2 Copyright (C) 2003-2013 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 #include "tree-data-ref.h"
26 #include "hash-table.h"
28 /* Used for naming of new temporaries. */
35 /* Defines type of operation. */
42 /* Define type of available alignment support. */
43 enum dr_alignment_support
{
44 dr_unaligned_unsupported
,
45 dr_unaligned_supported
,
47 dr_explicit_realign_optimized
,
51 /* Define type of def-use cross-iteration cycle. */
53 vect_uninitialized_def
= 0,
54 vect_constant_def
= 1,
59 vect_double_reduction_def
,
64 #define VECTORIZABLE_CYCLE_DEF(D) (((D) == vect_reduction_def) \
65 || ((D) == vect_double_reduction_def) \
66 || ((D) == vect_nested_cycle))
68 /* Structure to encapsulate information about a group of like
69 instructions to be presented to the target cost model. */
70 typedef struct _stmt_info_for_cost
{
72 enum vect_cost_for_stmt kind
;
78 typedef vec
<stmt_info_for_cost
> stmt_vector_for_cost
;
81 add_stmt_info_to_vec (stmt_vector_for_cost
*stmt_cost_vec
, int count
,
82 enum vect_cost_for_stmt kind
, gimple stmt
, int misalign
)
84 stmt_info_for_cost si
;
88 si
.misalign
= misalign
;
89 stmt_cost_vec
->safe_push (si
);
92 /************************************************************************
94 ************************************************************************/
95 typedef struct _slp_tree
*slp_tree
;
97 /* A computation tree of an SLP instance. Each node corresponds to a group of
98 stmts to be packed in a SIMD stmt. */
100 /* Nodes that contain def-stmts of this node statements operands. */
101 vec
<slp_tree
> children
;
102 /* A group of scalar stmts to be vectorized together. */
104 /* Load permutation relative to the stores, NULL if there is no
106 vec
<unsigned> load_permutation
;
107 /* Vectorized stmt/s. */
108 vec
<gimple
> vec_stmts
;
109 /* Number of vector stmts that are created to replace the group of scalar
110 stmts. It is calculated during the transformation phase as the number of
111 scalar elements in one scalar iteration (GROUP_SIZE) multiplied by VF
112 divided by vector size. */
113 unsigned int vec_stmts_size
;
117 /* SLP instance is a sequence of stmts in a loop that can be packed into
119 typedef struct _slp_instance
{
120 /* The root of SLP tree. */
123 /* Size of groups of scalar stmts that will be replaced by SIMD stmt/s. */
124 unsigned int group_size
;
126 /* The unrolling factor required to vectorized this SLP instance. */
127 unsigned int unrolling_factor
;
129 /* Vectorization costs associated with SLP instance. */
130 stmt_vector_for_cost body_cost_vec
;
132 /* The group of nodes that contain loads of this SLP instance. */
135 /* The first scalar load of the instance. The created vector loads will be
136 inserted before this statement. */
141 /* Access Functions. */
142 #define SLP_INSTANCE_TREE(S) (S)->root
143 #define SLP_INSTANCE_GROUP_SIZE(S) (S)->group_size
144 #define SLP_INSTANCE_UNROLLING_FACTOR(S) (S)->unrolling_factor
145 #define SLP_INSTANCE_BODY_COST_VEC(S) (S)->body_cost_vec
146 #define SLP_INSTANCE_LOADS(S) (S)->loads
147 #define SLP_INSTANCE_FIRST_LOAD_STMT(S) (S)->first_load
149 #define SLP_TREE_CHILDREN(S) (S)->children
150 #define SLP_TREE_SCALAR_STMTS(S) (S)->stmts
151 #define SLP_TREE_VEC_STMTS(S) (S)->vec_stmts
152 #define SLP_TREE_NUMBER_OF_VEC_STMTS(S) (S)->vec_stmts_size
153 #define SLP_TREE_LOAD_PERMUTATION(S) (S)->load_permutation
155 /* This structure is used in creation of an SLP tree. Each instance
156 corresponds to the same operand in a group of scalar stmts in an SLP
158 typedef struct _slp_oprnd_info
160 /* Def-stmts for the operands. */
161 vec
<gimple
> def_stmts
;
162 /* Information about the first statement, its vector def-type, type, the
163 operand itself in case it's constant, and an indication if it's a pattern
165 enum vect_def_type first_dt
;
172 /* This struct is used to store the information of a data reference,
173 including the data ref itself, the access offset (calculated by summing its
174 offset and init) and the segment length for aliasing checks.
