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"
30 /* Used for naming of new temporaries. */
38 /* Defines type of operation. */
45 /* Define type of available alignment support. */
46 enum dr_alignment_support
{
47 dr_unaligned_unsupported
,
48 dr_unaligned_supported
,
50 dr_explicit_realign_optimized
,
54 /* Define type of def-use cross-iteration cycle. */
56 vect_uninitialized_def
= 0,
57 vect_constant_def
= 1,
62 vect_double_reduction_def
,
67 /* Define type of reduction. */
68 enum vect_reduction_type
{
71 INTEGER_INDUC_COND_REDUCTION
,
74 /* Retain a scalar phi and use a FOLD_EXTRACT_LAST within the loop
77 for (int i = 0; i < VF; ++i)
78 res = cond[i] ? val[i] : res; */
79 EXTRACT_LAST_REDUCTION
,
81 /* Use a folding reduction within the loop to implement:
83 for (int i = 0; i < VF; ++i)
86 (with no reassocation). */
90 #define VECTORIZABLE_CYCLE_DEF(D) (((D) == vect_reduction_def) \
91 || ((D) == vect_double_reduction_def) \
92 || ((D) == vect_nested_cycle))
94 /* Structure to encapsulate information about a group of like
95 instructions to be presented to the target cost model. */
96 struct stmt_info_for_cost
{
98 enum vect_cost_for_stmt kind
;
99 enum vect_cost_model_location where
;
100 stmt_vec_info stmt_info
;
104 typedef vec
<stmt_info_for_cost
> stmt_vector_for_cost
;
106 /* Maps base addresses to an innermost_loop_behavior that gives the maximum
107 known alignment for that base. */
108 typedef hash_map
<tree_operand_hash
,
109 innermost_loop_behavior
*> vec_base_alignments
;
111 /************************************************************************
113 ************************************************************************/
114 typedef struct _slp_tree
*slp_tree
;
116 /* A computation tree of an SLP instance. Each node corresponds to a group of
117 stmts to be packed in a SIMD stmt. */
119 /* Nodes that contain def-stmts of this node statements operands. */
120 vec
<slp_tree
> children
;
121 /* A group of scalar stmts to be vectorized together. */
122 vec
<stmt_vec_info
> stmts
;
123 /* Load permutation relative to the stores, NULL if there is no
125 vec
<unsigned> load_permutation
;
126 /* Vectorized stmt/s. */
127 vec
<stmt_vec_info
> vec_stmts
;
128 /* Number of vector stmts that are created to replace the group of scalar
129 stmts. It is calculated during the transformation phase as the number of
130 scalar elements in one scalar iteration (GROUP_SIZE) multiplied by VF
131 divided by vector size. */
132 unsigned int vec_stmts_size
;
133 /* Reference count in the SLP graph. */
135 /* The maximum number of vector elements for the subtree rooted
137 poly_uint64 max_nunits
;
138 /* Whether the scalar computations use two different operators. */
140 /* The DEF type of this node. */
141 enum vect_def_type def_type
;
145 /* SLP instance is a sequence of stmts in a loop that can be packed into
147 typedef class _slp_instance
{
149 /* The root of SLP tree. */
152 /* Size of groups of scalar stmts that will be replaced by SIMD stmt/s. */
153 unsigned int group_size
;
155 /* The unrolling factor required to vectorized this SLP instance. */
156 poly_uint64 unrolling_factor
;
158 /* The group of nodes that contain loads of this SLP instance. */
161 /* The SLP node containing the reduction PHIs. */
166 /* Access Functions. */
167 #define SLP_INSTANCE_TREE(S) (S)->root
168 #define SLP_INSTANCE_GROUP_SIZE(S) (S)->group_size
169 #define SLP_INSTANCE_UNROLLING_FACTOR(S) (S)->unrolling_factor
170 #define SLP_INSTANCE_LOADS(S) (S)->loads
172 #define SLP_TREE_CHILDREN(S) (S)->children
173 #define SLP_TREE_SCALAR_STMTS(S) (S)->stmts
174 #define SLP_TREE_VEC_STMTS(S) (S)->vec_stmts
175 #define SLP_TREE_NUMBER_OF_VEC_STMTS(S) (S)->vec_stmts_size
176 #define SLP_TREE_LOAD_PERMUTATION(S) (S)->load_permutation
177 #define SLP_TREE_TWO_OPERATORS(S) (S)->two_operators
178 #define SLP_TREE_DEF_TYPE(S) (S)->def_type
182 /* Describes two objects whose addresses must be unequal for the vectorized
184 typedef std::pair
<tree
, tree
> vec_object_pair
;
186 /* Records that vectorization is only possible if abs (EXPR) >= MIN_VALUE.
187 UNSIGNED_P is true if we can assume that abs (EXPR) == EXPR. */
188 class vec_lower_bound
{
190 vec_lower_bound () {}
191 vec_lower_bound (tree e
, bool u
, poly_uint64 m
)
192 : expr (e
), unsigned_p (u
), min_value (m
) {}
196 poly_uint64 min_value
;
199 /* Vectorizer state shared between different analyses like vector sizes
200 of the same CFG region. */
201 class vec_info_shared
{
206 void save_datarefs();
207 void check_datarefs();
209 /* All data references. Freed by free_data_refs, so not an auto_vec. */
210 vec
<data_reference_p
> datarefs
;
211 vec
<data_reference
> datarefs_copy
;
213 /* The loop nest in which the data dependences are computed. */
214 auto_vec
<loop_p
> loop_nest
;
216 /* All data dependences. Freed by free_dependence_relations, so not
221 /* Vectorizer state common between loop and basic-block vectorization. */
224 enum vec_kind
{ bb
, loop
};
226 vec_info (vec_kind
, void *, vec_info_shared
*);
229 stmt_vec_info
add_stmt (gimple
*);
230 stmt_vec_info
lookup_stmt (gimple
*);
231 stmt_vec_info
lookup_def (tree
);
232 stmt_vec_info
lookup_single_use (tree
);
233 class dr_vec_info
*lookup_dr (data_reference
*);
234 void move_dr (stmt_vec_info
, stmt_vec_info
);
235 void remove_stmt (stmt_vec_info
);
236 void replace_stmt (gimple_stmt_iterator
*, stmt_vec_info
, gimple
*);
238 /* The type of vectorization. */
241 /* Shared vectorizer state. */
242 vec_info_shared
*shared
;
244 /* The mapping of GIMPLE UID to stmt_vec_info. */
245 vec
<stmt_vec_info
> stmt_vec_infos
;
247 /* All SLP instances. */
248 auto_vec
<slp_instance
> slp_instances
;
250 /* Maps base addresses to an innermost_loop_behavior that gives the maximum
251 known alignment for that base. */
252 vec_base_alignments base_alignments
;
254 /* All interleaving chains of stores, represented by the first
255 stmt in the chain. */
256 auto_vec
<stmt_vec_info
> grouped_stores
;
258 /* Cost data used by the target cost model. */
259 void *target_cost_data
;
262 stmt_vec_info
new_stmt_vec_info (gimple
*stmt
);
263 void set_vinfo_for_stmt (gimple
*, stmt_vec_info
);
264 void free_stmt_vec_infos ();
265 void free_stmt_vec_info (stmt_vec_info
);
268 class _loop_vec_info
;
274 is_a_helper
<_loop_vec_info
*>::test (vec_info
*i
)
276 return i
->kind
== vec_info::loop
;
282 is_a_helper
<_bb_vec_info
*>::test (vec_info
*i
)
284 return i
->kind
== vec_info::bb
;
288 /* In general, we can divide the vector statements in a vectorized loop
289 into related groups ("rgroups") and say that for each rgroup there is
290 some nS such that the rgroup operates on nS values from one scalar
291 iteration followed by nS values from the next. That is, if VF is the
292 vectorization factor of the loop, the rgroup operates on a sequence:
294 (1,1) (1,2) ... (1,nS) (2,1) ... (2,nS) ... (VF,1) ... (VF,nS)
296 where (i,j) represents a scalar value with index j in a scalar
297 iteration with index i.
299 [ We use the term "rgroup" to emphasise that this grouping isn't
300 necessarily the same as the grouping of statements used elsewhere.
301 For example, if we implement a group of scalar loads using gather
302 loads, we'll use a separate gather load for each scalar load, and
303 thus each gather load will belong to its own rgroup. ]
305 In general this sequence will occupy nV vectors concatenated
306 together. If these vectors have nL lanes each, the total number
307 of scalar values N is given by:
309 N = nS * VF = nV * nL
311 None of nS, VF, nV and nL are required to be a power of 2. nS and nV
312 are compile-time constants but VF and nL can be variable (if the target
313 supports variable-length vectors).
