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 /* Whether the scalar computations use two different operators. */
137 /* The DEF type of this node. */
138 enum vect_def_type def_type
;
142 /* SLP instance is a sequence of stmts in a loop that can be packed into
144 typedef class _slp_instance
{
146 /* The root of SLP tree. */
149 /* Size of groups of scalar stmts that will be replaced by SIMD stmt/s. */
150 unsigned int group_size
;
152 /* The unrolling factor required to vectorized this SLP instance. */
153 poly_uint64 unrolling_factor
;
155 /* The group of nodes that contain loads of this SLP instance. */
158 /* The SLP node containing the reduction PHIs. */
163 /* Access Functions. */
164 #define SLP_INSTANCE_TREE(S) (S)->root
165 #define SLP_INSTANCE_GROUP_SIZE(S) (S)->group_size
166 #define SLP_INSTANCE_UNROLLING_FACTOR(S) (S)->unrolling_factor
167 #define SLP_INSTANCE_LOADS(S) (S)->loads
169 #define SLP_TREE_CHILDREN(S) (S)->children
170 #define SLP_TREE_SCALAR_STMTS(S) (S)->stmts
171 #define SLP_TREE_VEC_STMTS(S) (S)->vec_stmts
172 #define SLP_TREE_NUMBER_OF_VEC_STMTS(S) (S)->vec_stmts_size
173 #define SLP_TREE_LOAD_PERMUTATION(S) (S)->load_permutation
174 #define SLP_TREE_TWO_OPERATORS(S) (S)->two_operators
175 #define SLP_TREE_DEF_TYPE(S) (S)->def_type
179 /* Describes two objects whose addresses must be unequal for the vectorized
181 typedef std::pair
<tree
, tree
> vec_object_pair
;
183 /* Records that vectorization is only possible if abs (EXPR) >= MIN_VALUE.
184 UNSIGNED_P is true if we can assume that abs (EXPR) == EXPR. */
185 class vec_lower_bound
{
187 vec_lower_bound () {}
188 vec_lower_bound (tree e
, bool u
, poly_uint64 m
)
189 : expr (e
), unsigned_p (u
), min_value (m
) {}
193 poly_uint64 min_value
;
196 /* Vectorizer state shared between different analyses like vector sizes
197 of the same CFG region. */
198 class vec_info_shared
{
203 void save_datarefs();
204 void check_datarefs();
206 /* All data references. Freed by free_data_refs, so not an auto_vec. */
207 vec
<data_reference_p
> datarefs
;
208 vec
<data_reference
> datarefs_copy
;
210 /* The loop nest in which the data dependences are computed. */
211 auto_vec
<loop_p
> loop_nest
;
213 /* All data dependences. Freed by free_dependence_relations, so not
218 /* Vectorizer state common between loop and basic-block vectorization. */
221 enum vec_kind
{ bb
, loop
};
223 vec_info (vec_kind
, void *, vec_info_shared
*);
226 stmt_vec_info
add_stmt (gimple
*);
227 stmt_vec_info
lookup_stmt (gimple
*);
228 stmt_vec_info
lookup_def (tree
);
229 stmt_vec_info
lookup_single_use (tree
);
230 class dr_vec_info
*lookup_dr (data_reference
*);
231 void move_dr (stmt_vec_info
, stmt_vec_info
);
232 void remove_stmt (stmt_vec_info
);
233 void replace_stmt (gimple_stmt_iterator
*, stmt_vec_info
, gimple
*);
235 /* The type of vectorization. */
238 /* Shared vectorizer state. */
239 vec_info_shared
*shared
;
241 /* The mapping of GIMPLE UID to stmt_vec_info. */
242 vec
<stmt_vec_info
> stmt_vec_infos
;
244 /* All SLP instances. */
245 auto_vec
<slp_instance
> slp_instances
;
247 /* Maps base addresses to an innermost_loop_behavior that gives the maximum
248 known alignment for that base. */
249 vec_base_alignments base_alignments
;
251 /* All interleaving chains of stores, represented by the first
252 stmt in the chain. */
253 auto_vec
<stmt_vec_info
> grouped_stores
;
255 /* Cost data used by the target cost model. */
256 void *target_cost_data
;
259 stmt_vec_info
new_stmt_vec_info (gimple
*stmt
);
260 void set_vinfo_for_stmt (gimple
*, stmt_vec_info
);
261 void free_stmt_vec_infos ();
262 void free_stmt_vec_info (stmt_vec_info
);
265 class _loop_vec_info
;
271 is_a_helper
<_loop_vec_info
*>::test (vec_info
*i
)
273 return i
->kind
== vec_info::loop
;
279 is_a_helper
<_bb_vec_info
*>::test (vec_info
*i
)
281 return i
->kind
== vec_info::bb
;
285 /* In general, we can divide the vector statements in a vectorized loop
286 into related groups ("rgroups") and say that for each rgroup there is
287 some nS such that the rgroup operates on nS values from one scalar
288 iteration followed by nS values from the next. That is, if VF is the
289 vectorization factor of the loop, the rgroup operates on a sequence:
291 (1,1) (1,2) ... (1,nS) (2,1) ... (2,nS) ... (VF,1) ... (VF,nS)
293 where (i,j) represents a scalar value with index j in a scalar
294 iteration with index i.
296 [ We use the term "rgroup" to emphasise that this grouping isn't
297 necessarily the same as the grouping of statements used elsewhere.
298 For example, if we implement a group of scalar loads using gather
299 loads, we'll use a separate gather load for each scalar load, and
300 thus each gather load will belong to its own rgroup. ]
302 In general this sequence will occupy nV vectors concatenated
303 together. If these vectors have nL lanes each, the total number
304 of scalar values N is given by:
306 N = nS * VF = nV * nL
308 None of nS, VF, nV and nL are required to be a power of 2. nS and nV
309 are compile-time constants but VF and nL can be variable (if the target
310 supports variable-length vectors).
312 In classical vectorization, each iteration of the vector loop would
313 handle exactly VF iterations of the original scalar loop. However,
314 in a fully-masked loop, a particular iteration of the vector loop
315 might handle fewer than VF iterations of the scalar loop. The vector
316 lanes that correspond to iterations of the scalar loop are said to be
317 "active" and the other lanes are said to be "inactive".
319 In a fully-masked loop, many rgroups need to be masked to ensure that
320 they have no effect for the inactive lanes. Each such rgroup needs a
321 sequence of booleans in the same order as above, but with each (i,j)
322 replaced by a boolean that indicates whether iteration i is active.
323 This sequence occupies nV vector masks that again have nL lanes each.
324 Thus the mask sequence as a whole consists of VF independent booleans
325 that are each repeated nS times.
