1 /* Vectorizer
2    Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011, 2012
3    Free Software Foundation, Inc.
4    Contributed by Dorit Naishlos <dorit@il.ibm.com>
5 
6 This file is part of GCC.
7 
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
11 version.
12 
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
16 for more details.
17 
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3.  If not see
20 <http://www.gnu.org/licenses/>.  */
21 
22 #ifndef GCC_TREE_VECTORIZER_H
23 #define GCC_TREE_VECTORIZER_H
24 
25 #include "tree-data-ref.h"
26 
27 typedef source_location LOC;
28 #define UNKNOWN_LOC UNKNOWN_LOCATION
29 #define EXPR_LOC(e) EXPR_LOCATION(e)
30 #define LOC_FILE(l) LOCATION_FILE (l)
31 #define LOC_LINE(l) LOCATION_LINE (l)
32 
33 /* Used for naming of new temporaries.  */
34 enum vect_var_kind {
35   vect_simple_var,
36   vect_pointer_var,
37   vect_scalar_var
38 };
39 
40 /* Defines type of operation.  */
41 enum operation_type {
42   unary_op = 1,
43   binary_op,
44   ternary_op
45 };
46 
47 /* Define type of available alignment support.  */
48 enum dr_alignment_support {
49   dr_unaligned_unsupported,
50   dr_unaligned_supported,
51   dr_explicit_realign,
52   dr_explicit_realign_optimized,
53   dr_aligned
54 };
55 
56 /* Define type of def-use cross-iteration cycle.  */
57 enum vect_def_type {
58   vect_uninitialized_def = 0,
59   vect_constant_def = 1,
60   vect_external_def,
61   vect_internal_def,
62   vect_induction_def,
63   vect_reduction_def,
64   vect_double_reduction_def,
65   vect_nested_cycle,
66   vect_unknown_def_type
67 };
68 
69 #define VECTORIZABLE_CYCLE_DEF(D) (((D) == vect_reduction_def)           \
70                                    || ((D) == vect_double_reduction_def) \
71                                    || ((D) == vect_nested_cycle))
72 
73 /************************************************************************
74   SLP
75  ************************************************************************/
76 typedef void *slp_void_p;
77 DEF_VEC_P (slp_void_p);
78 DEF_VEC_ALLOC_P (slp_void_p, heap);
79 
80 /* A computation tree of an SLP instance.  Each node corresponds to a group of
81    stmts to be packed in a SIMD stmt.  */
82 typedef struct _slp_tree {
83   /* Nodes that contain def-stmts of this node statements operands.  */
84   VEC (slp_void_p, heap) *children;
85   /* A group of scalar stmts to be vectorized together.  */
86   VEC (gimple, heap) *stmts;
87   /* Vectorized stmt/s.  */
88   VEC (gimple, heap) *vec_stmts;
89   /* Number of vector stmts that are created to replace the group of scalar
90      stmts. It is calculated during the transformation phase as the number of
91      scalar elements in one scalar iteration (GROUP_SIZE) multiplied by VF
92      divided by vector size.  */
93   unsigned int vec_stmts_size;
94   /* Vectorization costs associated with SLP node.  */
95   struct
96   {
97     int outside_of_loop;     /* Statements generated outside loop.  */
98     int inside_of_loop;      /* Statements generated inside loop.  */
99   } cost;
100 } *slp_tree;
101 
102 DEF_VEC_P(slp_tree);
103 DEF_VEC_ALLOC_P(slp_tree, heap);
104 
105 /* SLP instance is a sequence of stmts in a loop that can be packed into
106    SIMD stmts.  */
107 typedef struct _slp_instance {
108   /* The root of SLP tree.  */
109   slp_tree root;
110 
111   /* Size of groups of scalar stmts that will be replaced by SIMD stmt/s.  */
112   unsigned int group_size;
113 
114   /* The unrolling factor required to vectorized this SLP instance.  */
115   unsigned int unrolling_factor;
116 
117   /* Vectorization costs associated with SLP instance.  */
118   struct
119   {
120     int outside_of_loop;     /* Statements generated outside loop.  */
121     int inside_of_loop;      /* Statements generated inside loop.  */
122   } cost;
123 
124   /* Loads permutation relatively to the stores, NULL if there is no
125      permutation.  */
126   VEC (int, heap) *load_permutation;
127 
128   /* The group of nodes that contain loads of this SLP instance.  */
129   VEC (slp_tree, heap) *loads;
130 
131   /* The first scalar load of the instance. The created vector loads will be
132      inserted before this statement.  */
133   gimple first_load;
134 } *slp_instance;
135 
136 DEF_VEC_P(slp_instance);
137 DEF_VEC_ALLOC_P(slp_instance, heap);
138 
139 /* Access Functions.  */
140 #define SLP_INSTANCE_TREE(S)                     (S)->root
141 #define SLP_INSTANCE_GROUP_SIZE(S)               (S)->group_size
142 #define SLP_INSTANCE_UNROLLING_FACTOR(S)         (S)->unrolling_factor
143 #define SLP_INSTANCE_OUTSIDE_OF_LOOP_COST(S)     (S)->cost.outside_of_loop
144 #define SLP_INSTANCE_INSIDE_OF_LOOP_COST(S)      (S)->cost.inside_of_loop
145 #define SLP_INSTANCE_LOAD_PERMUTATION(S)         (S)->load_permutation
146 #define SLP_INSTANCE_LOADS(S)                    (S)->loads
147 #define SLP_INSTANCE_FIRST_LOAD_STMT(S)          (S)->first_load
148 
149 #define SLP_TREE_CHILDREN(S)                     (S)->children
150 #define SLP_TREE_SCALAR_STMTS(S)                 (S)->stmts
151 #define SLP_TREE_VEC_STMTS(S)                    (S)->vec_stmts
152 #define SLP_TREE_NUMBER_OF_VEC_STMTS(S)          (S)->vec_stmts_size
