2 Copyright (C) 2003-2018 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
27 stmt_vec_info (struct _stmt_vec_info
*ptr
) : m_ptr (ptr
) {}
28 struct _stmt_vec_info
*operator-> () const { return m_ptr
; }
29 struct _stmt_vec_info
&operator* () const;
30 operator struct _stmt_vec_info
* () const { return m_ptr
; }
31 operator gimple
* () const;
32 operator void * () const { return m_ptr
; }
33 operator bool () const { return m_ptr
; }
34 bool operator == (const stmt_vec_info
&x
) { return x
.m_ptr
== m_ptr
; }
35 bool operator == (_stmt_vec_info
*x
) { return x
== m_ptr
; }
36 bool operator != (const stmt_vec_info
&x
) { return x
.m_ptr
!= m_ptr
; }
37 bool operator != (_stmt_vec_info
*x
) { return x
!= m_ptr
; }
40 struct _stmt_vec_info
*m_ptr
;
43 #define NULL_STMT_VEC_INFO (stmt_vec_info (NULL))
45 #include "tree-data-ref.h"
46 #include "tree-hash-traits.h"
49 /* Used for naming of new temporaries. */
57 /* Defines type of operation. */
64 /* Define type of available alignment support. */
65 enum dr_alignment_support
{
66 dr_unaligned_unsupported
,
67 dr_unaligned_supported
,
69 dr_explicit_realign_optimized
,
73 /* Define type of def-use cross-iteration cycle. */
75 vect_uninitialized_def
= 0,
76 vect_constant_def
= 1,
81 vect_double_reduction_def
,
86 /* Define type of reduction. */
87 enum vect_reduction_type
{
90 INTEGER_INDUC_COND_REDUCTION
,
93 /* Retain a scalar phi and use a FOLD_EXTRACT_LAST within the loop
96 for (int i = 0; i < VF; ++i)
97 res = cond[i] ? val[i] : res; */
98 EXTRACT_LAST_REDUCTION
,
100 /* Use a folding reduction within the loop to implement:
102 for (int i = 0; i < VF; ++i)
105 (with no reassocation). */
109 #define VECTORIZABLE_CYCLE_DEF(D) (((D) == vect_reduction_def) \
110 || ((D) == vect_double_reduction_def) \
111 || ((D) == vect_nested_cycle))
113 /* Structure to encapsulate information about a group of like
114 instructions to be presented to the target cost model. */
115 struct stmt_info_for_cost
{
117 enum vect_cost_for_stmt kind
;
118 enum vect_cost_model_location where
;
119 stmt_vec_info stmt_info
;
123 typedef vec
<stmt_info_for_cost
> stmt_vector_for_cost
;
125 /* Maps base addresses to an innermost_loop_behavior that gives the maximum
126 known alignment for that base. */
127 typedef hash_map
<tree_operand_hash
,
128 innermost_loop_behavior
*> vec_base_alignments
;
130 /************************************************************************
132 ************************************************************************/
133 typedef struct _slp_tree
*slp_tree
;
135 /* A computation tree of an SLP instance. Each node corresponds to a group of
136 stmts to be packed in a SIMD stmt. */
138 /* Nodes that contain def-stmts of this node statements operands. */
139 vec
<slp_tree
> children
;
140 /* A group of scalar stmts to be vectorized together. */
141 vec
<stmt_vec_info
> stmts
;
142 /* Load permutation relative to the stores, NULL if there is no
144 vec
<unsigned> load_permutation
;
145 /* Vectorized stmt/s. */
146 vec
<stmt_vec_info
> vec_stmts
;
147 /* Number of vector stmts that are created to replace the group of scalar
148 stmts. It is calculated during the transformation phase as the number of
149 scalar elements in one scalar iteration (GROUP_SIZE) multiplied by VF
150 divided by vector size. */
151 unsigned int vec_stmts_size
;
152 /* Whether the scalar computations use two different operators. */
154 /* The DEF type of this node. */
155 enum vect_def_type def_type
;
159 /* SLP instance is a sequence of stmts in a loop that can be packed into
161 typedef struct _slp_instance
{
162 /* The root of SLP tree. */
165 /* Size of groups of scalar stmts that will be replaced by SIMD stmt/s. */
166 unsigned int group_size
;
168 /* The unrolling factor required to vectorized this SLP instance. */
169 poly_uint64 unrolling_factor
;
171 /* The group of nodes that contain loads of this SLP instance. */
174 /* The SLP node containing the reduction PHIs. */
179 /* Access Functions. */
180 #define SLP_INSTANCE_TREE(S) (S)->root
181 #define SLP_INSTANCE_GROUP_SIZE(S) (S)->group_size
182 #define SLP_INSTANCE_UNROLLING_FACTOR(S) (S)->unrolling_factor
183 #define SLP_INSTANCE_LOADS(S) (S)->loads
185 #define SLP_TREE_CHILDREN(S) (S)->children
186 #define SLP_TREE_SCALAR_STMTS(S) (S)->stmts
187 #define SLP_TREE_VEC_STMTS(S) (S)->vec_stmts
188 #define SLP_TREE_NUMBER_OF_VEC_STMTS(S) (S)->vec_stmts_size
189 #define SLP_TREE_LOAD_PERMUTATION(S) (S)->load_permutation
190 #define SLP_TREE_TWO_OPERATORS(S) (S)->two_operators
191 #define SLP_TREE_DEF_TYPE(S) (S)->def_type
195 /* Describes two objects whose addresses must be unequal for the vectorized
197 typedef std::pair
<tree
, tree
> vec_object_pair
;
199 /* Records that vectorization is only possible if abs (EXPR) >= MIN_VALUE.
200 UNSIGNED_P is true if we can assume that abs (EXPR) == EXPR. */
201 struct vec_lower_bound
{
202 vec_lower_bound () {}
203 vec_lower_bound (tree e
, bool u
, poly_uint64 m
)
204 : expr (e
), unsigned_p (u
), min_value (m
) {}
208 poly_uint64 min_value
;
211 /* Vectorizer state shared between different analyses like vector sizes
212 of the same CFG region. */
213 struct vec_info_shared
{
217 void save_datarefs();
218 void check_datarefs();
220 /* All data references. Freed by free_data_refs, so not an auto_vec. */
221 vec
<data_reference_p
> datarefs
;
222 vec
<data_reference
> datarefs_copy
;
224 /* The loop nest in which the data dependences are computed. */
225 auto_vec
<loop_p
> loop_nest
;
227 /* All data dependences. Freed by free_dependence_relations, so not
232 /* Vectorizer state common between loop and basic-block vectorization. */
234 enum vec_kind
{ bb
, loop
};
236 vec_info (vec_kind
, void *, vec_info_shared
*);
239 stmt_vec_info
add_stmt (gimple
*);
240 stmt_vec_info
lookup_stmt (gimple
*);
241 stmt_vec_info
lookup_def (tree
);
242 stmt_vec_info
lookup_single_use (tree
);
243 struct dr_vec_info
*lookup_dr (data_reference
*);
244 void move_dr (stmt_vec_info
, stmt_vec_info
);
245 void remove_stmt (stmt_vec_info
);
246 void replace_stmt (gimple_stmt_iterator
*, stmt_vec_info
, gimple
*);
248 /* The type of vectorization. */
251 /* Shared vectorizer state. */
252 vec_info_shared
*shared
;
254 /* The mapping of GIMPLE UID to stmt_vec_info. */
255 vec
<stmt_vec_info
> stmt_vec_infos
;
257 /* All SLP instances. */
258 auto_vec
<slp_instance
> slp_instances
;
260 /* Maps base addresses to an innermost_loop_behavior that gives the maximum
261 known alignment for that base. */
262 vec_base_alignments base_alignments
;
264 /* All interleaving chains of stores, represented by the first
265 stmt in the chain. */
266 auto_vec
<stmt_vec_info
> grouped_stores
;
268 /* Cost data used by the target cost model. */
269 void *target_cost_data
;
272 stmt_vec_info
new_stmt_vec_info (gimple
*stmt
);
273 void set_vinfo_for_stmt (gimple
*, stmt_vec_info
);
274 void free_stmt_vec_infos ();
275 void free_stmt_vec_info (stmt_vec_info
);
278 struct _loop_vec_info
;
284 is_a_helper
<_loop_vec_info
*>::test (vec_info
*i
)
286 return i
->kind
== vec_info::loop
;
292 is_a_helper
<_bb_vec_info
*>::test (vec_info
*i
)
294 return i
->kind
== vec_info::bb
;
298 /* In general, we can divide the vector statements in a vectorized loop
299 into related groups ("rgroups") and say that for each rgroup there is
300 some nS such that the rgroup operates on nS values from one scalar
301 iteration followed by nS values from the next. That is, if VF is the
302 vectorization factor of the loop, the rgroup operates on a sequence:
304 (1,1) (1,2) ... (1,nS) (2,1) ... (2,nS) ... (VF,1) ... (VF,nS)
306 where (i,j) represents a scalar value with index j in a scalar
307 iteration with index i.
