1/* Vectorizer 2 Copyright (C) 2003-2015 Free Software Foundation, Inc. 3 Contributed by Dorit Naishlos <dorit@il.ibm.com> 4 5This file is part of GCC. 6 7GCC is free software; you can redistribute it and/or modify it under 8the terms of the GNU General Public License as published by the Free 9Software Foundation; either version 3, or (at your option) any later 10version. 11 12GCC is distributed in the hope that it will be useful, but WITHOUT ANY 13WARRANTY; without even the implied warranty of MERCHANTABILITY or 14FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 15for more details. 16 17You should have received a copy of the GNU General Public License 18along with GCC; see the file COPYING3. If not see 19<http://www.gnu.org/licenses/>. */ 20 21#ifndef GCC_TREE_VECTORIZER_H 22#define GCC_TREE_VECTORIZER_H 23 24#include "tree-data-ref.h" 25#include "target.h" 26#include "hash-table.h" 27 28/* Used for naming of new temporaries. */ 29enum vect_var_kind { 30 vect_simple_var, 31 vect_pointer_var, 32 vect_scalar_var 33}; 34 35/* Defines type of operation. */ 36enum operation_type { 37 unary_op = 1, 38 binary_op, 39 ternary_op 40}; 41 42/* Define type of available alignment support. */ 43enum dr_alignment_support { 44 dr_unaligned_unsupported, 45 dr_unaligned_supported, 46 dr_explicit_realign, 47 dr_explicit_realign_optimized, 48 dr_aligned 49}; 50 51/* Define type of def-use cross-iteration cycle. */ 52enum vect_def_type { 53 vect_uninitialized_def = 0, 54 vect_constant_def = 1, 55 vect_external_def, 56 vect_internal_def, 57 vect_induction_def, 58 vect_reduction_def, 59 vect_double_reduction_def, 60 vect_nested_cycle, 61 vect_unknown_def_type 62}; 63 64#define VECTORIZABLE_CYCLE_DEF(D) (((D) == vect_reduction_def) \ 65 || ((D) == vect_double_reduction_def) \ 66 || ((D) == vect_nested_cycle)) 67 68/* Structure to encapsulate information about a group of like 69 instructions to be presented to the target cost model. */ 70typedef struct _stmt_info_for_cost { 71 int count; 72 enum vect_cost_for_stmt kind; 73 gimple stmt; 74 int misalign; 75} stmt_info_for_cost; 76 77 78typedef vec<stmt_info_for_cost> stmt_vector_for_cost; 79 80static inline void 81add_stmt_info_to_vec (stmt_vector_for_cost *stmt_cost_vec, int count, 82 enum vect_cost_for_stmt kind, gimple stmt, int misalign) 83{ 84 stmt_info_for_cost si; 85 si.count = count; 86 si.kind = kind; 87 si.stmt = stmt; 88 si.misalign = misalign; 89 stmt_cost_vec->safe_push (si); 90} 91 92/************************************************************************ 93 SLP 94 ************************************************************************/ 95typedef struct _slp_tree *slp_tree; 96 97/* A computation tree of an SLP instance. Each node corresponds to a group of 98 stmts to be packed in a SIMD stmt. */ 99struct _slp_tree { 100 /* Nodes that contain def-stmts of this node statements operands. */ 101 vec<slp_tree> children; 102 /* A group of scalar stmts to be vectorized together. */ 103 vec<gimple> stmts; 104 /* Load permutation relative to the stores, NULL if there is no 105 permutation. */ 106 vec<unsigned> load_permutation; 107 /* Vectorized stmt/s. */ 108 vec<gimple> vec_stmts; 109 /* Number of vector stmts that are created to replace the group of scalar 110 stmts. It is calculated during the transformation phase as the number of 111 scalar elements in one scalar iteration (GROUP_SIZE) multiplied by VF 112 divided by vector size. */ 113 unsigned int vec_stmts_size; 114}; 115 116 117/* SLP instance is a sequence of stmts in a loop that can be packed into 118 SIMD stmts. */ 119typedef struct _slp_instance { 120 /* The root of SLP tree. */ 121 slp_tree root; 122 123 /* Size of groups of scalar stmts that will be replaced by SIMD stmt/s. */ 124 unsigned int group_size; 125 126 /* The unrolling factor required to vectorized this SLP instance. */ 127 unsigned int unrolling_factor; 128 129 /* Vectorization costs associated with SLP