153 #define SLP_TREE_OUTSIDE_OF_LOOP_COST(S)         (S)->cost.outside_of_loop
154 #define SLP_TREE_INSIDE_OF_LOOP_COST(S)          (S)->cost.inside_of_loop
155 
156 /* This structure is used in creation of an SLP tree.  Each instance
157    corresponds to the same operand in a group of scalar stmts in an SLP
158    node.  */
159 typedef struct _slp_oprnd_info
160 {
161   /* Def-stmts for the operands.  */
162   VEC (gimple, heap) *def_stmts;
163   /* Information about the first statement, its vector def-type, type, the
164      operand itself in case it's constant, and an indication if it's a pattern
165      stmt.  */
166   enum vect_def_type first_dt;
167   tree first_def_type;
168   tree first_const_oprnd;
169   bool first_pattern;
170 } *slp_oprnd_info;
171 
172 DEF_VEC_P(slp_oprnd_info);
173 DEF_VEC_ALLOC_P(slp_oprnd_info, heap);
174 
175 
176 typedef struct _vect_peel_info
177 {
178   int npeel;
179   struct data_reference *dr;
180   unsigned int count;
181 } *vect_peel_info;
182 
183 typedef struct _vect_peel_extended_info
184 {
185   struct _vect_peel_info peel_info;
186   unsigned int inside_cost;
187   unsigned int outside_cost;
188 } *vect_peel_extended_info;
189 
190 /*-----------------------------------------------------------------*/
191 /* Info on vectorized loops.                                       */
192 /*-----------------------------------------------------------------*/
193 typedef struct _loop_vec_info {
194 
195   /* The loop to which this info struct refers to.  */
196   struct loop *loop;
197 
198   /* The loop basic blocks.  */
199   basic_block *bbs;
200 
201   /* Number of iterations.  */
202   tree num_iters;
203   tree num_iters_unchanged;
204 
205   /* Minimum number of iterations below which vectorization is expected to
206      not be profitable (as estimated by the cost model).
207      -1 indicates that vectorization will not be profitable.
208      FORNOW: This field is an int. Will be a tree in the future, to represent
209 	     values unknown at compile time.  */
210   int min_profitable_iters;
211 
212   /* Is the loop vectorizable? */
213   bool vectorizable;
214 
215   /* Unrolling factor  */
216   int vectorization_factor;
217 
218   /* The loop location in the source.  */
219   LOC loop_line_number;
220 
221   /* Unknown DRs according to which loop was peeled.  */
222   struct data_reference *unaligned_dr;
223 
224   /* peeling_for_alignment indicates whether peeling for alignment will take
225      place, and what the peeling factor should be:
226      peeling_for_alignment = X means:
227         If X=0: Peeling for alignment will not be applied.
228         If X>0: Peel first X iterations.
229         If X=-1: Generate a runtime test to calculate the number of iterations
230                  to be peeled, using the dataref recorded in the field
231                  unaligned_dr.  */
232   int peeling_for_alignment;
233 
234   /* The mask used to check the alignment of pointers or arrays.  */
235   int ptr_mask;
236 
237   /* The loop nest in which the data dependences are computed.  */
238   VEC (loop_p, heap) *loop_nest;
239 
240   /* All data references in the loop.  */
241   VEC (data_reference_p, heap) *datarefs;
242 
243   /* All data dependences in the loop.  */
244   VEC (ddr_p, heap) *ddrs;
245 
246   /* Data Dependence Relations defining address ranges that are candidates
247      for a run-time aliasing check.  */
248   VEC (ddr_p, heap) *may_alias_ddrs;
249 
250   /* Statements in the loop that have data references that are candidates for a
251      runtime (loop versioning) misalignment check.  */
252   VEC(gimple,heap) *may_misalign_stmts;
253 
254   /* All interleaving chains of stores in the loop, represented by the first
255      stmt in the chain.  */
256   VEC(gimple, heap) *strided_stores;
257 
258   /* All SLP instances in the loop. This is a subset of the set of STRIDED_STORES
259      of the loop.  */
260   VEC(slp_instance, heap) *slp_instances;
261 
262   /* The unrolling factor needed to SLP the loop. In case of that pure SLP is
263      applied to the loop, i.e., no unrolling is needed, this is 1.  */
264   unsigned slp_unrolling_factor;
265 
266   /* Reduction cycles detected in the loop. Used in loop-aware SLP.  */
267   VEC (gimple, heap) *reductions;
268 
269   /* All reduction chains in the loop, represented by the first
270      stmt in the chain.  */
271   VEC (gimple, heap) *reduction_chains;
272 
273   /* Hash table used to choose the best peeling option.  */
274   htab_t peeling_htab;
275 
276   /* When we have strided data accesses with gaps, we may introduce invalid
277      memory accesses.  We peel the last iteration of the loop to prevent
278      this.  */
279   bool peeling_for_gaps;
280 
281 } *loop_vec_info;
282 
283 /* Access Functions.  */
284 #define LOOP_VINFO_LOOP(L)                 (L)->loop
285 #define LOOP_VINFO_BBS(L)                  (L)->bbs
286 #define LOOP_VINFO_NITERS(L)               (L)->num_iters
287 /* Since LOOP_VINFO_NITERS can change after prologue peeling
288    retain total unchanged scalar loop iterations for cost model.  */
