2013-10-30 Balaji V. Iyer <balaji.v.iyer@intel.com>
[official-gcc.git] / gcc / tree-vectorizer.h
blob8b7b345150993da5d79ed125883af7ea732f16dc
1 /* Vectorizer
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
10 version.
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
15 for more details.
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"
25 #include "target.h"
26 #include "hash-table.h"
28 typedef source_location LOC;
29 #define UNKNOWN_LOC UNKNOWN_LOCATION
30 #define EXPR_LOC(e) EXPR_LOCATION (e)
31 #define LOC_FILE(l) LOCATION_FILE (l)
32 #define LOC_LINE(l) LOCATION_LINE (l)
34 /* Used for naming of new temporaries. */
35 enum vect_var_kind {
36 vect_simple_var,
37 vect_pointer_var,
38 vect_scalar_var
41 /* Defines type of operation. */
42 enum operation_type {
43 unary_op = 1,
44 binary_op,
45 ternary_op
48 /* Define type of available alignment support. */
49 enum dr_alignment_support {
50 dr_unaligned_unsupported,
51 dr_unaligned_supported,
52 dr_explicit_realign,
53 dr_explicit_realign_optimized,
54 dr_aligned
57 /* Define type of def-use cross-iteration cycle. */
58 enum vect_def_type {
59 vect_uninitialized_def = 0,
60 vect_constant_def = 1,
61 vect_external_def,
62 vect_internal_def,
63 vect_induction_def,
64 vect_reduction_def,
65 vect_double_reduction_def,
66 vect_nested_cycle,
67 vect_unknown_def_type
70 #define VECTORIZABLE_CYCLE_DEF(D) (((D) == vect_reduction_def) \
71 || ((D) == vect_double_reduction_def) \
72 || ((D) == vect_nested_cycle))
74 /* Structure to encapsulate information about a group of like
75 instructions to be presented to the target cost model. */
76 typedef struct _stmt_info_for_cost {
77 int count;
78 enum vect_cost_for_stmt kind;
79 gimple stmt;
80 int misalign;
81 } stmt_info_for_cost;
84 typedef vec<stmt_info_for_cost> stmt_vector_for_cost;
86 static inline void
87 add_stmt_info_to_vec (stmt_vector_for_cost *stmt_cost_vec, int count,
88 enum vect_cost_for_stmt kind, gimple stmt, int misalign)
90 stmt_info_for_cost si;
91 si.count = count;
92 si.kind = kind;
93 si.stmt = stmt;
94 si.misalign = misalign;
95 stmt_cost_vec->safe_push (si);
98 /************************************************************************
99 SLP
100 ************************************************************************/
101 typedef struct _slp_tree *slp_tree;
103 /* A computation tree of an SLP instance. Each node corresponds to a group of
104 stmts to be packed in a SIMD stmt. */
105 struct _slp_tree {
106 /* Nodes that contain def-stmts of this node statements operands. */
107 vec<slp_tree> children;
108 /* A group of scalar stmts to be vectorized together. */
109 vec<gimple> stmts;
110 /* Load permutation relative to the stores, NULL if there is no
111 permutation. */
112 vec<unsigned> load_permutation;
113 /* Vectorized stmt/s. */
114 vec<gimple> vec_stmts;
115 /* Number of vector stmts that are created to replace the group of scalar
116 stmts. It is calculated during the transformation phase as the number of
117 scalar elements in one scalar iteration (GROUP_SIZE) multiplied by VF
118 divided by vector size. */
119 unsigned int vec_stmts_size;
123 /* SLP instance is a sequence of stmts in a loop that can be packed into
124 SIMD stmts. */
125 typedef struct _slp_instance {
126 /* The root of SLP tree. */
127 slp_tree root;
129 /* Size of groups of scalar stmts that will be replaced by SIMD stmt/s. */
130 unsigned int group_size;
132 /* The unrolling factor required to vectorized this SLP instance. */
133 unsigned int unrolling_factor;
135 /* Vectorization costs associated with SLP instance. */
136 stmt_vector_for_cost body_cost_vec;
138 /* The group of nodes that contain loads of this SLP instance. */
139 vec<slp_tree> loads;
141 /* The first scalar load of the instance. The created vector loads will be
142 inserted before this statement. */
143 gimple first_load;
144 } *slp_instance;
147 /* Access Functions. */
148 #define SLP_INSTANCE_TREE(S) (S)->root
149 #define SLP_INSTANCE_GROUP_SIZE(S) (S)->group_size
150 #define SLP_INSTANCE_UNROLLING_FACTOR(S) (S)->unrolling_factor
151 #define SLP_INSTANCE_BODY_COST_VEC(S) (S)->body_cost_vec
152 #define SLP_INSTANCE_LOADS(S) (S)->loads
153 #define SLP_INSTANCE_FIRST_LOAD_STMT(S) (S)->first_load
155 #define SLP_TREE_CHILDREN(S) (S)->children
156 #define SLP_TREE_SCALAR_STMTS(S) (S)->stmts
157 #define SLP_TREE_VEC_STMTS(S) (S)->vec_stmts
158 #define SLP_TREE_NUMBER_OF_VEC_STMTS(S) (S)->vec_stmts_size
159 #define SLP_TREE_LOAD_PERMUTATION(S) (S)->load_permutation
161 /* This structure is used in creation of an SLP tree. Each instance
162 corresponds to the same operand in a group of scalar stmts in an SLP
163 node. */
164 typedef struct _slp_oprnd_info
166 /* Def-stmts for the operands. */
167 vec<gimple> def_stmts;
168 /* Information about the first statement, its vector def-type, type, the