175 This is used to merge alias checks. */
177 struct dr_with_seg_len
179 dr_with_seg_len (data_reference_p d
, tree len
)
181 offset (size_binop (PLUS_EXPR
, DR_OFFSET (d
), DR_INIT (d
))),
189 /* This struct contains two dr_with_seg_len objects with aliasing data
190 refs. Two comparisons are generated from them. */
192 struct dr_with_seg_len_pair_t
194 dr_with_seg_len_pair_t (const dr_with_seg_len
& d1
,
195 const dr_with_seg_len
& d2
)
196 : first (d1
), second (d2
) {}
198 dr_with_seg_len first
;
199 dr_with_seg_len second
;
203 typedef struct _vect_peel_info
206 struct data_reference
*dr
;
210 typedef struct _vect_peel_extended_info
212 struct _vect_peel_info peel_info
;
213 unsigned int inside_cost
;
214 unsigned int outside_cost
;
215 stmt_vector_for_cost body_cost_vec
;
216 } *vect_peel_extended_info
;
219 /* Peeling hashtable helpers. */
221 struct peel_info_hasher
: typed_free_remove
<_vect_peel_info
>
223 typedef _vect_peel_info value_type
;
224 typedef _vect_peel_info compare_type
;
225 static inline hashval_t
hash (const value_type
*);
226 static inline bool equal (const value_type
*, const compare_type
*);
230 peel_info_hasher::hash (const value_type
*peel_info
)
232 return (hashval_t
) peel_info
->npeel
;
236 peel_info_hasher::equal (const value_type
*a
, const compare_type
*b
)
238 return (a
->npeel
== b
->npeel
);
242 /*-----------------------------------------------------------------*/
243 /* Info on vectorized loops. */
244 /*-----------------------------------------------------------------*/
245 typedef struct _loop_vec_info
{
247 /* The loop to which this info struct refers to. */
250 /* The loop basic blocks. */
253 /* Number of iterations. */
255 tree num_iters_unchanged
;
257 /* Minimum number of iterations below which vectorization is expected to
258 not be profitable (as estimated by the cost model).
259 -1 indicates that vectorization will not be profitable.
260 FORNOW: This field is an int. Will be a tree in the future, to represent
261 values unknown at compile time. */
262 int min_profitable_iters
;
264 /* Is the loop vectorizable? */
267 /* Unrolling factor */
268 int vectorization_factor
;
270 /* Unknown DRs according to which loop was peeled. */
271 struct data_reference
*unaligned_dr
;
273 /* peeling_for_alignment indicates whether peeling for alignment will take
274 place, and what the peeling factor should be:
275 peeling_for_alignment = X means:
276 If X=0: Peeling for alignment will not be applied.
277 If X>0: Peel first X iterations.
278 If X=-1: Generate a runtime test to calculate the number of iterations
279 to be peeled, using the dataref recorded in the field
281 int peeling_for_alignment
;
283 /* The mask used to check the alignment of pointers or arrays. */
286 /* The loop nest in which the data dependences are computed. */
287 vec
<loop_p
> loop_nest
;
289 /* All data references in the loop. */
290 vec
<data_reference_p
> datarefs
;
292 /* All data dependences in the loop. */
295 /* Data Dependence Relations defining address ranges that are candidates
296 for a run-time aliasing check. */
297 vec
<ddr_p
> may_alias_ddrs
;
299 /* Data Dependence Relations defining address ranges together with segment
300 lengths from which the run-time aliasing check is built. */
301 vec
<dr_with_seg_len_pair_t
> comp_alias_ddrs
;
303 /* Statements in the loop that have data references that are candidates for a
304 runtime (loop versioning) misalignment check. */
305 vec
<gimple
> may_misalign_stmts
;
307 /* All interleaving chains of stores in the loop, represented by the first
308 stmt in the chain. */
309 vec
<gimple
> grouped_stores
;
311 /* All SLP instances in the loop. This is a subset of the set of GROUP_STORES
313 vec
<slp_instance
> slp_instances
;
315 /* The unrolling factor needed to SLP the loop. In case of that pure SLP is
316 applied to the loop, i.e., no unrolling is needed, this is 1. */
317 unsigned slp_unrolling_factor
;
319 /* Reduction cycles detected in the loop. Used in loop-aware SLP. */
320 vec
<gimple
> reductions
;
322 /* All reduction chains in the loop, represented by the first
323 stmt in the chain. */
324 vec
<gimple
> reduction_chains
;
326 /* Hash table used to choose the best peeling option. */
327 hash_table
<peel_info_hasher
> peeling_htab
;
329 /* Cost data used by the target cost model. */
330 void *target_cost_data
;
332 /* When we have grouped data accesses with gaps, we may introduce invalid
333 memory accesses. We peel the last iteration of the loop to prevent
335 bool peeling_for_gaps
;
337 /* When the number of iterations is not a multiple of the vector size
338 we need to peel off iterations at the end to form an epilogue loop. */
339 bool peeling_for_niter
;
341 /* Reductions are canonicalized so that the last operand is the reduction
342 operand. If this places a constant into RHS1, this decanonicalizes
343 GIMPLE for other phases, so we must track when this has occurred and