289 #define LOOP_VINFO_NITERS_UNCHANGED(L)     (L)->num_iters_unchanged
290 #define LOOP_VINFO_COST_MODEL_MIN_ITERS(L) (L)->min_profitable_iters
291 #define LOOP_VINFO_VECTORIZABLE_P(L)       (L)->vectorizable
292 #define LOOP_VINFO_VECT_FACTOR(L)          (L)->vectorization_factor
293 #define LOOP_VINFO_PTR_MASK(L)             (L)->ptr_mask
294 #define LOOP_VINFO_LOOP_NEST(L)            (L)->loop_nest
295 #define LOOP_VINFO_DATAREFS(L)             (L)->datarefs
296 #define LOOP_VINFO_DDRS(L)                 (L)->ddrs
297 #define LOOP_VINFO_INT_NITERS(L)           (TREE_INT_CST_LOW ((L)->num_iters))
298 #define LOOP_PEELING_FOR_ALIGNMENT(L)      (L)->peeling_for_alignment
299 #define LOOP_VINFO_UNALIGNED_DR(L)         (L)->unaligned_dr
300 #define LOOP_VINFO_MAY_MISALIGN_STMTS(L)   (L)->may_misalign_stmts
301 #define LOOP_VINFO_LOC(L)                  (L)->loop_line_number
302 #define LOOP_VINFO_MAY_ALIAS_DDRS(L)       (L)->may_alias_ddrs
303 #define LOOP_VINFO_STRIDED_STORES(L)       (L)->strided_stores
304 #define LOOP_VINFO_SLP_INSTANCES(L)        (L)->slp_instances
305 #define LOOP_VINFO_SLP_UNROLLING_FACTOR(L) (L)->slp_unrolling_factor
306 #define LOOP_VINFO_REDUCTIONS(L)           (L)->reductions
307 #define LOOP_VINFO_REDUCTION_CHAINS(L)     (L)->reduction_chains
308 #define LOOP_VINFO_PEELING_HTAB(L)         (L)->peeling_htab
309 #define LOOP_VINFO_PEELING_FOR_GAPS(L)     (L)->peeling_for_gaps
310 
311 #define LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT(L) \
312 VEC_length (gimple, (L)->may_misalign_stmts) > 0
313 #define LOOP_REQUIRES_VERSIONING_FOR_ALIAS(L)     \
314 VEC_length (ddr_p, (L)->may_alias_ddrs) > 0
315 
316 #define NITERS_KNOWN_P(n)                     \
317 (host_integerp ((n),0)                        \
318 && TREE_INT_CST_LOW ((n)) > 0)
319 
320 #define LOOP_VINFO_NITERS_KNOWN_P(L)          \
321 NITERS_KNOWN_P((L)->num_iters)
322 
323 static inline loop_vec_info
324 loop_vec_info_for_loop (struct loop *loop)
325 {
326   return (loop_vec_info) loop->aux;
327 }
328 
329 static inline bool
330 nested_in_vect_loop_p (struct loop *loop, gimple stmt)
331 {
332   return (loop->inner
333           && (loop->inner == (gimple_bb (stmt))->loop_father));
334 }
335 
336 typedef struct _bb_vec_info {
337 
338   basic_block bb;
339   /* All interleaving chains of stores in the basic block, represented by the
340      first stmt in the chain.  */
341   VEC(gimple, heap) *strided_stores;
342 
343   /* All SLP instances in the basic block. This is a subset of the set of
344      STRIDED_STORES of the basic block.  */
345   VEC(slp_instance, heap) *slp_instances;
346 
347   /* All data references in the basic block.  */
348   VEC (data_reference_p, heap) *datarefs;
349 
350   /* All data dependences in the basic block.  */
351   VEC (ddr_p, heap) *ddrs;
352 } *bb_vec_info;
353 
354 #define BB_VINFO_BB(B)              (B)->bb
355 #define BB_VINFO_STRIDED_STORES(B)  (B)->strided_stores
356 #define BB_VINFO_SLP_INSTANCES(B)   (B)->slp_instances
357 #define BB_VINFO_DATAREFS(B)        (B)->datarefs
358 #define BB_VINFO_DDRS(B)            (B)->ddrs
359 
360 static inline bb_vec_info
361 vec_info_for_bb (basic_block bb)
362 {
363   return (bb_vec_info) bb->aux;
364 }
365 
366 /*-----------------------------------------------------------------*/
367 /* Info on vectorized defs.                                        */
368 /*-----------------------------------------------------------------*/
369 enum stmt_vec_info_type {
370   undef_vec_info_type = 0,
371   load_vec_info_type,
372   store_vec_info_type,
373   shift_vec_info_type,
374   op_vec_info_type,
375   call_vec_info_type,
376   assignment_vec_info_type,
377   condition_vec_info_type,
378   reduc_vec_info_type,
379   induc_vec_info_type,
380   type_promotion_vec_info_type,
381   type_demotion_vec_info_type,
382   type_conversion_vec_info_type,
383   loop_exit_ctrl_vec_info_type
384 };
385 
386 /* Indicates whether/how a variable is used in the scope of loop/basic
387    block.  */
388 enum vect_relevant {
389   vect_unused_in_scope = 0,
390   /* The def is in the inner loop, and the use is in the outer loop, and the
391      use is a reduction stmt.  */
392   vect_used_in_outer_by_reduction,
393   /* The def is in the inner loop, and the use is in the outer loop (and is
394      not part of reduction).  */
395   vect_used_in_outer,
396 
397   /* defs that feed computations that end up (only) in a reduction. These
398      defs may be used by non-reduction stmts, but eventually, any
399      computations/values that are affected by these defs are used to compute
400      a reduction (i.e. don't get stored to memory, for example). We use this
401      to identify computations that we can change the order in which they are
402      computed.  */
403   vect_used_by_reduction,
404 
405   vect_used_in_scope
406 };
407 
408 /* The type of vectorization that can be applied to the stmt: regular loop-based
409    vectorization; pure SLP - the stmt is a part of SLP instances and does not
410    have uses outside SLP instances; or hybrid SLP and loop-based - the stmt is
411    a part of SLP instance and also must be loop-based vectorized, since it has
412    uses outside SLP sequences.