315 In classical vectorization, each iteration of the vector loop would
316 handle exactly VF iterations of the original scalar loop. However,
317 in a fully-masked loop, a particular iteration of the vector loop
318 might handle fewer than VF iterations of the scalar loop. The vector
319 lanes that correspond to iterations of the scalar loop are said to be
320 "active" and the other lanes are said to be "inactive".
322 In a fully-masked loop, many rgroups need to be masked to ensure that
323 they have no effect for the inactive lanes. Each such rgroup needs a
324 sequence of booleans in the same order as above, but with each (i,j)
325 replaced by a boolean that indicates whether iteration i is active.
326 This sequence occupies nV vector masks that again have nL lanes each.
327 Thus the mask sequence as a whole consists of VF independent booleans
328 that are each repeated nS times.
330 We make the simplifying assumption that if a sequence of nV masks is
331 suitable for one (nS,nL) pair, we can reuse it for (nS/2,nL/2) by
332 VIEW_CONVERTing it. This holds for all current targets that support
333 fully-masked loops. For example, suppose the scalar loop is:
337 for (int i = 0; i < n; ++i)
339 f[i * 2 + 0] += 1.0f;
340 f[i * 2 + 1] += 2.0f;
344 and suppose that vectors have 256 bits. The vectorized f accesses
345 will belong to one rgroup and the vectorized d access to another:
347 f rgroup: nS = 2, nV = 1, nL = 8
348 d rgroup: nS = 1, nV = 1, nL = 4
351 [ In this simple example the rgroups do correspond to the normal
352 SLP grouping scheme. ]
354 If only the first three lanes are active, the masks we need are:
356 f rgroup: 1 1 | 1 1 | 1 1 | 0 0
357 d rgroup: 1 | 1 | 1 | 0
359 Here we can use a mask calculated for f's rgroup for d's, but not
362 Thus for each value of nV, it is enough to provide nV masks, with the
363 mask being calculated based on the highest nL (or, equivalently, based
364 on the highest nS) required by any rgroup with that nV. We therefore
365 represent the entire collection of masks as a two-level table, with the
366 first level being indexed by nV - 1 (since nV == 0 doesn't exist) and
367 the second being indexed by the mask index 0 <= i < nV. */
369 /* The masks needed by rgroups with nV vectors, according to the
370 description above. */
371 struct rgroup_masks
{
372 /* The largest nS for all rgroups that use these masks. */
373 unsigned int max_nscalars_per_iter
;
375 /* The type of mask to use, based on the highest nS recorded above. */
378 /* A vector of nV masks, in iteration order. */
382 typedef auto_vec
<rgroup_masks
> vec_loop_masks
;
384 /*-----------------------------------------------------------------*/
385 /* Info on vectorized loops. */
386 /*-----------------------------------------------------------------*/
387 typedef class _loop_vec_info
: public vec_info
{
389 _loop_vec_info (class loop
*, vec_info_shared
*);
392 /* The loop to which this info struct refers to. */
395 /* The loop basic blocks. */
398 /* Number of latch executions. */
400 /* Number of iterations. */
402 /* Number of iterations of the original loop. */
403 tree num_iters_unchanged
;
404 /* Condition under which this loop is analyzed and versioned. */
405 tree num_iters_assumptions
;
407 /* Threshold of number of iterations below which vectorization will not be
408 performed. It is calculated from MIN_PROFITABLE_ITERS and
409 PARAM_MIN_VECT_LOOP_BOUND. */
412 /* When applying loop versioning, the vector form should only be used
413 if the number of scalar iterations is >= this value, on top of all
414 the other requirements. Ignored when loop versioning is not being
416 poly_uint64 versioning_threshold
;
418 /* Unrolling factor */
419 poly_uint64 vectorization_factor
;
421 /* Maximum runtime vectorization factor, or MAX_VECTORIZATION_FACTOR
422 if there is no particular limit. */
423 unsigned HOST_WIDE_INT max_vectorization_factor
;
425 /* The masks that a fully-masked loop should use to avoid operating
426 on inactive scalars. */
427 vec_loop_masks masks
;
429 /* If we are using a loop mask to align memory addresses, this variable
430 contains the number of vector elements that we should skip in the
431 first iteration of the vector loop (i.e. the number of leading
432 elements that should be false in the first mask). */
433 tree mask_skip_niters
;
435 /* Type of the variables to use in the WHILE_ULT call for fully-masked
437 tree mask_compare_type
;
439 /* For #pragma omp simd if (x) loops the x expression. If constant 0,
440 the loop should not be vectorized, if constant non-zero, simd_if_cond
441 shouldn't be set and loop vectorized normally, if SSA_NAME, the loop
442 should be versioned on that condition, using scalar loop if the condition
443 is false and vectorized loop otherwise. */
446 /* Type of the IV to use in the WHILE_ULT call for fully-masked
450 /* Unknown DRs according to which loop was peeled. */
451 class dr_vec_info
*unaligned_dr
;
453 /* peeling_for_alignment indicates whether peeling for alignment will take
454 place, and what the peeling factor should be:
455 peeling_for_alignment = X means:
456 If X=0: Peeling for alignment will not be applied.
457 If X>0: Peel first X iterations.
458 If X=-1: Generate a runtime test to calculate the number of iterations
459 to be peeled, using the dataref recorded in the field
461 int peeling_for_alignment
;
463 /* The mask used to check the alignment of pointers or arrays. */
466 /* Data Dependence Relations defining address ranges that are candidates
467 for a run-time aliasing check. */
468 auto_vec
<ddr_p
> may_alias_ddrs
;
470 /* Data Dependence Relations defining address ranges together with segment
471 lengths from which the run-time aliasing check is built. */
472 auto_vec
<dr_with_seg_len_pair_t
> comp_alias_ddrs
;
474 /* Check that the addresses of each pair of objects is unequal. */
475 auto_vec
<vec_object_pair
> check_unequal_addrs
;
477 /* List of values that are required to be nonzero. This is used to check
478 whether things like "x[i * n] += 1;" are safe and eventually gets added
479 to the checks for lower bounds below. */
480 auto_vec
<tree
> check_nonzero
;
482 /* List of values that need to be checked for a minimum value. */
483 auto_vec
<vec_lower_bound
> lower_bounds
;
485 /* Statements in the loop that have data references that are candidates for a
486 runtime (loop versioning) misalignment check. */
487 auto_vec
<stmt_vec_info
> may_misalign_stmts
;
489 /* Reduction cycles detected in the loop. Used in loop-aware SLP. */
490 auto_vec
<stmt_vec_info
> reductions
;
492 /* All reduction chains in the loop, represented by the first
493 stmt in the chain. */
494 auto_vec
<stmt_vec_info
> reduction_chains
;
496 /* Cost vector for a single scalar iteration. */
497 auto_vec
<stmt_info_for_cost
> scalar_cost_vec
;
499 /* Map of IV base/step expressions to inserted name in the preheader. */
500 hash_map
<tree_operand_hash
, tree
> *ivexpr_map
;
502 /* Map of OpenMP "omp simd array" scan variables to corresponding
503 rhs of the store of the initializer. */
504 hash_map
<tree
, tree
> *scan_map
;
506 /* The unrolling factor needed to SLP the loop. In case of that pure SLP is
507 applied to the loop, i.e., no unrolling is needed, this is 1. */
508 poly_uint64 slp_unrolling_factor
;
510 /* Cost of a single scalar iteration. */
511 int single_scalar_iteration_cost
;
513 /* Is the loop vectorizable? */
516 /* Records whether we still have the option of using a fully-masked loop. */
517 bool can_fully_mask_p
;
519 /* True if have decided to use a fully-masked loop. */
522 /* When we have grouped data accesses with gaps, we may introduce invalid
523 memory accesses. We peel the last iteration of the loop to prevent
525 bool peeling_for_gaps
;
527 /* When the number of iterations is not a multiple of the vector size
528 we need to peel off iterations at the end to form an epilogue loop. */
529 bool peeling_for_niter
;
531 /* True if there are no loop carried data dependencies in the loop.
532 If loop->safelen <= 1, then this is always true, either the loop
533 didn't have any loop carried data dependencies, or the loop is being
534 vectorized guarded with some runtime alias checks, or couldn't
535 be vectorized at all, but then this field shouldn't be used.