169 operand itself in case it's constant, and an indication if it's a pattern
170 stmt. */
171 enum vect_def_type first_dt;
172 tree first_op_type;
173 bool first_pattern;
174 } *slp_oprnd_info;
178 typedef struct _vect_peel_info
180 int npeel;
181 struct data_reference *dr;
182 unsigned int count;
183 } *vect_peel_info;
185 typedef struct _vect_peel_extended_info
187 struct _vect_peel_info peel_info;
188 unsigned int inside_cost;
189 unsigned int outside_cost;
190 stmt_vector_for_cost body_cost_vec;
191 } *vect_peel_extended_info;
194 /* Peeling hashtable helpers. */
196 struct peel_info_hasher : typed_free_remove <_vect_peel_info>
198 typedef _vect_peel_info value_type;
199 typedef _vect_peel_info compare_type;
200 static inline hashval_t hash (const value_type *);
201 static inline bool equal (const value_type *, const compare_type *);
204 inline hashval_t
205 peel_info_hasher::hash (const value_type *peel_info)
207 return (hashval_t) peel_info->npeel;
210 inline bool
211 peel_info_hasher::equal (const value_type *a, const compare_type *b)
213 return (a->npeel == b->npeel);
217 /*-----------------------------------------------------------------*/
218 /* Info on vectorized loops. */
219 /*-----------------------------------------------------------------*/
220 typedef struct _loop_vec_info {
222 /* The loop to which this info struct refers to. */
223 struct loop *loop;
225 /* The loop basic blocks. */
226 basic_block *bbs;
228 /* Number of iterations. */
229 tree num_iters;
230 tree num_iters_unchanged;
232 /* Minimum number of iterations below which vectorization is expected to
233 not be profitable (as estimated by the cost model).
234 -1 indicates that vectorization will not be profitable.
235 FORNOW: This field is an int. Will be a tree in the future, to represent
236 values unknown at compile time. */
237 int min_profitable_iters;
239 /* Is the loop vectorizable? */
240 bool vectorizable;
242 /* Unrolling factor */
243 int vectorization_factor;
245 /* The loop location in the source. */
246 LOC loop_line_number;
248 /* Unknown DRs according to which loop was peeled. */
249 struct data_reference *unaligned_dr;
251 /* peeling_for_alignment indicates whether peeling for alignment will take
252 place, and what the peeling factor should be:
253 peeling_for_alignment = X means:
254 If X=0: Peeling for alignment will not be applied.
255 If X>0: Peel first X iterations.
256 If X=-1: Generate a runtime test to calculate the number of iterations
257 to be peeled, using the dataref recorded in the field
258 unaligned_dr. */
259 int peeling_for_alignment;
261 /* The mask used to check the alignment of pointers or arrays. */
262 int ptr_mask;
264 /* The loop nest in which the data dependences are computed. */
265 vec<loop_p> loop_nest;
267 /* All data references in the loop. */
268 vec<data_reference_p> datarefs;
270 /* All data dependences in the loop. */
271 vec<ddr_p> ddrs;
273 /* Data Dependence Relations defining address ranges that are candidates
274 for a run-time aliasing check. */
275 vec<ddr_p> may_alias_ddrs;
277 /* Statements in the loop that have data references that are candidates for a
278 runtime (loop versioning) misalignment check. */
279 vec<gimple> may_misalign_stmts;
281 /* All interleaving chains of stores in the loop, represented by the first
282 stmt in the chain. */
283 vec<gimple> grouped_stores;
285 /* All SLP instances in the loop. This is a subset of the set of GROUP_STORES
286 of the loop. */
287 vec<slp_instance> slp_instances;
289 /* The unrolling factor needed to SLP the loop. In case of that pure SLP is
290 applied to the loop, i.e., no unrolling is needed, this is 1. */
291 unsigned slp_unrolling_factor;
293 /* Reduction cycles detected in the loop. Used in loop-aware SLP. */
294 vec<gimple> reductions;
296 /* All reduction chains in the loop, represented by the first
297 stmt in the chain. */
298 vec<gimple> reduction_chains;
300 /* Hash table used to choose the best peeling option. */
301 hash_table <peel_info_hasher> peeling_htab;
303 /* Cost data used by the target cost model. */
304 void *target_cost_data;
306 /* When we have grouped data accesses with gaps, we may introduce invalid
307 memory accesses. We peel the last iteration of the loop to prevent
308 this. */
309 bool peeling_for_gaps;
311 /* Reductions are canonicalized so that the last operand is the reduction
312 operand. If this places a constant into RHS1, this decanonicalizes
313 GIMPLE for other phases, so we must track when this has occurred and
314 fix it up. */
315 bool operands_swapped;
317 } *loop_vec_info;
319 /* Access Functions. */
320 #define LOOP_VINFO_LOOP(L) (L)->loop
321 #define LOOP_VINFO_BBS(L) (L)->bbs
322 #define LOOP_VINFO_NITERS(L) (L)->num_iters
323 /* Since LOOP_VINFO_NITERS can change after prologue peeling
324 retain total unchanged scalar loop iterations for cost model. */