413 
414    In the loop context the meanings of pure and hybrid SLP are slightly
415    different. By saying that pure SLP is applied to the loop, we mean that we
416    exploit only intra-iteration parallelism in the loop; i.e., the loop can be
417    vectorized without doing any conceptual unrolling, cause we don't pack
418    together stmts from different iterations, only within a single iteration.
419    Loop hybrid SLP means that we exploit both intra-iteration and
420    inter-iteration parallelism (e.g., number of elements in the vector is 4
421    and the slp-group-size is 2, in which case we don't have enough parallelism
422    within an iteration, so we obtain the rest of the parallelism from subsequent
423    iterations by unrolling the loop by 2).  */
424 enum slp_vect_type {
425   loop_vect = 0,
426   pure_slp,
427   hybrid
428 };
429 
430 
431 typedef struct data_reference *dr_p;
432 DEF_VEC_P(dr_p);
433 DEF_VEC_ALLOC_P(dr_p,heap);
434 
435 typedef struct _stmt_vec_info {
436 
437   enum stmt_vec_info_type type;
438 
439   /* Indicates whether this stmts is part of a computation whose result is
440      used outside the loop.  */
441   bool live;
442 
443   /* Stmt is part of some pattern (computation idiom)  */
444   bool in_pattern_p;
445 
446   /* For loads only, if there is a store with the same location, this field is
447      TRUE.  */
448   bool read_write_dep;
449 
450   /* The stmt to which this info struct refers to.  */
451   gimple stmt;
452 
453   /* The loop_vec_info with respect to which STMT is vectorized.  */
454   loop_vec_info loop_vinfo;
455 
456   /* The vector type to be used for the LHS of this statement.  */
457   tree vectype;
458 
459   /* The vectorized version of the stmt.  */
460   gimple vectorized_stmt;
461 
462 
463   /** The following is relevant only for stmts that contain a non-scalar
464      data-ref (array/pointer/struct access). A GIMPLE stmt is expected to have
465      at most one such data-ref.  **/
466 
467   /* Information about the data-ref (access function, etc),
468      relative to the inner-most containing loop.  */
469   struct data_reference *data_ref_info;
470 
471   /* Information about the data-ref relative to this loop
472      nest (the loop that is being considered for vectorization).  */
473   tree dr_base_address;
474   tree dr_init;
475   tree dr_offset;
476   tree dr_step;
477   tree dr_aligned_to;
478 
479   /* For loop PHI nodes, the evolution part of it.  This makes sure
480      this information is still available in vect_update_ivs_after_vectorizer
481      where we may not be able to re-analyze the PHI nodes evolution as
482      peeling for the prologue loop can make it unanalyzable.  The evolution
483      part is still correct though.  */
484   tree loop_phi_evolution_part;
485 
486   /* Used for various bookkeeping purposes, generally holding a pointer to
487      some other stmt S that is in some way "related" to this stmt.