536 For loop->safelen >= 2, the user has asserted that there are no
537 backward dependencies, but there still could be loop carried forward
538 dependencies in such loops. This flag will be false if normal
539 vectorizer data dependency analysis would fail or require versioning
540 for alias, but because of loop->safelen >= 2 it has been vectorized
541 even without versioning for alias. E.g. in:
543 for (int i = 0; i < m; i++)
545 (or #pragma simd or #pragma ivdep) we can vectorize this and it will
546 DTRT even for k > 0 && k < m, but without safelen we would not
547 vectorize this, so this field would be false. */
548 bool no_data_dependencies
;
550 /* Mark loops having masked stores. */
553 /* Queued scaling factor for the scalar loop. */
554 profile_probability scalar_loop_scaling
;
556 /* If if-conversion versioned this loop before conversion, this is the
557 loop version without if-conversion. */
558 class loop
*scalar_loop
;
560 /* For loops being epilogues of already vectorized loops
561 this points to the original vectorized loop. Otherwise NULL. */
562 _loop_vec_info
*orig_loop_info
;
566 /* Access Functions. */
567 #define LOOP_VINFO_LOOP(L) (L)->loop
568 #define LOOP_VINFO_BBS(L) (L)->bbs
569 #define LOOP_VINFO_NITERSM1(L) (L)->num_itersm1
570 #define LOOP_VINFO_NITERS(L) (L)->num_iters
571 /* Since LOOP_VINFO_NITERS and LOOP_VINFO_NITERSM1 can change after
572 prologue peeling retain total unchanged scalar loop iterations for
574 #define LOOP_VINFO_NITERS_UNCHANGED(L) (L)->num_iters_unchanged
575 #define LOOP_VINFO_NITERS_ASSUMPTIONS(L) (L)->num_iters_assumptions
576 #define LOOP_VINFO_COST_MODEL_THRESHOLD(L) (L)->th
577 #define LOOP_VINFO_VERSIONING_THRESHOLD(L) (L)->versioning_threshold
578 #define LOOP_VINFO_VECTORIZABLE_P(L) (L)->vectorizable
579 #define LOOP_VINFO_CAN_FULLY_MASK_P(L) (L)->can_fully_mask_p
580 #define LOOP_VINFO_FULLY_MASKED_P(L) (L)->fully_masked_p
581 #define LOOP_VINFO_VECT_FACTOR(L) (L)->vectorization_factor
582 #define LOOP_VINFO_MAX_VECT_FACTOR(L) (L)->max_vectorization_factor
583 #define LOOP_VINFO_MASKS(L) (L)->masks
584 #define LOOP_VINFO_MASK_SKIP_NITERS(L) (L)->mask_skip_niters
585 #define LOOP_VINFO_MASK_COMPARE_TYPE(L) (L)->mask_compare_type
586 #define LOOP_VINFO_MASK_IV_TYPE(L) (L)->iv_type
587 #define LOOP_VINFO_PTR_MASK(L) (L)->ptr_mask
588 #define LOOP_VINFO_LOOP_NEST(L) (L)->shared->loop_nest
589 #define LOOP_VINFO_DATAREFS(L) (L)->shared->datarefs
590 #define LOOP_VINFO_DDRS(L) (L)->shared->ddrs
591 #define LOOP_VINFO_INT_NITERS(L) (TREE_INT_CST_LOW ((L)->num_iters))
592 #define LOOP_VINFO_PEELING_FOR_ALIGNMENT(L) (L)->peeling_for_alignment
593 #define LOOP_VINFO_UNALIGNED_DR(L) (L)->unaligned_dr
594 #define LOOP_VINFO_MAY_MISALIGN_STMTS(L) (L)->may_misalign_stmts
595 #define LOOP_VINFO_MAY_ALIAS_DDRS(L) (L)->may_alias_ddrs
596 #define LOOP_VINFO_COMP_ALIAS_DDRS(L) (L)->comp_alias_ddrs
597 #define LOOP_VINFO_CHECK_UNEQUAL_ADDRS(L) (L)->check_unequal_addrs
598 #define LOOP_VINFO_CHECK_NONZERO(L) (L)->check_nonzero
599 #define LOOP_VINFO_LOWER_BOUNDS(L) (L)->lower_bounds
600 #define LOOP_VINFO_GROUPED_STORES(L) (L)->grouped_stores
601 #define LOOP_VINFO_SLP_INSTANCES(L) (L)->slp_instances
602 #define LOOP_VINFO_SLP_UNROLLING_FACTOR(L) (L)->slp_unrolling_factor
603 #define LOOP_VINFO_REDUCTIONS(L) (L)->reductions
604 #define LOOP_VINFO_REDUCTION_CHAINS(L) (L)->reduction_chains
605 #define LOOP_VINFO_TARGET_COST_DATA(L) (L)->target_cost_data
606 #define LOOP_VINFO_PEELING_FOR_GAPS(L) (L)->peeling_for_gaps
607 #define LOOP_VINFO_PEELING_FOR_NITER(L) (L)->peeling_for_niter
608 #define LOOP_VINFO_NO_DATA_DEPENDENCIES(L) (L)->no_data_dependencies
609 #define LOOP_VINFO_SCALAR_LOOP(L) (L)->scalar_loop
610 #define LOOP_VINFO_SCALAR_LOOP_SCALING(L) (L)->scalar_loop_scaling
611 #define LOOP_VINFO_HAS_MASK_STORE(L) (L)->has_mask_store
612 #define LOOP_VINFO_SCALAR_ITERATION_COST(L) (L)->scalar_cost_vec
613 #define LOOP_VINFO_SINGLE_SCALAR_ITERATION_COST(L) (L)->single_scalar_iteration_cost
614 #define LOOP_VINFO_ORIG_LOOP_INFO(L) (L)->orig_loop_info
615 #define LOOP_VINFO_SIMD_IF_COND(L) (L)->simd_if_cond
617 #define LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT(L) \
618 ((L)->may_misalign_stmts.length () > 0)
619 #define LOOP_REQUIRES_VERSIONING_FOR_ALIAS(L) \
620 ((L)->comp_alias_ddrs.length () > 0 \
621 || (L)->check_unequal_addrs.length () > 0 \
622 || (L)->lower_bounds.length () > 0)
623 #define LOOP_REQUIRES_VERSIONING_FOR_NITERS(L) \
624 (LOOP_VINFO_NITERS_ASSUMPTIONS (L))
625 #define LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND(L) \
626 (LOOP_VINFO_SIMD_IF_COND (L))
627 #define LOOP_REQUIRES_VERSIONING(L) \
628 (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (L) \
629 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (L) \
630 || LOOP_REQUIRES_VERSIONING_FOR_NITERS (L) \
631 || LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND (L))
633 #define LOOP_VINFO_NITERS_KNOWN_P(L) \
634 (tree_fits_shwi_p ((L)->num_iters) && tree_to_shwi ((L)->num_iters) > 0)
636 #define LOOP_VINFO_EPILOGUE_P(L) \
637 (LOOP_VINFO_ORIG_LOOP_INFO (L) != NULL)
639 #define LOOP_VINFO_ORIG_MAX_VECT_FACTOR(L) \
640 (LOOP_VINFO_MAX_VECT_FACTOR (LOOP_VINFO_ORIG_LOOP_INFO (L)))
642 /* Wrapper for loop_vec_info, for tracking success/failure, where a non-NULL
643 value signifies success, and a NULL value signifies failure, supporting
644 propagating an opt_problem * describing the failure back up the call
646 typedef opt_pointer_wrapper
<loop_vec_info
> opt_loop_vec_info
;
648 static inline loop_vec_info
649 loop_vec_info_for_loop (class loop
*loop
)
651 return (loop_vec_info
) loop
->aux
;
654 typedef class _bb_vec_info
: public vec_info
657 _bb_vec_info (gimple_stmt_iterator
, gimple_stmt_iterator
, vec_info_shared
*);
661 gimple_stmt_iterator region_begin
;
662 gimple_stmt_iterator region_end
;
665 #define BB_VINFO_BB(B) (B)->bb
666 #define BB_VINFO_GROUPED_STORES(B) (B)->grouped_stores
667 #define BB_VINFO_SLP_INSTANCES(B) (B)->slp_instances
668 #define BB_VINFO_DATAREFS(B) (B)->shared->datarefs
669 #define BB_VINFO_DDRS(B) (B)->shared->ddrs
670 #define BB_VINFO_TARGET_COST_DATA(B) (B)->target_cost_data
672 static inline bb_vec_info
673 vec_info_for_bb (basic_block bb
)
675 return (bb_vec_info
) bb
->aux
;
678 /*-----------------------------------------------------------------*/
679 /* Info on vectorized defs. */
680 /*-----------------------------------------------------------------*/
681 enum stmt_vec_info_type
{
682 undef_vec_info_type
= 0,
688 call_simd_clone_vec_info_type
,
689 assignment_vec_info_type
,
690 condition_vec_info_type
,
691 comparison_vec_info_type
,
694 type_promotion_vec_info_type
,
695 type_demotion_vec_info_type
,
696 type_conversion_vec_info_type
,
699 loop_exit_ctrl_vec_info_type
702 /* Indicates whether/how a variable is used in the scope of loop/basic
705 vect_unused_in_scope
= 0,
707 /* The def is only used outside the loop. */
709 /* The def is in the inner loop, and the use is in the outer loop, and the
710 use is a reduction stmt. */
711 vect_used_in_outer_by_reduction
,
712 /* The def is in the inner loop, and the use is in the outer loop (and is
713 not part of reduction). */
716 /* defs that feed computations that end up (only) in a reduction. These
717 defs may be used by non-reduction stmts, but eventually, any
718 computations/values that are affected by these defs are used to compute
719 a reduction (i.e. don't get stored to memory, for example). We use this
720 to identify computations that we can change the order in which they are
722 vect_used_by_reduction
,
727 /* The type of vectorization that can be applied to the stmt: regular loop-based
728 vectorization; pure SLP - the stmt is a part of SLP instances and does not
729 have uses outside SLP instances; or hybrid SLP and loop-based - the stmt is
730 a part of SLP instance and also must be loop-based vectorized, since it has
731 uses outside SLP sequences.