309 [ We use the term "rgroup" to emphasise that this grouping isn't
310 necessarily the same as the grouping of statements used elsewhere.
311 For example, if we implement a group of scalar loads using gather
312 loads, we'll use a separate gather load for each scalar load, and
313 thus each gather load will belong to its own rgroup. ]
315 In general this sequence will occupy nV vectors concatenated
316 together. If these vectors have nL lanes each, the total number
317 of scalar values N is given by:
319 N = nS * VF = nV * nL
321 None of nS, VF, nV and nL are required to be a power of 2. nS and nV
322 are compile-time constants but VF and nL can be variable (if the target
323 supports variable-length vectors).
325 In classical vectorization, each iteration of the vector loop would
326 handle exactly VF iterations of the original scalar loop. However,
327 in a fully-masked loop, a particular iteration of the vector loop
328 might handle fewer than VF iterations of the scalar loop. The vector
329 lanes that correspond to iterations of the scalar loop are said to be
330 "active" and the other lanes are said to be "inactive".
332 In a fully-masked loop, many rgroups need to be masked to ensure that
333 they have no effect for the inactive lanes. Each such rgroup needs a
334 sequence of booleans in the same order as above, but with each (i,j)
335 replaced by a boolean that indicates whether iteration i is active.
336 This sequence occupies nV vector masks that again have nL lanes each.
337 Thus the mask sequence as a whole consists of VF independent booleans
338 that are each repeated nS times.
340 We make the simplifying assumption that if a sequence of nV masks is
341 suitable for one (nS,nL) pair, we can reuse it for (nS/2,nL/2) by
342 VIEW_CONVERTing it. This holds for all current targets that support
343 fully-masked loops. For example, suppose the scalar loop is:
347 for (int i = 0; i < n; ++i)
349 f[i * 2 + 0] += 1.0f;
350 f[i * 2 + 1] += 2.0f;
354 and suppose that vectors have 256 bits. The vectorized f accesses
355 will belong to one rgroup and the vectorized d access to another:
357 f rgroup: nS = 2, nV = 1, nL = 8
358 d rgroup: nS = 1, nV = 1, nL = 4
361 [ In this simple example the rgroups do correspond to the normal
362 SLP grouping scheme. ]
364 If only the first three lanes are active, the masks we need are:
366 f rgroup: 1 1 | 1 1 | 1 1 | 0 0
367 d rgroup: 1 | 1 | 1 | 0
369 Here we can use a mask calculated for f's rgroup for d's, but not
372 Thus for each value of nV, it is enough to provide nV masks, with the
373 mask being calculated based on the highest nL (or, equivalently, based
374 on the highest nS) required by any rgroup with that nV. We therefore
375 represent the entire collection of masks as a two-level table, with the
376 first level being indexed by nV - 1 (since nV == 0 doesn't exist) and
377 the second being indexed by the mask index 0 <= i < nV. */
379 /* The masks needed by rgroups with nV vectors, according to the
380 description above. */
381 struct rgroup_masks
{
382 /* The largest nS for all rgroups that use these masks. */
383 unsigned int max_nscalars_per_iter
;
385 /* The type of mask to use, based on the highest nS recorded above. */
388 /* A vector of nV masks, in iteration order. */
392 typedef auto_vec
<rgroup_masks
> vec_loop_masks
;
394 /*-----------------------------------------------------------------*/
395 /* Info on vectorized loops. */
396 /*-----------------------------------------------------------------*/
397 typedef struct _loop_vec_info
: public vec_info
{
398 _loop_vec_info (struct loop
*, vec_info_shared
*);
401 /* The loop to which this info struct refers to. */
404 /* The loop basic blocks. */
407 /* Number of latch executions. */
409 /* Number of iterations. */
411 /* Number of iterations of the original loop. */
412 tree num_iters_unchanged
;
413 /* Condition under which this loop is analyzed and versioned. */
414 tree num_iters_assumptions
;
416 /* Threshold of number of iterations below which vectorzation will not be
417 performed. It is calculated from MIN_PROFITABLE_ITERS and
418 PARAM_MIN_VECT_LOOP_BOUND. */
421 /* When applying loop versioning, the vector form should only be used
422 if the number of scalar iterations is >= this value, on top of all
423 the other requirements. Ignored when loop versioning is not being
425 poly_uint64 versioning_threshold
;
427 /* Unrolling factor */
428 poly_uint64 vectorization_factor
;
430 /* Maximum runtime vectorization factor, or MAX_VECTORIZATION_FACTOR
431 if there is no particular limit. */
432 unsigned HOST_WIDE_INT max_vectorization_factor
;
434 /* The masks that a fully-masked loop should use to avoid operating
435 on inactive scalars. */
436 vec_loop_masks masks
;
438 /* If we are using a loop mask to align memory addresses, this variable
439 contains the number of vector elements that we should skip in the
440 first iteration of the vector loop (i.e. the number of leading
441 elements that should be false in the first mask). */
442 tree mask_skip_niters
;
444 /* Type of the variables to use in the WHILE_ULT call for fully-masked
446 tree mask_compare_type
;
448 /* Unknown DRs according to which loop was peeled. */
449 struct dr_vec_info
*unaligned_dr
;
451 /* peeling_for_alignment indicates whether peeling for alignment will take
452 place, and what the peeling factor should be:
453 peeling_for_alignment = X means:
454 If X=0: Peeling for alignment will not be applied.
455 If X>0: Peel first X iterations.
456 If X=-1: Generate a runtime test to calculate the number of iterations
457 to be peeled, using the dataref recorded in the field
459 int peeling_for_alignment
;
461 /* The mask used to check the alignment of pointers or arrays. */
464 /* Data Dependence Relations defining address ranges that are candidates
465 for a run-time aliasing check. */
466 auto_vec
<ddr_p
> may_alias_ddrs
;
468 /* Data Dependence Relations defining address ranges together with segment
469 lengths from which the run-time aliasing check is built. */
470 auto_vec
<dr_with_seg_len_pair_t
> comp_alias_ddrs
;
472 /* Check that the addresses of each pair of objects is unequal. */
473 auto_vec
<vec_object_pair
> check_unequal_addrs
;
475 /* List of values that are required to be nonzero. This is used to check
476 whether things like "x[i * n] += 1;" are safe and eventually gets added
477 to the checks for lower bounds below. */
478 auto_vec
<tree
> check_nonzero
;
480 /* List of values that need to be checked for a minimum value. */
481 auto_vec
<vec_lower_bound
> lower_bounds
;
483 /* Statements in the loop that have data references that are candidates for a
484 runtime (loop versioning) misalignment check. */
485 auto_vec
<stmt_vec_info
> may_misalign_stmts
;
487 /* Reduction cycles detected in the loop. Used in loop-aware SLP. */
488 auto_vec
<stmt_vec_info
> reductions
;
490 /* All reduction chains in the loop, represented by the first
491 stmt in the chain. */
492 auto_vec
<stmt_vec_info
> reduction_chains
;
494 /* Cost vector for a single scalar iteration. */
495 auto_vec
<stmt_info_for_cost
> scalar_cost_vec
;
497 /* Map of IV base/step expressions to inserted name in the preheader. */
498 hash_map
<tree_operand_hash
, tree
> *ivexpr_map
;
500 /* The unrolling factor needed to SLP the loop. In case of that pure SLP is
501 applied to the loop, i.e., no unrolling is needed, this is 1. */
502 poly_uint64 slp_unrolling_factor
;
504 /* Cost of a single scalar iteration. */
505 int single_scalar_iteration_cost
;
507 /* Is the loop vectorizable? */
510 /* Records whether we still have the option of using a fully-masked loop. */
511 bool can_fully_mask_p
;
513 /* True if have decided to use a fully-masked loop. */
516 /* When we have grouped data accesses with gaps, we may introduce invalid
517 memory accesses. We peel the last iteration of the loop to prevent
519 bool peeling_for_gaps
;
521 /* When the number of iterations is not a multiple of the vector size
522 we need to peel off iterations at the end to form an epilogue loop. */
523 bool peeling_for_niter
;
525 /* Reductions are canonicalized so that the last operand is the reduction
526 operand. If this places a constant into RHS1, this decanonicalizes
527 GIMPLE for other phases, so we must track when this has occurred and
529 bool operands_swapped
;
531 /* True if there are no loop carried data dependencies in the loop.