instance. */ 130 stmt_vector_for_cost body_cost_vec; 131 132 /* The group of nodes that contain loads of this SLP instance. */ 133 vec<slp_tree> loads; 134 135 /* The first scalar load of the instance. The created vector loads will be 136 inserted before this statement. */ 137 gimple first_load; 138} *slp_instance; 139 140 141/* Access Functions. */ 142#define SLP_INSTANCE_TREE(S) (S)->root 143#define SLP_INSTANCE_GROUP_SIZE(S) (S)->group_size 144#define SLP_INSTANCE_UNROLLING_FACTOR(S) (S)->unrolling_factor 145#define SLP_INSTANCE_BODY_COST_VEC(S) (S)->body_cost_vec 146#define SLP_INSTANCE_LOADS(S) (S)->loads 147#define SLP_INSTANCE_FIRST_LOAD_STMT(S) (S)->first_load 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_LOAD_PERMUTATION(S) (S)->load_permutation 154 155/* This structure is used in creation of an SLP tree. Each instance 156 corresponds to the same operand in a group of scalar stmts in an SLP 157 node. */ 158typedef struct _slp_oprnd_info 159{ 160 /* Def-stmts for the operands. */ 161 vec<gimple> def_stmts; 162 /* Information about the first statement, its vector def-type, type, the 163 operand itself in case it's constant, and an indication if it's a pattern 164 stmt. */ 165 enum vect_def_type first_dt; 166 tree first_op_type; 167 bool first_pattern; 168} *slp_oprnd_info; 169 170 171 172/* This struct is used to store the information of a data reference, 173 including the data ref itself, the access offset (calculated by summing its 174 offset and init) and the segment length for aliasing checks. 175 This is used to merge alias checks. */ 176 177struct dr_with_seg_len 178{ 179 dr_with_seg_len (data_reference_p d, tree len) 180 : dr (d), 181 offset (size_binop (PLUS_EXPR, DR_OFFSET (d), DR_INIT (d))), 182 seg_len (len) {} 183 184 data_reference_p dr; 185 tree offset; 186 tree seg_len; 187}; 188 189/* This struct contains two dr_with_seg_len objects with aliasing data 190 refs. Two comparisons are generated from them. */ 191 192struct dr_with_seg_len_pair_t 193{ 194 dr_with_seg_len_pair_t (const dr_with_seg_len& d1, 195 const dr_with_seg_len& d2) 196 : first (d1), second (d2) {} 197 198 dr_with_seg_len first; 199 dr_with_seg_len second; 200}; 201 202 203typedef struct _vect_peel_info 204{ 205 int npeel; 206 struct data_reference *dr; 207 unsigned int count; 208} *vect_peel_info; 209 210typedef struct _vect_peel_extended_info 211{ 212 struct _vect_peel_info peel_info; 213 unsigned int inside_cost; 214 unsigned int outside_cost; 215 stmt_vector_for_cost body_cost_vec; 216} *vect_peel_extended_info; 217 218 219/* Peeling hashtable helpers. */ 220 221struct peel_info_hasher : typed_free_remove <_vect_peel_info> 222{ 223 typedef _vect_peel_info value_type; 224 typedef _vect_peel_info compare_type; 225 static inline hashval_t hash (const value_type *); 226 static inline bool equal (const value_type *, const compare_type *); 227}; 228 229inline hashval_t 230peel_info_hasher::hash (const value_type *peel_info) 231{ 232 return (hashval_t) peel_info->npeel; 233} 234 235inline bool 236peel_info_hasher::equal (const value_type *a, const compare_type *b) 237{ 238 return (a->npeel == b->npeel); 239} 240 241 242/*-----------------------------------------------------------------*/ 243/* Info on vectorized loops. */ 244/*-----------------------------------------------------------------*/ 245typedef struct _loop_vec_info { 246 247 /* The loop to which this info struct refers to. */ 248 struct loop *loop; 249 250 /* The loop basic blocks. */ 251 basic_block *bbs; 252 253 /* Number of latch executions. */ 254 tree num_itersm1; 255 /* Number of iterations. */ 256 tree num_iters; 257 /* Number of iterations of the original loop. */ 258 tree num_iters_unchanged; 259 260 /* Minimum number of iterations below which vectorization is expected to 261 not be profitable (as estimated by the cost model). 262 -1 indicates that vectorization will not be profitable. 