327 We make the simplifying assumption that if a sequence of nV masks is
328 suitable for one (nS,nL) pair, we can reuse it for (nS/2,nL/2) by
329 VIEW_CONVERTing it. This holds for all current targets that support
330 fully-masked loops. For example, suppose the scalar loop is:
334 for (int i = 0; i < n; ++i)
336 f[i * 2 + 0] += 1.0f;
337 f[i * 2 + 1] += 2.0f;
341 and suppose that vectors have 256 bits. The vectorized f accesses
342 will belong to one rgroup and the vectorized d access to another:
344 f rgroup: nS = 2, nV = 1, nL = 8
345 d rgroup: nS = 1, nV = 1, nL = 4
348 [ In this simple example the rgroups do correspond to the normal
349 SLP grouping scheme. ]
351 If only the first three lanes are active, the masks we need are:
353 f rgroup: 1 1 | 1 1 | 1 1 | 0 0
354 d rgroup: 1 | 1 | 1 | 0
356 Here we can use a mask calculated for f's rgroup for d's, but not
359 Thus for each value of nV, it is enough to provide nV masks, with the
360 mask being calculated based on the highest nL (or, equivalently, based
361 on the highest nS) required by any rgroup with that nV. We therefore
362 represent the entire collection of masks as a two-level table, with the
363 first level being indexed by nV - 1 (since nV == 0 doesn't exist) and
364 the second being indexed by the mask index 0 <= i < nV. */
366 /* The masks needed by rgroups with nV vectors, according to the
367 description above. */
368 struct rgroup_masks
{
369 /* The largest nS for all rgroups that use these masks. */
370 unsigned int max_nscalars_per_iter
;
372 /* The type of mask to use, based on the highest nS recorded above. */
375 /* A vector of nV masks, in iteration order. */
379 typedef auto_vec
<rgroup_masks
> vec_loop_masks
;
381 /*-----------------------------------------------------------------*/
382 /* Info on vectorized loops. */
383 /*-----------------------------------------------------------------*/
384 typedef class _loop_vec_info
: public vec_info
{
386 _loop_vec_info (class loop
*, vec_info_shared
*);
389 /* The loop to which this info struct refers to. */
392 /* The loop basic blocks. */
395 /* Number of latch executions. */
397 /* Number of iterations. */
399 /* Number of iterations of the original loop. */
400 tree num_iters_unchanged
;
401 /* Condition under which this loop is analyzed and versioned. */
402 tree num_iters_assumptions
;
404 /* Threshold of number of iterations below which vectorization will not be
405 performed. It is calculated from MIN_PROFITABLE_ITERS and
406 PARAM_MIN_VECT_LOOP_BOUND. */
409 /* When applying loop versioning, the vector form should only be used
410 if the number of scalar iterations is >= this value, on top of all
411 the other requirements. Ignored when loop versioning is not being
413 poly_uint64 versioning_threshold
;
415 /* Unrolling factor */
416 poly_uint64 vectorization_factor
;
418 /* Maximum runtime vectorization factor, or MAX_VECTORIZATION_FACTOR
419 if there is no particular limit. */
420 unsigned HOST_WIDE_INT max_vectorization_factor
;
422 /* The masks that a fully-masked loop should use to avoid operating
423 on inactive scalars. */
424 vec_loop_masks masks
;
426 /* If we are using a loop mask to align memory addresses, this variable
427 contains the number of vector elements that we should skip in the
428 first iteration of the vector loop (i.e. the number of leading
429 elements that should be false in the first mask). */
430 tree mask_skip_niters
;
432 /* Type of the variables to use in the WHILE_ULT call for fully-masked
434 tree mask_compare_type
;
436 /* For #pragma omp simd if (x) loops the x expression. If constant 0,
437 the loop should not be vectorized, if constant non-zero, simd_if_cond
438 shouldn't be set and loop vectorized normally, if SSA_NAME, the loop
439 should be versioned on that condition, using scalar loop if the condition
440 is false and vectorized loop otherwise. */
443 /* Type of the IV to use in the WHILE_ULT call for fully-masked
447 /* Unknown DRs according to which loop was peeled. */
448 class dr_vec_info
*unaligned_dr
;
450 /* peeling_for_alignment indicates whether peeling for alignment will take
451 place, and what the peeling factor should be:
452 peeling_for_alignment = X means:
453 If X=0: Peeling for alignment will not be applied.
454 If X>0: Peel first X iterations.
455 If X=-1: Generate a runtime test to calculate the number of iterations
456 to be peeled, using the dataref recorded in the field
458 int peeling_for_alignment
;
460 /* The mask used to check the alignment of pointers or arrays. */
463 /* Data Dependence Relations defining address ranges that are candidates
464 for a run-time aliasing check. */
465 auto_vec
<ddr_p
> may_alias_ddrs
;
467 /* Data Dependence Relations defining address ranges together with segment
468 lengths from which the run-time aliasing check is built. */
469 auto_vec
<dr_with_seg_len_pair_t
> comp_alias_ddrs
;
471 /* Check that the addresses of each pair of objects is unequal. */
472 auto_vec
<vec_object_pair
> check_unequal_addrs
;
474 /* List of values that are required to be nonzero. This is used to check
475 whether things like "x[i * n] += 1;" are safe and eventually gets added
476 to the checks for lower bounds below. */
477 auto_vec
<tree
> check_nonzero
;
479 /* List of values that need to be checked for a minimum value. */
480 auto_vec
<vec_lower_bound
> lower_bounds
;
482 /* Statements in the loop that have data references that are candidates for a
483 runtime (loop versioning) misalignment check. */
484 auto_vec
<stmt_vec_info
> may_misalign_stmts
;
486 /* Reduction cycles detected in the loop. Used in loop-aware SLP. */
487 auto_vec
<stmt_vec_info
> reductions
;
489 /* All reduction chains in the loop, represented by the first
490 stmt in the chain. */
491 auto_vec
<stmt_vec_info
> reduction_chains
;
493 /* Cost vector for a single scalar iteration. */
494 auto_vec
<stmt_info_for_cost
> scalar_cost_vec
;
496 /* Map of IV base/step expressions to inserted name in the preheader. */
497 hash_map
<tree_operand_hash
, tree
> *ivexpr_map
;
499 /* Map of OpenMP "omp simd array" scan variables to corresponding
500 rhs of the store of the initializer. */
501 hash_map
<tree
, tree
> *scan_map
;
503 /* The unrolling factor needed to SLP the loop. In case of that pure SLP is
504 applied to the loop, i.e., no unrolling is needed, this is 1. */
505 poly_uint64 slp_unrolling_factor
;
507 /* Cost of a single scalar iteration. */
508 int single_scalar_iteration_cost
;
510 /* Is the loop vectorizable? */
513 /* Records whether we still have the option of using a fully-masked loop. */
514 bool can_fully_mask_p
;
516 /* True if have decided to use a fully-masked loop. */
519 /* When we have grouped data accesses with gaps, we may introduce invalid
520 memory accesses. We peel the last iteration of the loop to prevent
522 bool peeling_for_gaps
;
524 /* When the number of iterations is not a multiple of the vector size
525 we need to peel off iterations at the end to form an epilogue loop. */
526 bool peeling_for_niter
;
528 /* Reductions are canonicalized so that the last operand is the reduction
529 operand. If this places a constant into RHS1, this decanonicalizes
530 GIMPLE for other phases, so we must track when this has occurred and
532 bool operands_swapped
;
534 /* True if there are no loop carried data dependencies in the loop.
535 If loop->safelen <= 1, then this is always true, either the loop
536 didn't have any loop carried data dependencies, or the loop is being
537 vectorized guarded with some runtime alias checks, or couldn't
538 be vectorized at all, but then this field shouldn't be used.