325 #define LOOP_VINFO_NITERS_UNCHANGED(L) (L)->num_iters_unchanged
326 #define LOOP_VINFO_COST_MODEL_MIN_ITERS(L) (L)->min_profitable_iters
327 #define LOOP_VINFO_VECTORIZABLE_P(L) (L)->vectorizable
328 #define LOOP_VINFO_VECT_FACTOR(L) (L)->vectorization_factor
329 #define LOOP_VINFO_PTR_MASK(L) (L)->ptr_mask
330 #define LOOP_VINFO_LOOP_NEST(L) (L)->loop_nest
331 #define LOOP_VINFO_DATAREFS(L) (L)->datarefs
332 #define LOOP_VINFO_DDRS(L) (L)->ddrs
333 #define LOOP_VINFO_INT_NITERS(L) (TREE_INT_CST_LOW ((L)->num_iters))
334 #define LOOP_PEELING_FOR_ALIGNMENT(L) (L)->peeling_for_alignment
335 #define LOOP_VINFO_UNALIGNED_DR(L) (L)->unaligned_dr
336 #define LOOP_VINFO_MAY_MISALIGN_STMTS(L) (L)->may_misalign_stmts
337 #define LOOP_VINFO_LOC(L) (L)->loop_line_number
338 #define LOOP_VINFO_MAY_ALIAS_DDRS(L) (L)->may_alias_ddrs
339 #define LOOP_VINFO_GROUPED_STORES(L) (L)->grouped_stores
340 #define LOOP_VINFO_SLP_INSTANCES(L) (L)->slp_instances
341 #define LOOP_VINFO_SLP_UNROLLING_FACTOR(L) (L)->slp_unrolling_factor
342 #define LOOP_VINFO_REDUCTIONS(L) (L)->reductions
343 #define LOOP_VINFO_REDUCTION_CHAINS(L) (L)->reduction_chains
344 #define LOOP_VINFO_PEELING_HTAB(L) (L)->peeling_htab
345 #define LOOP_VINFO_TARGET_COST_DATA(L) (L)->target_cost_data
346 #define LOOP_VINFO_PEELING_FOR_GAPS(L) (L)->peeling_for_gaps
347 #define LOOP_VINFO_OPERANDS_SWAPPED(L) (L)->operands_swapped
349 #define LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT(L) \
350 (L)->may_misalign_stmts.length () > 0
351 #define LOOP_REQUIRES_VERSIONING_FOR_ALIAS(L) \
352 (L)->may_alias_ddrs.length () > 0
354 #define NITERS_KNOWN_P(n) \
355 (host_integerp ((n),0) \
356 && TREE_INT_CST_LOW ((n)) > 0)
358 #define LOOP_VINFO_NITERS_KNOWN_P(L) \
359 NITERS_KNOWN_P ((L)->num_iters)
361 static inline loop_vec_info
362 loop_vec_info_for_loop (struct loop *loop)
364 return (loop_vec_info) loop->aux;
367 static inline bool
368 nested_in_vect_loop_p (struct loop *loop, gimple stmt)
370 return (loop->inner
371 && (loop->inner == (gimple_bb (stmt))->loop_father));
374 typedef struct _bb_vec_info {
376 basic_block bb;
377 /* All interleaving chains of stores in the basic block, represented by the
378 first stmt in the chain. */
379 vec<gimple> grouped_stores;
381 /* All SLP instances in the basic block. This is a subset of the set of
382 GROUP_STORES of the basic block. */
383 vec<slp_instance> slp_instances;
385 /* All data references in the basic block. */
386 vec<data_reference_p> datarefs;
388 /* All data dependences in the basic block. */
389 vec<ddr_p> ddrs;
391 /* Cost data used by the target cost model. */
392 void *target_cost_data;
394 } *bb_vec_info;
396 #define BB_VINFO_BB(B) (B)->bb
397 #define BB_VINFO_GROUPED_STORES(B) (B)->grouped_stores
398 #define BB_VINFO_SLP_INSTANCES(B) (B)->slp_instances
399 #define BB_VINFO_DATAREFS(B) (B)->datarefs
400 #define BB_VINFO_DDRS(B) (B)->ddrs
401 #define BB_VINFO_TARGET_COST_DATA(B) (B)->target_cost_data
403 static inline bb_vec_info
404 vec_info_for_bb (basic_block bb)
406 return (bb_vec_info) bb->aux;
409 /*-----------------------------------------------------------------*/
410 /* Info on vectorized defs. */
411 /*-----------------------------------------------------------------*/
412 enum stmt_vec_info_type {
413 undef_vec_info_type = 0,
414 load_vec_info_type,
415 store_vec_info_type,
416 shift_vec_info_type,
417 op_vec_info_type,
418 call_vec_info_type,
419 assignment_vec_info_type,
420 condition_vec_info_type,
421 reduc_vec_info_type,
422 induc_vec_info_type,
423 type_promotion_vec_info_type,
424 type_demotion_vec_info_type,
425 type_conversion_vec_info_type,
426 loop_exit_ctrl_vec_info_type
429 /* Indicates whether/how a variable is used in the scope of loop/basic
430 block. */
431 enum vect_relevant {
432 vect_unused_in_scope = 0,
433 /* The def is in the inner loop, and the use is in the outer loop, and the
434 use is a reduction stmt. */
435 vect_used_in_outer_by_reduction,
436 /* The def is in the inner loop, and the use is in the outer loop (and is
437 not part of reduction). */
438 vect_used_in_outer,
440 /* defs that feed computations that end up (only) in a reduction. These
441 defs may be used by non-reduction stmts, but eventually, any
442 computations/values that are affected by these defs are used to compute
443 a reduction (i.e. don't get stored to memory, for example). We use this
444 to identify computations that we can change the order in which they are
445 computed. */
446 vect_used_by_reduction,
448 vect_used_in_scope
451 /* The type of vectorization that can be applied to the stmt: regular loop-based
452 vectorization; pure SLP - the stmt is a part of SLP instances and does not
453 have uses outside SLP instances; or hybrid SLP and loop-based - the stmt is
454 a part of SLP instance and also must be loop-based vectorized, since it has
455 uses outside SLP sequences.