488      Current use of this field is:
489         If this stmt is part of a pattern (i.e. the field 'in_pattern_p' is
490         true): S is the "pattern stmt" that represents (and replaces) the
491         sequence of stmts that constitutes the pattern.  Similarly, the
492         related_stmt of the "pattern stmt" points back to this stmt (which is
493         the last stmt in the original sequence of stmts that constitutes the
494         pattern).  */
495   gimple related_stmt;
496 
497   /* Used to keep a sequence of def stmts of a pattern stmt if such exists.  */
498   gimple_seq pattern_def_seq;
499 
500   /* List of datarefs that are known to have the same alignment as the dataref
501      of this stmt.  */
502   VEC(dr_p,heap) *same_align_refs;
503 
504   /* Classify the def of this stmt.  */
505   enum vect_def_type def_type;
506 
507   /*  Whether the stmt is SLPed, loop-based vectorized, or both.  */
508   enum slp_vect_type slp_type;
509 
510   /* Interleaving and reduction chains info.  */
511   /* First element in the group.  */
512   gimple first_element;
513   /* Pointer to the next element in the group.  */
514   gimple next_element;
515   /* For data-refs, in case that two or more stmts share data-ref, this is the
516      pointer to the previously detected stmt with the same dr.  */
517   gimple same_dr_stmt;
518   /* The size of the group.  */
519   unsigned int size;
520   /* For stores, number of stores from this group seen. We vectorize the last
521      one.  */
522   unsigned int store_count;
523   /* For loads only, the gap from the previous load. For consecutive loads, GAP
524      is 1.  */
525   unsigned int gap;
526 
527   /* Not all stmts in the loop need to be vectorized. e.g, the increment
528      of the loop induction variable and computation of array indexes. relevant
529      indicates whether the stmt needs to be vectorized.  */
530   enum vect_relevant relevant;
531 
532   /* Vectorization costs associated with statement.  */
533   struct
534   {
535     int outside_of_loop;     /* Statements generated outside loop.  */
536     int inside_of_loop;      /* Statements generated inside loop.  */
537   } cost;
538 
539   /* The bb_vec_info with respect to which STMT is vectorized.  */
540   bb_vec_info bb_vinfo;
541 
542   /* Is this statement vectorizable or should it be skipped in (partial)
543      vectorization.  */
544   bool vectorizable;
545 
546   /* For loads only, true if this is a gather load.  */
547   bool gather_p;
548 } *stmt_vec_info;
549 
550 /* Access Functions.  */
551 #define STMT_VINFO_TYPE(S)                 (S)->type
552 #define STMT_VINFO_STMT(S)                 (S)->stmt
553 #define STMT_VINFO_LOOP_VINFO(S)           (S)->loop_vinfo
554 #define STMT_VINFO_BB_VINFO(S)             (S)->bb_vinfo
555 #define STMT_VINFO_RELEVANT(S)             (S)->relevant
556 #define STMT_VINFO_LIVE_P(S)               (S)->live
557 #define STMT_VINFO_VECTYPE(S)              (S)->vectype
558 #define STMT_VINFO_VEC_STMT(S)             (S)->vectorized_stmt
559 #define STMT_VINFO_VECTORIZABLE(S)         (S)->vectorizable
560 #define STMT_VINFO_DATA_REF(S)             (S)->data_ref_info
561 #define STMT_VINFO_GATHER_P(S)		   (S)->gather_p
562 
563 #define STMT_VINFO_DR_BASE_ADDRESS(S)      (S)->dr_base_address
564 #define STMT_VINFO_DR_INIT(S)              (S)->dr_init
565 #define STMT_VINFO_DR_OFFSET(S)            (S)->dr_offset
566 #define STMT_VINFO_DR_STEP(S)              (S)->dr_step
567 #define STMT_VINFO_DR_ALIGNED_TO(S)        (S)->dr_aligned_to
568 
569 #define STMT_VINFO_IN_PATTERN_P(S)         (S)->in_pattern_p
570 #define STMT_VINFO_RELATED_STMT(S)         (S)->related_stmt
571 #define STMT_VINFO_PATTERN_DEF_SEQ(S)      (S)->pattern_def_seq
572 #define STMT_VINFO_SAME_ALIGN_REFS(S)      (S)->same_align_refs
573 #define STMT_VINFO_DEF_TYPE(S)             (S)->def_type
574 #define STMT_VINFO_GROUP_FIRST_ELEMENT(S)  (S)->first_element
575 #define STMT_VINFO_GROUP_NEXT_ELEMENT(S)   (S)->next_element
576 #define STMT_VINFO_GROUP_SIZE(S)           (S)->size
577 #define STMT_VINFO_GROUP_STORE_COUNT(S)    (S)->store_count
578 #define STMT_VINFO_GROUP_GAP(S)            (S)->gap
579 #define STMT_VINFO_GROUP_SAME_DR_STMT(S)   (S)->same_dr_stmt
580 #define STMT_VINFO_GROUP_READ_WRITE_DEPENDENCE(S)  (S)->read_write_dep
581 #define STMT_VINFO_STRIDED_ACCESS(S)      ((S)->first_element != NULL && (S)->data_ref_info)
582 #define STMT_VINFO_LOOP_PHI_EVOLUTION_PART(S) (S)->loop_phi_evolution_part
583 
584 #define GROUP_FIRST_ELEMENT(S)          (S)->first_element
585 #define GROUP_NEXT_ELEMENT(S)           (S)->next_element
586 #define GROUP_SIZE(S)                   (S)->size
587 #define GROUP_STORE_COUNT(S)            (S)->store_count
588 #define GROUP_GAP(S)                    (S)->gap
589 #define GROUP_SAME_DR_STMT(S)           (S)->same_dr_stmt
590 #define GROUP_READ_WRITE_DEPENDENCE(S)  (S)->read_write_dep
591 
592 #define STMT_VINFO_RELEVANT_P(S)          ((S)->relevant != vect_unused_in_scope)