539 For loop->safelen >= 2, the user has asserted that there are no
540 backward dependencies, but there still could be loop carried forward
541 dependencies in such loops. This flag will be false if normal
542 vectorizer data dependency analysis would fail or require versioning
543 for alias, but because of loop->safelen >= 2 it has been vectorized
544 even without versioning for alias. E.g. in:
546 for (int i = 0; i < m; i++)
548 (or #pragma simd or #pragma ivdep) we can vectorize this and it will
549 DTRT even for k > 0 && k < m, but without safelen we would not
550 vectorize this, so this field would be false. */
551 bool no_data_dependencies
;
553 /* Mark loops having masked stores. */
556 /* Queued scaling factor for the scalar loop. */
557 profile_probability scalar_loop_scaling
;
559 /* If if-conversion versioned this loop before conversion, this is the
560 loop version without if-conversion. */
561 class loop
*scalar_loop
;
563 /* For loops being epilogues of already vectorized loops
564 this points to the original vectorized loop. Otherwise NULL. */
565 _loop_vec_info
*orig_loop_info
;
569 /* Access Functions. */
570 #define LOOP_VINFO_LOOP(L) (L)->loop
571 #define LOOP_VINFO_BBS(L) (L)->bbs
572 #define LOOP_VINFO_NITERSM1(L) (L)->num_itersm1
573 #define LOOP_VINFO_NITERS(L) (L)->num_iters
574 /* Since LOOP_VINFO_NITERS and LOOP_VINFO_NITERSM1 can change after
575 prologue peeling retain total unchanged scalar loop iterations for
577 #define LOOP_VINFO_NITERS_UNCHANGED(L) (L)->num_iters_unchanged
578 #define LOOP_VINFO_NITERS_ASSUMPTIONS(L) (L)->num_iters_assumptions
579 #define LOOP_VINFO_COST_MODEL_THRESHOLD(L) (L)->th
580 #define LOOP_VINFO_VERSIONING_THRESHOLD(L) (L)->versioning_threshold
581 #define LOOP_VINFO_VECTORIZABLE_P(L) (L)->vectorizable
582 #define LOOP_VINFO_CAN_FULLY_MASK_P(L) (L)->can_fully_mask_p
583 #define LOOP_VINFO_FULLY_MASKED_P(L) (L)->fully_masked_p
584 #define LOOP_VINFO_VECT_FACTOR(L) (L)->vectorization_factor
585 #define LOOP_VINFO_MAX_VECT_FACTOR(L) (L)->max_vectorization_factor
586 #define LOOP_VINFO_MASKS(L) (L)->masks
587 #define LOOP_VINFO_MASK_SKIP_NITERS(L) (L)->mask_skip_niters
588 #define LOOP_VINFO_MASK_COMPARE_TYPE(L) (L)->mask_compare_type
589 #define LOOP_VINFO_MASK_IV_TYPE(L) (L)->iv_type
590 #define LOOP_VINFO_PTR_MASK(L) (L)->ptr_mask
591 #define LOOP_VINFO_LOOP_NEST(L) (L)->shared->loop_nest
592 #define LOOP_VINFO_DATAREFS(L) (L)->shared->datarefs
593 #define LOOP_VINFO_DDRS(L) (L)->shared->ddrs
594 #define LOOP_VINFO_INT_NITERS(L) (TREE_INT_CST_LOW ((L)->num_iters))
595 #define LOOP_VINFO_PEELING_FOR_ALIGNMENT(L) (L)->peeling_for_alignment
596 #define LOOP_VINFO_UNALIGNED_DR(L) (L)->unaligned_dr
597 #define LOOP_VINFO_MAY_MISALIGN_STMTS(L) (L)->may_misalign_stmts
598 #define LOOP_VINFO_MAY_ALIAS_DDRS(L) (L)->may_alias_ddrs
599 #define LOOP_VINFO_COMP_ALIAS_DDRS(L) (L)->comp_alias_ddrs
600 #define LOOP_VINFO_CHECK_UNEQUAL_ADDRS(L) (L)->check_unequal_addrs
601 #define LOOP_VINFO_CHECK_NONZERO(L) (L)->check_nonzero
602 #define LOOP_VINFO_LOWER_BOUNDS(L) (L)->lower_bounds
603 #define LOOP_VINFO_GROUPED_STORES(L) (L)->grouped_stores
604 #define LOOP_VINFO_SLP_INSTANCES(L) (L)->slp_instances
605 #define LOOP_VINFO_SLP_UNROLLING_FACTOR(L) (L)->slp_unrolling_factor
606 #define LOOP_VINFO_REDUCTIONS(L) (L)->reductions
607 #define LOOP_VINFO_REDUCTION_CHAINS(L) (L)->reduction_chains
608 #define LOOP_VINFO_TARGET_COST_DATA(L) (L)->target_cost_data
609 #define LOOP_VINFO_PEELING_FOR_GAPS(L) (L)->peeling_for_gaps
610 #define LOOP_VINFO_OPERANDS_SWAPPED(L) (L)->operands_swapped
611 #define LOOP_VINFO_PEELING_FOR_NITER(L) (L)->peeling_for_niter
612 #define LOOP_VINFO_NO_DATA_DEPENDENCIES(L) (L)->no_data_dependencies
613 #define LOOP_VINFO_SCALAR_LOOP(L) (L)->scalar_loop
614 #define LOOP_VINFO_SCALAR_LOOP_SCALING(L) (L)->scalar_loop_scaling
615 #define LOOP_VINFO_HAS_MASK_STORE(L) (L)->has_mask_store
616 #define LOOP_VINFO_SCALAR_ITERATION_COST(L) (L)->scalar_cost_vec
617 #define LOOP_VINFO_SINGLE_SCALAR_ITERATION_COST(L) (L)->single_scalar_iteration_cost
618 #define LOOP_VINFO_ORIG_LOOP_INFO(L) (L)->orig_loop_info
619 #define LOOP_VINFO_SIMD_IF_COND(L) (L)->simd_if_cond
621 #define LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT(L) \
622 ((L)->may_misalign_stmts.length () > 0)
623 #define LOOP_REQUIRES_VERSIONING_FOR_ALIAS(L) \
624 ((L)->comp_alias_ddrs.length () > 0 \
625 || (L)->check_unequal_addrs.length () > 0 \
626 || (L)->lower_bounds.length () > 0)
627 #define LOOP_REQUIRES_VERSIONING_FOR_NITERS(L) \
628 (LOOP_VINFO_NITERS_ASSUMPTIONS (L))
629 #define LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND(L) \
630 (LOOP_VINFO_SIMD_IF_COND (L))
631 #define LOOP_REQUIRES_VERSIONING(L) \
632 (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (L) \
633 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (L) \
634 || LOOP_REQUIRES_VERSIONING_FOR_NITERS (L) \
635 || LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND (L))
637 #define LOOP_VINFO_NITERS_KNOWN_P(L) \
638 (tree_fits_shwi_p ((L)->num_iters) && tree_to_shwi ((L)->num_iters) > 0)
640 #define LOOP_VINFO_EPILOGUE_P(L) \
641 (LOOP_VINFO_ORIG_LOOP_INFO (L) != NULL)
643 #define LOOP_VINFO_ORIG_MAX_VECT_FACTOR(L) \
644 (LOOP_VINFO_MAX_VECT_FACTOR (LOOP_VINFO_ORIG_LOOP_INFO (L)))
646 /* Wrapper for loop_vec_info, for tracking success/failure, where a non-NULL
647 value signifies success, and a NULL value signifies failure, supporting
648 propagating an opt_problem * describing the failure back up the call
650 typedef opt_pointer_wrapper
<loop_vec_info
> opt_loop_vec_info
;
652 static inline loop_vec_info
653 loop_vec_info_for_loop (class loop
*loop
)
655 return (loop_vec_info
) loop
->aux
;
658 typedef class _bb_vec_info
: public vec_info
661 _bb_vec_info (gimple_stmt_iterator
, gimple_stmt_iterator
, vec_info_shared
*);
665 gimple_stmt_iterator region_begin
;
666 gimple_stmt_iterator region_end
;
669 #define BB_VINFO_BB(B) (B)->bb
670 #define BB_VINFO_GROUPED_STORES(B) (B)->grouped_stores
671 #define BB_VINFO_SLP_INSTANCES(B) (B)->slp_instances
672 #define BB_VINFO_DATAREFS(B) (B)->shared->datarefs
673 #define BB_VINFO_DDRS(B) (B)->shared->ddrs
674 #define BB_VINFO_TARGET_COST_DATA(B) (B)->target_cost_data
676 static inline bb_vec_info
677 vec_info_for_bb (basic_block bb
)
679 return (bb_vec_info
) bb
->aux
;
682 /*-----------------------------------------------------------------*/
683 /* Info on vectorized defs. */
684 /*-----------------------------------------------------------------*/
685 enum stmt_vec_info_type
{
686 undef_vec_info_type
= 0,
692 call_simd_clone_vec_info_type
,
693 assignment_vec_info_type
,
694 condition_vec_info_type
,
695 comparison_vec_info_type
,
698 type_promotion_vec_info_type
,
699 type_demotion_vec_info_type
,
700 type_conversion_vec_info_type
,
701 loop_exit_ctrl_vec_info_type
704 /* Indicates whether/how a variable is used in the scope of loop/basic
707 vect_unused_in_scope
= 0,
709 /* The def is only used outside the loop. */
711 /* The def is in the inner loop, and the use is in the outer loop, and the
712 use is a reduction stmt. */
713 vect_used_in_outer_by_reduction
,
714 /* The def is in the inner loop, and the use is in the outer loop (and is
715 not part of reduction). */
718 /* defs that feed computations that end up (only) in a reduction. These
719 defs may be used by non-reduction stmts, but eventually, any
720 computations/values that are affected by these defs are used to compute
721 a reduction (i.e. don't get stored to memory, for example). We use this
722 to identify computations that we can change the order in which they are
724 vect_used_by_reduction
,
729 /* The type of vectorization that can be applied to the stmt: regular loop-based
730 vectorization; pure SLP - the stmt is a part of SLP instances and does not
731 have uses outside SLP instances; or hybrid SLP and loop-based - the stmt is
732 a part of SLP instance and also must be loop-based vectorized, since it has
733 uses outside SLP sequences.
735 In the loop context the meanings of pure and hybrid SLP are slightly
736 different. By saying that pure SLP is applied to the loop, we mean that we
737 exploit only intra-iteration parallelism in the loop; i.e., the loop can be
738 vectorized without doing any conceptual unrolling, cause we don't pack
739 together stmts from different iterations, only within a single iteration.