345 bool operands_swapped
;
349 /* Access Functions. */
350 #define LOOP_VINFO_LOOP(L) (L)->loop
351 #define LOOP_VINFO_BBS(L) (L)->bbs
352 #define LOOP_VINFO_NITERS(L) (L)->num_iters
353 /* Since LOOP_VINFO_NITERS can change after prologue peeling
354 retain total unchanged scalar loop iterations for cost model. */
355 #define LOOP_VINFO_NITERS_UNCHANGED(L) (L)->num_iters_unchanged
356 #define LOOP_VINFO_COST_MODEL_MIN_ITERS(L) (L)->min_profitable_iters
357 #define LOOP_VINFO_VECTORIZABLE_P(L) (L)->vectorizable
358 #define LOOP_VINFO_VECT_FACTOR(L) (L)->vectorization_factor
359 #define LOOP_VINFO_PTR_MASK(L) (L)->ptr_mask
360 #define LOOP_VINFO_LOOP_NEST(L) (L)->loop_nest
361 #define LOOP_VINFO_DATAREFS(L) (L)->datarefs
362 #define LOOP_VINFO_DDRS(L) (L)->ddrs
363 #define LOOP_VINFO_INT_NITERS(L) (TREE_INT_CST_LOW ((L)->num_iters))
364 #define LOOP_VINFO_PEELING_FOR_ALIGNMENT(L) (L)->peeling_for_alignment
365 #define LOOP_VINFO_UNALIGNED_DR(L) (L)->unaligned_dr
366 #define LOOP_VINFO_MAY_MISALIGN_STMTS(L) (L)->may_misalign_stmts
367 #define LOOP_VINFO_MAY_ALIAS_DDRS(L) (L)->may_alias_ddrs
368 #define LOOP_VINFO_COMP_ALIAS_DDRS(L) (L)->comp_alias_ddrs
369 #define LOOP_VINFO_GROUPED_STORES(L) (L)->grouped_stores
370 #define LOOP_VINFO_SLP_INSTANCES(L) (L)->slp_instances
371 #define LOOP_VINFO_SLP_UNROLLING_FACTOR(L) (L)->slp_unrolling_factor
372 #define LOOP_VINFO_REDUCTIONS(L) (L)->reductions
373 #define LOOP_VINFO_REDUCTION_CHAINS(L) (L)->reduction_chains
374 #define LOOP_VINFO_PEELING_HTAB(L) (L)->peeling_htab
375 #define LOOP_VINFO_TARGET_COST_DATA(L) (L)->target_cost_data
376 #define LOOP_VINFO_PEELING_FOR_GAPS(L) (L)->peeling_for_gaps
377 #define LOOP_VINFO_OPERANDS_SWAPPED(L) (L)->operands_swapped
378 #define LOOP_VINFO_PEELING_FOR_NITER(L) (L)->peeling_for_niter
380 #define LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT(L) \
381 (L)->may_misalign_stmts.length () > 0
382 #define LOOP_REQUIRES_VERSIONING_FOR_ALIAS(L) \
383 (L)->may_alias_ddrs.length () > 0
385 #define LOOP_VINFO_NITERS_KNOWN_P(L) \
386 (tree_fits_shwi_p ((L)->num_iters) && tree_to_shwi ((L)->num_iters) > 0)
388 static inline loop_vec_info
389 loop_vec_info_for_loop (struct loop
*loop
)
391 return (loop_vec_info
) loop
->aux
;
395 nested_in_vect_loop_p (struct loop
*loop
, gimple stmt
)
398 && (loop
->inner
== (gimple_bb (stmt
))->loop_father
));
401 typedef struct _bb_vec_info
{
404 /* All interleaving chains of stores in the basic block, represented by the
405 first stmt in the chain. */
406 vec
<gimple
> grouped_stores
;
408 /* All SLP instances in the basic block. This is a subset of the set of
409 GROUP_STORES of the basic block. */
410 vec
<slp_instance
> slp_instances
;
412 /* All data references in the basic block. */
413 vec
<data_reference_p
> datarefs
;
415 /* All data dependences in the basic block. */
418 /* Cost data used by the target cost model. */
419 void *target_cost_data
;
423 #define BB_VINFO_BB(B) (B)->bb
424 #define BB_VINFO_GROUPED_STORES(B) (B)->grouped_stores
425 #define BB_VINFO_SLP_INSTANCES(B) (B)->slp_instances
426 #define BB_VINFO_DATAREFS(B) (B)->datarefs
427 #define BB_VINFO_DDRS(B) (B)->ddrs
428 #define BB_VINFO_TARGET_COST_DATA(B) (B)->target_cost_data
430 static inline bb_vec_info
431 vec_info_for_bb (basic_block bb
)
433 return (bb_vec_info
) bb
->aux
;
436 /*-----------------------------------------------------------------*/
437 /* Info on vectorized defs. */
438 /*-----------------------------------------------------------------*/
439 enum stmt_vec_info_type
{
440 undef_vec_info_type
= 0,
446 call_simd_clone_vec_info_type
,
447 assignment_vec_info_type
,
448 condition_vec_info_type
,
451 type_promotion_vec_info_type
,
452 type_demotion_vec_info_type
,
453 type_conversion_vec_info_type
,
454 loop_exit_ctrl_vec_info_type
457 /* Indicates whether/how a variable is used in the scope of loop/basic
460 vect_unused_in_scope
= 0,
461 /* The def is in the inner loop, and the use is in the outer loop, and the
462 use is a reduction stmt. */
463 vect_used_in_outer_by_reduction
,
464 /* The def is in the inner loop, and the use is in the outer loop (and is
465 not part of reduction). */
468 /* defs that feed computations that end up (only) in a reduction. These
469 defs may be used by non-reduction stmts, but eventually, any
470 computations/values that are affected by these defs are used to compute
471 a reduction (i.e. don't get stored to memory, for example). We use this
472 to identify computations that we can change the order in which they are
474 vect_used_by_reduction
,
479 /* The type of vectorization that can be applied to the stmt: regular loop-based
480 vectorization; pure SLP - the stmt is a part of SLP instances and does not
481 have uses outside SLP instances; or hybrid SLP and loop-based - the stmt is
482 a part of SLP instance and also must be loop-based vectorized, since it has
483 uses outside SLP sequences.