457 In the loop context the meanings of pure and hybrid SLP are slightly
458 different. By saying that pure SLP is applied to the loop, we mean that we
459 exploit only intra-iteration parallelism in the loop; i.e., the loop can be
460 vectorized without doing any conceptual unrolling, cause we don't pack
461 together stmts from different iterations, only within a single iteration.
462 Loop hybrid SLP means that we exploit both intra-iteration and
463 inter-iteration parallelism (e.g., number of elements in the vector is 4
464 and the slp-group-size is 2, in which case we don't have enough parallelism
465 within an iteration, so we obtain the rest of the parallelism from subsequent
466 iterations by unrolling the loop by 2). */
467 enum slp_vect_type {
468 loop_vect = 0,
469 pure_slp,
470 hybrid
474 typedef struct data_reference *dr_p;
476 typedef struct _stmt_vec_info {
478 enum stmt_vec_info_type type;
480 /* Indicates whether this stmts is part of a computation whose result is
481 used outside the loop. */
482 bool live;
484 /* Stmt is part of some pattern (computation idiom) */
485 bool in_pattern_p;
487 /* The stmt to which this info struct refers to. */
488 gimple stmt;
490 /* The loop_vec_info with respect to which STMT is vectorized. */
491 loop_vec_info loop_vinfo;
493 /* The vector type to be used for the LHS of this statement. */
494 tree vectype;
496 /* The vectorized version of the stmt. */
497 gimple vectorized_stmt;
500 /** The following is relevant only for stmts that contain a non-scalar
501 data-ref (array/pointer/struct access). A GIMPLE stmt is expected to have
502 at most one such data-ref. **/
504 /* Information about the data-ref (access function, etc),
505 relative to the inner-most containing loop. */
506 struct data_reference *data_ref_info;
508 /* Information about the data-ref relative to this loop
509 nest (the loop that is being considered for vectorization). */
510 tree dr_base_address;
511 tree dr_init;
512 tree dr_offset;
513 tree dr_step;
514 tree dr_aligned_to;
516 /* For loop PHI nodes, the evolution part of it. This makes sure
517 this information is still available in vect_update_ivs_after_vectorizer
518 where we may not be able to re-analyze the PHI nodes evolution as
519 peeling for the prologue loop can make it unanalyzable. The evolution
520 part is still correct though. */
521 tree loop_phi_evolution_part;
523 /* Used for various bookkeeping purposes, generally holding a pointer to
524 some other stmt S that is in some way "related" to this stmt.
525 Current use of this field is:
526 If this stmt is part of a pattern (i.e. the field 'in_pattern_p' is
527 true): S is the "pattern stmt" that represents (and replaces) the
528 sequence of stmts that constitutes the pattern. Similarly, the
529 related_stmt of the "pattern stmt" points back to this stmt (which is
530 the last stmt in the original sequence of stmts that constitutes the
531 pattern). */
532 gimple related_stmt;
534 /* Used to keep a sequence of def stmts of a pattern stmt if such exists. */
535 gimple_seq pattern_def_seq;
537 /* List of datarefs that are known to have the same alignment as the dataref
538 of this stmt. */
539 vec<dr_p> same_align_refs;
541 /* Classify the def of this stmt. */
542 enum vect_def_type def_type;
544 /* Whether the stmt is SLPed, loop-based vectorized, or both. */
545 enum slp_vect_type slp_type;
547 /* Interleaving and reduction chains info. */
548 /* First element in the group. */
549 gimple first_element;
550 /* Pointer to the next element in the group. */
551 gimple next_element;
552 /* For data-refs, in case that two or more stmts share data-ref, this is the
553 pointer to the previously detected stmt with the same dr. */
554 gimple same_dr_stmt;
555 /* The size of the group. */
556 unsigned int size;
557 /* For stores, number of stores from this group seen. We vectorize the last
558 one. */
559 unsigned int store_count;
560 /* For loads only, the gap from the previous load. For consecutive loads, GAP
561 is 1. */
562 unsigned int gap;
564 /* Not all stmts in the loop need to be vectorized. e.g, the increment
565 of the loop induction variable and computation of array indexes. relevant
566 indicates whether the stmt needs to be vectorized. */