532 If loop->safelen <= 1, then this is always true, either the loop
533 didn't have any loop carried data dependencies, or the loop is being
534 vectorized guarded with some runtime alias checks, or couldn't
535 be vectorized at all, but then this field shouldn't be used.
536 For loop->safelen >= 2, the user has asserted that there are no
537 backward dependencies, but there still could be loop carried forward
538 dependencies in such loops. This flag will be false if normal
539 vectorizer data dependency analysis would fail or require versioning
540 for alias, but because of loop->safelen >= 2 it has been vectorized
541 even without versioning for alias. E.g. in:
543 for (int i = 0; i < m; i++)
545 (or #pragma simd or #pragma ivdep) we can vectorize this and it will
546 DTRT even for k > 0 && k < m, but without safelen we would not
547 vectorize this, so this field would be false. */
548 bool no_data_dependencies
;
550 /* Mark loops having masked stores. */
553 /* If if-conversion versioned this loop before conversion, this is the
554 loop version without if-conversion. */
555 struct loop
*scalar_loop
;
557 /* For loops being epilogues of already vectorized loops
558 this points to the original vectorized loop. Otherwise NULL. */
559 _loop_vec_info
*orig_loop_info
;
563 /* Access Functions. */
564 #define LOOP_VINFO_LOOP(L) (L)->loop
565 #define LOOP_VINFO_BBS(L) (L)->bbs
566 #define LOOP_VINFO_NITERSM1(L) (L)->num_itersm1
567 #define LOOP_VINFO_NITERS(L) (L)->num_iters
568 /* Since LOOP_VINFO_NITERS and LOOP_VINFO_NITERSM1 can change after
569 prologue peeling retain total unchanged scalar loop iterations for
571 #define LOOP_VINFO_NITERS_UNCHANGED(L) (L)->num_iters_unchanged
572 #define LOOP_VINFO_NITERS_ASSUMPTIONS(L) (L)->num_iters_assumptions
573 #define LOOP_VINFO_COST_MODEL_THRESHOLD(L) (L)->th
574 #define LOOP_VINFO_VERSIONING_THRESHOLD(L) (L)->versioning_threshold
575 #define LOOP_VINFO_VECTORIZABLE_P(L) (L)->vectorizable
576 #define LOOP_VINFO_CAN_FULLY_MASK_P(L) (L)->can_fully_mask_p
577 #define LOOP_VINFO_FULLY_MASKED_P(L) (L)->fully_masked_p
578 #define LOOP_VINFO_VECT_FACTOR(L) (L)->vectorization_factor
579 #define LOOP_VINFO_MAX_VECT_FACTOR(L) (L)->max_vectorization_factor
580 #define LOOP_VINFO_MASKS(L) (L)->masks
581 #define LOOP_VINFO_MASK_SKIP_NITERS(L) (L)->mask_skip_niters
582 #define LOOP_VINFO_MASK_COMPARE_TYPE(L) (L)->mask_compare_type
583 #define LOOP_VINFO_PTR_MASK(L) (L)->ptr_mask
584 #define LOOP_VINFO_LOOP_NEST(L) (L)->shared->loop_nest
585 #define LOOP_VINFO_DATAREFS(L) (L)->shared->datarefs
586 #define LOOP_VINFO_DDRS(L) (L)->shared->ddrs
587 #define LOOP_VINFO_INT_NITERS(L) (TREE_INT_CST_LOW ((L)->num_iters))
588 #define LOOP_VINFO_PEELING_FOR_ALIGNMENT(L) (L)->peeling_for_alignment
589 #define LOOP_VINFO_UNALIGNED_DR(L) (L)->unaligned_dr
590 #define LOOP_VINFO_MAY_MISALIGN_STMTS(L) (L)->may_misalign_stmts
591 #define LOOP_VINFO_MAY_ALIAS_DDRS(L) (L)->may_alias_ddrs
592 #define LOOP_VINFO_COMP_ALIAS_DDRS(L) (L)->comp_alias_ddrs
593 #define LOOP_VINFO_CHECK_UNEQUAL_ADDRS(L) (L)->check_unequal_addrs
594 #define LOOP_VINFO_CHECK_NONZERO(L) (L)->check_nonzero
595 #define LOOP_VINFO_LOWER_BOUNDS(L) (L)->lower_bounds
596 #define LOOP_VINFO_GROUPED_STORES(L) (L)->grouped_stores
597 #define LOOP_VINFO_SLP_INSTANCES(L) (L)->slp_instances
598 #define LOOP_VINFO_SLP_UNROLLING_FACTOR(L) (L)->slp_unrolling_factor
599 #define LOOP_VINFO_REDUCTIONS(L) (L)->reductions
600 #define LOOP_VINFO_REDUCTION_CHAINS(L) (L)->reduction_chains
601 #define LOOP_VINFO_TARGET_COST_DATA(L) (L)->target_cost_data
602 #define LOOP_VINFO_PEELING_FOR_GAPS(L) (L)->peeling_for_gaps
603 #define LOOP_VINFO_OPERANDS_SWAPPED(L) (L)->operands_swapped
604 #define LOOP_VINFO_PEELING_FOR_NITER(L) (L)->peeling_for_niter
605 #define LOOP_VINFO_NO_DATA_DEPENDENCIES(L) (L)->no_data_dependencies
606 #define LOOP_VINFO_SCALAR_LOOP(L) (L)->scalar_loop
607 #define LOOP_VINFO_HAS_MASK_STORE(L) (L)->has_mask_store
608 #define LOOP_VINFO_SCALAR_ITERATION_COST(L) (L)->scalar_cost_vec
609 #define LOOP_VINFO_SINGLE_SCALAR_ITERATION_COST(L) (L)->single_scalar_iteration_cost
610 #define LOOP_VINFO_ORIG_LOOP_INFO(L) (L)->orig_loop_info
612 #define LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT(L) \
613 ((L)->may_misalign_stmts.length () > 0)
614 #define LOOP_REQUIRES_VERSIONING_FOR_ALIAS(L) \
615 ((L)->comp_alias_ddrs.length () > 0 \
616 || (L)->check_unequal_addrs.length () > 0 \
617 || (L)->lower_bounds.length () > 0)
618 #define LOOP_REQUIRES_VERSIONING_FOR_NITERS(L) \
619 (LOOP_VINFO_NITERS_ASSUMPTIONS (L))
620 #define LOOP_REQUIRES_VERSIONING(L) \
621 (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (L) \
622 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (L) \
623 || LOOP_REQUIRES_VERSIONING_FOR_NITERS (L))
625 #define LOOP_VINFO_NITERS_KNOWN_P(L) \
626 (tree_fits_shwi_p ((L)->num_iters) && tree_to_shwi ((L)->num_iters) > 0)
628 #define LOOP_VINFO_EPILOGUE_P(L) \
629 (LOOP_VINFO_ORIG_LOOP_INFO (L) != NULL)
631 #define LOOP_VINFO_ORIG_MAX_VECT_FACTOR(L) \
632 (LOOP_VINFO_MAX_VECT_FACTOR (LOOP_VINFO_ORIG_LOOP_INFO (L)))
634 static inline loop_vec_info
635 loop_vec_info_for_loop (struct loop
*loop
)
637 return (loop_vec_info
) loop
->aux
;
640 typedef struct _bb_vec_info
: public vec_info
642 _bb_vec_info (gimple_stmt_iterator
, gimple_stmt_iterator
, vec_info_shared
*);
646 gimple_stmt_iterator region_begin
;
647 gimple_stmt_iterator region_end
;
650 #define BB_VINFO_BB(B) (B)->bb
651 #define BB_VINFO_GROUPED_STORES(B) (B)->grouped_stores
652 #define BB_VINFO_SLP_INSTANCES(B) (B)->slp_instances
653 #define BB_VINFO_DATAREFS(B) (B)->shared->datarefs
654 #define BB_VINFO_DDRS(B) (B)->shared->ddrs
655 #define BB_VINFO_TARGET_COST_DATA(B) (B)->target_cost_data
657 static inline bb_vec_info
658 vec_info_for_bb (basic_block bb
)
660 return (bb_vec_info
) bb
->aux
;
663 /*-----------------------------------------------------------------*/
664 /* Info on vectorized defs. */
665 /*-----------------------------------------------------------------*/
666 enum stmt_vec_info_type
{
667 undef_vec_info_type
= 0,
673 call_simd_clone_vec_info_type
,
674 assignment_vec_info_type
,
675 condition_vec_info_type
,
676 comparison_vec_info_type
,
679 type_promotion_vec_info_type
,
680 type_demotion_vec_info_type
,
681 type_conversion_vec_info_type
,
682 loop_exit_ctrl_vec_info_type
685 /* Indicates whether/how a variable is used in the scope of loop/basic
688 vect_unused_in_scope
= 0,
690 /* The def is only used outside the loop. */
692 /* The def is in the inner loop, and the use is in the outer loop, and the
693 use is a reduction stmt. */
694 vect_used_in_outer_by_reduction
,
695 /* The def is in the inner loop, and the use is in the outer loop (and is
696 not part of reduction). */
699 /* defs that feed computations that end up (only) in a reduction. These
700 defs may be used by non-reduction stmts, but eventually, any
701 computations/values that are affected by these defs are used to compute
702 a reduction (i.e. don't get stored to memory, for example). We use this
703 to identify computations that we can change the order in which they are
705 vect_used_by_reduction
,
710 /* The type of vectorization that can be applied to the stmt: regular loop-based
711 vectorization; pure SLP - the stmt is a part of SLP instances and does not
712 have uses outside SLP instances; or hybrid SLP and loop-based - the stmt is
713 a part of SLP instance and also must be loop-based vectorized, since it has
714 uses outside SLP sequences.