740 Loop hybrid SLP means that we exploit both intra-iteration and
741 inter-iteration parallelism (e.g., number of elements in the vector is 4
742 and the slp-group-size is 2, in which case we don't have enough parallelism
743 within an iteration, so we obtain the rest of the parallelism from subsequent
744 iterations by unrolling the loop by 2). */
751 /* Says whether a statement is a load, a store of a vectorized statement
752 result, or a store of an invariant value. */
753 enum vec_load_store_type
{
759 /* Describes how we're going to vectorize an individual load or store,
760 or a group of loads or stores. */
761 enum vect_memory_access_type
{
762 /* An access to an invariant address. This is used only for loads. */
765 /* A simple contiguous access. */
768 /* A contiguous access that goes down in memory rather than up,
769 with no additional permutation. This is used only for stores
771 VMAT_CONTIGUOUS_DOWN
,
773 /* A simple contiguous access in which the elements need to be permuted
774 after loading or before storing. Only used for loop vectorization;
775 SLP uses separate permutes. */
776 VMAT_CONTIGUOUS_PERMUTE
,
778 /* A simple contiguous access in which the elements need to be reversed
779 after loading or before storing. */
780 VMAT_CONTIGUOUS_REVERSE
,
782 /* An access that uses IFN_LOAD_LANES or IFN_STORE_LANES. */
783 VMAT_LOAD_STORE_LANES
,
785 /* An access in which each scalar element is loaded or stored
789 /* A hybrid of VMAT_CONTIGUOUS and VMAT_ELEMENTWISE, used for grouped
790 SLP accesses. Each unrolled iteration uses a contiguous load
791 or store for the whole group, but the groups from separate iterations
792 are combined in the same way as for VMAT_ELEMENTWISE. */
795 /* The access uses gather loads or scatter stores. */
801 /* The data reference itself. */
803 /* The statement that contains the data reference. */
805 /* The misalignment in bytes of the reference, or -1 if not known. */
807 /* The byte alignment that we'd ideally like the reference to have,
808 and the value that misalignment is measured against. */
809 poly_uint64 target_alignment
;
810 /* If true the alignment of base_decl needs to be increased. */
811 bool base_misaligned
;
815 typedef struct data_reference
*dr_p
;
817 class _stmt_vec_info
{
820 enum stmt_vec_info_type type
;
822 /* Indicates whether this stmts is part of a computation whose result is
823 used outside the loop. */
826 /* Stmt is part of some pattern (computation idiom) */
829 /* True if the statement was created during pattern recognition as
830 part of the replacement for RELATED_STMT. This implies that the
831 statement isn't part of any basic block, although for convenience
832 its gimple_bb is the same as for RELATED_STMT. */
835 /* Is this statement vectorizable or should it be skipped in (partial)
839 /* The stmt to which this info struct refers to. */
842 /* The vec_info with respect to which STMT is vectorized. */
845 /* The vector type to be used for the LHS of this statement. */
848 /* The vectorized version of the stmt. */
849 stmt_vec_info vectorized_stmt
;
852 /* The following is relevant only for stmts that contain a non-scalar
853 data-ref (array/pointer/struct access). A GIMPLE stmt is expected to have
854 at most one such data-ref. */
858 /* Information about the data-ref relative to this loop
859 nest (the loop that is being considered for vectorization). */
860 innermost_loop_behavior dr_wrt_vec_loop
;
862 /* For loop PHI nodes, the base and evolution part of it. This makes sure
863 this information is still available in vect_update_ivs_after_vectorizer
864 where we may not be able to re-analyze the PHI nodes evolution as
865 peeling for the prologue loop can make it unanalyzable. The evolution
866 part is still correct after peeling, but the base may have changed from
868 tree loop_phi_evolution_base_unchanged
;
869 tree loop_phi_evolution_part
;
871 /* Used for various bookkeeping purposes, generally holding a pointer to
872 some other stmt S that is in some way "related" to this stmt.
873 Current use of this field is:
874 If this stmt is part of a pattern (i.e. the field 'in_pattern_p' is
875 true): S is the "pattern stmt" that represents (and replaces) the
876 sequence of stmts that constitutes the pattern. Similarly, the
877 related_stmt of the "pattern stmt" points back to this stmt (which is
878 the last stmt in the original sequence of stmts that constitutes the
880 stmt_vec_info related_stmt
;
882 /* Used to keep a sequence of def stmts of a pattern stmt if such exists.
883 The sequence is attached to the original statement rather than the
884 pattern statement. */
885 gimple_seq pattern_def_seq
;
887 /* List of datarefs that are known to have the same alignment as the dataref
889 vec
<dr_p
> same_align_refs
;
891 /* Selected SIMD clone's function info. First vector element
892 is SIMD clone's function decl, followed by a pair of trees (base + step)
893 for linear arguments (pair of NULLs for other arguments). */
894 vec
<tree
> simd_clone_info
;
896 /* Classify the def of this stmt. */
897 enum vect_def_type def_type
;
899 /* Whether the stmt is SLPed, loop-based vectorized, or both. */
900 enum slp_vect_type slp_type
;
902 /* Interleaving and reduction chains info. */
903 /* First element in the group. */
904 stmt_vec_info first_element
;
905 /* Pointer to the next element in the group. */
906 stmt_vec_info next_element
;
907 /* The size of the group. */
909 /* For stores, number of stores from this group seen. We vectorize the last
911 unsigned int store_count
;
912 /* For loads only, the gap from the previous load. For consecutive loads, GAP
916 /* The minimum negative dependence distance this stmt participates in
918 unsigned int min_neg_dist
;
920 /* Not all stmts in the loop need to be vectorized. e.g, the increment
921 of the loop induction variable and computation of array indexes. relevant
922 indicates whether the stmt needs to be vectorized. */
923 enum vect_relevant relevant
;
925 /* For loads if this is a gather, for stores if this is a scatter. */
926 bool gather_scatter_p
;
928 /* True if this is an access with loop-invariant stride. */
931 /* For both loads and stores. */
932 unsigned simd_lane_access_p
: 3;
934 /* Classifies how the load or store is going to be implemented
935 for loop vectorization. */
936 vect_memory_access_type memory_access_type
;
938 /* For reduction loops, this is the type of reduction. */
939 enum vect_reduction_type v_reduc_type
;
941 /* For CONST_COND_REDUCTION, record the reduc code. */
942 enum tree_code const_cond_reduc_code
;
944 /* On a reduction PHI the reduction type as detected by
945 vect_force_simple_reduction. */
946 enum vect_reduction_type reduc_type
;
948 /* On a reduction PHI the def returned by vect_force_simple_reduction.