567 enum vect_relevant relevant;
569 /* The bb_vec_info with respect to which STMT is vectorized. */
570 bb_vec_info bb_vinfo;
572 /* Is this statement vectorizable or should it be skipped in (partial)
573 vectorization. */
574 bool vectorizable;
576 /* For loads only, true if this is a gather load. */
577 bool gather_p;
578 bool stride_load_p;
580 /* For both loads and stores. */
581 bool simd_lane_access_p;
582 } *stmt_vec_info;
584 /* Access Functions. */
585 #define STMT_VINFO_TYPE(S) (S)->type
586 #define STMT_VINFO_STMT(S) (S)->stmt
587 #define STMT_VINFO_LOOP_VINFO(S) (S)->loop_vinfo
588 #define STMT_VINFO_BB_VINFO(S) (S)->bb_vinfo
589 #define STMT_VINFO_RELEVANT(S) (S)->relevant
590 #define STMT_VINFO_LIVE_P(S) (S)->live
591 #define STMT_VINFO_VECTYPE(S) (S)->vectype
592 #define STMT_VINFO_VEC_STMT(S) (S)->vectorized_stmt
593 #define STMT_VINFO_VECTORIZABLE(S) (S)->vectorizable
594 #define STMT_VINFO_DATA_REF(S) (S)->data_ref_info
595 #define STMT_VINFO_GATHER_P(S) (S)->gather_p
596 #define STMT_VINFO_STRIDE_LOAD_P(S) (S)->stride_load_p
597 #define STMT_VINFO_SIMD_LANE_ACCESS_P(S) (S)->simd_lane_access_p
599 #define STMT_VINFO_DR_BASE_ADDRESS(S) (S)->dr_base_address
600 #define STMT_VINFO_DR_INIT(S) (S)->dr_init
601 #define STMT_VINFO_DR_OFFSET(S) (S)->dr_offset
602 #define STMT_VINFO_DR_STEP(S) (S)->dr_step
603 #define STMT_VINFO_DR_ALIGNED_TO(S) (S)->dr_aligned_to
605 #define STMT_VINFO_IN_PATTERN_P(S) (S)->in_pattern_p
606 #define STMT_VINFO_RELATED_STMT(S) (S)->related_stmt
607 #define STMT_VINFO_PATTERN_DEF_SEQ(S) (S)->pattern_def_seq
608 #define STMT_VINFO_SAME_ALIGN_REFS(S) (S)->same_align_refs
609 #define STMT_VINFO_DEF_TYPE(S) (S)->def_type
610 #define STMT_VINFO_GROUP_FIRST_ELEMENT(S) (S)->first_element
611 #define STMT_VINFO_GROUP_NEXT_ELEMENT(S) (S)->next_element
612 #define STMT_VINFO_GROUP_SIZE(S) (S)->size
613 #define STMT_VINFO_GROUP_STORE_COUNT(S) (S)->store_count
614 #define STMT_VINFO_GROUP_GAP(S) (S)->gap
615 #define STMT_VINFO_GROUP_SAME_DR_STMT(S) (S)->same_dr_stmt
616 #define STMT_VINFO_GROUPED_ACCESS(S) ((S)->first_element != NULL && (S)->data_ref_info)
617 #define STMT_VINFO_LOOP_PHI_EVOLUTION_PART(S) (S)->loop_phi_evolution_part
619 #define GROUP_FIRST_ELEMENT(S) (S)->first_element
620 #define GROUP_NEXT_ELEMENT(S) (S)->next_element
621 #define GROUP_SIZE(S) (S)->size
622 #define GROUP_STORE_COUNT(S) (S)->store_count
623 #define GROUP_GAP(S) (S)->gap
624 #define GROUP_SAME_DR_STMT(S) (S)->same_dr_stmt
626 #define STMT_VINFO_RELEVANT_P(S) ((S)->relevant != vect_unused_in_scope)
628 #define HYBRID_SLP_STMT(S) ((S)->slp_type == hybrid)
629 #define PURE_SLP_STMT(S) ((S)->slp_type == pure_slp)
630 #define STMT_SLP_TYPE(S) (S)->slp_type
632 struct dataref_aux {
633 tree base_decl;
634 bool base_misaligned;
635 int misalignment;
638 #define VECT_MAX_COST 1000
640 /* The maximum number of intermediate steps required in multi-step type
641 conversion. */
642 #define MAX_INTERM_CVT_STEPS 3
644 /* The maximum vectorization factor supported by any target (V32QI). */
645 #define MAX_VECTORIZATION_FACTOR 32
647 /* Avoid GTY(()) on stmt_vec_info. */
648 typedef void *vec_void_p;
650 extern vec<vec_void_p> stmt_vec_info_vec;
652 void init_stmt_vec_info_vec (void);
653 void free_stmt_vec_info_vec (void);
655 /* Return a stmt_vec_info corresponding to STMT. */
657 static inline stmt_vec_info
658 vinfo_for_stmt (gimple stmt)
660 unsigned int uid = gimple_uid (stmt);
661 if (uid == 0)
662 return NULL;
664 return (stmt_vec_info) stmt_vec_info_vec[uid - 1];
667 /* Set vectorizer information INFO for STMT. */
669 static inline void
670 set_vinfo_for_stmt (gimple stmt, stmt_vec_info info)
672 unsigned int uid = gimple_uid (stmt);
673 if (uid == 0)
675 gcc_checking_assert (info);
676 uid = stmt_vec_info_vec.length () + 1;
677 gimple_set_uid (stmt, uid);
678 stmt_vec_info_vec.safe_push ((vec_void_p) info);
680 else
681 stmt_vec_info_vec[uid - 1] = (vec_void_p) info;
684 /* Return the earlier statement between STMT1 and STMT2. */
686 static inline gimple
687 get_earlier_stmt (gimple stmt1, gimple stmt2)
689 unsigned int uid1, uid2;
691 if (stmt1 == NULL)
692 return stmt2;
694 if (stmt2 == NULL)
695 return stmt1;