716 In the loop context the meanings of pure and hybrid SLP are slightly
717 different. By saying that pure SLP is applied to the loop, we mean that we
718 exploit only intra-iteration parallelism in the loop; i.e., the loop can be
719 vectorized without doing any conceptual unrolling, cause we don't pack
720 together stmts from different iterations, only within a single iteration.
721 Loop hybrid SLP means that we exploit both intra-iteration and
722 inter-iteration parallelism (e.g., number of elements in the vector is 4
723 and the slp-group-size is 2, in which case we don't have enough parallelism
724 within an iteration, so we obtain the rest of the parallelism from subsequent
725 iterations by unrolling the loop by 2). */
732 /* Says whether a statement is a load, a store of a vectorized statement
733 result, or a store of an invariant value. */
734 enum vec_load_store_type
{
740 /* Describes how we're going to vectorize an individual load or store,
741 or a group of loads or stores. */
742 enum vect_memory_access_type
{
743 /* An access to an invariant address. This is used only for loads. */
746 /* A simple contiguous access. */
749 /* A contiguous access that goes down in memory rather than up,
750 with no additional permutation. This is used only for stores
752 VMAT_CONTIGUOUS_DOWN
,
754 /* A simple contiguous access in which the elements need to be permuted
755 after loading or before storing. Only used for loop vectorization;
756 SLP uses separate permutes. */
757 VMAT_CONTIGUOUS_PERMUTE
,
759 /* A simple contiguous access in which the elements need to be reversed
760 after loading or before storing. */
761 VMAT_CONTIGUOUS_REVERSE
,
763 /* An access that uses IFN_LOAD_LANES or IFN_STORE_LANES. */
764 VMAT_LOAD_STORE_LANES
,
766 /* An access in which each scalar element is loaded or stored
770 /* A hybrid of VMAT_CONTIGUOUS and VMAT_ELEMENTWISE, used for grouped
771 SLP accesses. Each unrolled iteration uses a contiguous load
772 or store for the whole group, but the groups from separate iterations
773 are combined in the same way as for VMAT_ELEMENTWISE. */
776 /* The access uses gather loads or scatter stores. */
781 /* The data reference itself. */
783 /* The statement that contains the data reference. */
785 /* The misalignment in bytes of the reference, or -1 if not known. */
787 /* The byte alignment that we'd ideally like the reference to have,
788 and the value that misalignment is measured against. */
789 int target_alignment
;
790 /* If true the alignment of base_decl needs to be increased. */
791 bool base_misaligned
;
795 typedef struct data_reference
*dr_p
;
797 struct _stmt_vec_info
{
799 enum stmt_vec_info_type type
;
801 /* Indicates whether this stmts is part of a computation whose result is
802 used outside the loop. */
805 /* Stmt is part of some pattern (computation idiom) */
808 /* True if the statement was created during pattern recognition as
809 part of the replacement for RELATED_STMT. This implies that the
810 statement isn't part of any basic block, although for convenience
811 its gimple_bb is the same as for RELATED_STMT. */
814 /* Is this statement vectorizable or should it be skipped in (partial)
818 /* The stmt to which this info struct refers to. */
821 /* The vec_info with respect to which STMT is vectorized. */
824 /* The vector type to be used for the LHS of this statement. */
827 /* The vectorized version of the stmt. */
828 stmt_vec_info vectorized_stmt
;
831 /* The following is relevant only for stmts that contain a non-scalar
832 data-ref (array/pointer/struct access). A GIMPLE stmt is expected to have
833 at most one such data-ref. */
837 /* Information about the data-ref relative to this loop
838 nest (the loop that is being considered for vectorization). */
839 innermost_loop_behavior dr_wrt_vec_loop
;
841 /* For loop PHI nodes, the base and evolution part of it. This makes sure
842 this information is still available in vect_update_ivs_after_vectorizer
843 where we may not be able to re-analyze the PHI nodes evolution as
844 peeling for the prologue loop can make it unanalyzable. The evolution
845 part is still correct after peeling, but the base may have changed from
847 tree loop_phi_evolution_base_unchanged
;
848 tree loop_phi_evolution_part
;
850 /* Used for various bookkeeping purposes, generally holding a pointer to
851 some other stmt S that is in some way "related" to this stmt.
852 Current use of this field is:
853 If this stmt is part of a pattern (i.e. the field 'in_pattern_p' is
854 true): S is the "pattern stmt" that represents (and replaces) the
855 sequence of stmts that constitutes the pattern. Similarly, the
856 related_stmt of the "pattern stmt" points back to this stmt (which is
857 the last stmt in the original sequence of stmts that constitutes the
859 stmt_vec_info related_stmt
;
861 /* Used to keep a sequence of def stmts of a pattern stmt if such exists.
862 The sequence is attached to the original statement rather than the
863 pattern statement. */
864 gimple_seq pattern_def_seq
;
866 /* List of datarefs that are known to have the same alignment as the dataref
868 vec
<dr_p
> same_align_refs
;
870 /* Selected SIMD clone's function info. First vector element
871 is SIMD clone's function decl, followed by a pair of trees (base + step)
872 for linear arguments (pair of NULLs for other arguments). */
873 vec
<tree
> simd_clone_info
;
875 /* Classify the def of this stmt. */
876 enum vect_def_type def_type
;
878 /* Whether the stmt is SLPed, loop-based vectorized, or both. */
879 enum slp_vect_type slp_type
;
881 /* Interleaving and reduction chains info. */
882 /* First element in the group. */
883 stmt_vec_info first_element
;
884 /* Pointer to the next element in the group. */
885 stmt_vec_info next_element
;
886 /* For data-refs, in case that two or more stmts share data-ref, this is the
887 pointer to the previously detected stmt with the same dr. */
888 stmt_vec_info same_dr_stmt
;
889 /* The size of the group. */
891 /* For stores, number of stores from this group seen. We vectorize the last
893 unsigned int store_count
;
894 /* For loads only, the gap from the previous load. For consecutive loads, GAP
898 /* The minimum negative dependence distance this stmt participates in
900 unsigned int min_neg_dist
;
902 /* Not all stmts in the loop need to be vectorized. e.g, the increment
903 of the loop induction variable and computation of array indexes. relevant
904 indicates whether the stmt needs to be vectorized. */
905 enum vect_relevant relevant
;
907 /* For loads if this is a gather, for stores if this is a scatter. */
908 bool gather_scatter_p
;
910 /* True if this is an access with loop-invariant stride. */
913 /* For both loads and stores. */
914 bool simd_lane_access_p
;
916 /* Classifies how the load or store is going to be implemented
917 for loop vectorization. */
918 vect_memory_access_type memory_access_type
;
920 /* For reduction loops, this is the type of reduction. */
921 enum vect_reduction_type v_reduc_type
;
923 /* For CONST_COND_REDUCTION, record the reduc code. */
924 enum tree_code const_cond_reduc_code
;
926 /* On a reduction PHI the reduction type as detected by
927 vect_force_simple_reduction. */
928 enum vect_reduction_type reduc_type
;
930 /* On a reduction PHI the def returned by vect_force_simple_reduction.