593 #define STMT_VINFO_OUTSIDE_OF_LOOP_COST(S) (S)->cost.outside_of_loop
594 #define STMT_VINFO_INSIDE_OF_LOOP_COST(S)  (S)->cost.inside_of_loop
595 
596 #define HYBRID_SLP_STMT(S)                ((S)->slp_type == hybrid)
597 #define PURE_SLP_STMT(S)                  ((S)->slp_type == pure_slp)
598 #define STMT_SLP_TYPE(S)                   (S)->slp_type
599 
600 #define VECT_MAX_COST 1000
601 
602 /* The maximum number of intermediate steps required in multi-step type
603    conversion.  */
604 #define MAX_INTERM_CVT_STEPS         3
605 
606 /* The maximum vectorization factor supported by any target (V32QI).  */
607 #define MAX_VECTORIZATION_FACTOR 32
608 
609 /* Avoid GTY(()) on stmt_vec_info.  */
610 typedef void *vec_void_p;
611 DEF_VEC_P (vec_void_p);
612 DEF_VEC_ALLOC_P (vec_void_p, heap);
613 
614 extern VEC(vec_void_p,heap) *stmt_vec_info_vec;
615 
616 void init_stmt_vec_info_vec (void);
617 void free_stmt_vec_info_vec (void);
618 
619 /* Return a stmt_vec_info corresponding to STMT.  */
620 
621 static inline stmt_vec_info
622 vinfo_for_stmt (gimple stmt)
623 {
624   unsigned int uid = gimple_uid (stmt);
625   if (uid == 0)
626     return NULL;
627 
628   return (stmt_vec_info) VEC_index (vec_void_p, stmt_vec_info_vec, uid - 1);
629 }
630 
631 /* Set vectorizer information INFO for STMT.  */
632 
633 static inline void
634 set_vinfo_for_stmt (gimple stmt, stmt_vec_info info)
635 {
636   unsigned int uid = gimple_uid (stmt);
637   if (uid == 0)
638     {
639       gcc_checking_assert (info);
640       uid = VEC_length (vec_void_p, stmt_vec_info_vec) + 1;
641       gimple_set_uid (stmt, uid);
642       VEC_safe_push (vec_void_p, heap, stmt_vec_info_vec, (vec_void_p) info);
643     }
644   else
645     VEC_replace (vec_void_p, stmt_vec_info_vec, uid - 1, (vec_void_p) info);
646 }
647 
648 /* Return the earlier statement between STMT1 and STMT2.  */
649 
650 static inline gimple
651 get_earlier_stmt (gimple stmt1, gimple stmt2)
652 {
653   unsigned int uid1, uid2;
654 
655   if (stmt1 == NULL)
656     return stmt2;
657 
658   if (stmt2 == NULL)
659     return stmt1;
660 
661   uid1 = gimple_uid (stmt1);
662   uid2 = gimple_uid (stmt2);
663 
664   if (uid1 == 0 || uid2 == 0)
665     return NULL;
666 
667   gcc_checking_assert (uid1 <= VEC_length (vec_void_p, stmt_vec_info_vec)
668 		       && uid2 <= VEC_length (vec_void_p, stmt_vec_info_vec));
669 
670   if (uid1 < uid2)
671     return stmt1;
672   else
673     return stmt2;
674 }
675 
676 /* Return the later statement between STMT1 and STMT2.  */
677 
678 static inline gimple
679 get_later_stmt (gimple stmt1, gimple stmt2)
680 {
681   unsigned int uid1, uid2;
682 
683   if (stmt1 == NULL)
684     return stmt2;
685 
686   if (stmt2 == NULL)
687     return stmt1;
688 
689   uid1 = gimple_uid (stmt1);
690   uid2 = gimple_uid (stmt2);
691 
692   if (uid1 == 0 || uid2 == 0)
693     return NULL;
694 
695   gcc_assert (uid1 <= VEC_length (vec_void_p, stmt_vec_info_vec));
696   gcc_assert (uid2 <= VEC_length (vec_void_p, stmt_vec_info_vec));
697 
698   if (uid1 > uid2)
699     return stmt1;
700   else
701     return stmt2;
702 }
703 
704 /* Return TRUE if a statement represented by STMT_INFO is a part of a
705    pattern.  */
706 
707 static inline bool
708 is_pattern_stmt_p (stmt_vec_info stmt_info)
709 {
710   gimple related_stmt;
711   stmt_vec_info related_stmt_info;
712 
713   related_stmt = STMT_VINFO_RELATED_STMT (stmt_info);
714   if (related_stmt
715       && (related_stmt_info = vinfo_for_stmt (related_stmt))
716       && STMT_VINFO_IN_PATTERN_P (related_stmt_info))
717     return true;
718 
719   return false;
720 }
721 
722 /* Return true if BB is a loop header.  */
723 
724 static inline bool
725 is_loop_header_bb_p (basic_block bb)
726 {
727   if (bb == (bb->loop_father)->header)
728     return true;
729   gcc_checking_assert (EDGE_COUNT (bb->preds) == 1);
730   return false;
731 }
732 
733 /* Set inside loop vectorization cost.  */
734 
735 static inline void
736 stmt_vinfo_set_inside_of_loop_cost (stmt_vec_info stmt_info, slp_tree slp_node,
737 				    int cost)
738 {
739   if (slp_node)
740     SLP_TREE_INSIDE_OF_LOOP_COST (slp_node) = cost;
741   else
742     STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info) = cost;
743 }
744 
745 /* Set inside loop vectorization cost.  */
746 
747 static inline void
748 stmt_vinfo_set_outside_of_loop_cost (stmt_vec_info stmt_info, slp_tree slp_node,
749 				     int cost)
750 {
751   if (slp_node)
752     SLP_TREE_OUTSIDE_OF_LOOP_COST (slp_node) = cost;
753   else
754     STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info) = cost;
755 }
756 
757 /* Return pow2 (X).  */
758 
759 static inline int
760 vect_pow2 (int x)
761 {
762   int i, res = 1;
763 
764   for (i = 0; i < x; i++)
765     res *= 2;
766 
767   return res;
768 }
769 
770 /*-----------------------------------------------------------------*/
771 /* Info on data references alignment.                              */