697 uid1 = gimple_uid (stmt1);
698 uid2 = gimple_uid (stmt2);
700 if (uid1 == 0 || uid2 == 0)
701 return NULL;
703 gcc_checking_assert (uid1 <= stmt_vec_info_vec.length ()
704 && uid2 <= stmt_vec_info_vec.length ());
706 if (uid1 < uid2)
707 return stmt1;
708 else
709 return stmt2;
712 /* Return the later statement between STMT1 and STMT2. */
714 static inline gimple
715 get_later_stmt (gimple stmt1, gimple stmt2)
717 unsigned int uid1, uid2;
719 if (stmt1 == NULL)
720 return stmt2;
722 if (stmt2 == NULL)
723 return stmt1;
725 uid1 = gimple_uid (stmt1);
726 uid2 = gimple_uid (stmt2);
728 if (uid1 == 0 || uid2 == 0)
729 return NULL;
731 gcc_assert (uid1 <= stmt_vec_info_vec.length ());
732 gcc_assert (uid2 <= stmt_vec_info_vec.length ());
734 if (uid1 > uid2)
735 return stmt1;
736 else
737 return stmt2;
740 /* Return TRUE if a statement represented by STMT_INFO is a part of a
741 pattern. */
743 static inline bool
744 is_pattern_stmt_p (stmt_vec_info stmt_info)
746 gimple related_stmt;
747 stmt_vec_info related_stmt_info;
749 related_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
750 if (related_stmt
751 && (related_stmt_info = vinfo_for_stmt (related_stmt))
752 && STMT_VINFO_IN_PATTERN_P (related_stmt_info))
753 return true;
755 return false;
758 /* Return true if BB is a loop header. */
760 static inline bool
761 is_loop_header_bb_p (basic_block bb)
763 if (bb == (bb->loop_father)->header)
764 return true;
765 gcc_checking_assert (EDGE_COUNT (bb->preds) == 1);
766 return false;
769 /* Return pow2 (X). */
771 static inline int
772 vect_pow2 (int x)
774 int i, res = 1;
776 for (i = 0; i < x; i++)
777 res *= 2;
779 return res;
782 /* Alias targetm.vectorize.builtin_vectorization_cost. */
784 static inline int
785 builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost,
786 tree vectype, int misalign)
788 return targetm.vectorize.builtin_vectorization_cost (type_of_cost,
789 vectype, misalign);
792 /* Get cost by calling cost target builtin. */
794 static inline
795 int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost)
797 return builtin_vectorization_cost (type_of_cost, NULL, 0);
800 /* Alias targetm.vectorize.init_cost. */
802 static inline void *
803 init_cost (struct loop *loop_info)
805 return targetm.vectorize.init_cost (loop_info);
808 /* Alias targetm.vectorize.add_stmt_cost. */
810 static inline unsigned
811 add_stmt_cost (void *data, int count, enum vect_cost_for_stmt kind,
812 stmt_vec_info stmt_info, int misalign,
813 enum vect_cost_model_location where)
815 return targetm.vectorize.add_stmt_cost (data, count, kind,
816 stmt_info, misalign, where);
819 /* Alias targetm.vectorize.finish_cost. */
821 static inline void
822 finish_cost (void *data, unsigned *prologue_cost,
823 unsigned *body_cost, unsigned *epilogue_cost)
825 targetm.vectorize.finish_cost (data, prologue_cost, body_cost, epilogue_cost);
828 /* Alias targetm.vectorize.destroy_cost_data. */
830 static inline void
831 destroy_cost_data (void *data)
833 targetm.vectorize.destroy_cost_data (data);
837 /*-----------------------------------------------------------------*/
838 /* Info on data references alignment. */
839 /*-----------------------------------------------------------------*/
840 inline void
841 set_dr_misalignment (struct data_reference *dr, int val)
843 dataref_aux *data_aux = (dataref_aux *) dr->aux;
845 if (!data_aux)
847 data_aux = XCNEW (dataref_aux);
848 dr->aux = data_aux;
851 data_aux->misalignment = val;
854 inline int
855 dr_misalignment (struct data_reference *dr)
857 gcc_assert (dr->aux);
858 return ((dataref_aux *) dr->aux)->misalignment;
861 /* Reflects actual alignment of first access in the vectorized loop,
862 taking into account peeling/versioning if applied. */
863 #define DR_MISALIGNMENT(DR) dr_misalignment (DR)
864 #define SET_DR_MISALIGNMENT(DR, VAL) set_dr_misalignment (DR, VAL)
866 /* Return TRUE if the data access is aligned, and FALSE otherwise. */
868 static inline bool
869 aligned_access_p (struct data_reference *data_ref_info)
871 return (DR_MISALIGNMENT (data_ref_info) == 0);
874 /* Return TRUE if the alignment of the data access is known, and FALSE
875 otherwise. */
877 static inline bool
878 known_alignment_for_access_p (struct data_reference *data_ref_info)