931 On the def returned by vect_force_simple_reduction the
932 corresponding PHI. */
933 stmt_vec_info reduc_def
;
935 /* The number of scalar stmt references from active SLP instances. */
936 unsigned int num_slp_uses
;
938 /* If nonzero, the lhs of the statement could be truncated to this
939 many bits without affecting any users of the result. */
940 unsigned int min_output_precision
;
942 /* If nonzero, all non-boolean input operands have the same precision,
943 and they could each be truncated to this many bits without changing
945 unsigned int min_input_precision
;
947 /* If OPERATION_BITS is nonzero, the statement could be performed on
948 an integer with the sign and number of bits given by OPERATION_SIGN
949 and OPERATION_BITS without changing the result. */
950 unsigned int operation_precision
;
951 signop operation_sign
;
954 /* Information about a gather/scatter call. */
955 struct gather_scatter_info
{
956 /* The internal function to use for the gather/scatter operation,
957 or IFN_LAST if a built-in function should be used instead. */
960 /* The FUNCTION_DECL for the built-in gather/scatter function,
961 or null if an internal function should be used instead. */
964 /* The loop-invariant base value. */
967 /* The original scalar offset, which is a non-loop-invariant SSA_NAME. */
970 /* Each offset element should be multiplied by this amount before
971 being added to the base. */
974 /* The definition type for the vectorized offset. */
975 enum vect_def_type offset_dt
;
977 /* The type of the vectorized offset. */
980 /* The type of the scalar elements after loading or before storing. */
983 /* The type of the scalar elements being loaded or stored. */
987 /* Access Functions. */
988 #define STMT_VINFO_TYPE(S) (S)->type
989 #define STMT_VINFO_STMT(S) (S)->stmt
991 STMT_VINFO_LOOP_VINFO (stmt_vec_info stmt_vinfo
)
993 if (loop_vec_info loop_vinfo
= dyn_cast
<loop_vec_info
> (stmt_vinfo
->vinfo
))
998 STMT_VINFO_BB_VINFO (stmt_vec_info stmt_vinfo
)
1000 if (bb_vec_info bb_vinfo
= dyn_cast
<bb_vec_info
> (stmt_vinfo
->vinfo
))
1004 #define STMT_VINFO_RELEVANT(S) (S)->relevant
1005 #define STMT_VINFO_LIVE_P(S) (S)->live
1006 #define STMT_VINFO_VECTYPE(S) (S)->vectype
1007 #define STMT_VINFO_VEC_STMT(S) (S)->vectorized_stmt
1008 #define STMT_VINFO_VECTORIZABLE(S) (S)->vectorizable
1009 #define STMT_VINFO_DATA_REF(S) ((S)->dr_aux.dr + 0)
1010 #define STMT_VINFO_GATHER_SCATTER_P(S) (S)->gather_scatter_p
1011 #define STMT_VINFO_STRIDED_P(S) (S)->strided_p
1012 #define STMT_VINFO_MEMORY_ACCESS_TYPE(S) (S)->memory_access_type
1013 #define STMT_VINFO_SIMD_LANE_ACCESS_P(S) (S)->simd_lane_access_p
1014 #define STMT_VINFO_VEC_REDUCTION_TYPE(S) (S)->v_reduc_type
1015 #define STMT_VINFO_VEC_CONST_COND_REDUC_CODE(S) (S)->const_cond_reduc_code
1017 #define STMT_VINFO_DR_WRT_VEC_LOOP(S) (S)->dr_wrt_vec_loop
1018 #define STMT_VINFO_DR_BASE_ADDRESS(S) (S)->dr_wrt_vec_loop.base_address
1019 #define STMT_VINFO_DR_INIT(S) (S)->dr_wrt_vec_loop.init
1020 #define STMT_VINFO_DR_OFFSET(S) (S)->dr_wrt_vec_loop.offset
1021 #define STMT_VINFO_DR_STEP(S) (S)->dr_wrt_vec_loop.step
1022 #define STMT_VINFO_DR_BASE_ALIGNMENT(S) (S)->dr_wrt_vec_loop.base_alignment
1023 #define STMT_VINFO_DR_BASE_MISALIGNMENT(S) \
1024 (S)->dr_wrt_vec_loop.base_misalignment
1025 #define STMT_VINFO_DR_OFFSET_ALIGNMENT(S) \
1026 (S)->dr_wrt_vec_loop.offset_alignment
1027 #define STMT_VINFO_DR_STEP_ALIGNMENT(S) \
1028 (S)->dr_wrt_vec_loop.step_alignment
1030 #define STMT_VINFO_DR_INFO(S) \
1031 (gcc_checking_assert ((S)->dr_aux.stmt == (S)), &(S)->dr_aux)
1033 #define STMT_VINFO_IN_PATTERN_P(S) (S)->in_pattern_p
1034 #define STMT_VINFO_RELATED_STMT(S) (S)->related_stmt
1035 #define STMT_VINFO_PATTERN_DEF_SEQ(S) (S)->pattern_def_seq
1036 #define STMT_VINFO_SAME_ALIGN_REFS(S) (S)->same_align_refs
1037 #define STMT_VINFO_SIMD_CLONE_INFO(S) (S)->simd_clone_info
1038 #define STMT_VINFO_DEF_TYPE(S) (S)->def_type
1039 #define STMT_VINFO_GROUPED_ACCESS(S) \
1040 ((S)->dr_aux.dr && DR_GROUP_FIRST_ELEMENT(S))
1041 #define STMT_VINFO_LOOP_PHI_EVOLUTION_BASE_UNCHANGED(S) (S)->loop_phi_evolution_base_unchanged
1042 #define STMT_VINFO_LOOP_PHI_EVOLUTION_PART(S) (S)->loop_phi_evolution_part
1043 #define STMT_VINFO_MIN_NEG_DIST(S) (S)->min_neg_dist
1044 #define STMT_VINFO_NUM_SLP_USES(S) (S)->num_slp_uses
1045 #define STMT_VINFO_REDUC_TYPE(S) (S)->reduc_type
1046 #define STMT_VINFO_REDUC_DEF(S) (S)->reduc_def
1048 #define DR_GROUP_FIRST_ELEMENT(S) \
1049 (gcc_checking_assert ((S)->dr_aux.dr), (S)->first_element)
1050 #define DR_GROUP_NEXT_ELEMENT(S) \
1051 (gcc_checking_assert ((S)->dr_aux.dr), (S)->next_element)
1052 #define DR_GROUP_SIZE(S) \
1053 (gcc_checking_assert ((S)->dr_aux.dr), (S)->size)
1054 #define DR_GROUP_STORE_COUNT(S) \
1055 (gcc_checking_assert ((S)->dr_aux.dr), (S)->store_count)
1056 #define DR_GROUP_GAP(S) \
1057 (gcc_checking_assert ((S)->dr_aux.dr), (S)->gap)
1058 #define DR_GROUP_SAME_DR_STMT(S) \
1059 (gcc_checking_assert ((S)->dr_aux.dr), (S)->same_dr_stmt)
1061 #define REDUC_GROUP_FIRST_ELEMENT(S) \
1062 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->first_element)
1063 #define REDUC_GROUP_NEXT_ELEMENT(S) \
1064 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->next_element)
1065 #define REDUC_GROUP_SIZE(S) \
1066 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->size)
1068 #define STMT_VINFO_RELEVANT_P(S) ((S)->relevant != vect_unused_in_scope)
1070 #define HYBRID_SLP_STMT(S) ((S)->slp_type == hybrid)
1071 #define PURE_SLP_STMT(S) ((S)->slp_type == pure_slp)
1072 #define STMT_SLP_TYPE(S) (S)->slp_type
1074 #define VECT_MAX_COST 1000
1076 /* The maximum number of intermediate steps required in multi-step type
1078 #define MAX_INTERM_CVT_STEPS 3
1080 #define MAX_VECTORIZATION_FACTOR INT_MAX
1082 /* Nonzero if TYPE represents a (scalar) boolean type or type
1083 in the middle-end compatible with it (unsigned precision 1 integral
1084 types). Used to determine which types should be vectorized as
1085 VECTOR_BOOLEAN_TYPE_P. */
1087 #define VECT_SCALAR_BOOLEAN_TYPE_P(TYPE) \
1088 (TREE_CODE (TYPE) == BOOLEAN_TYPE \
1089 || ((TREE_CODE (TYPE) == INTEGER_TYPE \
1090 || TREE_CODE (TYPE) == ENUMERAL_TYPE) \
1091 && TYPE_PRECISION (TYPE) == 1 \
1092 && TYPE_UNSIGNED (TYPE)))
1094 inline _stmt_vec_info
&
1095 stmt_vec_info::operator* () const
1100 inline stmt_vec_info::operator gimple
* () const
1102 return m_ptr
? m_ptr
->stmt
: NULL
;
1106 nested_in_vect_loop_p (struct loop
*loop
, stmt_vec_info stmt_info
)
1109 && (loop
->inner
== (gimple_bb (stmt_info
->stmt
))->loop_father
));
1112 /* Return the earlier statement between STMT1_INFO and STMT2_INFO. */
1114 static inline stmt_vec_info
1115 get_earlier_stmt (stmt_vec_info stmt1_info
, stmt_vec_info stmt2_info
)
1117 gcc_checking_assert ((STMT_VINFO_IN_PATTERN_P (stmt1_info
)
1118 || !STMT_VINFO_RELATED_STMT (stmt1_info
))
1119 && (STMT_VINFO_IN_PATTERN_P (stmt2_info
)
1120 || !STMT_VINFO_RELATED_STMT (stmt2_info
)));
1122 if (gimple_uid (stmt1_info
->stmt
) < gimple_uid (stmt2_info
->stmt
))
1128 /* Return the later statement between STMT1_INFO and STMT2_INFO. */