485 In the loop context the meanings of pure and hybrid SLP are slightly
486 different. By saying that pure SLP is applied to the loop, we mean that we
487 exploit only intra-iteration parallelism in the loop; i.e., the loop can be
488 vectorized without doing any conceptual unrolling, cause we don't pack
489 together stmts from different iterations, only within a single iteration.
490 Loop hybrid SLP means that we exploit both intra-iteration and
491 inter-iteration parallelism (e.g., number of elements in the vector is 4
492 and the slp-group-size is 2, in which case we don't have enough parallelism
493 within an iteration, so we obtain the rest of the parallelism from subsequent
494 iterations by unrolling the loop by 2). */
502 typedef struct data_reference
*dr_p
;
504 typedef struct _stmt_vec_info
{
506 enum stmt_vec_info_type type
;
508 /* Indicates whether this stmts is part of a computation whose result is
509 used outside the loop. */
512 /* Stmt is part of some pattern (computation idiom) */
515 /* The stmt to which this info struct refers to. */
518 /* The loop_vec_info with respect to which STMT is vectorized. */
519 loop_vec_info loop_vinfo
;
521 /* The vector type to be used for the LHS of this statement. */
524 /* The vectorized version of the stmt. */
525 gimple vectorized_stmt
;
528 /** The following is relevant only for stmts that contain a non-scalar
529 data-ref (array/pointer/struct access). A GIMPLE stmt is expected to have
530 at most one such data-ref. **/
532 /* Information about the data-ref (access function, etc),
533 relative to the inner-most containing loop. */
534 struct data_reference
*data_ref_info
;
536 /* Information about the data-ref relative to this loop
537 nest (the loop that is being considered for vectorization). */
538 tree dr_base_address
;
544 /* For loop PHI nodes, the evolution part of it. This makes sure
545 this information is still available in vect_update_ivs_after_vectorizer
546 where we may not be able to re-analyze the PHI nodes evolution as
547 peeling for the prologue loop can make it unanalyzable. The evolution
548 part is still correct though. */
549 tree loop_phi_evolution_part
;
551 /* Used for various bookkeeping purposes, generally holding a pointer to
552 some other stmt S that is in some way "related" to this stmt.
553 Current use of this field is:
554 If this stmt is part of a pattern (i.e. the field 'in_pattern_p' is
555 true): S is the "pattern stmt" that represents (and replaces) the
556 sequence of stmts that constitutes the pattern. Similarly, the
557 related_stmt of the "pattern stmt" points back to this stmt (which is
558 the last stmt in the original sequence of stmts that constitutes the
562 /* Used to keep a sequence of def stmts of a pattern stmt if such exists. */
563 gimple_seq pattern_def_seq
;
565 /* List of datarefs that are known to have the same alignment as the dataref
567 vec
<dr_p
> same_align_refs
;
569 /* Selected SIMD clone's function decl. */
570 tree simd_clone_fndecl
;
572 /* Classify the def of this stmt. */
573 enum vect_def_type def_type
;
575 /* Whether the stmt is SLPed, loop-based vectorized, or both. */
576 enum slp_vect_type slp_type
;
578 /* Interleaving and reduction chains info. */
579 /* First element in the group. */
580 gimple first_element
;
581 /* Pointer to the next element in the group. */
583 /* For data-refs, in case that two or more stmts share data-ref, this is the
584 pointer to the previously detected stmt with the same dr. */
586 /* The size of the group. */
588 /* For stores, number of stores from this group seen. We vectorize the last
590 unsigned int store_count
;
591 /* For loads only, the gap from the previous load. For consecutive loads, GAP
595 /* Not all stmts in the loop need to be vectorized. e.g, the increment
596 of the loop induction variable and computation of array indexes. relevant
597 indicates whether the stmt needs to be vectorized. */
598 enum vect_relevant relevant
;