949 On the def returned by vect_force_simple_reduction the
950 corresponding PHI. */
951 stmt_vec_info reduc_def
;
953 /* The number of scalar stmt references from active SLP instances. */
954 unsigned int num_slp_uses
;
956 /* If nonzero, the lhs of the statement could be truncated to this
957 many bits without affecting any users of the result. */
958 unsigned int min_output_precision
;
960 /* If nonzero, all non-boolean input operands have the same precision,
961 and they could each be truncated to this many bits without changing
963 unsigned int min_input_precision
;
965 /* If OPERATION_BITS is nonzero, the statement could be performed on
966 an integer with the sign and number of bits given by OPERATION_SIGN
967 and OPERATION_BITS without changing the result. */
968 unsigned int operation_precision
;
969 signop operation_sign
;
971 /* True if this is only suitable for SLP vectorization. */
972 bool slp_vect_only_p
;
975 /* Information about a gather/scatter call. */
976 struct gather_scatter_info
{
977 /* The internal function to use for the gather/scatter operation,
978 or IFN_LAST if a built-in function should be used instead. */
981 /* The FUNCTION_DECL for the built-in gather/scatter function,
982 or null if an internal function should be used instead. */
985 /* The loop-invariant base value. */
988 /* The original scalar offset, which is a non-loop-invariant SSA_NAME. */
991 /* Each offset element should be multiplied by this amount before
992 being added to the base. */
995 /* The definition type for the vectorized offset. */
996 enum vect_def_type offset_dt
;
998 /* The type of the vectorized offset. */
1001 /* The type of the scalar elements after loading or before storing. */
1004 /* The type of the scalar elements being loaded or stored. */
1008 /* Access Functions. */
1009 #define STMT_VINFO_TYPE(S) (S)->type
1010 #define STMT_VINFO_STMT(S) (S)->stmt
1011 inline loop_vec_info
1012 STMT_VINFO_LOOP_VINFO (stmt_vec_info stmt_vinfo
)
1014 if (loop_vec_info loop_vinfo
= dyn_cast
<loop_vec_info
> (stmt_vinfo
->vinfo
))
1019 STMT_VINFO_BB_VINFO (stmt_vec_info stmt_vinfo
)
1021 if (bb_vec_info bb_vinfo
= dyn_cast
<bb_vec_info
> (stmt_vinfo
->vinfo
))
1025 #define STMT_VINFO_RELEVANT(S) (S)->relevant
1026 #define STMT_VINFO_LIVE_P(S) (S)->live
1027 #define STMT_VINFO_VECTYPE(S) (S)->vectype
1028 #define STMT_VINFO_VEC_STMT(S) (S)->vectorized_stmt
1029 #define STMT_VINFO_VECTORIZABLE(S) (S)->vectorizable
1030 #define STMT_VINFO_DATA_REF(S) ((S)->dr_aux.dr + 0)
1031 #define STMT_VINFO_GATHER_SCATTER_P(S) (S)->gather_scatter_p
1032 #define STMT_VINFO_STRIDED_P(S) (S)->strided_p
1033 #define STMT_VINFO_MEMORY_ACCESS_TYPE(S) (S)->memory_access_type
1034 #define STMT_VINFO_SIMD_LANE_ACCESS_P(S) (S)->simd_lane_access_p
1035 #define STMT_VINFO_VEC_REDUCTION_TYPE(S) (S)->v_reduc_type
1036 #define STMT_VINFO_VEC_CONST_COND_REDUC_CODE(S) (S)->const_cond_reduc_code
1038 #define STMT_VINFO_DR_WRT_VEC_LOOP(S) (S)->dr_wrt_vec_loop
1039 #define STMT_VINFO_DR_BASE_ADDRESS(S) (S)->dr_wrt_vec_loop.base_address
1040 #define STMT_VINFO_DR_INIT(S) (S)->dr_wrt_vec_loop.init
1041 #define STMT_VINFO_DR_OFFSET(S) (S)->dr_wrt_vec_loop.offset
1042 #define STMT_VINFO_DR_STEP(S) (S)->dr_wrt_vec_loop.step
1043 #define STMT_VINFO_DR_BASE_ALIGNMENT(S) (S)->dr_wrt_vec_loop.base_alignment
1044 #define STMT_VINFO_DR_BASE_MISALIGNMENT(S) \
1045 (S)->dr_wrt_vec_loop.base_misalignment
1046 #define STMT_VINFO_DR_OFFSET_ALIGNMENT(S) \
1047 (S)->dr_wrt_vec_loop.offset_alignment
1048 #define STMT_VINFO_DR_STEP_ALIGNMENT(S) \
1049 (S)->dr_wrt_vec_loop.step_alignment
1051 #define STMT_VINFO_DR_INFO(S) \
1052 (gcc_checking_assert ((S)->dr_aux.stmt == (S)), &(S)->dr_aux)
1054 #define STMT_VINFO_IN_PATTERN_P(S) (S)->in_pattern_p
1055 #define STMT_VINFO_RELATED_STMT(S) (S)->related_stmt
1056 #define STMT_VINFO_PATTERN_DEF_SEQ(S) (S)->pattern_def_seq
1057 #define STMT_VINFO_SAME_ALIGN_REFS(S) (S)->same_align_refs
1058 #define STMT_VINFO_SIMD_CLONE_INFO(S) (S)->simd_clone_info
1059 #define STMT_VINFO_DEF_TYPE(S) (S)->def_type
1060 #define STMT_VINFO_GROUPED_ACCESS(S) \
1061 ((S)->dr_aux.dr && DR_GROUP_FIRST_ELEMENT(S))
1062 #define STMT_VINFO_LOOP_PHI_EVOLUTION_BASE_UNCHANGED(S) (S)->loop_phi_evolution_base_unchanged
1063 #define STMT_VINFO_LOOP_PHI_EVOLUTION_PART(S) (S)->loop_phi_evolution_part
1064 #define STMT_VINFO_MIN_NEG_DIST(S) (S)->min_neg_dist
1065 #define STMT_VINFO_NUM_SLP_USES(S) (S)->num_slp_uses
1066 #define STMT_VINFO_REDUC_TYPE(S) (S)->reduc_type
1067 #define STMT_VINFO_REDUC_DEF(S) (S)->reduc_def
1068 #define STMT_VINFO_SLP_VECT_ONLY(S) (S)->slp_vect_only_p
1070 #define DR_GROUP_FIRST_ELEMENT(S) \
1071 (gcc_checking_assert ((S)->dr_aux.dr), (S)->first_element)
1072 #define DR_GROUP_NEXT_ELEMENT(S) \
1073 (gcc_checking_assert ((S)->dr_aux.dr), (S)->next_element)
1074 #define DR_GROUP_SIZE(S) \
1075 (gcc_checking_assert ((S)->dr_aux.dr), (S)->size)
1076 #define DR_GROUP_STORE_COUNT(S) \
1077 (gcc_checking_assert ((S)->dr_aux.dr), (S)->store_count)
1078 #define DR_GROUP_GAP(S) \
1079 (gcc_checking_assert ((S)->dr_aux.dr), (S)->gap)
1081 #define REDUC_GROUP_FIRST_ELEMENT(S) \
1082 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->first_element)
1083 #define REDUC_GROUP_NEXT_ELEMENT(S) \
1084 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->next_element)
1085 #define REDUC_GROUP_SIZE(S) \
1086 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->size)
1088 #define STMT_VINFO_RELEVANT_P(S) ((S)->relevant != vect_unused_in_scope)
1090 #define HYBRID_SLP_STMT(S) ((S)->slp_type == hybrid)
1091 #define PURE_SLP_STMT(S) ((S)->slp_type == pure_slp)
1092 #define STMT_SLP_TYPE(S) (S)->slp_type
1094 #define VECT_MAX_COST 1000
1096 /* The maximum number of intermediate steps required in multi-step type
1098 #define MAX_INTERM_CVT_STEPS 3
1100 #define MAX_VECTORIZATION_FACTOR INT_MAX
1102 /* Nonzero if TYPE represents a (scalar) boolean type or type
1103 in the middle-end compatible with it (unsigned precision 1 integral
1104 types). Used to determine which types should be vectorized as
1105 VECTOR_BOOLEAN_TYPE_P. */
1107 #define VECT_SCALAR_BOOLEAN_TYPE_P(TYPE) \
1108 (TREE_CODE (TYPE) == BOOLEAN_TYPE \
1109 || ((TREE_CODE (TYPE) == INTEGER_TYPE \
1110 || TREE_CODE (TYPE) == ENUMERAL_TYPE) \
1111 && TYPE_PRECISION (TYPE) == 1 \
1112 && TYPE_UNSIGNED (TYPE)))
1115 nested_in_vect_loop_p (class loop
*loop
, stmt_vec_info stmt_info
)
1118 && (loop
->inner
== (gimple_bb (stmt_info
->stmt
))->loop_father
));
1121 /* Return TRUE if a statement represented by STMT_INFO is a part of a
1125 is_pattern_stmt_p (stmt_vec_info stmt_info
)
1127 return stmt_info
->pattern_stmt_p
;
1130 /* If STMT_INFO is a pattern statement, return the statement that it
1131 replaces, otherwise return STMT_INFO itself. */
1133 inline stmt_vec_info
1134 vect_orig_stmt (stmt_vec_info stmt_info
)
1136 if (is_pattern_stmt_p (stmt_info
))
1137 return STMT_VINFO_RELATED_STMT (stmt_info
);
1141 /* Return the later statement between STMT1_INFO and STMT2_INFO. */
1143 static inline stmt_vec_info
1144 get_later_stmt (stmt_vec_info stmt1_info
, stmt_vec_info stmt2_info
)
1146 if (gimple_uid (vect_orig_stmt (stmt1_info
)->stmt
)