1130 static inline stmt_vec_info
1131 get_later_stmt (stmt_vec_info stmt1_info
, stmt_vec_info stmt2_info
)
1133 gcc_checking_assert ((STMT_VINFO_IN_PATTERN_P (stmt1_info
)
1134 || !STMT_VINFO_RELATED_STMT (stmt1_info
))
1135 && (STMT_VINFO_IN_PATTERN_P (stmt2_info
)
1136 || !STMT_VINFO_RELATED_STMT (stmt2_info
)));
1138 if (gimple_uid (stmt1_info
->stmt
) > gimple_uid (stmt2_info
->stmt
))
1144 /* Return TRUE if a statement represented by STMT_INFO is a part of a
1148 is_pattern_stmt_p (stmt_vec_info stmt_info
)
1150 return stmt_info
->pattern_stmt_p
;
1153 /* Return true if BB is a loop header. */
1156 is_loop_header_bb_p (basic_block bb
)
1158 if (bb
== (bb
->loop_father
)->header
)
1160 gcc_checking_assert (EDGE_COUNT (bb
->preds
) == 1);
1164 /* Return pow2 (X). */
1171 for (i
= 0; i
< x
; i
++)
1177 /* Alias targetm.vectorize.builtin_vectorization_cost. */
1180 builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost
,
1181 tree vectype
, int misalign
)
1183 return targetm
.vectorize
.builtin_vectorization_cost (type_of_cost
,
1187 /* Get cost by calling cost target builtin. */
1190 int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost
)
1192 return builtin_vectorization_cost (type_of_cost
, NULL
, 0);
1195 /* Alias targetm.vectorize.init_cost. */
1197 static inline void *
1198 init_cost (struct loop
*loop_info
)
1200 return targetm
.vectorize
.init_cost (loop_info
);
1203 extern void dump_stmt_cost (FILE *, void *, int, enum vect_cost_for_stmt
,
1204 stmt_vec_info
, int, enum vect_cost_model_location
);
1206 /* Alias targetm.vectorize.add_stmt_cost. */
1208 static inline unsigned
1209 add_stmt_cost (void *data
, int count
, enum vect_cost_for_stmt kind
,
1210 stmt_vec_info stmt_info
, int misalign
,
1211 enum vect_cost_model_location where
)
1213 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1214 dump_stmt_cost (dump_file
, data
, count
, kind
, stmt_info
, misalign
, where
);
1215 return targetm
.vectorize
.add_stmt_cost (data
, count
, kind
,
1216 stmt_info
, misalign
, where
);
1219 /* Alias targetm.vectorize.finish_cost. */
1222 finish_cost (void *data
, unsigned *prologue_cost
,
1223 unsigned *body_cost
, unsigned *epilogue_cost
)
1225 targetm
.vectorize
.finish_cost (data
, prologue_cost
, body_cost
, epilogue_cost
);
1228 /* Alias targetm.vectorize.destroy_cost_data. */
1231 destroy_cost_data (void *data
)
1233 targetm
.vectorize
.destroy_cost_data (data
);
1237 add_stmt_costs (void *data
, stmt_vector_for_cost
*cost_vec
)
1239 stmt_info_for_cost
*cost
;
1241 FOR_EACH_VEC_ELT (*cost_vec
, i
, cost
)
1242 add_stmt_cost (data
, cost
->count
, cost
->kind
, cost
->stmt_info
,
1243 cost
->misalign
, cost
->where
);
1246 /*-----------------------------------------------------------------*/
1247 /* Info on data references alignment. */
1248 /*-----------------------------------------------------------------*/
1249 #define DR_MISALIGNMENT_UNKNOWN (-1)
1250 #define DR_MISALIGNMENT_UNINITIALIZED (-2)
1253 set_dr_misalignment (dr_vec_info
*dr_info
, int val
)
1255 dr_info
->misalignment
= val
;
1259 dr_misalignment (dr_vec_info
*dr_info
)
1261 int misalign
= dr_info
->misalignment
;
1262 gcc_assert (misalign
!= DR_MISALIGNMENT_UNINITIALIZED
);
1266 /* Reflects actual alignment of first access in the vectorized loop,
1267 taking into account peeling/versioning if applied. */
1268 #define DR_MISALIGNMENT(DR) dr_misalignment (DR)
1269 #define SET_DR_MISALIGNMENT(DR, VAL) set_dr_misalignment (DR, VAL)
1271 /* Only defined once DR_MISALIGNMENT is defined. */
1272 #define DR_TARGET_ALIGNMENT(DR) ((DR)->target_alignment)
1274 /* Return true if data access DR_INFO is aligned to its target alignment
1275 (which may be less than a full vector). */
1278 aligned_access_p (dr_vec_info
*dr_info
)
1280 return (DR_MISALIGNMENT (dr_info
) == 0);
1283 /* Return TRUE if the alignment of the data access is known, and FALSE
1287 known_alignment_for_access_p (dr_vec_info
*dr_info
)
1289 return (DR_MISALIGNMENT (dr_info
) != DR_MISALIGNMENT_UNKNOWN
);
1292 /* Return the minimum alignment in bytes that the vectorized version
1293 of DR_INFO is guaranteed to have. */
1295 static inline unsigned int
1296 vect_known_alignment_in_bytes (dr_vec_info
*dr_info
)
1298 if (DR_MISALIGNMENT (dr_info
) == DR_MISALIGNMENT_UNKNOWN
)
1299 return TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr_info
->dr
)));
1300 if (DR_MISALIGNMENT (dr_info
) == 0)
1301 return DR_TARGET_ALIGNMENT (dr_info
);
1302 return DR_MISALIGNMENT (dr_info
) & -DR_MISALIGNMENT (dr_info
);
1305 /* Return the behavior of DR_INFO with respect to the vectorization context
1306 (which for outer loop vectorization might not be the behavior recorded
1307 in DR_INFO itself). */
1309 static inline innermost_loop_behavior
*
1310 vect_dr_behavior (dr_vec_info
*dr_info
)
1312 stmt_vec_info stmt_info
= dr_info
->stmt
;
1313 loop_vec_info loop_vinfo
= STMT_VINFO_LOOP_VINFO (stmt_info
);
1314 if (loop_vinfo
== NULL
1315 || !nested_in_vect_loop_p (LOOP_VINFO_LOOP (loop_vinfo
), stmt_info
))
1316 return &DR_INNERMOST (dr_info
->dr
);
1318 return &STMT_VINFO_DR_WRT_VEC_LOOP (stmt_info
);
1321 /* Return true if the vect cost model is unlimited. */
1323 unlimited_cost_model (loop_p loop
)
1325 if (loop
!= NULL
&& loop
->force_vectorize
1326 && flag_simd_cost_model
!= VECT_COST_MODEL_DEFAULT
)
1327 return flag_simd_cost_model
== VECT_COST_MODEL_UNLIMITED
;
1328 return (flag_vect_cost_model
== VECT_COST_MODEL_UNLIMITED
);
1331 /* Return true if the loop described by LOOP_VINFO is fully-masked and
1332 if the first iteration should use a partial mask in order to achieve
1336 vect_use_loop_mask_for_alignment_p (loop_vec_info loop_vinfo
)
1338 return (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo
)
1339 && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo
));
1342 /* Return the number of vectors of type VECTYPE that are needed to get
1343 NUNITS elements. NUNITS should be based on the vectorization factor,
1344 so it is always a known multiple of the number of elements in VECTYPE. */
1346 static inline unsigned int
1347 vect_get_num_vectors (poly_uint64 nunits
, tree vectype
)
1349 return exact_div (nunits
, TYPE_VECTOR_SUBPARTS (vectype
)).to_constant ();
1352 /* Return the number of copies needed for loop vectorization when
1353 a statement operates on vectors of type VECTYPE. This is the
1354 vectorization factor divided by the number of elements in
1355 VECTYPE and is always known at compile time. */
1357 static inline unsigned int
1358 vect_get_num_copies (loop_vec_info loop_vinfo
, tree vectype
)
1360 return vect_get_num_vectors (LOOP_VINFO_VECT_FACTOR (loop_vinfo
), vectype
);
1363 /* Update maximum unit count *MAX_NUNITS so that it accounts for
1364 the number of units in vector type VECTYPE. *MAX_NUNITS can be 1
1365 if we haven't yet recorded any vector types. */
1368 vect_update_max_nunits (poly_uint64
*max_nunits
, tree vectype
)
1370 /* All unit counts have the form current_vector_size * X for some
1371 rational X, so two unit sizes must have a common multiple.