600 /* The bb_vec_info with respect to which STMT is vectorized. */
601 bb_vec_info bb_vinfo
;
603 /* Is this statement vectorizable or should it be skipped in (partial)
607 /* For loads only, true if this is a gather load. */
611 /* For both loads and stores. */
612 bool simd_lane_access_p
;
615 /* Access Functions. */
616 #define STMT_VINFO_TYPE(S) (S)->type
617 #define STMT_VINFO_STMT(S) (S)->stmt
618 #define STMT_VINFO_LOOP_VINFO(S) (S)->loop_vinfo
619 #define STMT_VINFO_BB_VINFO(S) (S)->bb_vinfo
620 #define STMT_VINFO_RELEVANT(S) (S)->relevant
621 #define STMT_VINFO_LIVE_P(S) (S)->live
622 #define STMT_VINFO_VECTYPE(S) (S)->vectype
623 #define STMT_VINFO_VEC_STMT(S) (S)->vectorized_stmt
624 #define STMT_VINFO_VECTORIZABLE(S) (S)->vectorizable
625 #define STMT_VINFO_DATA_REF(S) (S)->data_ref_info
626 #define STMT_VINFO_GATHER_P(S) (S)->gather_p
627 #define STMT_VINFO_STRIDE_LOAD_P(S) (S)->stride_load_p
628 #define STMT_VINFO_SIMD_LANE_ACCESS_P(S) (S)->simd_lane_access_p
630 #define STMT_VINFO_DR_BASE_ADDRESS(S) (S)->dr_base_address
631 #define STMT_VINFO_DR_INIT(S) (S)->dr_init
632 #define STMT_VINFO_DR_OFFSET(S) (S)->dr_offset
633 #define STMT_VINFO_DR_STEP(S) (S)->dr_step
634 #define STMT_VINFO_DR_ALIGNED_TO(S) (S)->dr_aligned_to
636 #define STMT_VINFO_IN_PATTERN_P(S) (S)->in_pattern_p
637 #define STMT_VINFO_RELATED_STMT(S) (S)->related_stmt
638 #define STMT_VINFO_PATTERN_DEF_SEQ(S) (S)->pattern_def_seq
639 #define STMT_VINFO_SAME_ALIGN_REFS(S) (S)->same_align_refs
640 #define STMT_VINFO_SIMD_CLONE_FNDECL(S) (S)->simd_clone_fndecl
641 #define STMT_VINFO_DEF_TYPE(S) (S)->def_type
642 #define STMT_VINFO_GROUP_FIRST_ELEMENT(S) (S)->first_element
643 #define STMT_VINFO_GROUP_NEXT_ELEMENT(S) (S)->next_element
644 #define STMT_VINFO_GROUP_SIZE(S) (S)->size
645 #define STMT_VINFO_GROUP_STORE_COUNT(S) (S)->store_count
646 #define STMT_VINFO_GROUP_GAP(S) (S)->gap
647 #define STMT_VINFO_GROUP_SAME_DR_STMT(S) (S)->same_dr_stmt
648 #define STMT_VINFO_GROUPED_ACCESS(S) ((S)->first_element != NULL && (S)->data_ref_info)
649 #define STMT_VINFO_LOOP_PHI_EVOLUTION_PART(S) (S)->loop_phi_evolution_part
651 #define GROUP_FIRST_ELEMENT(S) (S)->first_element
652 #define GROUP_NEXT_ELEMENT(S) (S)->next_element
653 #define GROUP_SIZE(S) (S)->size
654 #define GROUP_STORE_COUNT(S) (S)->store_count
655 #define GROUP_GAP(S) (S)->gap
656 #define GROUP_SAME_DR_STMT(S) (S)->same_dr_stmt
658 #define STMT_VINFO_RELEVANT_P(S) ((S)->relevant != vect_unused_in_scope)
660 #define HYBRID_SLP_STMT(S) ((S)->slp_type == hybrid)
661 #define PURE_SLP_STMT(S) ((S)->slp_type == pure_slp)
662 #define STMT_SLP_TYPE(S) (S)->slp_type
666 bool base_misaligned
;
670 #define VECT_MAX_COST 1000
672 /* The maximum number of intermediate steps required in multi-step type
674 #define MAX_INTERM_CVT_STEPS 3
676 /* The maximum vectorization factor supported by any target (V32QI). */
677 #define MAX_VECTORIZATION_FACTOR 32
679 /* Avoid GTY(()) on stmt_vec_info. */
680 typedef void *vec_void_p
;
682 extern vec
<vec_void_p
> stmt_vec_info_vec
;
684 void init_stmt_vec_info_vec (void);
685 void free_stmt_vec_info_vec (void);
687 /* Return a stmt_vec_info corresponding to STMT. */
689 static inline stmt_vec_info
690 vinfo_for_stmt (gimple stmt
)
692 unsigned int uid
= gimple_uid (stmt
);
696 return (stmt_vec_info
) stmt_vec_info_vec
[uid
- 1];
699 /* Set vectorizer information INFO for STMT. */
702 set_vinfo_for_stmt (gimple stmt
, stmt_vec_info info
)
704 unsigned int uid
= gimple_uid (stmt
);
707 gcc_checking_assert (info
);
708 uid
= stmt_vec_info_vec
.length () + 1;
709 gimple_set_uid (stmt
, uid
);
710 stmt_vec_info_vec
.safe_push ((vec_void_p
) info
);
713 stmt_vec_info_vec
[uid
- 1] = (vec_void_p
) info
;
716 /* Return the earlier statement between STMT1 and STMT2. */
719 get_earlier_stmt (gimple stmt1
, gimple stmt2
)
721 unsigned int uid1
, uid2
;
729 uid1
= gimple_uid (stmt1
);
730 uid2
= gimple_uid (stmt2
);
732 if (uid1
== 0 || uid2
== 0)
735 gcc_checking_assert (uid1
<= stmt_vec_info_vec