733 In the loop context the meanings of pure and hybrid SLP are slightly
734 different. By saying that pure SLP is applied to the loop, we mean that we
735 exploit only intra-iteration parallelism in the loop; i.e., the loop can be
736 vectorized without doing any conceptual unrolling, cause we don't pack
737 together stmts from different iterations, only within a single iteration.
738 Loop hybrid SLP means that we exploit both intra-iteration and
739 inter-iteration parallelism (e.g., number of elements in the vector is 4
740 and the slp-group-size is 2, in which case we don't have enough parallelism
741 within an iteration, so we obtain the rest of the parallelism from subsequent
742 iterations by unrolling the loop by 2). */
749 /* Says whether a statement is a load, a store of a vectorized statement
750 result, or a store of an invariant value. */
751 enum vec_load_store_type
{
757 /* Describes how we're going to vectorize an individual load or store,
758 or a group of loads or stores. */
759 enum vect_memory_access_type
{
760 /* An access to an invariant address. This is used only for loads. */
763 /* A simple contiguous access. */
766 /* A contiguous access that goes down in memory rather than up,
767 with no additional permutation. This is used only for stores
769 VMAT_CONTIGUOUS_DOWN
,
771 /* A simple contiguous access in which the elements need to be permuted
772 after loading or before storing. Only used for loop vectorization;
773 SLP uses separate permutes. */
774 VMAT_CONTIGUOUS_PERMUTE
,
776 /* A simple contiguous access in which the elements need to be reversed
777 after loading or before storing. */
778 VMAT_CONTIGUOUS_REVERSE
,
780 /* An access that uses IFN_LOAD_LANES or IFN_STORE_LANES. */
781 VMAT_LOAD_STORE_LANES
,
783 /* An access in which each scalar element is loaded or stored
787 /* A hybrid of VMAT_CONTIGUOUS and VMAT_ELEMENTWISE, used for grouped
788 SLP accesses. Each unrolled iteration uses a contiguous load
789 or store for the whole group, but the groups from separate iterations
790 are combined in the same way as for VMAT_ELEMENTWISE. */
793 /* The access uses gather loads or scatter stores. */
799 /* The data reference itself. */
801 /* The statement that contains the data reference. */
803 /* The misalignment in bytes of the reference, or -1 if not known. */
805 /* The byte alignment that we'd ideally like the reference to have,
806 and the value that misalignment is measured against. */
807 poly_uint64 target_alignment
;
808 /* If true the alignment of base_decl needs to be increased. */
809 bool base_misaligned
;
813 typedef struct data_reference
*dr_p
;
815 class _stmt_vec_info
{
818 enum stmt_vec_info_type type
;
820 /* Indicates whether this stmts is part of a computation whose result is
821 used outside the loop. */
824 /* Stmt is part of some pattern (computation idiom) */
827 /* True if the statement was created during pattern recognition as
828 part of the replacement for RELATED_STMT. This implies that the
829 statement isn't part of any basic block, although for convenience
830 its gimple_bb is the same as for RELATED_STMT. */
833 /* Is this statement vectorizable or should it be skipped in (partial)
837 /* The stmt to which this info struct refers to. */
840 /* The vec_info with respect to which STMT is vectorized. */
843 /* The vector type to be used for the LHS of this statement. */
846 /* The vectorized version of the stmt. */
847 stmt_vec_info vectorized_stmt
;
850 /* The following is relevant only for stmts that contain a non-scalar
851 data-ref (array/pointer/struct access). A GIMPLE stmt is expected to have
852 at most one such data-ref. */
856 /* Information about the data-ref relative to this loop
857 nest (the loop that is being considered for vectorization). */
858 innermost_loop_behavior dr_wrt_vec_loop
;
860 /* For loop PHI nodes, the base and evolution part of it. This makes sure
861 this information is still available in vect_update_ivs_after_vectorizer
862 where we may not be able to re-analyze the PHI nodes evolution as
863 peeling for the prologue loop can make it unanalyzable. The evolution
864 part is still correct after peeling, but the base may have changed from
866 tree loop_phi_evolution_base_unchanged
;
867 tree loop_phi_evolution_part
;
869 /* Used for various bookkeeping purposes, generally holding a pointer to
870 some other stmt S that is in some way "related" to this stmt.
871 Current use of this field is:
872 If this stmt is part of a pattern (i.e. the field 'in_pattern_p' is
873 true): S is the "pattern stmt" that represents (and replaces) the
874 sequence of stmts that constitutes the pattern. Similarly, the
875 related_stmt of the "pattern stmt" points back to this stmt (which is
876 the last stmt in the original sequence of stmts that constitutes the
878 stmt_vec_info related_stmt
;
880 /* Used to keep a sequence of def stmts of a pattern stmt if such exists.
881 The sequence is attached to the original statement rather than the
882 pattern statement. */
883 gimple_seq pattern_def_seq
;
885 /* List of datarefs that are known to have the same alignment as the dataref
887 vec
<dr_p
> same_align_refs
;
889 /* Selected SIMD clone's function info. First vector element
890 is SIMD clone's function decl, followed by a pair of trees (base + step)
891 for linear arguments (pair of NULLs for other arguments). */
892 vec
<tree
> simd_clone_info
;
894 /* Classify the def of this stmt. */
895 enum vect_def_type def_type
;
897 /* Whether the stmt is SLPed, loop-based vectorized, or both. */
898 enum slp_vect_type slp_type
;
900 /* Interleaving and reduction chains info. */
901 /* First element in the group. */
902 stmt_vec_info first_element
;
903 /* Pointer to the next element in the group. */
904 stmt_vec_info next_element
;
905 /* The size of the group. */
907 /* For stores, number of stores from this group seen. We vectorize the last
909 unsigned int store_count
;
910 /* For loads only, the gap from the previous load. For consecutive loads, GAP
914 /* The minimum negative dependence distance this stmt participates in
916 unsigned int min_neg_dist
;
918 /* Not all stmts in the loop need to be vectorized. e.g, the increment
919 of the loop induction variable and computation of array indexes. relevant
920 indicates whether the stmt needs to be vectorized. */
921 enum vect_relevant relevant
;
923 /* For loads if this is a gather, for stores if this is a scatter. */
924 bool gather_scatter_p
;
926 /* True if this is an access with loop-invariant stride. */
929 /* For both loads and stores. */
930 unsigned simd_lane_access_p
: 3;
932 /* Classifies how the load or store is going to be implemented
933 for loop vectorization. */
934 vect_memory_access_type memory_access_type
;
936 /* For CONST_COND_REDUCTION and INTEGER_INDUC_COND_REDUCTION, the
938 enum tree_code cond_reduc_code
;
940 /* For INTEGER_INDUC_COND_REDUCTION, the initial value to be used. */
941 tree induc_cond_initial_val
;
943 /* If not NULL the value to be added to compute final reduction value. */
944 tree reduc_epilogue_adjustment
;
946 /* On a reduction PHI the reduction type as detected by
947 vect_is_simple_reduction and vectorizable_reduction. */
948 enum vect_reduction_type reduc_type
;
950 /* The original reduction code, to be used in the epilogue. */
951 enum tree_code reduc_code
;
952 /* An internal function we should use in the epilogue. */
953 internal_fn reduc_fn
;
955 /* On a stmt participating in the reduction the index of the operand
956 on the reduction SSA cycle. */
959 /* On a reduction PHI the def returned by vect_force_simple_reduction.