1147 > gimple_uid (vect_orig_stmt (stmt2_info
)->stmt
))
1153 /* If STMT_INFO has been replaced by a pattern statement, return the
1154 replacement statement, otherwise return STMT_INFO itself. */
1156 inline stmt_vec_info
1157 vect_stmt_to_vectorize (stmt_vec_info stmt_info
)
1159 if (STMT_VINFO_IN_PATTERN_P (stmt_info
))
1160 return STMT_VINFO_RELATED_STMT (stmt_info
);
1164 /* Return true if BB is a loop header. */
1167 is_loop_header_bb_p (basic_block bb
)
1169 if (bb
== (bb
->loop_father
)->header
)
1171 gcc_checking_assert (EDGE_COUNT (bb
->preds
) == 1);
1175 /* Return pow2 (X). */
1182 for (i
= 0; i
< x
; i
++)
1188 /* Alias targetm.vectorize.builtin_vectorization_cost. */
1191 builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost
,
1192 tree vectype
, int misalign
)
1194 return targetm
.vectorize
.builtin_vectorization_cost (type_of_cost
,
1198 /* Get cost by calling cost target builtin. */
1201 int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost
)
1203 return builtin_vectorization_cost (type_of_cost
, NULL
, 0);
1206 /* Alias targetm.vectorize.init_cost. */
1208 static inline void *
1209 init_cost (class loop
*loop_info
)
1211 return targetm
.vectorize
.init_cost (loop_info
);
1214 extern void dump_stmt_cost (FILE *, void *, int, enum vect_cost_for_stmt
,
1215 stmt_vec_info
, int, unsigned,
1216 enum vect_cost_model_location
);
1218 /* Alias targetm.vectorize.add_stmt_cost. */
1220 static inline unsigned
1221 add_stmt_cost (void *data
, int count
, enum vect_cost_for_stmt kind
,
1222 stmt_vec_info stmt_info
, int misalign
,
1223 enum vect_cost_model_location where
)
1225 unsigned cost
= targetm
.vectorize
.add_stmt_cost (data
, count
, kind
,
1226 stmt_info
, misalign
, where
);
1227 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1228 dump_stmt_cost (dump_file
, data
, count
, kind
, stmt_info
, misalign
,
1233 /* Alias targetm.vectorize.finish_cost. */
1236 finish_cost (void *data
, unsigned *prologue_cost
,
1237 unsigned *body_cost
, unsigned *epilogue_cost
)
1239 targetm
.vectorize
.finish_cost (data
, prologue_cost
, body_cost
, epilogue_cost
);
1242 /* Alias targetm.vectorize.destroy_cost_data. */
1245 destroy_cost_data (void *data
)
1247 targetm
.vectorize
.destroy_cost_data (data
);
1251 add_stmt_costs (void *data
, stmt_vector_for_cost
*cost_vec
)
1253 stmt_info_for_cost
*cost
;
1255 FOR_EACH_VEC_ELT (*cost_vec
, i
, cost
)
1256 add_stmt_cost (data
, cost
->count
, cost
->kind
, cost
->stmt_info
,
1257 cost
->misalign
, cost
->where
);
1260 /*-----------------------------------------------------------------*/
1261 /* Info on data references alignment. */
1262 /*-----------------------------------------------------------------*/
1263 #define DR_MISALIGNMENT_UNKNOWN (-1)
1264 #define DR_MISALIGNMENT_UNINITIALIZED (-2)
1267 set_dr_misalignment (dr_vec_info
*dr_info
, int val
)
1269 dr_info
->misalignment
= val
;
1273 dr_misalignment (dr_vec_info
*dr_info
)
1275 int misalign
= dr_info
->misalignment
;
1276 gcc_assert (misalign
!= DR_MISALIGNMENT_UNINITIALIZED
);
1280 /* Reflects actual alignment of first access in the vectorized loop,
1281 taking into account peeling/versioning if applied. */
1282 #define DR_MISALIGNMENT(DR) dr_misalignment (DR)
1283 #define SET_DR_MISALIGNMENT(DR, VAL) set_dr_misalignment (DR, VAL)
1285 /* Only defined once DR_MISALIGNMENT is defined. */
1286 #define DR_TARGET_ALIGNMENT(DR) ((DR)->target_alignment)
1288 /* Return true if data access DR_INFO is aligned to its target alignment
1289 (which may be less than a full vector). */
1292 aligned_access_p (dr_vec_info
*dr_info
)
1294 return (DR_MISALIGNMENT (dr_info
) == 0);
1297 /* Return TRUE if the alignment of the data access is known, and FALSE
1301 known_alignment_for_access_p (dr_vec_info
*dr_info
)
1303 return (DR_MISALIGNMENT (dr_info
) != DR_MISALIGNMENT_UNKNOWN
);
1306 /* Return the minimum alignment in bytes that the vectorized version
1307 of DR_INFO is guaranteed to have. */
1309 static inline unsigned int
1310 vect_known_alignment_in_bytes (dr_vec_info
*dr_info
)
1312 if (DR_MISALIGNMENT (dr_info
) == DR_MISALIGNMENT_UNKNOWN
)
1313 return TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr_info
->dr
)));
1314 if (DR_MISALIGNMENT (dr_info
) == 0)
1315 return known_alignment (DR_TARGET_ALIGNMENT (dr_info
));
1316 return DR_MISALIGNMENT (dr_info
) & -DR_MISALIGNMENT (dr_info
);
1319 /* Return the behavior of DR_INFO with respect to the vectorization context
1320 (which for outer loop vectorization might not be the behavior recorded
1321 in DR_INFO itself). */
1323 static inline innermost_loop_behavior
*
1324 vect_dr_behavior (dr_vec_info
*dr_info
)
1326 stmt_vec_info stmt_info
= dr_info
->stmt
;
1327 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
1328 if (loop_vinfo
== NULL
1329 || !nested_in_vect_loop_p (LOOP_VINFO_LOOP (loop_vinfo
), stmt_info
))
1330 return &DR_INNERMOST (dr_info
->dr
);
1332 return &STMT_VINFO_DR_WRT_VEC_LOOP (stmt_info
);
1335 /* Return true if the vect cost model is unlimited. */
1337 unlimited_cost_model (loop_p loop
)
1339 if (loop
!= NULL
&& loop
->force_vectorize
1340 && flag_simd_cost_model
!= VECT_COST_MODEL_DEFAULT
)
1341 return flag_simd_cost_model
== VECT_COST_MODEL_UNLIMITED
;
1342 return (flag_vect_cost_model
== VECT_COST_MODEL_UNLIMITED
);
1345 /* Return true if the loop described by LOOP_VINFO is fully-masked and
1346 if the first iteration should use a partial mask in order to achieve
1350 vect_use_loop_mask_for_alignment_p (loop_vec_info loop_vinfo
)
1352 return (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo
)
1353 && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo
));
1356 /* Return the number of vectors of type VECTYPE that are needed to get
1357 NUNITS elements. NUNITS should be based on the vectorization factor,
1358 so it is always a known multiple of the number of elements in VECTYPE. */
1360 static inline unsigned int
1361 vect_get_num_vectors (poly_uint64 nunits
, tree vectype
)
1363 return exact_div (nunits
, TYPE_VECTOR_SUBPARTS (vectype
)).to_constant ();
1366 /* Return the number of copies needed for loop vectorization when
1367 a statement operates on vectors of type VECTYPE. This is the
1368 vectorization factor divided by the number of elements in
1369 VECTYPE and is always known at compile time. */
1371 static inline unsigned int
1372 vect_get_num_copies (loop_vec_info loop_vinfo
, tree vectype
)
1374 return vect_get_num_vectors (LOOP_VINFO_VECT_FACTOR (loop_vinfo
), vectype
);
1377 /* Update maximum unit count *MAX_NUNITS so that it accounts for
1378 the number of units in vector type VECTYPE. *MAX_NUNITS can be 1
1379 if we haven't yet recorded any vector types. */
1382 vect_update_max_nunits (poly_uint64
*max_nunits
, tree vectype
)
1384 /* All unit counts have the form current_vector_size * X for some
1385 rational X, so two unit sizes must have a common multiple.
1386 Everything is a multiple of the initial value of 1. */
1387 poly_uint64 nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
1388 *max_nunits
= force_common_multiple (*max_nunits
, nunits
);
1391 /* Return the vectorization factor that should be used for costing
1392 purposes while vectorizing the loop described by LOOP_VINFO.