.length ()
736 && uid2
<= stmt_vec_info_vec
.length ());
744 /* Return the later statement between STMT1 and STMT2. */
747 get_later_stmt (gimple stmt1
, gimple stmt2
)
749 unsigned int uid1
, uid2
;
757 uid1
= gimple_uid (stmt1
);
758 uid2
= gimple_uid (stmt2
);
760 if (uid1
== 0 || uid2
== 0)
763 gcc_assert (uid1
<= stmt_vec_info_vec
.length ());
764 gcc_assert (uid2
<= stmt_vec_info_vec
.length ());
772 /* Return TRUE if a statement represented by STMT_INFO is a part of a
776 is_pattern_stmt_p (stmt_vec_info stmt_info
)
779 stmt_vec_info related_stmt_info
;
781 related_stmt
= STMT_VINFO_RELATED_STMT (stmt_info
);
783 && (related_stmt_info
= vinfo_for_stmt (related_stmt
))
784 && STMT_VINFO_IN_PATTERN_P (related_stmt_info
))
790 /* Return true if BB is a loop header. */
793 is_loop_header_bb_p (basic_block bb
)
795 if (bb
== (bb
->loop_father
)->header
)
797 gcc_checking_assert (EDGE_COUNT (bb
->preds
) == 1);
801 /* Return pow2 (X). */
808 for (i
= 0; i
< x
; i
++)
814 /* Alias targetm.vectorize.builtin_vectorization_cost. */
817 builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost
,
818 tree vectype
, int misalign
)
820 return targetm
.vectorize
.builtin_vectorization_cost (type_of_cost
,
824 /* Get cost by calling cost target builtin. */
827 int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost
)
829 return builtin_vectorization_cost (type_of_cost
, NULL
, 0);
832 /* Alias targetm.vectorize.init_cost. */
835 init_cost (struct loop
*loop_info
)
837 return targetm
.vectorize
.init_cost (loop_info
);
840 /* Alias targetm.vectorize.add_stmt_cost. */
842 static inline unsigned
843 add_stmt_cost (void *data
, int count
, enum vect_cost_for_stmt kind
,
844 stmt_vec_info stmt_info
, int misalign
,
845 enum vect_cost_model_location where
)
847 return targetm
.vectorize
.add_stmt_cost (data
, count
, kind
,
848 stmt_info
, misalign
, where
);
851 /* Alias targetm.vectorize.finish_cost. */
854 finish_cost (void *data
, unsigned *prologue_cost
,
855 unsigned *body_cost
, unsigned *epilogue_cost
)
857 targetm
.vectorize
.finish_cost (data
, prologue_cost
, body_cost
, epilogue_cost
);
860 /* Alias targetm.vectorize.destroy_cost_data. */
863 destroy_cost_data (void *data
)
865 targetm
.vectorize
.destroy_cost_data (data
);
869 /*-----------------------------------------------------------------*/
870 /* Info on data references alignment. */
871 /*-----------------------------------------------------------------*/
873 set_dr_misalignment (struct data_reference
*dr
, int val
)
875 dataref_aux
*data_aux
= (dataref_aux
*) dr
->aux
;
879 data_aux
= XCNEW (dataref_aux
);
883 data_aux
->misalignment
= val
;
887 dr_misalignment (struct data_reference
*dr
)
889 gcc_assert (dr
->aux
);
890 return ((dataref_aux
*) dr
->aux
)->misalignment
;
893 /* Reflects actual alignment of first access in the vectorized loop,
894 taking into account peeling/versioning if applied. */
895 #define DR_MISALIGNMENT(DR) dr_misalignment (DR)
896 #define SET_DR_MISALIGNMENT(DR, VAL) set_dr_misalignment (DR, VAL)
898 /* Return TRUE if the data access is aligned, and FALSE otherwise. */
901 aligned_access_p (struct data_reference
*data_ref_info
)
903 return (DR_MISALIGNMENT (data_ref_info
) == 0);
906 /* Return TRUE if the alignment of the data access is known, and FALSE
910 known_alignment_for_access_p (struct data_reference
*data_ref_info
)
912 return (DR_MISALIGNMENT (data_ref_info
) != -1);
916 /* Return true if the vect cost model is unlimited. */
918 unlimited_cost_model (loop_p loop
)
920 if (loop
!= NULL
&& loop
->force_vect
921 && flag_simd_cost_model
!= VECT_COST_MODEL_DEFAULT
)
922 return flag_simd_cost_model
== VECT_COST_MODEL_UNLIMITED
;
923 return (flag_vect_cost_model
== VECT_COST_MODEL_UNLIMITED
);
926 /* Source location */
927 extern source_location vect_location
;
929 /*-----------------------------------------------------------------*/
930 /* Function prototypes. */
931 /*-----------------------------------------------------------------*/