960 On the def returned by vect_force_simple_reduction the
961 corresponding PHI. */
962 stmt_vec_info reduc_def
;
964 /* The vector input type relevant for reduction vectorization. */
965 tree reduc_vectype_in
;
967 /* Whether we force a single cycle PHI during reduction vectorization. */
968 bool force_single_cycle
;
970 /* Whether on this stmt reduction meta is recorded. */
973 /* The number of scalar stmt references from active SLP instances. */
974 unsigned int num_slp_uses
;
976 /* If nonzero, the lhs of the statement could be truncated to this
977 many bits without affecting any users of the result. */
978 unsigned int min_output_precision
;
980 /* If nonzero, all non-boolean input operands have the same precision,
981 and they could each be truncated to this many bits without changing
983 unsigned int min_input_precision
;
985 /* If OPERATION_BITS is nonzero, the statement could be performed on
986 an integer with the sign and number of bits given by OPERATION_SIGN
987 and OPERATION_BITS without changing the result. */
988 unsigned int operation_precision
;
989 signop operation_sign
;
991 /* True if this is only suitable for SLP vectorization. */
992 bool slp_vect_only_p
;
995 /* Information about a gather/scatter call. */
996 struct gather_scatter_info
{
997 /* The internal function to use for the gather/scatter operation,
998 or IFN_LAST if a built-in function should be used instead. */
1001 /* The FUNCTION_DECL for the built-in gather/scatter function,
1002 or null if an internal function should be used instead. */
1005 /* The loop-invariant base value. */
1008 /* The original scalar offset, which is a non-loop-invariant SSA_NAME. */
1011 /* Each offset element should be multiplied by this amount before
1012 being added to the base. */
1015 /* The definition type for the vectorized offset. */
1016 enum vect_def_type offset_dt
;
1018 /* The type of the vectorized offset. */
1019 tree offset_vectype
;
1021 /* The type of the scalar elements after loading or before storing. */
1024 /* The type of the scalar elements being loaded or stored. */
1028 /* Access Functions. */
1029 #define STMT_VINFO_TYPE(S) (S)->type
1030 #define STMT_VINFO_STMT(S) (S)->stmt
1031 inline loop_vec_info
1032 STMT_VINFO_LOOP_VINFO (stmt_vec_info stmt_vinfo
)
1034 if (loop_vec_info loop_vinfo
= dyn_cast
<loop_vec_info
> (stmt_vinfo
->vinfo
))
1039 STMT_VINFO_BB_VINFO (stmt_vec_info stmt_vinfo
)
1041 if (bb_vec_info bb_vinfo
= dyn_cast
<bb_vec_info
> (stmt_vinfo
->vinfo
))
1045 #define STMT_VINFO_RELEVANT(S) (S)->relevant
1046 #define STMT_VINFO_LIVE_P(S) (S)->live
1047 #define STMT_VINFO_VECTYPE(S) (S)->vectype
1048 #define STMT_VINFO_VEC_STMT(S) (S)->vectorized_stmt
1049 #define STMT_VINFO_VECTORIZABLE(S) (S)->vectorizable
1050 #define STMT_VINFO_DATA_REF(S) ((S)->dr_aux.dr + 0)
1051 #define STMT_VINFO_GATHER_SCATTER_P(S) (S)->gather_scatter_p
1052 #define STMT_VINFO_STRIDED_P(S) (S)->strided_p
1053 #define STMT_VINFO_MEMORY_ACCESS_TYPE(S) (S)->memory_access_type
1054 #define STMT_VINFO_SIMD_LANE_ACCESS_P(S) (S)->simd_lane_access_p
1055 #define STMT_VINFO_VEC_COND_REDUC_CODE(S) (S)->cond_reduc_code
1056 #define STMT_VINFO_VEC_INDUC_COND_INITIAL_VAL(S) (S)->induc_cond_initial_val
1057 #define STMT_VINFO_REDUC_EPILOGUE_ADJUSTMENT(S) (S)->reduc_epilogue_adjustment
1058 #define STMT_VINFO_REDUC_IDX(S) (S)->reduc_idx
1059 #define STMT_VINFO_FORCE_SINGLE_CYCLE(S) (S)->force_single_cycle
1061 #define STMT_VINFO_DR_WRT_VEC_LOOP(S) (S)->dr_wrt_vec_loop
1062 #define STMT_VINFO_DR_BASE_ADDRESS(S) (S)->dr_wrt_vec_loop.base_address
1063 #define STMT_VINFO_DR_INIT(S) (S)->dr_wrt_vec_loop.init
1064 #define STMT_VINFO_DR_OFFSET(S) (S)->dr_wrt_vec_loop.offset
1065 #define STMT_VINFO_DR_STEP(S) (S)->dr_wrt_vec_loop.step
1066 #define STMT_VINFO_DR_BASE_ALIGNMENT(S) (S)->dr_wrt_vec_loop.base_alignment
1067 #define STMT_VINFO_DR_BASE_MISALIGNMENT(S) \
1068 (S)->dr_wrt_vec_loop.base_misalignment
1069 #define STMT_VINFO_DR_OFFSET_ALIGNMENT(S) \
1070 (S)->dr_wrt_vec_loop.offset_alignment
1071 #define STMT_VINFO_DR_STEP_ALIGNMENT(S) \
1072 (S)->dr_wrt_vec_loop.step_alignment
1074 #define STMT_VINFO_DR_INFO(S) \
1075 (gcc_checking_assert ((S)->dr_aux.stmt == (S)), &(S)->dr_aux)
1077 #define STMT_VINFO_IN_PATTERN_P(S) (S)->in_pattern_p
1078 #define STMT_VINFO_RELATED_STMT(S) (S)->related_stmt
1079 #define STMT_VINFO_PATTERN_DEF_SEQ(S) (S)->pattern_def_seq
1080 #define STMT_VINFO_SAME_ALIGN_REFS(S) (S)->same_align_refs
1081 #define STMT_VINFO_SIMD_CLONE_INFO(S) (S)->simd_clone_info
1082 #define STMT_VINFO_DEF_TYPE(S) (S)->def_type
1083 #define STMT_VINFO_GROUPED_ACCESS(S) \
1084 ((S)->dr_aux.dr && DR_GROUP_FIRST_ELEMENT(S))
1085 #define STMT_VINFO_LOOP_PHI_EVOLUTION_BASE_UNCHANGED(S) (S)->loop_phi_evolution_base_unchanged
1086 #define STMT_VINFO_LOOP_PHI_EVOLUTION_PART(S) (S)->loop_phi_evolution_part
1087 #define STMT_VINFO_MIN_NEG_DIST(S) (S)->min_neg_dist
1088 #define STMT_VINFO_NUM_SLP_USES(S) (S)->num_slp_uses
1089 #define STMT_VINFO_REDUC_TYPE(S) (S)->reduc_type
1090 #define STMT_VINFO_REDUC_CODE(S) (S)->reduc_code
1091 #define STMT_VINFO_REDUC_FN(S) (S)->reduc_fn
1092 #define STMT_VINFO_REDUC_DEF(S) (S)->reduc_def
1093 #define STMT_VINFO_REDUC_VECTYPE_IN(S) (S)->reduc_vectype_in
1094 #define STMT_VINFO_SLP_VECT_ONLY(S) (S)->slp_vect_only_p
1096 #define DR_GROUP_FIRST_ELEMENT(S) \
1097 (gcc_checking_assert ((S)->dr_aux.dr), (S)->first_element)
1098 #define DR_GROUP_NEXT_ELEMENT(S) \
1099 (gcc_checking_assert ((S)->dr_aux.dr), (S)->next_element)
1100 #define DR_GROUP_SIZE(S) \
1101 (gcc_checking_assert ((S)->dr_aux.dr), (S)->size)
1102 #define DR_GROUP_STORE_COUNT(S) \
1103 (gcc_checking_assert ((S)->dr_aux.dr), (S)->store_count)
1104 #define DR_GROUP_GAP(S) \
1105 (gcc_checking_assert ((S)->dr_aux.dr), (S)->gap)
1107 #define REDUC_GROUP_FIRST_ELEMENT(S) \
1108 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->first_element)
1109 #define REDUC_GROUP_NEXT_ELEMENT(S) \
1110 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->next_element)
1111 #define REDUC_GROUP_SIZE(S) \
1112 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->size)
1114 #define STMT_VINFO_RELEVANT_P(S) ((S)->relevant != vect_unused_in_scope)
1116 #define HYBRID_SLP_STMT(S) ((S)->slp_type == hybrid)
1117 #define PURE_SLP_STMT(S) ((S)->slp_type == pure_slp)
1118 #define STMT_SLP_TYPE(S) (S)->slp_type
1120 #define VECT_MAX_COST 1000
1122 /* The maximum number of intermediate steps required in multi-step type
1124 #define MAX_INTERM_CVT_STEPS 3
1126 #define MAX_VECTORIZATION_FACTOR INT_MAX
1128 /* Nonzero if TYPE represents a (scalar) boolean type or type
1129 in the middle-end compatible with it (unsigned precision 1 integral
1130 types). Used to determine which types should be vectorized as
1131 VECTOR_BOOLEAN_TYPE_P. */
1133 #define VECT_SCALAR_BOOLEAN_TYPE_P(TYPE) \
1134 (TREE_CODE (TYPE) == BOOLEAN_TYPE \
1135 || ((TREE_CODE (TYPE) == INTEGER_TYPE \
1136 || TREE_CODE (TYPE) == ENUMERAL_TYPE) \
1137 && TYPE_PRECISION (TYPE) == 1 \
1138 && TYPE_UNSIGNED (TYPE)))
1141 nested_in_vect_loop_p (class loop
*loop
, stmt_vec_info stmt_info
)
1144 && (loop
->inner
== (gimple_bb (stmt_info
->stmt
))->loop_father
));
1147 /* Return TRUE if a statement represented by STMT_INFO is a part of a
1151 is_pattern_stmt_p (stmt_vec_info stmt_info
)
1153 return stmt_info
->pattern_stmt_p
;