1393 Pick a reasonable estimate if the vectorization factor isn't
1394 known at compile time. */
1396 static inline unsigned int
1397 vect_vf_for_cost (loop_vec_info loop_vinfo
)
1399 return estimated_poly_value (LOOP_VINFO_VECT_FACTOR (loop_vinfo
));
1402 /* Estimate the number of elements in VEC_TYPE for costing purposes.
1403 Pick a reasonable estimate if the exact number isn't known at
1406 static inline unsigned int
1407 vect_nunits_for_cost (tree vec_type
)
1409 return estimated_poly_value (TYPE_VECTOR_SUBPARTS (vec_type
));
1412 /* Return the maximum possible vectorization factor for LOOP_VINFO. */
1414 static inline unsigned HOST_WIDE_INT
1415 vect_max_vf (loop_vec_info loop_vinfo
)
1417 unsigned HOST_WIDE_INT vf
;
1418 if (LOOP_VINFO_VECT_FACTOR (loop_vinfo
).is_constant (&vf
))
1420 return MAX_VECTORIZATION_FACTOR
;
1423 /* Return the size of the value accessed by unvectorized data reference
1424 DR_INFO. This is only valid once STMT_VINFO_VECTYPE has been calculated
1425 for the associated gimple statement, since that guarantees that DR_INFO
1426 accesses either a scalar or a scalar equivalent. ("Scalar equivalent"
1427 here includes things like V1SI, which can be vectorized in the same way
1431 vect_get_scalar_dr_size (dr_vec_info
*dr_info
)
1433 return tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_info
->dr
))));
1436 /* Source location + hotness information. */
1437 extern dump_user_location_t vect_location
;
1439 /* A macro for calling:
1440 dump_begin_scope (MSG, vect_location);
1441 via an RAII object, thus printing "=== MSG ===\n" to the dumpfile etc,
1444 once the object goes out of scope, thus capturing the nesting of
1447 These scopes affect dump messages within them: dump messages at the
1448 top level implicitly default to MSG_PRIORITY_USER_FACING, whereas those
1449 in a nested scope implicitly default to MSG_PRIORITY_INTERNALS. */
1451 #define DUMP_VECT_SCOPE(MSG) \
1452 AUTO_DUMP_SCOPE (MSG, vect_location)
1454 /* A sentinel class for ensuring that the "vect_location" global gets
1455 reset at the end of a scope.
1457 The "vect_location" global is used during dumping and contains a
1458 location_t, which could contain references to a tree block via the
1459 ad-hoc data. This data is used for tracking inlining information,
1460 but it's not a GC root; it's simply assumed that such locations never
1461 get accessed if the blocks are optimized away.
1463 Hence we need to ensure that such locations are purged at the end
1464 of any operations using them (e.g. via this class). */
1466 class auto_purge_vect_location
1469 ~auto_purge_vect_location ();
1472 /*-----------------------------------------------------------------*/
1473 /* Function prototypes. */
1474 /*-----------------------------------------------------------------*/
1476 /* Simple loop peeling and versioning utilities for vectorizer's purposes -
1477 in tree-vect-loop-manip.c. */
1478 extern void vect_set_loop_condition (class loop
*, loop_vec_info
,
1479 tree
, tree
, tree
, bool);
1480 extern bool slpeel_can_duplicate_loop_p (const class loop
*, const_edge
);
1481 class loop
*slpeel_tree_duplicate_loop_to_edge_cfg (class loop
*,
1482 class loop
*, edge
);
1483 class loop
*vect_loop_versioning (loop_vec_info
, unsigned int, bool,
1485 extern class loop
*vect_do_peeling (loop_vec_info
, tree
, tree
,
1486 tree
*, tree
*, tree
*, int, bool, bool);
1487 extern void vect_prepare_for_masked_peels (loop_vec_info
);
1488 extern dump_user_location_t
find_loop_location (class loop
*);
1489 extern bool vect_can_advance_ivs_p (loop_vec_info
);
1491 /* In tree-vect-stmts.c. */
1492 extern poly_uint64 current_vector_size
;
1493 extern tree
get_vectype_for_scalar_type (tree
);
1494 extern tree
get_vectype_for_scalar_type_and_size (tree
, poly_uint64
);
1495 extern tree
get_mask_type_for_scalar_type (tree
);
1496 extern tree
get_same_sized_vectype (tree
, tree
);
1497 extern bool vect_get_loop_mask_type (loop_vec_info
);
1498 extern bool vect_is_simple_use (tree
, vec_info
*, enum vect_def_type
*,
1499 stmt_vec_info
* = NULL
, gimple
** = NULL
);
1500 extern bool vect_is_simple_use (tree
, vec_info
*, enum vect_def_type
*,
1501 tree
*, stmt_vec_info
* = NULL
,
1503 extern bool supportable_widening_operation (enum tree_code
, stmt_vec_info
,
1504 tree
, tree
, enum tree_code
*,
1505 enum tree_code
*, int *,
1507 extern bool supportable_narrowing_operation (enum tree_code
, tree
, tree
,
1509 int *, vec
<tree
> *);
1510 extern unsigned record_stmt_cost (stmt_vector_for_cost
*, int,
1511 enum vect_cost_for_stmt
, stmt_vec_info
,
1512 int, enum vect_cost_model_location
);
1513 extern stmt_vec_info
vect_finish_replace_stmt (stmt_vec_info
, gimple
*);
1514 extern stmt_vec_info
vect_finish_stmt_generation (stmt_vec_info
, gimple
*,
1515 gimple_stmt_iterator
*);
1516 extern opt_result
vect_mark_stmts_to_be_vectorized (loop_vec_info
, bool *);
1517 extern tree
vect_get_store_rhs (stmt_vec_info
);
1518 extern tree
vect_get_vec_def_for_operand_1 (stmt_vec_info
, enum vect_def_type
);
1519 extern tree
vect_get_vec_def_for_operand (tree
, stmt_vec_info
, tree
= NULL
);
1520 extern void vect_get_vec_defs (tree
, tree
, stmt_vec_info
, vec
<tree
> *,
1521 vec
<tree
> *, slp_tree
);
1522 extern void vect_get_vec_defs_for_stmt_copy (vec_info
*,
1523 vec
<tree
> *, vec
<tree
> *);
1524 extern tree
vect_init_vector (stmt_vec_info
, tree
, tree
,
1525 gimple_stmt_iterator
*);
1526 extern tree
vect_get_vec_def_for_stmt_copy (vec_info
*, tree
);
1527 extern bool vect_transform_stmt (stmt_vec_info
, gimple_stmt_iterator
*,
1528 slp_tree
, slp_instance
);
1529 extern void vect_remove_stores (stmt_vec_info
);
1530 extern opt_result
vect_analyze_stmt (stmt_vec_info
, bool *, slp_tree
,
1531 slp_instance
, stmt_vector_for_cost
*);
1532 extern bool vectorizable_condition (stmt_vec_info
, gimple_stmt_iterator
*,
1533 stmt_vec_info
*, bool, slp_tree
,
1534 stmt_vector_for_cost
*);
1535 extern bool vectorizable_shift (stmt_vec_info
, gimple_stmt_iterator
*,
1536 stmt_vec_info
*, slp_tree
,
1537 stmt_vector_for_cost
*);
1538 extern void vect_get_load_cost (stmt_vec_info
, int, bool,
1539 unsigned int *, unsigned int *,
1540 stmt_vector_for_cost
*,
1541 stmt_vector_for_cost
*, bool);
1542 extern void vect_get_store_cost (stmt_vec_info
, int,
1543 unsigned int *, stmt_vector_for_cost
*);
1544 extern bool vect_supportable_shift (enum tree_code
, tree
);
1545 extern tree
vect_gen_perm_mask_any (tree
, const vec_perm_indices
&);
1546 extern tree
vect_gen_perm_mask_checked (tree
, const vec_perm_indices
&);
1547 extern void optimize_mask_stores (class loop
*);
1548 extern gcall
*vect_gen_while (tree