880 return (DR_MISALIGNMENT (data_ref_info) != -1);
884 /* Return true if the vect cost model is unlimited. */
885 static inline bool
886 unlimited_cost_model ()
888 return flag_vect_cost_model == VECT_COST_MODEL_UNLIMITED;
891 /* Source location */
892 extern LOC vect_location;
894 /*-----------------------------------------------------------------*/
895 /* Function prototypes. */
896 /*-----------------------------------------------------------------*/
898 /* Simple loop peeling and versioning utilities for vectorizer's purposes -
899 in tree-vect-loop-manip.c. */
900 extern void slpeel_make_loop_iterate_ntimes (struct loop *, tree);
901 extern bool slpeel_can_duplicate_loop_p (const struct loop *, const_edge);
902 struct loop *slpeel_tree_duplicate_loop_to_edge_cfg (struct loop *, edge);
903 extern void vect_loop_versioning (loop_vec_info, unsigned int, bool);
904 extern void vect_do_peeling_for_loop_bound (loop_vec_info, tree *,
905 unsigned int, bool);
906 extern void vect_do_peeling_for_alignment (loop_vec_info, unsigned int, bool);
907 extern LOC find_loop_location (struct loop *);
908 extern bool vect_can_advance_ivs_p (loop_vec_info);
910 /* In tree-vect-stmts.c. */
911 extern unsigned int current_vector_size;
912 extern tree get_vectype_for_scalar_type (tree);
913 extern tree get_same_sized_vectype (tree, tree);
914 extern bool vect_is_simple_use (tree, gimple, loop_vec_info,
915 bb_vec_info, gimple *,
916 tree *, enum vect_def_type *);
917 extern bool vect_is_simple_use_1 (tree, gimple, loop_vec_info,
918 bb_vec_info, gimple *,
919 tree *, enum vect_def_type *, tree *);
920 extern bool supportable_widening_operation (enum tree_code, gimple, tree, tree,
921 enum tree_code *, enum tree_code *,
922 int *, vec<tree> *);
923 extern bool supportable_narrowing_operation (enum tree_code, tree, tree,
924 enum tree_code *,
925 int *, vec<tree> *);
926 extern stmt_vec_info new_stmt_vec_info (gimple stmt, loop_vec_info,
927 bb_vec_info);
928 extern void free_stmt_vec_info (gimple stmt);
929 extern tree vectorizable_function (gimple, tree, tree);
930 extern void vect_model_simple_cost (stmt_vec_info, int, enum vect_def_type *,
931 stmt_vector_for_cost *,
932 stmt_vector_for_cost *);
933 extern void vect_model_store_cost (stmt_vec_info, int, bool,
934 enum vect_def_type, slp_tree,
935 stmt_vector_for_cost *,
936 stmt_vector_for_cost *);
937 extern void vect_model_load_cost (stmt_vec_info, int, bool, slp_tree,
938 stmt_vector_for_cost *,
939 stmt_vector_for_cost *);
940 extern unsigned record_stmt_cost (stmt_vector_for_cost *, int,
941 enum vect_cost_for_stmt, stmt_vec_info,
942 int, enum vect_cost_model_location);
943 extern void vect_finish_stmt_generation (gimple, gimple,
944 gimple_stmt_iterator *);
945 extern bool vect_mark_stmts_to_be_vectorized (loop_vec_info);
946 extern tree vect_get_vec_def_for_operand (tree, gimple, tree *);
947 extern tree vect_init_vector (gimple, tree, tree,
948 gimple_stmt_iterator *);
949 extern tree vect_get_vec_def_for_stmt_copy (enum vect_def_type, tree);
950 extern bool vect_transform_stmt (gimple, gimple_stmt_iterator *,
951 bool *, slp_tree, slp_instance);
952 extern void vect_remove_stores (gimple);
953 extern bool vect_analyze_stmt (gimple, bool *, slp_tree);
954 extern bool vectorizable_condition (gimple, gimple_stmt_iterator *, gimple *,
955 tree, int, slp_tree);
956 extern void vect_get_load_cost (struct data_reference *, int, bool,
957 unsigned int *, unsigned int *,
958 stmt_vector_for_cost *,
959 stmt_vector_for_cost *, bool);
960 extern void vect_get_store_cost (struct data_reference *, int,
961 unsigned int *, stmt_vector_for_cost *);
962 extern bool vect_supportable_shift (enum tree_code, tree);
963 extern void vect_get_vec_defs (tree, tree, gimple, vec<tree> *,
964 vec<tree> *, slp_tree, int);
965 extern tree vect_gen_perm_mask (tree, unsigned char *);
967 /* In tree-vect-data-refs.c. */
968 extern bool vect_can_force_dr_alignment_p (const_tree, unsigned int);
969 extern enum dr_alignment_support vect_supportable_dr_alignment
970 (struct data_reference *, bool);
971 extern tree vect_get_smallest_scalar_type (gimple, HOST_WIDE_INT *,
972 HOST_WIDE_INT *);