1156 /* If STMT_INFO is a pattern statement, return the statement that it
1157 replaces, otherwise return STMT_INFO itself. */
1159 inline stmt_vec_info
1160 vect_orig_stmt (stmt_vec_info stmt_info
)
1162 if (is_pattern_stmt_p (stmt_info
))
1163 return STMT_VINFO_RELATED_STMT (stmt_info
);
1167 /* Return the later statement between STMT1_INFO and STMT2_INFO. */
1169 static inline stmt_vec_info
1170 get_later_stmt (stmt_vec_info stmt1_info
, stmt_vec_info stmt2_info
)
1172 if (gimple_uid (vect_orig_stmt (stmt1_info
)->stmt
)
1173 > gimple_uid (vect_orig_stmt (stmt2_info
)->stmt
))
1179 /* If STMT_INFO has been replaced by a pattern statement, return the
1180 replacement statement, otherwise return STMT_INFO itself. */
1182 inline stmt_vec_info
1183 vect_stmt_to_vectorize (stmt_vec_info stmt_info
)
1185 if (STMT_VINFO_IN_PATTERN_P (stmt_info
))
1186 return STMT_VINFO_RELATED_STMT (stmt_info
);
1190 /* Return true if BB is a loop header. */
1193 is_loop_header_bb_p (basic_block bb
)
1195 if (bb
== (bb
->loop_father
)->header
)
1197 gcc_checking_assert (EDGE_COUNT (bb
->preds
) == 1);
1201 /* Return pow2 (X). */
1208 for (i
= 0; i
< x
; i
++)
1214 /* Alias targetm.vectorize.builtin_vectorization_cost. */
1217 builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost
,
1218 tree vectype
, int misalign
)
1220 return targetm
.vectorize
.builtin_vectorization_cost (type_of_cost
,
1224 /* Get cost by calling cost target builtin. */
1227 int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost
)
1229 return builtin_vectorization_cost (type_of_cost
, NULL
, 0);
1232 /* Alias targetm.vectorize.init_cost. */
1234 static inline void *
1235 init_cost (class loop
*loop_info
)
1237 return targetm
.vectorize
.init_cost (loop_info
);
1240 extern void dump_stmt_cost (FILE *, void *, int, enum vect_cost_for_stmt
,
1241 stmt_vec_info
, int, unsigned,
1242 enum vect_cost_model_location
);
1244 /* Alias targetm.vectorize.add_stmt_cost. */
1246 static inline unsigned
1247 add_stmt_cost (void *data
, int count
, enum vect_cost_for_stmt kind
,
1248 stmt_vec_info stmt_info
, int misalign
,
1249 enum vect_cost_model_location where
)
1251 unsigned cost
= targetm
.vectorize
.add_stmt_cost (data
, count
, kind
,
1252 stmt_info
, misalign
, where
);
1253 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1254 dump_stmt_cost (dump_file
, data
, count
, kind
, stmt_info
, misalign
,
1259 /* Alias targetm.vectorize.finish_cost. */
1262 finish_cost (void *data
, unsigned *prologue_cost
,
1263 unsigned *body_cost
, unsigned *epilogue_cost
)
1265 targetm
.vectorize
.finish_cost (data
, prologue_cost
, body_cost
, epilogue_cost
);
1268 /* Alias targetm.vectorize.destroy_cost_data. */
1271 destroy_cost_data (void *data
)
1273 targetm
.vectorize
.destroy_cost_data (data
);
1277 add_stmt_costs (void *data
, stmt_vector_for_cost
*cost_vec
)
1279 stmt_info_for_cost
*cost
;
1281 FOR_EACH_VEC_ELT (*cost_vec
, i
, cost
)
1282 add_stmt_cost (data
, cost
->count
, cost
->kind
, cost
->stmt_info
,
1283 cost
->misalign
, cost
->where
);
1286 /*-----------------------------------------------------------------*/
1287 /* Info on data references alignment. */
1288 /*-----------------------------------------------------------------*/
1289 #define DR_MISALIGNMENT_UNKNOWN (-1)
1290 #define DR_MISALIGNMENT_UNINITIALIZED (-2)
1293 set_dr_misalignment (dr_vec_info
*dr_info
, int val
)
1295 dr_info
->misalignment
= val
;
1299 dr_misalignment (dr_vec_info
*dr_info
)
1301 int misalign
= dr_info
->misalignment
;
1302 gcc_assert (misalign
!= DR_MISALIGNMENT_UNINITIALIZED
);
1306 /* Reflects actual alignment of first access in the vectorized loop,
1307 taking into account peeling/versioning if applied. */
1308 #define DR_MISALIGNMENT(DR) dr_misalignment (DR)
1309 #define SET_DR_MISALIGNMENT(DR, VAL) set_dr_misalignment (DR, VAL)
1311 /* Only defined once DR_MISALIGNMENT is defined. */
1312 #define DR_TARGET_ALIGNMENT(DR) ((DR)->target_alignment)
1314 /* Return true if data access DR_INFO is aligned to its target alignment
1315 (which may be less than a full vector). */
1318 aligned_access_p (dr_vec_info
*dr_info
)
1320 return (DR_MISALIGNMENT (dr_info
) == 0);
1323 /* Return TRUE if the alignment of the data access is known, and FALSE
1327 known_alignment_for_access_p (dr_vec_info
*dr_info
)
1329 return (DR_MISALIGNMENT (dr_info
) != DR_MISALIGNMENT_UNKNOWN
);
1332 /* Return the minimum alignment in bytes that the vectorized version
1333 of DR_INFO is guaranteed to have. */
1335 static inline unsigned int
1336 vect_known_alignment_in_bytes (dr_vec_info
*dr_info
)
1338 if (DR_MISALIGNMENT (dr_info
) == DR_MISALIGNMENT_UNKNOWN
)
1339 return TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr_info
->dr
)));
1340 if (DR_MISALIGNMENT (dr_info
) == 0)
1341 return known_alignment (DR_TARGET_ALIGNMENT (dr_info
));
1342 return DR_MISALIGNMENT (dr_info
) & -DR_MISALIGNMENT (dr_info
);
1345 /* Return the behavior of DR_INFO with respect to the vectorization context
1346 (which for outer loop vectorization might not be the behavior recorded
1347 in DR_INFO itself). */
1349 static inline innermost_loop_behavior
*
1350 vect_dr_behavior (dr_vec_info
*dr_info
)
1352 stmt_vec_info stmt_info
= dr_info
->stmt
;
1353 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
1354 if (loop_vinfo
== NULL
1355 || !nested_in_vect_loop_p (LOOP_VINFO_LOOP (loop_vinfo
), stmt_info
))
1356 return &DR_INNERMOST (dr_info
->dr
);
1358 return &STMT_VINFO_DR_WRT_VEC_LOOP (stmt_info
);
1361 /* Return true if the vect cost model is unlimited. */
1363 unlimited_cost_model (loop_p loop
)
1365 if (loop
!= NULL
&& loop
->force_vectorize
1366 && flag_simd_cost_model
!= VECT_COST_MODEL_DEFAULT
)
1367 return flag_simd_cost_model
== VECT_COST_MODEL_UNLIMITED
;
1368 return (flag_vect_cost_model
== VECT_COST_MODEL_UNLIMITED
);
1371 /* Return true if the loop described by LOOP_VINFO is fully-masked and
1372 if the first iteration should use a partial mask in order to achieve
1376 vect_use_loop_mask_for_alignment_p (loop_vec_info loop_vinfo
)
1378 return (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo
)
1379 && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo
));
1382 /* Return the number of vectors of type VECTYPE that are needed to get
1383 NUNITS elements. NUNITS should be based on the vectorization factor,
1384 so it is always a known multiple of the number of elements in VECTYPE. */
1386 static inline unsigned int
1387 vect_get_num_vectors (poly_uint64 nunits
, tree vectype
)
1389 return exact_div (nunits
, TYPE_VECTOR_SUBPARTS (vectype
)).to_constant ();
1392 /* Return the number of copies needed for loop vectorization when
1393 a statement operates on vectors of type VECTYPE. This is the
1394 vectorization factor divided by the number of elements in
1395 VECTYPE and is always known at compile time. */
1397 static inline unsigned int
1398 vect_get_num_copies (loop_vec_info loop_vinfo
, tree vectype
)
1400 return vect_get_num_vectors (LOOP_VINFO_VECT_FACTOR (loop_vinfo
), vectype
);
1403 /* Update maximum unit count *MAX_NUNITS so that it accounts for
1404 NUNITS. *MAX_NUNITS can be 1 if we haven't yet recorded anything. */
1407 vect_update_max_nunits (poly_uint64
*max_nunits
, poly_uint64 nunits
)
1409 /* All unit counts have the form current_vector_size * X for some
1410 rational X, so two unit sizes must have a common multiple.