933 /* Simple loop peeling and versioning utilities for vectorizer's purposes -
934 in tree-vect-loop-manip.c. */
935 extern void slpeel_make_loop_iterate_ntimes (struct loop
*, tree
);
936 extern bool slpeel_can_duplicate_loop_p (const struct loop
*, const_edge
);
937 struct loop
*slpeel_tree_duplicate_loop_to_edge_cfg (struct loop
*, edge
);
938 extern void vect_loop_versioning (loop_vec_info
, unsigned int, bool);
939 extern void vect_do_peeling_for_loop_bound (loop_vec_info
, tree
, tree
,
941 extern void vect_do_peeling_for_alignment (loop_vec_info
, tree
,
943 extern source_location
find_loop_location (struct loop
*);
944 extern bool vect_can_advance_ivs_p (loop_vec_info
);
946 /* In tree-vect-stmts.c. */
947 extern unsigned int current_vector_size
;
948 extern tree
get_vectype_for_scalar_type (tree
);
949 extern tree
get_same_sized_vectype (tree
, tree
);
950 extern bool vect_is_simple_use (tree
, gimple
, loop_vec_info
,
951 bb_vec_info
, gimple
*,
952 tree
*, enum vect_def_type
*);
953 extern bool vect_is_simple_use_1 (tree
, gimple
, loop_vec_info
,
954 bb_vec_info
, gimple
*,
955 tree
*, enum vect_def_type
*, tree
*);
956 extern bool supportable_widening_operation (enum tree_code
, gimple
, tree
, tree
,
957 enum tree_code
*, enum tree_code
*,
959 extern bool supportable_narrowing_operation (enum tree_code
, tree
, tree
,
962 extern stmt_vec_info
new_stmt_vec_info (gimple stmt
, loop_vec_info
,
964 extern void free_stmt_vec_info (gimple stmt
);
965 extern tree
vectorizable_function (gimple
, tree
, tree
);
966 extern void vect_model_simple_cost (stmt_vec_info
, int, enum vect_def_type
*,
967 stmt_vector_for_cost
*,
968 stmt_vector_for_cost
*);
969 extern void vect_model_store_cost (stmt_vec_info
, int, bool,
970 enum vect_def_type
, slp_tree
,
971 stmt_vector_for_cost
*,
972 stmt_vector_for_cost
*);
973 extern void vect_model_load_cost (stmt_vec_info
, int, bool, slp_tree
,
974 stmt_vector_for_cost
*,
975 stmt_vector_for_cost
*);
976 extern unsigned record_stmt_cost (stmt_vector_for_cost
*, int,
977 enum vect_cost_for_stmt
, stmt_vec_info
,
978 int, enum vect_cost_model_location
);
979 extern void vect_finish_stmt_generation (gimple
, gimple
,
980 gimple_stmt_iterator
*);
981 extern bool vect_mark_stmts_to_be_vectorized (loop_vec_info
);
982 extern tree
vect_get_vec_def_for_operand (tree
, gimple
, tree
*);
983 extern tree
vect_init_vector (gimple
, tree
, tree
,
984 gimple_stmt_iterator
*);
985 extern tree
vect_get_vec_def_for_stmt_copy (enum vect_def_type
, tree
);
986 extern bool vect_transform_stmt (gimple
, gimple_stmt_iterator
*,
987 bool *, slp_tree
, slp_instance
);
988 extern void vect_remove_stores (gimple
);
989 extern bool vect_analyze_stmt (gimple
, bool *, slp_tree
);
990 extern bool vectorizable_condition (gimple
, gimple_stmt_iterator
*, gimple
*,
991 tree
, int, slp_tree
);
992 extern void vect_get_load_cost (struct data_reference
*, int, bool,
993 unsigned int *, unsigned int *,
994 stmt_vector_for_cost
*,
995 stmt_vector_for_cost
*, bool);
996 extern void vect_get_store_cost (struct data_reference
*, int,
997 unsigned int *, stmt_vector_for_cost
*);
998 extern bool vect_supportable_shift (enum tree_code
, tree
);
999 extern void vect_get_vec_defs (tree
, tree
, gimple
, vec
<tree
> *,
1000 vec
<tree
> *, slp_tree
, int);
1001 extern tree
vect_gen_perm_mask (tree
, unsigned char *);
1003 /* In tree-vect-data-refs.c. */
1004 extern bool vect_can_force_dr_alignment_p (const_tree
, unsigned int);
1005 extern enum dr_alignment_support vect_supportable_dr_alignment
1006 (struct data_reference
*, bool);
1007 extern tree
vect_get_smallest_scalar_type (gimple
, HOST_WIDE_INT
*,
1009 extern bool vect_analyze_data_ref_dependences (loop_vec_info
, int *);
1010 extern bool vect_slp_analyze_data_ref_dependences (bb_vec_info
);
1011 extern bool vect_enhance_data_refs_alignment (loop_vec_info
);
1012 extern bool vect_analyze_data_refs_alignment (loop_vec_info
, bb_vec_info
);