772 /*-----------------------------------------------------------------*/
773 
774 /* Reflects actual alignment of first access in the vectorized loop,
775    taking into account peeling/versioning if applied.  */
776 #define DR_MISALIGNMENT(DR)   ((int) (size_t) (DR)->aux)
777 #define SET_DR_MISALIGNMENT(DR, VAL)   ((DR)->aux = (void *) (size_t) (VAL))
778 
779 /* Return TRUE if the data access is aligned, and FALSE otherwise.  */
780 
781 static inline bool
782 aligned_access_p (struct data_reference *data_ref_info)
783 {
784   return (DR_MISALIGNMENT (data_ref_info) == 0);
785 }
786 
787 /* Return TRUE if the alignment of the data access is known, and FALSE
788    otherwise.  */
789 
790 static inline bool
791 known_alignment_for_access_p (struct data_reference *data_ref_info)
792 {
793   return (DR_MISALIGNMENT (data_ref_info) != -1);
794 }
795 
796 /* vect_dump will be set to stderr or dump_file if exist.  */
797 extern FILE *vect_dump;
798 extern LOC vect_loop_location;
799 
800 /*-----------------------------------------------------------------*/
801 /* Function prototypes.                                            */
802 /*-----------------------------------------------------------------*/
803 
804 /* Simple loop peeling and versioning utilities for vectorizer's purposes -
805    in tree-vect-loop-manip.c.  */
806 extern void slpeel_make_loop_iterate_ntimes (struct loop *, tree);
807 extern bool slpeel_can_duplicate_loop_p (const struct loop *, const_edge);
808 extern void vect_loop_versioning (loop_vec_info, bool, tree *, gimple_seq *);
809 extern void vect_do_peeling_for_loop_bound (loop_vec_info, tree *,
810                                             tree, gimple_seq);
811 extern void vect_do_peeling_for_alignment (loop_vec_info);
812 extern LOC find_loop_location (struct loop *);
813 extern bool vect_can_advance_ivs_p (loop_vec_info);
814 
815 /* In tree-vect-stmts.c.  */
816 extern unsigned int current_vector_size;
817 extern tree get_vectype_for_scalar_type (tree);
818 extern tree get_same_sized_vectype (tree, tree);
819 extern bool vect_is_simple_use (tree, gimple, loop_vec_info,
820 			        bb_vec_info, gimple *,
821                                 tree *,  enum vect_def_type *);
822 extern bool vect_is_simple_use_1 (tree, gimple, loop_vec_info,
823 				  bb_vec_info, gimple *,
824 				  tree *,  enum vect_def_type *, tree *);
825 extern bool supportable_widening_operation (enum tree_code, gimple, tree, tree,
826                                             tree *, tree *, enum tree_code *,
827                                             enum tree_code *, int *,
828                                             VEC (tree, heap) **);
829 extern bool supportable_narrowing_operation (enum tree_code, tree, tree,
830 					     enum tree_code *,
831 					     int *, VEC (tree, heap) **);
832 extern stmt_vec_info new_stmt_vec_info (gimple stmt, loop_vec_info,
833                                         bb_vec_info);
834 extern void free_stmt_vec_info (gimple stmt);
835 extern tree vectorizable_function (gimple, tree, tree);
836 extern void vect_model_simple_cost (stmt_vec_info, int, enum vect_def_type *,
837                                     slp_tree);
838 extern void vect_model_store_cost (stmt_vec_info, int, bool,
839 				   enum vect_def_type, slp_tree);
840 extern void vect_model_load_cost (stmt_vec_info, int, bool, slp_tree);
841 extern void vect_finish_stmt_generation (gimple, gimple,
842                                          gimple_stmt_iterator *);
843 extern bool vect_mark_stmts_to_be_vectorized (loop_vec_info);
844 extern int cost_for_stmt (gimple);
845 extern tree vect_get_vec_def_for_operand (tree, gimple, tree *);
846 extern tree vect_init_vector (gimple, tree, tree,
847                               gimple_stmt_iterator *);
848 extern tree vect_get_vec_def_for_stmt_copy (enum vect_def_type, tree);
849 extern bool vect_transform_stmt (gimple, gimple_stmt_iterator *,
850                                  bool *, slp_tree, slp_instance);
851 extern void vect_remove_stores (gimple);
852 extern bool vect_analyze_stmt (gimple, bool *, slp_tree);
853 extern bool vectorizable_condition (gimple, gimple_stmt_iterator *, gimple *,
854                                     tree, int, slp_tree);
855 extern void vect_get_load_cost (struct data_reference *, int, bool,
856                                 unsigned int *, unsigned int *);
857 extern void vect_get_store_cost (struct data_reference *, int, unsigned int *);
858 extern bool vect_supportable_shift (enum tree_code, tree);
859 extern void vect_get_vec_defs (tree, tree, gimple, VEC (tree, heap) **,
860 			       VEC (tree, heap) **, slp_tree, int);
861 extern tree vect_gen_perm_mask (tree, unsigned char *);
862 
863 /* In tree-vect-data-refs.c.  */
864 extern bool vect_can_force_dr_alignment_p (const_tree, unsigned int);
865 extern enum dr_alignment_support vect_supportable_dr_alignment