1411 Everything is a multiple of the initial value of 1. */
1412 *max_nunits
= force_common_multiple (*max_nunits
, nunits
);
1415 /* Update maximum unit count *MAX_NUNITS so that it accounts for
1416 the number of units in vector type VECTYPE. *MAX_NUNITS can be 1
1417 if we haven't yet recorded any vector types. */
1420 vect_update_max_nunits (poly_uint64
*max_nunits
, tree vectype
)
1422 vect_update_max_nunits (max_nunits
, TYPE_VECTOR_SUBPARTS (vectype
));
1425 /* Return the vectorization factor that should be used for costing
1426 purposes while vectorizing the loop described by LOOP_VINFO.
1427 Pick a reasonable estimate if the vectorization factor isn't
1428 known at compile time. */
1430 static inline unsigned int
1431 vect_vf_for_cost (loop_vec_info loop_vinfo
)
1433 return estimated_poly_value (LOOP_VINFO_VECT_FACTOR (loop_vinfo
));
1436 /* Estimate the number of elements in VEC_TYPE for costing purposes.
1437 Pick a reasonable estimate if the exact number isn't known at
1440 static inline unsigned int
1441 vect_nunits_for_cost (tree vec_type
)
1443 return estimated_poly_value (TYPE_VECTOR_SUBPARTS (vec_type
));
1446 /* Return the maximum possible vectorization factor for LOOP_VINFO. */
1448 static inline unsigned HOST_WIDE_INT
1449 vect_max_vf (loop_vec_info loop_vinfo
)
1451 unsigned HOST_WIDE_INT vf
;
1452 if (LOOP_VINFO_VECT_FACTOR (loop_vinfo
).is_constant (&vf
))
1454 return MAX_VECTORIZATION_FACTOR
;
1457 /* Return the size of the value accessed by unvectorized data reference
1458 DR_INFO. This is only valid once STMT_VINFO_VECTYPE has been calculated
1459 for the associated gimple statement, since that guarantees that DR_INFO
1460 accesses either a scalar or a scalar equivalent. ("Scalar equivalent"
1461 here includes things like V1SI, which can be vectorized in the same way
1465 vect_get_scalar_dr_size (dr_vec_info
*dr_info
)
1467 return tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_info
->dr
))));
1470 /* Source location + hotness information. */
1471 extern dump_user_location_t vect_location
;
1473 /* A macro for calling:
1474 dump_begin_scope (MSG, vect_location);
1475 via an RAII object, thus printing "=== MSG ===\n" to the dumpfile etc,
1478 once the object goes out of scope, thus capturing the nesting of
1481 These scopes affect dump messages within them: dump messages at the
1482 top level implicitly default to MSG_PRIORITY_USER_FACING, whereas those
1483 in a nested scope implicitly default to MSG_PRIORITY_INTERNALS. */
1485 #define DUMP_VECT_SCOPE(MSG) \
1486 AUTO_DUMP_SCOPE (MSG, vect_location)
1488 /* A sentinel class for ensuring that the "vect_location" global gets
1489 reset at the end of a scope.
1491 The "vect_location" global is used during dumping and contains a
1492 location_t, which could contain references to a tree block via the
1493 ad-hoc data. This data is used for tracking inlining information,
1494 but it's not a GC root; it's simply assumed that such locations never
1495 get accessed if the blocks are optimized away.
1497 Hence we need to ensure that such locations are purged at the end
1498 of any operations using them (e.g. via this class). */
1500 class auto_purge_vect_location
1503 ~auto_purge_vect_location ();
1506 /*-----------------------------------------------------------------*/
1507 /* Function prototypes. */
1508 /*-----------------------------------------------------------------*/
1510 /* Simple loop peeling and versioning utilities for vectorizer's purposes -
1511 in tree-vect-loop-manip.c. */
1512 extern void vect_set_loop_condition (class loop
*, loop_vec_info
,
1513 tree
, tree
, tree
, bool);
1514 extern bool slpeel_can_duplicate_loop_p (const class loop
*, const_edge
);
1515 class loop
*slpeel_tree_duplicate_loop_to_edge_cfg (class loop
*,
1516 class loop
*, edge
);
1517 class loop
*vect_loop_versioning (loop_vec_info
);
1518 extern class loop
*vect_do_peeling (loop_vec_info
, tree
, tree
,
1519 tree
*, tree
*, tree
*, int, bool, bool);
1520 extern void vect_prepare_for_masked_peels (loop_vec_info
);
1521 extern dump_user_location_t
find_loop_location (class loop
*);
1522 extern bool vect_can_advance_ivs_p (loop_vec_info
);
1524 /* In tree-vect-stmts.c. */
1525 extern poly_uint64 current_vector_size
;
1526 extern tree
get_vectype_for_scalar_type (tree
);
1527 extern tree
get_vectype_for_scalar_type_and_size (tree
, poly_uint64
);
1528 extern tree
get_mask_type_for_scalar_type (tree
);
1529 extern tree
get_same_sized_vectype (tree
, tree
);
1530 extern bool vect_get_loop_mask_type (loop_vec_info
);
1531 extern bool vect_is_simple_use (tree
, vec_info
*, enum vect_def_type
*,
1532 stmt_vec_info
* = NULL
, gimple
** = NULL
);
1533 extern bool vect_is_simple_use (tree
, vec_info
*, enum vect_def_type
*,
1534 tree
*, stmt_vec_info
* = NULL
,
1536 extern bool supportable_widening_operation (enum tree_code
, stmt_vec_info
,
1537 tree
, tree
, enum tree_code
*,
1538 enum tree_code
*, int *,
1540 extern bool supportable_narrowing_operation (enum tree_code
, tree
, tree
,
1542 int *, vec
<tree
> *);
1543 extern unsigned record_stmt_cost (stmt_vector_for_cost
*, int,
1544 enum vect_cost_for_stmt
, stmt_vec_info
,
1545 int, enum vect_cost_model_location
);
1546 extern stmt_vec_info
vect_finish_replace_stmt (stmt_vec_info
, gimple
*);
1547 extern stmt_vec_info
vect_finish_stmt_generation (stmt_vec_info
, gimple
*,
1548 gimple_stmt_iterator
*);
1549 extern opt_result
vect_mark_stmts_to_be_vectorized (loop_vec_info
, bool *);
1550 extern tree
vect_get_store_rhs (stmt_vec_info
);
1551 extern tree
vect_get_vec_def_for_operand_1 (stmt_vec_info
, enum vect_def_type
);
1552 extern tree
vect_get_vec_def_for_operand (tree
, stmt_vec_info
, tree
= NULL
);
1553 extern void vect_get_vec_defs (tree
, tree
, stmt_vec_info
, vec
<tree
> *,
1554 vec
<tree
> *, slp_tree
);
1555 extern void vect_get_vec_defs_for_stmt_copy (vec_info
*,
1556 vec
<tree
> *, vec
<tree
> *);
1557 extern tree
vect_init_vector (stmt_vec_info
, tree
, tree
,
1558 gimple_stmt_iterator
*);
1559 extern tree
vect_get_vec_def_for_stmt_copy (vec_info
*, tree
);
1560 extern bool vect_transform_stmt (stmt_vec_info
, gimple_stmt_iterator
*,
1561 slp_tree
, slp_instance
);
1562 extern void vect_remove_stores (stmt_vec_info
);
1563 extern opt_result
vect_analyze_stmt (stmt_vec_info
, bool *, slp_tree
,
1564 slp_instance
, stmt_vector_for_cost
*);
1565 extern void vect_get_load_cost (stmt_vec_info
, int, bool,
1566 unsigned int *, unsigned int *,
1567 stmt_vector_for_cost
*,
1568 stmt_vector_for_cost
*, bool);
1569 extern void vect_get_store_cost (stmt_vec_info
, int,
1570 unsigned int *, stmt_vector_for_cost
*);
1571 extern bool vect_supportable_shift (enum tree_code
, tree
);
1572 extern tree
vect_gen_perm_mask_any (tree
, const vec_perm_indices
&);
1573 extern tree
vect_gen_perm_mask_checked (tree
, const vec_perm_indices
&);
1574 extern void optimize_mask_stores (class loop