, tree
, tree
);
1549 extern tree
vect_gen_while_not (gimple_seq
*, tree
, tree
, tree
);
1550 extern opt_result
vect_get_vector_types_for_stmt (stmt_vec_info
, tree
*,
1552 extern opt_tree
vect_get_mask_type_for_stmt (stmt_vec_info
);
1554 /* In tree-vect-data-refs.c. */
1555 extern bool vect_can_force_dr_alignment_p (const_tree
, poly_uint64
);
1556 extern enum dr_alignment_support vect_supportable_dr_alignment
1557 (dr_vec_info
*, bool);
1558 extern tree
vect_get_smallest_scalar_type (stmt_vec_info
, HOST_WIDE_INT
*,
1560 extern opt_result
vect_analyze_data_ref_dependences (loop_vec_info
, unsigned int *);
1561 extern bool vect_slp_analyze_instance_dependence (slp_instance
);
1562 extern opt_result
vect_enhance_data_refs_alignment (loop_vec_info
);
1563 extern opt_result
vect_analyze_data_refs_alignment (loop_vec_info
);
1564 extern opt_result
vect_verify_datarefs_alignment (loop_vec_info
);
1565 extern bool vect_slp_analyze_and_verify_instance_alignment (slp_instance
);
1566 extern opt_result
vect_analyze_data_ref_accesses (vec_info
*);
1567 extern opt_result
vect_prune_runtime_alias_test_list (loop_vec_info
);
1568 extern bool vect_gather_scatter_fn_p (bool, bool, tree
, tree
, unsigned int,
1569 signop
, int, internal_fn
*, tree
*);
1570 extern bool vect_check_gather_scatter (stmt_vec_info
, loop_vec_info
,
1571 gather_scatter_info
*);
1572 extern opt_result
vect_find_stmt_data_reference (loop_p
, gimple
*,
1573 vec
<data_reference_p
> *);
1574 extern opt_result
vect_analyze_data_refs (vec_info
*, poly_uint64
*, bool *);
1575 extern void vect_record_base_alignments (vec_info
*);
1576 extern tree
vect_create_data_ref_ptr (stmt_vec_info
, tree
, class loop
*, tree
,
1577 tree
*, gimple_stmt_iterator
*,
1579 tree
= NULL_TREE
, tree
= NULL_TREE
);
1580 extern tree
bump_vector_ptr (tree
, gimple
*, gimple_stmt_iterator
*,
1581 stmt_vec_info
, tree
);
1582 extern void vect_copy_ref_info (tree
, tree
);
1583 extern tree
vect_create_destination_var (tree
, tree
);
1584 extern bool vect_grouped_store_supported (tree
, unsigned HOST_WIDE_INT
);
1585 extern bool vect_store_lanes_supported (tree
, unsigned HOST_WIDE_INT
, bool);
1586 extern bool vect_grouped_load_supported (tree
, bool, unsigned HOST_WIDE_INT
);
1587 extern bool vect_load_lanes_supported (tree
, unsigned HOST_WIDE_INT
, bool);
1588 extern void vect_permute_store_chain (vec
<tree
> ,unsigned int, stmt_vec_info
,
1589 gimple_stmt_iterator
*, vec
<tree
> *);
1590 extern tree
vect_setup_realignment (stmt_vec_info
, gimple_stmt_iterator
*,
1591 tree
*, enum dr_alignment_support
, tree
,
1593 extern void vect_transform_grouped_load (stmt_vec_info
, vec
<tree
> , int,
1594 gimple_stmt_iterator
*);
1595 extern void vect_record_grouped_load_vectors (stmt_vec_info
, vec
<tree
>);
1596 extern tree
vect_get_new_vect_var (tree
, enum vect_var_kind
, const char *);
1597 extern tree
vect_get_new_ssa_name (tree
, enum vect_var_kind
,
1598 const char * = NULL
);
1599 extern tree
vect_create_addr_base_for_vector_ref (stmt_vec_info
, gimple_seq
*,
1600 tree
, tree
= NULL_TREE
);
1602 /* In tree-vect-loop.c. */
1603 /* FORNOW: Used in tree-parloops.c. */
1604 extern stmt_vec_info
vect_force_simple_reduction (loop_vec_info
, stmt_vec_info
,
1606 extern widest_int
vect_iv_limit_for_full_masking (loop_vec_info loop_vinfo
);
1607 /* Used in gimple-loop-interchange.c. */
1608 extern bool check_reduction_path (dump_user_location_t
, loop_p
, gphi
*, tree
,
1610 /* Drive for loop analysis stage. */
1611 extern opt_loop_vec_info
vect_analyze_loop (class loop
*,
1614 extern tree
vect_build_loop_niters (loop_vec_info
, bool * = NULL
);
1615 extern void vect_gen_vector_loop_niters (loop_vec_info
, tree
, tree
*,
1617 extern tree
vect_halve_mask_nunits (tree
);
1618 extern tree
vect_double_mask_nunits (tree
);
1619 extern void vect_record_loop_mask (loop_vec_info
, vec_loop_masks
*,
1620 unsigned int, tree
);
1621 extern tree
vect_get_loop_mask (gimple_stmt_iterator
*, vec_loop_masks
*,
1622 unsigned int, tree
, unsigned int);
1624 /* Drive for loop transformation stage. */
1625 extern class loop
*vect_transform_loop (loop_vec_info
);
1626 extern opt_loop_vec_info
vect_analyze_loop_form (class loop
*,
1628 extern bool vectorizable_live_operation (stmt_vec_info
, gimple_stmt_iterator
*,
1629 slp_tree
, int, stmt_vec_info
*,
1630 stmt_vector_for_cost
*);
1631 extern bool vectorizable_reduction (stmt_vec_info
, gimple_stmt_iterator
*,
1632 stmt_vec_info
*, slp_tree
, slp_instance
,
1633 stmt_vector_for_cost
*);
1634 extern bool vectorizable_induction (stmt_vec_info
, gimple_stmt_iterator
*,
1635 stmt_vec_info
*, slp_tree
,
1636 stmt_vector_for_cost
*);
1637 extern tree
get_initial_def_for_reduction (stmt_vec_info
, tree
, tree
*);
1638 extern bool vect_worthwhile_without_simd_p (vec_info
*, tree_code
);
1639 extern int vect_get_known_peeling_cost (loop_vec_info
, int, int *,
1640 stmt_vector_for_cost
*,
1641 stmt_vector_for_cost
*,
1642 stmt_vector_for_cost
*);
1643 extern tree
cse_and_gimplify_to_preheader (loop_vec_info
, tree
);
1645 /* In tree-vect-slp.c. */
1646 extern void vect_free_slp_instance (slp_instance
, bool);
1647 extern bool vect_transform_slp_perm_load (slp_tree
, vec
<tree
> ,
1648 gimple_stmt_iterator
*, poly_uint64
,
1649 slp_instance
, bool, unsigned *);
1650 extern bool vect_slp_analyze_operations (vec_info
*);
1651 extern void vect_schedule_slp (vec_info
*);
1652 extern opt_result
vect_analyze_slp (vec_info
*, unsigned);
1653 extern bool vect_make_slp_decision (loop_vec_info
);
1654 extern void vect_detect_hybrid_slp (loop_vec_info
);
1655 extern void vect_get_slp_defs (vec
<tree
> , slp_tree
, vec
<vec
<tree
> > *);
1656 extern bool vect_slp_bb (basic_block
);
1657 extern stmt_vec_info
vect_find_last_scalar_stmt_in_slp (slp_tree
);
1658 extern bool is_simple_and_all_uses_invariant (stmt_vec_info
, loop_vec_info
);
1659 extern bool can_duplicate_and_interleave_p (unsigned int, machine_mode
,
1660 unsigned int * = NULL
,
1661 tree
* = NULL
, tree
* = NULL
);
1662 extern void duplicate_and_interleave (gimple_seq
*, tree
, vec
<tree
>,
1663 unsigned int, vec
<tree
> &);
1664 extern int vect_get_place_in_interleaving_chain (stmt_vec_info
, stmt_vec_info
);
1666 /* In tree-vect-patterns.c. */
1667 /* Pattern recognition functions.
1668 Additional pattern recognition functions can (and will) be added
1670 void vect_pattern_recog (vec_info
*);
1672 /* In tree-vectorizer.c. */
1673 unsigned vectorize_loops (void);
1674 void vect_free_loop_info_assumptions (class loop
*);
1675 gimple
*vect_loop_vectorized_call (class loop
*, gcond
**cond
= NULL
);
1678 #endif /* GCC_TREE_VECTORIZER_H */