1372 Everything is a multiple of the initial value of 1. */
1373 poly_uint64 nunits
= TYPE_VECTOR_SUBPARTS (vectype
);
1374 *max_nunits
= force_common_multiple (*max_nunits
, nunits
);
1377 /* Return the vectorization factor that should be used for costing
1378 purposes while vectorizing the loop described by LOOP_VINFO.
1379 Pick a reasonable estimate if the vectorization factor isn't
1380 known at compile time. */
1382 static inline unsigned int
1383 vect_vf_for_cost (loop_vec_info loop_vinfo
)
1385 return estimated_poly_value (LOOP_VINFO_VECT_FACTOR (loop_vinfo
));
1388 /* Estimate the number of elements in VEC_TYPE for costing purposes.
1389 Pick a reasonable estimate if the exact number isn't known at
1392 static inline unsigned int
1393 vect_nunits_for_cost (tree vec_type
)
1395 return estimated_poly_value (TYPE_VECTOR_SUBPARTS (vec_type
));
1398 /* Return the maximum possible vectorization factor for LOOP_VINFO. */
1400 static inline unsigned HOST_WIDE_INT
1401 vect_max_vf (loop_vec_info loop_vinfo
)
1403 unsigned HOST_WIDE_INT vf
;
1404 if (LOOP_VINFO_VECT_FACTOR (loop_vinfo
).is_constant (&vf
))
1406 return MAX_VECTORIZATION_FACTOR
;
1409 /* Return the size of the value accessed by unvectorized data reference
1410 DR_INFO. This is only valid once STMT_VINFO_VECTYPE has been calculated
1411 for the associated gimple statement, since that guarantees that DR_INFO
1412 accesses either a scalar or a scalar equivalent. ("Scalar equivalent"
1413 here includes things like V1SI, which can be vectorized in the same way
1417 vect_get_scalar_dr_size (dr_vec_info
*dr_info
)
1419 return tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_info
->dr
))));
1422 /* Source location + hotness information. */
1423 extern dump_user_location_t vect_location
;
1425 /* A macro for calling:
1426 dump_begin_scope (MSG, vect_location);
1427 via an RAII object, thus printing "=== MSG ===\n" to the dumpfile etc,
1430 once the object goes out of scope, thus capturing the nesting of
1433 #define DUMP_VECT_SCOPE(MSG) \
1434 AUTO_DUMP_SCOPE (MSG, vect_location)
1436 /*-----------------------------------------------------------------*/
1437 /* Function prototypes. */
1438 /*-----------------------------------------------------------------*/
1440 /* Simple loop peeling and versioning utilities for vectorizer's purposes -
1441 in tree-vect-loop-manip.c. */
1442 extern void vect_set_loop_condition (struct loop
*, loop_vec_info
,
1443 tree
, tree
, tree
, bool);
1444 extern bool slpeel_can_duplicate_loop_p (const struct loop
*, const_edge
);
1445 struct loop
*slpeel_tree_duplicate_loop_to_edge_cfg (struct loop
*,
1446 struct loop
*, edge
);
1447 extern void vect_loop_versioning (loop_vec_info
, unsigned int, bool,
1449 extern struct loop
*vect_do_peeling (loop_vec_info
, tree
, tree
,
1450 tree
*, tree
*, tree
*, int, bool, bool);
1451 extern void vect_prepare_for_masked_peels (loop_vec_info
);
1452 extern dump_user_location_t
find_loop_location (struct loop
*);
1453 extern bool vect_can_advance_ivs_p (loop_vec_info
);
1455 /* In tree-vect-stmts.c. */
1456 extern poly_uint64 current_vector_size
;
1457 extern tree
get_vectype_for_scalar_type (tree
);
1458 extern tree
get_vectype_for_scalar_type_and_size (tree
, poly_uint64
);
1459 extern tree
get_mask_type_for_scalar_type (tree
);
1460 extern tree
get_same_sized_vectype (tree
, tree
);
1461 extern bool vect_get_loop_mask_type (loop_vec_info
);
1462 extern bool vect_is_simple_use (tree
, vec_info
*, enum vect_def_type
*,
1463 stmt_vec_info
* = NULL
, gimple
** = NULL
);
1464 extern bool vect_is_simple_use (tree
, vec_info
*, enum vect_def_type
*,
1465 tree
*, stmt_vec_info
* = NULL
,
1467 extern bool supportable_widening_operation (enum tree_code
, stmt_vec_info
,
1468 tree
, tree
, enum tree_code
*,
1469 enum tree_code
*, int *,
1471 extern bool supportable_narrowing_operation (enum tree_code
, tree
, tree
,
1473 int *, vec
<tree
> *);
1474 extern unsigned record_stmt_cost (stmt_vector_for_cost
*, int,
1475 enum vect_cost_for_stmt
, stmt_vec_info
,
1476 int, enum vect_cost_model_location
);
1477 extern stmt_vec_info
vect_finish_replace_stmt (stmt_vec_info
, gimple
*);
1478 extern stmt_vec_info
vect_finish_stmt_generation (stmt_vec_info
, gimple
*,
1479 gimple_stmt_iterator
*);
1480 extern bool vect_mark_stmts_to_be_vectorized (loop_vec_info
);
1481 extern tree
vect_get_store_rhs (stmt_vec_info
);
1482 extern tree
vect_get_vec_def_for_operand_1 (stmt_vec_info
, enum vect_def_type
);
1483 extern tree
vect_get_vec_def_for_operand (tree
, stmt_vec_info
, tree
= NULL
);
1484 extern void vect_get_vec_defs (tree
, tree
, stmt_vec_info
, vec
<tree
> *,
1485 vec
<tree
> *, slp_tree
);
1486 extern void vect_get_vec_defs_for_stmt_copy (vec_info
*,
1487 vec
<tree
> *, vec
<tree
> *);
1488 extern tree
vect_init_vector (stmt_vec_info
, tree
, tree
,
1489 gimple_stmt_iterator
*);
1490 extern tree
vect_get_vec_def_for_stmt_copy (vec_info
*, tree
);
1491 extern bool vect_transform_stmt (stmt_vec_info
, gimple_stmt_iterator
*,
1492 bool *, slp_tree
, slp_instance
);
1493 extern void vect_remove_stores (stmt_vec_info
);
1494 extern bool vect_analyze_stmt (stmt_vec_info
, bool *, slp_tree
, slp_instance
,
1495 stmt_vector_for_cost
*);
1496 extern bool vectorizable_condition (stmt_vec_info
, gimple_stmt_iterator
*,
1497 stmt_vec_info
*, tree
, int, slp_tree
,
1498 stmt_vector_for_cost
*);
1499 extern void vect_get_load_cost (stmt_vec_info
, int, bool,
1500 unsigned int *, unsigned int *,
1501 stmt_vector_for_cost
*,
1502 stmt_vector_for_cost
*, bool);
1503 extern void vect_get_store_cost (stmt_vec_info
, int,
1504 unsigned int *, stmt_vector_for_cost
*);
1505 extern bool vect_supportable_shift (enum tree_code
, tree
);
1506 extern tree
vect_gen_perm_mask_any (tree
, const vec_perm_indices
&);
1507 extern tree
vect_gen_perm_mask_checked (tree
, const vec_perm_indices
&);
1508 extern void optimize_mask_stores (struct loop
*);
1509 extern gcall
*vect_gen_while (tree
, tree
, tree
);
1510 extern tree
vect_gen_while_not (gimple_seq
*, tree
, tree
, tree
);
1511 extern bool vect_get_vector_types_for_stmt (stmt_vec_info
, tree
*, tree
*);
1512 extern tree
vect_get_mask_type_for_stmt (stmt_vec_info
);
1514 /* In tree-vect-data-refs.c. */