1013 extern bool vect_verify_datarefs_alignment (loop_vec_info
, bb_vec_info
);
1014 extern bool vect_analyze_data_ref_accesses (loop_vec_info
, bb_vec_info
);
1015 extern bool vect_prune_runtime_alias_test_list (loop_vec_info
);
1016 extern tree
vect_check_gather (gimple
, loop_vec_info
, tree
*, tree
*,
1018 extern bool vect_analyze_data_refs (loop_vec_info
, bb_vec_info
, int *);
1019 extern tree
vect_create_data_ref_ptr (gimple
, tree
, struct loop
*, tree
,
1020 tree
*, gimple_stmt_iterator
*,
1021 gimple
*, bool, bool *);
1022 extern tree
bump_vector_ptr (tree
, gimple
, gimple_stmt_iterator
*, gimple
, tree
);
1023 extern tree
vect_create_destination_var (tree
, tree
);
1024 extern bool vect_grouped_store_supported (tree
, unsigned HOST_WIDE_INT
);
1025 extern bool vect_store_lanes_supported (tree
, unsigned HOST_WIDE_INT
);
1026 extern bool vect_grouped_load_supported (tree
, unsigned HOST_WIDE_INT
);
1027 extern bool vect_load_lanes_supported (tree
, unsigned HOST_WIDE_INT
);
1028 extern void vect_permute_store_chain (vec
<tree
> ,unsigned int, gimple
,
1029 gimple_stmt_iterator
*, vec
<tree
> *);
1030 extern tree
vect_setup_realignment (gimple
, gimple_stmt_iterator
*, tree
*,
1031 enum dr_alignment_support
, tree
,
1033 extern void vect_transform_grouped_load (gimple
, vec
<tree
> , int,
1034 gimple_stmt_iterator
*);
1035 extern void vect_record_grouped_load_vectors (gimple
, vec
<tree
> );
1036 extern tree
vect_get_new_vect_var (tree
, enum vect_var_kind
, const char *);
1037 extern tree
vect_create_addr_base_for_vector_ref (gimple
, gimple_seq
*,
1038 tree
, struct loop
*);
1040 /* In tree-vect-loop.c. */
1041 /* FORNOW: Used in tree-parloops.c. */
1042 extern void destroy_loop_vec_info (loop_vec_info
, bool);
1043 extern gimple
vect_force_simple_reduction (loop_vec_info
, gimple
, bool, bool *);
1044 /* Drive for loop analysis stage. */
1045 extern loop_vec_info
vect_analyze_loop (struct loop
*);
1046 /* Drive for loop transformation stage. */
1047 extern void vect_transform_loop (loop_vec_info
);
1048 extern loop_vec_info
vect_analyze_loop_form (struct loop
*);
1049 extern bool vectorizable_live_operation (gimple
, gimple_stmt_iterator
*,
1051 extern bool vectorizable_reduction (gimple
, gimple_stmt_iterator
*, gimple
*,
1053 extern bool vectorizable_induction (gimple
, gimple_stmt_iterator
*, gimple
*);
1054 extern tree
get_initial_def_for_reduction (gimple
, tree
, tree
*);
1055 extern int vect_min_worthwhile_factor (enum tree_code
);
1056 extern int vect_get_known_peeling_cost (loop_vec_info
, int, int *, int,
1057 stmt_vector_for_cost
*,
1058 stmt_vector_for_cost
*);
1059 extern int vect_get_single_scalar_iteration_cost (loop_vec_info
);
1061 /* In tree-vect-slp.c. */
1062 extern void vect_free_slp_instance (slp_instance
);
1063 extern bool vect_transform_slp_perm_load (slp_tree
, vec
<tree
> ,
1064 gimple_stmt_iterator
*, int,
1065 slp_instance
, bool);
1066 extern bool vect_schedule_slp (loop_vec_info
, bb_vec_info
);
1067 extern void vect_update_slp_costs_according_to_vf (loop_vec_info
);
1068 extern bool vect_analyze_slp (loop_vec_info
, bb_vec_info
);
1069 extern bool vect_make_slp_decision (loop_vec_info
);
1070 extern void vect_detect_hybrid_slp (loop_vec_info
);
1071 extern void vect_get_slp_defs (vec
<tree
> , slp_tree
,
1072 vec
<vec
<tree
> > *, int);
1074 extern source_location
find_bb_location (basic_block
);
1075 extern bb_vec_info
vect_slp_analyze_bb (basic_block
);
1076 extern void vect_slp_transform_bb (basic_block
);
1078 /* In tree-vect-patterns.c. */
1079 /* Pattern recognition functions.
1080 Additional pattern recognition functions can (and will) be added
1082 typedef gimple (* vect_recog_func_ptr
) (vec
<gimple
> *, tree
*, tree
*);
1083 #define NUM_PATTERNS 11
1084 void vect_pattern_recog (loop_vec_info
, bb_vec_info
);
1086 /* In tree-vectorizer.c. */
1087 unsigned vectorize_loops (void);
1088 void vect_destroy_datarefs (loop_vec_info
, bb_vec_info
);
1090 #endif /* GCC_TREE_VECTORIZER_H */