*);
1575 extern gcall
*vect_gen_while (tree
, tree
, tree
);
1576 extern tree
vect_gen_while_not (gimple_seq
*, tree
, tree
, tree
);
1577 extern opt_result
vect_get_vector_types_for_stmt (stmt_vec_info
, tree
*,
1579 extern opt_tree
vect_get_mask_type_for_stmt (stmt_vec_info
);
1581 /* In tree-vect-data-refs.c. */
1582 extern bool vect_can_force_dr_alignment_p (const_tree
, poly_uint64
);
1583 extern enum dr_alignment_support vect_supportable_dr_alignment
1584 (dr_vec_info
*, bool);
1585 extern tree
vect_get_smallest_scalar_type (stmt_vec_info
, HOST_WIDE_INT
*,
1587 extern opt_result
vect_analyze_data_ref_dependences (loop_vec_info
, unsigned int *);
1588 extern bool vect_slp_analyze_instance_dependence (slp_instance
);
1589 extern opt_result
vect_enhance_data_refs_alignment (loop_vec_info
);
1590 extern opt_result
vect_analyze_data_refs_alignment (loop_vec_info
);
1591 extern opt_result
vect_verify_datarefs_alignment (loop_vec_info
);
1592 extern bool vect_slp_analyze_and_verify_instance_alignment (slp_instance
);
1593 extern opt_result
vect_analyze_data_ref_accesses (vec_info
*);
1594 extern opt_result
vect_prune_runtime_alias_test_list (loop_vec_info
);
1595 extern bool vect_gather_scatter_fn_p (bool, bool, tree
, tree
, unsigned int,
1596 signop
, int, internal_fn
*, tree
*);
1597 extern bool vect_check_gather_scatter (stmt_vec_info
, loop_vec_info
,
1598 gather_scatter_info
*);
1599 extern opt_result
vect_find_stmt_data_reference (loop_p
, gimple
*,
1600 vec
<data_reference_p
> *);
1601 extern opt_result
vect_analyze_data_refs (vec_info
*, poly_uint64
*, bool *);
1602 extern void vect_record_base_alignments (vec_info
*);
1603 extern tree
vect_create_data_ref_ptr (stmt_vec_info
, tree
, class loop
*, tree
,
1604 tree
*, gimple_stmt_iterator
*,
1606 tree
= NULL_TREE
, tree
= NULL_TREE
);
1607 extern tree
bump_vector_ptr (tree
, gimple
*, gimple_stmt_iterator
*,
1608 stmt_vec_info
, tree
);
1609 extern void vect_copy_ref_info (tree
, tree
);
1610 extern tree
vect_create_destination_var (tree
, tree
);
1611 extern bool vect_grouped_store_supported (tree
, unsigned HOST_WIDE_INT
);
1612 extern bool vect_store_lanes_supported (tree
, unsigned HOST_WIDE_INT
, bool);
1613 extern bool vect_grouped_load_supported (tree
, bool, unsigned HOST_WIDE_INT
);
1614 extern bool vect_load_lanes_supported (tree
, unsigned HOST_WIDE_INT
, bool);
1615 extern void vect_permute_store_chain (vec
<tree
> ,unsigned int, stmt_vec_info
,
1616 gimple_stmt_iterator
*, vec
<tree
> *);
1617 extern tree
vect_setup_realignment (stmt_vec_info
, gimple_stmt_iterator
*,
1618 tree
*, enum dr_alignment_support
, tree
,
1620 extern void vect_transform_grouped_load (stmt_vec_info
, vec
<tree
> , int,
1621 gimple_stmt_iterator
*);
1622 extern void vect_record_grouped_load_vectors (stmt_vec_info
, vec
<tree
>);
1623 extern tree
vect_get_new_vect_var (tree
, enum vect_var_kind
, const char *);
1624 extern tree
vect_get_new_ssa_name (tree
, enum vect_var_kind
,
1625 const char * = NULL
);
1626 extern tree
vect_create_addr_base_for_vector_ref (stmt_vec_info
, gimple_seq
*,
1627 tree
, tree
= NULL_TREE
);
1629 /* In tree-vect-loop.c. */
1630 extern widest_int
vect_iv_limit_for_full_masking (loop_vec_info loop_vinfo
);
1631 /* Used in gimple-loop-interchange.c and tree-parloops.c. */
1632 extern bool check_reduction_path (dump_user_location_t
, loop_p
, gphi
*, tree
,
1634 extern bool needs_fold_left_reduction_p (tree
, tree_code
);
1635 /* Drive for loop analysis stage. */
1636 extern opt_loop_vec_info
vect_analyze_loop (class loop
*,
1639 extern tree
vect_build_loop_niters (loop_vec_info
, bool * = NULL
);
1640 extern void vect_gen_vector_loop_niters (loop_vec_info
, tree
, tree
*,
1642 extern tree
vect_halve_mask_nunits (tree
);
1643 extern tree
vect_double_mask_nunits (tree
);
1644 extern void vect_record_loop_mask (loop_vec_info
, vec_loop_masks
*,
1645 unsigned int, tree
);
1646 extern tree
vect_get_loop_mask (gimple_stmt_iterator
*, vec_loop_masks
*,
1647 unsigned int, tree
, unsigned int);
1648 extern stmt_vec_info
info_for_reduction (stmt_vec_info
);
1650 /* Drive for loop transformation stage. */
1651 extern class loop
*vect_transform_loop (loop_vec_info
);
1652 extern opt_loop_vec_info
vect_analyze_loop_form (class loop
*,
1654 extern bool vectorizable_live_operation (stmt_vec_info
, gimple_stmt_iterator
*,
1655 slp_tree
, slp_instance
, int,
1656 bool, stmt_vector_for_cost
*);
1657 extern bool vectorizable_reduction (stmt_vec_info
, slp_tree
, slp_instance
,
1658 stmt_vector_for_cost
*);
1659 extern bool vectorizable_induction (stmt_vec_info
, gimple_stmt_iterator
*,
1660 stmt_vec_info
*, slp_tree
,
1661 stmt_vector_for_cost
*);
1662 extern bool vect_transform_reduction (stmt_vec_info
, gimple_stmt_iterator
*,
1663 stmt_vec_info
*, slp_tree
);
1664 extern bool vect_transform_cycle_phi (stmt_vec_info
, stmt_vec_info
*,
1665 slp_tree
, slp_instance
);
1666 extern bool vectorizable_lc_phi (stmt_vec_info
, stmt_vec_info
*, slp_tree
);
1667 extern bool vect_worthwhile_without_simd_p (vec_info
*, tree_code
);
1668 extern int vect_get_known_peeling_cost (loop_vec_info
, int, int *,
1669 stmt_vector_for_cost
*,
1670 stmt_vector_for_cost
*,
1671 stmt_vector_for_cost
*);
1672 extern tree
cse_and_gimplify_to_preheader (loop_vec_info
, tree
);
1674 /* In tree-vect-slp.c. */
1675 extern void vect_free_slp_instance (slp_instance
, bool);
1676 extern bool vect_transform_slp_perm_load (slp_tree
, vec
<tree
> ,
1677 gimple_stmt_iterator
*, poly_uint64
,
1678 slp_instance
, bool, unsigned *);
1679 extern bool vect_slp_analyze_operations (vec_info
*);
1680 extern void vect_schedule_slp (vec_info
*);
1681 extern opt_result
vect_analyze_slp (vec_info
*, unsigned);
1682 extern bool vect_make_slp_decision (loop_vec_info
);
1683 extern void vect_detect_hybrid_slp (loop_vec_info
);
1684 extern void vect_get_slp_defs (vec
<tree
> , slp_tree
, vec
<vec
<tree
> > *);
1685 extern bool vect_slp_bb (basic_block
);
1686 extern stmt_vec_info
vect_find_last_scalar_stmt_in_slp (slp_tree
);
1687 extern bool is_simple_and_all_uses_invariant (stmt_vec_info
, loop_vec_info
);
1688 extern bool can_duplicate_and_interleave_p (unsigned int, machine_mode
,
1689 unsigned int * = NULL
,
1690 tree
* = NULL
, tree
* = NULL
);
1691 extern void duplicate_and_interleave (gimple_seq
*, tree
, vec
<tree
>,
1692 unsigned int, vec
<tree
> &);
1693 extern int vect_get_place_in_interleaving_chain (stmt_vec_info
, stmt_vec_info
);
1695 /* In tree-vect-patterns.c. */
1696 /* Pattern recognition functions.
1697 Additional pattern recognition functions can (and will) be added
1699 void vect_pattern_recog (vec_info
*);
1701 /* In tree-vectorizer.c. */
1702 unsigned vectorize_loops (void);
1703 void vect_free_loop_info_assumptions (class loop
*);
1704 gimple
*vect_loop_vectorized_call (class loop
*, gcond
**cond
= NULL
);
1707 #endif /* GCC_TREE_VECTORIZER_H */