1515 extern bool vect_can_force_dr_alignment_p (const_tree
, unsigned int);
1516 extern enum dr_alignment_support vect_supportable_dr_alignment
1517 (dr_vec_info
*, bool);
1518 extern tree
vect_get_smallest_scalar_type (stmt_vec_info
, HOST_WIDE_INT
*,
1520 extern bool vect_analyze_data_ref_dependences (loop_vec_info
, unsigned int *);
1521 extern bool vect_slp_analyze_instance_dependence (slp_instance
);
1522 extern bool vect_enhance_data_refs_alignment (loop_vec_info
);
1523 extern bool vect_analyze_data_refs_alignment (loop_vec_info
);
1524 extern bool vect_verify_datarefs_alignment (loop_vec_info
);
1525 extern bool vect_slp_analyze_and_verify_instance_alignment (slp_instance
);
1526 extern bool vect_analyze_data_ref_accesses (vec_info
*);
1527 extern bool vect_prune_runtime_alias_test_list (loop_vec_info
);
1528 extern bool vect_gather_scatter_fn_p (bool, bool, tree
, tree
, unsigned int,
1529 signop
, int, internal_fn
*, tree
*);
1530 extern bool vect_check_gather_scatter (stmt_vec_info
, loop_vec_info
,
1531 gather_scatter_info
*);
1532 extern bool vect_find_stmt_data_reference (loop_p
, gimple
*,
1533 vec
<data_reference_p
> *);
1534 extern bool vect_analyze_data_refs (vec_info
*, poly_uint64
*);
1535 extern void vect_record_base_alignments (vec_info
*);
1536 extern tree
vect_create_data_ref_ptr (stmt_vec_info
, tree
, struct loop
*, tree
,
1537 tree
*, gimple_stmt_iterator
*,
1538 gimple
**, bool, bool *,
1539 tree
= NULL_TREE
, tree
= NULL_TREE
);
1540 extern tree
bump_vector_ptr (tree
, gimple
*, gimple_stmt_iterator
*,
1541 stmt_vec_info
, tree
);
1542 extern void vect_copy_ref_info (tree
, tree
);
1543 extern tree
vect_create_destination_var (tree
, tree
);
1544 extern bool vect_grouped_store_supported (tree
, unsigned HOST_WIDE_INT
);
1545 extern bool vect_store_lanes_supported (tree
, unsigned HOST_WIDE_INT
, bool);
1546 extern bool vect_grouped_load_supported (tree
, bool, unsigned HOST_WIDE_INT
);
1547 extern bool vect_load_lanes_supported (tree
, unsigned HOST_WIDE_INT
, bool);
1548 extern void vect_permute_store_chain (vec
<tree
> ,unsigned int, stmt_vec_info
,
1549 gimple_stmt_iterator
*, vec
<tree
> *);
1550 extern tree
vect_setup_realignment (stmt_vec_info
, gimple_stmt_iterator
*,
1551 tree
*, enum dr_alignment_support
, tree
,
1553 extern void vect_transform_grouped_load (stmt_vec_info
, vec
<tree
> , int,
1554 gimple_stmt_iterator
*);
1555 extern void vect_record_grouped_load_vectors (stmt_vec_info
, vec
<tree
>);
1556 extern tree
vect_get_new_vect_var (tree
, enum vect_var_kind
, const char *);
1557 extern tree
vect_get_new_ssa_name (tree
, enum vect_var_kind
,
1558 const char * = NULL
);
1559 extern tree
vect_create_addr_base_for_vector_ref (stmt_vec_info
, gimple_seq
*,
1560 tree
, tree
= NULL_TREE
);
1562 /* In tree-vect-loop.c. */
1563 /* FORNOW: Used in tree-parloops.c. */
1564 extern stmt_vec_info
vect_force_simple_reduction (loop_vec_info
, stmt_vec_info
,
1566 /* Used in gimple-loop-interchange.c. */
1567 extern bool check_reduction_path (dump_user_location_t
, loop_p
, gphi
*, tree
,
1569 /* Drive for loop analysis stage. */
1570 extern loop_vec_info
vect_analyze_loop (struct loop
*, loop_vec_info
,
1572 extern tree
vect_build_loop_niters (loop_vec_info
, bool * = NULL
);
1573 extern void vect_gen_vector_loop_niters (loop_vec_info
, tree
, tree
*,
1575 extern tree
vect_halve_mask_nunits (tree
);
1576 extern tree
vect_double_mask_nunits (tree
);
1577 extern void vect_record_loop_mask (loop_vec_info
, vec_loop_masks
*,
1578 unsigned int, tree
);
1579 extern tree
vect_get_loop_mask (gimple_stmt_iterator
*, vec_loop_masks
*,
1580 unsigned int, tree
, unsigned int);
1582 /* Drive for loop transformation stage. */
1583 extern struct loop
*vect_transform_loop (loop_vec_info
);
1584 extern loop_vec_info
vect_analyze_loop_form (struct loop
*, vec_info_shared
*);
1585 extern bool vectorizable_live_operation (stmt_vec_info
, gimple_stmt_iterator
*,
1586 slp_tree
, int, stmt_vec_info
*,
1587 stmt_vector_for_cost
*);
1588 extern bool vectorizable_reduction (stmt_vec_info
, gimple_stmt_iterator
*,
1589 stmt_vec_info
*, slp_tree
, slp_instance
,
1590 stmt_vector_for_cost
*);
1591 extern bool vectorizable_induction (stmt_vec_info
, gimple_stmt_iterator
*,
1592 stmt_vec_info
*, slp_tree
,
1593 stmt_vector_for_cost
*);
1594 extern tree
get_initial_def_for_reduction (stmt_vec_info
, tree
, tree
*);
1595 extern bool vect_worthwhile_without_simd_p (vec_info
*, tree_code
);
1596 extern int vect_get_known_peeling_cost (loop_vec_info
, int, int *,
1597 stmt_vector_for_cost
*,
1598 stmt_vector_for_cost
*,
1599 stmt_vector_for_cost
*);
1600 extern tree
cse_and_gimplify_to_preheader (loop_vec_info
, tree
);
1602 /* In tree-vect-slp.c. */
1603 extern void vect_free_slp_instance (slp_instance
, bool);
1604 extern bool vect_transform_slp_perm_load (slp_tree
, vec
<tree
> ,
1605 gimple_stmt_iterator
*, poly_uint64
,
1606 slp_instance
, bool, unsigned *);
1607 extern bool vect_slp_analyze_operations (vec_info
*);
1608 extern bool vect_schedule_slp (vec_info
*);
1609 extern bool vect_analyze_slp (vec_info
*, unsigned);
1610 extern bool vect_make_slp_decision (loop_vec_info
);
1611 extern void vect_detect_hybrid_slp (loop_vec_info
);
1612 extern void vect_get_slp_defs (vec
<tree
> , slp_tree
, vec
<vec
<tree
> > *);
1613 extern bool vect_slp_bb (basic_block
);
1614 extern stmt_vec_info
vect_find_last_scalar_stmt_in_slp (slp_tree
);
1615 extern bool is_simple_and_all_uses_invariant (stmt_vec_info
, loop_vec_info
);
1616 extern bool can_duplicate_and_interleave_p (unsigned int, machine_mode
,
1617 unsigned int * = NULL
,
1618 tree
* = NULL
, tree
* = NULL
);
1619 extern void duplicate_and_interleave (gimple_seq
*, tree
, vec
<tree
>,
1620 unsigned int, vec
<tree
> &);
1621 extern int vect_get_place_in_interleaving_chain (stmt_vec_info
, stmt_vec_info
);
1623 /* In tree-vect-patterns.c. */
1624 /* Pattern recognition functions.
1625 Additional pattern recognition functions can (and will) be added
1627 void vect_pattern_recog (vec_info
*);
1629 /* In tree-vectorizer.c. */
1630 unsigned vectorize_loops (void);
1631 bool vect_stmt_in_region_p (vec_info
*, gimple
*);
1632 void vect_free_loop_info_assumptions (struct loop
*);
1634 #endif /* GCC_TREE_VECTORIZER_H */