263 FORNOW: This field is an int. Will be a tree in the future, to represent 264 values unknown at compile time. */ 265 int min_profitable_iters; 266 267 /* Threshold of number of iterations below which vectorzation will not be 268 performed. It is calculated from MIN_PROFITABLE_ITERS and 269 PARAM_MIN_VECT_LOOP_BOUND. */ 270 unsigned int th; 271 272 /* Is the loop vectorizable? */ 273 bool vectorizable; 274 275 /* Unrolling factor */ 276 int vectorization_factor; 277 278 /* Unknown DRs according to which loop was peeled. */ 279 struct data_reference *unaligned_dr; 280 281 /* peeling_for_alignment indicates whether peeling for alignment will take 282 place, and what the peeling factor should be: 283 peeling_for_alignment = X means: 284 If X=0: Peeling for alignment will not be applied. 285 If X>0: Peel first X iterations. 286 If X=-1: Generate a runtime test to calculate the number of iterations 287 to be peeled, using the dataref recorded in the field 288 unaligned_dr. */ 289 int peeling_for_alignment; 290 291 /* The mask used to check the alignment of pointers or arrays. */ 292 int ptr_mask; 293 294 /* The loop nest in which the data dependences are computed. */ 295 vec<loop_p> loop_nest; 296 297 /* All data references in the loop. */ 298 vec<data_reference_p> datarefs; 299 300 /* All data dependences in the loop. */ 301 vec<ddr_p> ddrs; 302 303 /* Data Dependence Relations defining address ranges that are candidates 304 for a run-time aliasing check. */ 305 vec<ddr_p> may_alias_ddrs; 306 307 /* Data Dependence Relations defining address ranges together with segment 308 lengths from which the run-time aliasing check is built. */ 309 vec<dr_with_seg_len_pair_t> comp_alias_ddrs; 310 311 /* Statements in the loop that have data references that are candidates for a 312 runtime (loop versioning) misalignment check. */ 313 vec<gimple> may_misalign_stmts; 314 315 /* All interleaving chains of stores in the loop, represented by the first 316 stmt in the chain. */ 317 vec<gimple> grouped_stores; 318 319 /* All SLP instances in the loop. This is a subset of the set of GROUP_STORES 320 of the loop. */ 321 vec<slp_instance> slp_instances; 322 323 /* The unrolling factor needed to SLP the loop. In case of that pure SLP is 324 applied to the loop, i.e., no unrolling is needed, this is 1. */ 325 unsigned slp_unrolling_factor; 326 327 /* Reduction cycles detected in the loop. Used in loop-aware SLP. */ 328 vec<gimple> reductions; 329 330 /* All reduction chains in the loop, represented by the first 331 stmt in the chain. */ 332 vec<gimple> reduction_chains; 333 334 /* Hash table used to choose the best peeling option. */ 335 hash_table<peel_info_hasher> *peeling_htab; 336 337 /* Cost data used by the target cost model. */ 338 void *target_cost_data; 339 340 /* When we have grouped data accesses with gaps, we may introduce invalid 341 memory accesses. We peel the last iteration of the loop to prevent 342 this. */ 343 bool peeling_for_gaps; 344 345 /* When the number of iterations is not a multiple of the vector size 346 we need to peel off iterations at the end to form an epilogue loop. */ 347 bool peeling_for_niter; 348 349 /* Reductions are canonicalized so that the last operand is the reduction 350 operand. If this places a constant into RHS1, this decanonicalizes 351 GIMPLE for other phases, so we must track when this has occurred and 352 fix it up. */ 353 bool operands_swapped; 354 355 /* True if there are no loop carried data dependencies in the loop. 356 If loop->safelen <= 1, then this is always true, either the loop 357 didn't have any loop carried data dependencies, or the loop is being 358 vectorized guarded with some runtime alias checks, or couldn't 359 be vectorized at all, but then this field shouldn't be used. 360 For loop->safelen >= 2, the user has asserted that there are no 361 backward dependencies, but there still could be loop carried forward 362 dependencies in such loops. This