973 extern bool vect_analyze_data_ref_dependences (loop_vec_info, int *);
974 extern bool vect_slp_analyze_data_ref_dependences (bb_vec_info);
975 extern bool vect_enhance_data_refs_alignment (loop_vec_info);
976 extern bool vect_analyze_data_refs_alignment (loop_vec_info, bb_vec_info);
977 extern bool vect_verify_datarefs_alignment (loop_vec_info, bb_vec_info);
978 extern bool vect_analyze_data_ref_accesses (loop_vec_info, bb_vec_info);
979 extern bool vect_prune_runtime_alias_test_list (loop_vec_info);
980 extern tree vect_check_gather (gimple, loop_vec_info, tree *, tree *,
981 int *);
982 extern bool vect_analyze_data_refs (loop_vec_info, bb_vec_info, int *);
983 extern tree vect_create_data_ref_ptr (gimple, tree, struct loop *, tree,
984 tree *, gimple_stmt_iterator *,
985 gimple *, bool, bool *);
986 extern tree bump_vector_ptr (tree, gimple, gimple_stmt_iterator *, gimple, tree);
987 extern tree vect_create_destination_var (tree, tree);
988 extern bool vect_grouped_store_supported (tree, unsigned HOST_WIDE_INT);
989 extern bool vect_store_lanes_supported (tree, unsigned HOST_WIDE_INT);
990 extern bool vect_grouped_load_supported (tree, unsigned HOST_WIDE_INT);
991 extern bool vect_load_lanes_supported (tree, unsigned HOST_WIDE_INT);
992 extern void vect_permute_store_chain (vec<tree> ,unsigned int, gimple,
993 gimple_stmt_iterator *, vec<tree> *);
994 extern tree vect_setup_realignment (gimple, gimple_stmt_iterator *, tree *,
995 enum dr_alignment_support, tree,
996 struct loop **);
997 extern void vect_transform_grouped_load (gimple, vec<tree> , int,
998 gimple_stmt_iterator *);
999 extern void vect_record_grouped_load_vectors (gimple, vec<tree> );
1000 extern tree vect_get_new_vect_var (tree, enum vect_var_kind, const char *);
1001 extern tree vect_create_addr_base_for_vector_ref (gimple, gimple_seq *,
1002 tree, struct loop *);
1004 /* In tree-vect-loop.c. */
1005 /* FORNOW: Used in tree-parloops.c. */
1006 extern void destroy_loop_vec_info (loop_vec_info, bool);
1007 extern gimple vect_force_simple_reduction (loop_vec_info, gimple, bool, bool *);
1008 /* Drive for loop analysis stage. */
1009 extern loop_vec_info vect_analyze_loop (struct loop *);
1010 /* Drive for loop transformation stage. */
1011 extern void vect_transform_loop (loop_vec_info);
1012 extern loop_vec_info vect_analyze_loop_form (struct loop *);
1013 extern bool vectorizable_live_operation (gimple, gimple_stmt_iterator *,
1014 gimple *);
1015 extern bool vectorizable_reduction (gimple, gimple_stmt_iterator *, gimple *,
1016 slp_tree);
1017 extern bool vectorizable_induction (gimple, gimple_stmt_iterator *, gimple *);
1018 extern tree get_initial_def_for_reduction (gimple, tree, tree *);
1019 extern int vect_min_worthwhile_factor (enum tree_code);
1020 extern int vect_get_known_peeling_cost (loop_vec_info, int, int *, int,
1021 stmt_vector_for_cost *,
1022 stmt_vector_for_cost *);
1023 extern int vect_get_single_scalar_iteration_cost (loop_vec_info);
1025 /* In tree-vect-slp.c. */
1026 extern void vect_free_slp_instance (slp_instance);
1027 extern bool vect_transform_slp_perm_load (slp_tree, vec<tree> ,
1028 gimple_stmt_iterator *, int,
1029 slp_instance, bool);
1030 extern bool vect_schedule_slp (loop_vec_info, bb_vec_info);
1031 extern void vect_update_slp_costs_according_to_vf (loop_vec_info);
1032 extern bool vect_analyze_slp (loop_vec_info, bb_vec_info);
1033 extern bool vect_make_slp_decision (loop_vec_info);
1034 extern void vect_detect_hybrid_slp (loop_vec_info);
1035 extern void vect_get_slp_defs (vec<tree> , slp_tree,
1036 vec<vec<tree> > *, int);
1038 extern LOC find_bb_location (basic_block);
1039 extern bb_vec_info vect_slp_analyze_bb (basic_block);
1040 extern void vect_slp_transform_bb (basic_block);
1042 /* In tree-vect-patterns.c. */
1043 /* Pattern recognition functions.
1044 Additional pattern recognition functions can (and will) be added
1045 in the future. */
1046 typedef gimple (* vect_recog_func_ptr) (vec<gimple> *, tree *, tree *);
1047 #define NUM_PATTERNS 11
1048 void vect_pattern_recog (loop_vec_info, bb_vec_info);
1050 /* In tree-vectorizer.c. */
1051 unsigned vectorize_loops (void);
1052 void vect_destroy_datarefs (loop_vec_info, bb_vec_info);
1054 #endif /* GCC_TREE_VECTORIZER_H */