866                                            (struct data_reference *, bool);
867 extern tree vect_get_smallest_scalar_type (gimple, HOST_WIDE_INT *,
868                                            HOST_WIDE_INT *);
869 extern bool vect_analyze_data_ref_dependences (loop_vec_info, bb_vec_info,
870 					       int *);
871 extern bool vect_enhance_data_refs_alignment (loop_vec_info);
872 extern bool vect_analyze_data_refs_alignment (loop_vec_info, bb_vec_info);
873 extern bool vect_verify_datarefs_alignment (loop_vec_info, bb_vec_info);
874 extern bool vect_analyze_data_ref_accesses (loop_vec_info, bb_vec_info);
875 extern bool vect_prune_runtime_alias_test_list (loop_vec_info);
876 extern tree vect_check_gather (gimple, loop_vec_info, tree *, tree *,
877 			       int *);
878 extern bool vect_analyze_data_refs (loop_vec_info, bb_vec_info, int *);
879 extern tree vect_create_data_ref_ptr (gimple, tree, struct loop *, tree,
880 				      tree *, gimple_stmt_iterator *,
881 				      gimple *, bool, bool *);
882 extern tree bump_vector_ptr (tree, gimple, gimple_stmt_iterator *, gimple, tree);
883 extern tree vect_create_destination_var (tree, tree);
884 extern bool vect_strided_store_supported (tree, unsigned HOST_WIDE_INT);
885 extern bool vect_store_lanes_supported (tree, unsigned HOST_WIDE_INT);
886 extern bool vect_strided_load_supported (tree, unsigned HOST_WIDE_INT);
887 extern bool vect_load_lanes_supported (tree, unsigned HOST_WIDE_INT);
888 extern void vect_permute_store_chain (VEC(tree,heap) *,unsigned int, gimple,
889                                     gimple_stmt_iterator *, VEC(tree,heap) **);
890 extern tree vect_setup_realignment (gimple, gimple_stmt_iterator *, tree *,
891                                     enum dr_alignment_support, tree,
892                                     struct loop **);
893 extern void vect_transform_strided_load (gimple, VEC(tree,heap) *, int,
894                                          gimple_stmt_iterator *);
895 extern void vect_record_strided_load_vectors (gimple, VEC(tree,heap) *);
896 extern int vect_get_place_in_interleaving_chain (gimple, gimple);
897 extern tree vect_get_new_vect_var (tree, enum vect_var_kind, const char *);
898 extern tree vect_create_addr_base_for_vector_ref (gimple, gimple_seq *,
899                                                   tree, struct loop *);
900 
901 /* In tree-vect-loop.c.  */
902 /* FORNOW: Used in tree-parloops.c.  */
903 extern void destroy_loop_vec_info (loop_vec_info, bool);
904 extern gimple vect_force_simple_reduction (loop_vec_info, gimple, bool, bool *);
905 /* Drive for loop analysis stage.  */
906 extern loop_vec_info vect_analyze_loop (struct loop *);
907 /* Drive for loop transformation stage.  */
908 extern void vect_transform_loop (loop_vec_info);
909 extern loop_vec_info vect_analyze_loop_form (struct loop *);
910 extern bool vectorizable_live_operation (gimple, gimple_stmt_iterator *,
911                                          gimple *);
912 extern bool vectorizable_reduction (gimple, gimple_stmt_iterator *, gimple *,
913                                     slp_tree);
914 extern bool vectorizable_induction (gimple, gimple_stmt_iterator *, gimple *);
915 extern int vect_estimate_min_profitable_iters (loop_vec_info);
916 extern tree get_initial_def_for_reduction (gimple, tree, tree *);
917 extern int vect_min_worthwhile_factor (enum tree_code);
918 extern int vect_get_known_peeling_cost (loop_vec_info, int, int *, int);
919 extern int vect_get_single_scalar_iteration_cost (loop_vec_info);
920 
921 /* In tree-vect-slp.c.  */
922 extern void vect_free_slp_instance (slp_instance);
923 extern bool vect_transform_slp_perm_load (gimple, VEC (tree, heap) *,
924                                           gimple_stmt_iterator *, int,
925                                           slp_instance, bool);
926 extern bool vect_schedule_slp (loop_vec_info, bb_vec_info);
927 extern void vect_update_slp_costs_according_to_vf (loop_vec_info);
928 extern bool vect_analyze_slp (loop_vec_info, bb_vec_info);
929 extern bool vect_make_slp_decision (loop_vec_info);
930 extern void vect_detect_hybrid_slp (loop_vec_info);
931 extern void vect_get_slp_defs (VEC (tree, heap) *, slp_tree,
932 			       VEC (slp_void_p, heap) **, int);
933 
934 extern LOC find_bb_location (basic_block);
935 extern bb_vec_info vect_slp_analyze_bb (basic_block);
936 extern void vect_slp_transform_bb (basic_block);
937 
938 /* In tree-vect-patterns.c.  */
939 /* Pattern recognition functions.
940    Additional pattern recognition functions can (and will) be added
941    in the future.  */
942 typedef gimple (* vect_recog_func_ptr) (VEC (gimple, heap) **, tree *, tree *);
943 #define NUM_PATTERNS 10
944 void vect_pattern_recog (loop_vec_info);
945 
946 /* In tree-vectorizer.c.  */
947 unsigned vectorize_loops (void);
948 /* Vectorization debug information */
949 extern bool vect_print_dump_info (enum vect_verbosity_levels);
950 
951 #endif  /* GCC_TREE_VECTORIZER_H  */
952