*, tree
*);
1687 extern bool vect_check_gather_scatter (stmt_vec_info
, loop_vec_info
,
1688 gather_scatter_info
*);
1689 extern opt_result
vect_find_stmt_data_reference (loop_p
, gimple
*,
1690 vec
<data_reference_p
> *);
1691 extern opt_result
vect_analyze_data_refs (vec_info
*, poly_uint64
*, bool *);
1692 extern void vect_record_base_alignments (vec_info
*);
1693 extern tree
vect_create_data_ref_ptr (stmt_vec_info
, tree
, class loop
*, tree
,
1694 tree
*, gimple_stmt_iterator
*,
1696 tree
= NULL_TREE
, tree
= NULL_TREE
);
1697 extern tree
bump_vector_ptr (tree
, gimple
*, gimple_stmt_iterator
*,
1698 stmt_vec_info
, tree
);
1699 extern void vect_copy_ref_info (tree
, tree
);
1700 extern tree
vect_create_destination_var (tree
, tree
);
1701 extern bool vect_grouped_store_supported (tree
, unsigned HOST_WIDE_INT
);
1702 extern bool vect_store_lanes_supported (tree
, unsigned HOST_WIDE_INT
, bool);
1703 extern bool vect_grouped_load_supported (tree
, bool, unsigned HOST_WIDE_INT
);
1704 extern bool vect_load_lanes_supported (tree
, unsigned HOST_WIDE_INT
, bool);
1705 extern void vect_permute_store_chain (vec
<tree
> ,unsigned int, stmt_vec_info
,
1706 gimple_stmt_iterator
*, vec
<tree
> *);
1707 extern tree
vect_setup_realignment (stmt_vec_info
, gimple_stmt_iterator
*,
1708 tree
*, enum dr_alignment_support
, tree
,
1710 extern void vect_transform_grouped_load (stmt_vec_info
, vec
<tree
> , int,
1711 gimple_stmt_iterator
*);
1712 extern void vect_record_grouped_load_vectors (stmt_vec_info
, vec
<tree
>);
1713 extern tree
vect_get_new_vect_var (tree
, enum vect_var_kind
, const char *);
1714 extern tree
vect_get_new_ssa_name (tree
, enum vect_var_kind
,
1715 const char * = NULL
);
1716 extern tree
vect_create_addr_base_for_vector_ref (stmt_vec_info
, gimple_seq
*,
1717 tree
, tree
= NULL_TREE
);
1719 /* In tree-vect-loop.c. */
1720 extern widest_int
vect_iv_limit_for_full_masking (loop_vec_info loop_vinfo
);
1721 /* Used in tree-vect-loop-manip.c */
1722 extern void determine_peel_for_niter (loop_vec_info
);
1723 /* Used in gimple-loop-interchange.c and tree-parloops.c. */
1724 extern bool check_reduction_path (dump_user_location_t
, loop_p
, gphi
*, tree
,
1726 extern bool needs_fold_left_reduction_p (tree
, tree_code
);
1727 /* Drive for loop analysis stage. */
1728 extern opt_loop_vec_info
vect_analyze_loop (class loop
*, vec_info_shared
*);
1729 extern tree
vect_build_loop_niters (loop_vec_info
, bool * = NULL
);
1730 extern void vect_gen_vector_loop_niters (loop_vec_info
, tree
, tree
*,
1732 extern tree
vect_halve_mask_nunits (vec_info
*, tree
);
1733 extern tree
vect_double_mask_nunits (vec_info
*, tree
);
1734 extern void vect_record_loop_mask (loop_vec_info
, vec_loop_masks
*,
1735 unsigned int, tree
, tree
);
1736 extern tree
vect_get_loop_mask (gimple_stmt_iterator
*, vec_loop_masks
*,
1737 unsigned int, tree
, unsigned int);
1738 extern stmt_vec_info
info_for_reduction (stmt_vec_info
);
1740 /* Drive for loop transformation stage. */
1741 extern class loop
*vect_transform_loop (loop_vec_info
);
1742 extern opt_loop_vec_info
vect_analyze_loop_form (class loop
*,
1744 extern bool vectorizable_live_operation (stmt_vec_info
, gimple_stmt_iterator
*,
1745 slp_tree
, slp_instance
, int,
1746 bool, stmt_vector_for_cost
*);
1747 extern bool vectorizable_reduction (stmt_vec_info
, slp_tree
, slp_instance
,
1748 stmt_vector_for_cost
*);
1749 extern bool vectorizable_induction (stmt_vec_info
, gimple_stmt_iterator
*,
1750 stmt_vec_info
*, slp_tree
,
1751 stmt_vector_for_cost
*);
1752 extern bool vect_transform_reduction (stmt_vec_info
, gimple_stmt_iterator
*,
1753 stmt_vec_info
*, slp_tree
);
1754 extern bool vect_transform_cycle_phi (stmt_vec_info
, stmt_vec_info
*,
1755 slp_tree
, slp_instance
);
1756 extern bool vectorizable_lc_phi (stmt_vec_info
, stmt_vec_info
*, slp_tree
);
1757 extern bool vect_worthwhile_without_simd_p (vec_info
*, tree_code
);
1758 extern int vect_get_known_peeling_cost (loop_vec_info
, int, int *,
1759 stmt_vector_for_cost
*,
1760 stmt_vector_for_cost
*,
1761 stmt_vector_for_cost
*);
1762 extern tree
cse_and_gimplify_to_preheader (loop_vec_info
, tree
);
1764 /* In tree-vect-slp.c. */
1765 extern void vect_free_slp_instance (slp_instance
, bool);
1766 extern bool vect_transform_slp_perm_load (slp_tree
, vec
<tree
> ,
1767 gimple_stmt_iterator
*, poly_uint64
,
1768 slp_instance
, bool, unsigned *);
1769 extern bool vect_slp_analyze_operations (vec_info
*);
1770 extern void vect_schedule_slp (vec_info
*);
1771 extern opt_result
vect_analyze_slp (vec_info
*, unsigned);
1772 extern bool vect_make_slp_decision (loop_vec_info
);
1773 extern void vect_detect_hybrid_slp (loop_vec_info
);
1774 extern void vect_get_slp_defs (slp_tree
, vec
<vec
<tree
> > *, unsigned n
= -1U);
1775 extern bool vect_slp_bb (basic_block
);
1776 extern stmt_vec_info
vect_find_last_scalar_stmt_in_slp (slp_tree
);
1777 extern bool is_simple_and_all_uses_invariant (stmt_vec_info
, loop_vec_info
);
1778 extern bool can_duplicate_and_interleave_p (vec_info
*, unsigned int,
1780 unsigned int * = NULL
,
1781 tree
* = NULL
, tree
* = NULL
);
1782 extern void duplicate_and_interleave (vec_info
*, gimple_seq
*, tree
,
1783 vec
<tree
>, unsigned int, vec
<tree
> &);
1784 extern int vect_get_place_in_interleaving_chain (stmt_vec_info
, stmt_vec_info
);
1786 /* In tree-vect-patterns.c. */
1787 /* Pattern recognition functions.
1788 Additional pattern recognition functions can (and will) be added
1790 void vect_pattern_recog (vec_info
*);
1792 /* In tree-vectorizer.c. */
1793 unsigned vectorize_loops (void);
1794 void vect_free_loop_info_assumptions (class loop
*);
1795 gimple
*vect_loop_vectorized_call (class loop
*, gcond
**cond
= NULL
);
1798 #endif /* GCC_TREE_VECTORIZER_H */