flag will be false if normal 363 vectorizer data dependency analysis would fail or require versioning 364 for alias, but because of loop->safelen >= 2 it has been vectorized 365 even without versioning for alias. E.g. in: 366 #pragma omp simd 367 for (int i = 0; i < m; i++) 368 a[i] = a[i + k] * c; 369 (or #pragma simd or #pragma ivdep) we can vectorize this and it will 370 DTRT even for k > 0 && k < m, but without safelen we would not 371 vectorize this, so this field would be false. */ 372 bool no_data_dependencies; 373 374 /* If if-conversion versioned this loop before conversion, this is the 375 loop version without if-conversion. */ 376 struct loop *scalar_loop; 377 378} *loop_vec_info; 379 380/* Access Functions. */ 381#define LOOP_VINFO_LOOP(L) (L)->loop 382#define LOOP_VINFO_BBS(L) (L)->bbs 383#define LOOP_VINFO_NITERSM1(L) (L)->num_itersm1 384#define LOOP_VINFO_NITERS(L) (L)->num_iters 385/* Since LOOP_VINFO_NITERS and LOOP_VINFO_NITERSM1 can change after 386 prologue peeling retain total unchanged scalar loop iterations for 387 cost model. */ 388#define LOOP_VINFO_NITERS_UNCHANGED(L) (L)->num_iters_unchanged 389#define LOOP_VINFO_COST_MODEL_MIN_ITERS(L) (L)->min_profitable_iters 390#define LOOP_VINFO_COST_MODEL_THRESHOLD(L) (L)->th 391#define LOOP_VINFO_VECTORIZABLE_P(L) (L)->vectorizable 392#define LOOP_VINFO_VECT_FACTOR(L) (L)->vectorization_factor 393#define LOOP_VINFO_PTR_MASK(L) (L)->ptr_mask 394#define LOOP_VINFO_LOOP_NEST(L) (L)->loop_nest 395#define LOOP_VINFO_DATAREFS(L) (L)->datarefs 396#define LOOP_VINFO_DDRS(L) (L)->ddrs 397#define LOOP_VINFO_INT_NITERS(L) (TREE_INT_CST_LOW ((L)->num_iters)) 398#define LOOP_VINFO_PEELING_FOR_ALIGNMENT(L) (L)->peeling_for_alignment 399#define LOOP_VINFO_UNALIGNED_DR(L) (L)->unaligned_dr 400#define LOOP_VINFO_MAY_MISALIGN_STMTS(L) (L)->may_misalign_stmts 401#define LOOP_VINFO_MAY_ALIAS_DDRS(L) (L)->may_alias_ddrs 402#define LOOP_VINFO_COMP_ALIAS_DDRS(L) (L)->comp_alias_ddrs 403#define LOOP_VINFO_GROUPED_STORES(L) (L)->grouped_stores 404#define LOOP_VINFO_SLP_INSTANCES(L) (L)->slp_instances 405#define LOOP_VINFO_SLP_UNROLLING_FACTOR(L) (L)->slp_unrolling_factor 406#define LOOP_VINFO_REDUCTIONS(L) (L)->reductions 407#define LOOP_VINFO_REDUCTION_CHAINS(L) (L)->reduction_chains 408#define LOOP_VINFO_PEELING_HTAB(L) (L)->peeling_htab 409#define LOOP_VINFO_TARGET_COST_DATA(L) (L)->target_cost_data 410#define LOOP_VINFO_PEELING_FOR_GAPS(L) (L)->peeling_for_gaps 411#define LOOP_VINFO_OPERANDS_SWAPPED(L) (L)->operands_swapped 412#define LOOP_VINFO_PEELING_FOR_NITER(L) (L)->peeling_for_niter 413#define LOOP_VINFO_NO_DATA_DEPENDENCIES(L) (L)->no_data_dependencies 414#define LOOP_VINFO_SCALAR_LOOP(L) (L)->scalar_loop 415 416#define LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT(L) \ 417 ((L)->may_misalign_stmts.length () > 0) 418#define LOOP_REQUIRES_VERSIONING_FOR_ALIAS(L) \ 419 ((L)->may_alias_ddrs.length () > 0) 420 421#define LOOP_VINFO_NITERS_KNOWN_P(L) \ 422 (tree_fits_shwi_p ((L)->num_iters) && tree_to_shwi ((L)->num_iters) > 0) 423 424static inline loop_vec_info 425loop_vec_info_for_loop (struct loop *loop) 426{ 427 return (loop_vec_info) loop->aux; 428} 429 430static inline bool 431nested_in_vect_loop_p (struct loop *loop, gimple stmt) 432{ 433 return (loop->inner 434 && (loop->inner == (gimple_bb (stmt))->loop_father)); 435} 436 437typedef struct _bb_vec_info { 438 439 basic_block bb; 440 /* All interleaving chains of stores in the basic block, represented by the 441 first stmt in the chain. */ 442 vec<gimple> grouped_stores; 443 444 /* All SLP instances in the basic block. This is a subset of the set of 445 GROUP_STORES of the basic block. */ 446 vec<slp_instance> slp_instances; 447 448 /* All data references in the basic block. */ 449 vec<data_reference_p> datarefs; 450 451 /* All data dependences in the basic block. */ 452 vec<ddr_p> ddrs; 453 454 /* Cost data used by the target cost model. */ 455 void *target_cost_data; 456 457} *bb_vec_info; 458 459#define BB_VINFO_BB(B) (B)->bb 460#define BB_VINFO_GROUPED_STORES(B) (B)->grouped_stores 461#define BB_VINFO_SLP_INSTANCES(B) (B)->slp_instances 462#define BB_VINFO_DATAREFS(B) (B)->datarefs 463#define BB_VINFO_DDRS(B) (B)->ddrs 464#define BB_VINFO_TARGET_COST_DATA(B) (B)->target_cost_data 465 466static inline bb_vec_info 467vec_info_for_bb (basic_block bb) 468{ 469 return (bb_vec_info) bb->aux; 470} 471 472/*-----------------------------------------------------------------*/ 473/* Info on vectorized defs. */ 474/*-----------------------------------------------------------------*/ 475enum stmt_vec_info_type { 476 undef_vec_info_type = 0, 477 load_vec_info_type, 478 store_vec_info_type, 479 shift_vec_info_type, 480 op_vec_info_type, 481 call_vec_info_type, 482 call_simd_clone_vec_info_type, 483 assignment_vec_info_type, 484 condition_vec_info_type, 485 reduc_vec_info_type, 486 induc_vec_info_type, 487 type_promotion_vec_info_type, 488 type_demotion_vec_info_type, 489 type_conversion_vec_info_type, 490 loop_exit_ctrl_vec_info_type 491}; 492 493/* Indicates whether/how a variable is used in the scope of loop/basic 494 block. */ 495enum vect_relevant { 496 vect_unused_in_scope = 0, 497 /* The def is in the inner loop, and the use is in the outer loop, and the 498 use is a reduction stmt. */ 499 vect_used_in_outer_by_reduction, 500 /* The def is in the inner loop, and the use is in the outer loop (and is 501 not part of reduction). */ 502 vect_used_in_outer, 503 504 /* defs that feed computations that end up (only) in a reduction. These 505 defs may be used by non-reduction stmts, but eventually, any 506 computations/values that are affected by these defs are used to compute 507 a reduction (i.e. don't get stored to memory, for example). We use this 508 to identify computations that we can change the order in which they are 509 computed. */ 510 vect_used_by_reduction, 511 512 vect_used_in_scope 513}; 514 515/* The type of vectorization that can be applied to the stmt: regular loop-based 516 vectorization; pure SLP - the stmt is a part of SLP instances and does not 517 have uses outside SLP instances; or hybrid SLP and loop-based - the stmt is 518 a part of SLP instance and also must be loop-based vectorized, since it has 519 uses outside SLP sequences. 520 521 In the loop context the meanings of pure and hybrid SLP are slightly 522 different. By saying that pure SLP is applied to the loop, we mean that we 523 exploit only intra-iteration parallelism in the loop; i.e., the loop can be 524 vectorized without doing any conceptual unrolling, cause we don't pack 525 together stmts from different iterations, only within a single iteration. 526 Loop hybrid SLP means that we exploit both intra-iteration and 527 inter-iteration parallelism (e.g., number of elements in the vector is 4 528 and the slp-group-size is 2, in which case we don't have enough parallelism 529 within an iteration, so we obtain the rest of the parallelism from subsequent 530 iterations by unrolling the loop by 2). */ 531enum slp_vect_type { 532 loop_vect = 0, 533 pure_slp, 534 hybrid 535}; 536 537 538typedef struct data_reference *dr_p; 539 540typedef struct _stmt_vec_info { 541 542 enum stmt_vec_info_type type; 543 544 /* Indicates whether this stmts is part of a computation whose result is 545 used outside the loop. */ 546 bool live; 547 548 /* Stmt is part of some pattern (computation idiom) */ 549 bool in_pattern_p; 550 551 /* The stmt to which this info struct refers to. */ 552 gimple stmt; 553 554 /* The loop_vec_info with respect to which STMT is vectorized. */ 555 loop_vec_info loop_vinfo; 556 557 /* The vector type to be used for the LHS of this statement. */ 558 tree vectype; 559 560 /* The vectorized version of the stmt. */ 561 gimple vectorized_stmt; 562 563 564 /** The following is relevant only for stmts that contain a non-scalar 565 data-ref (array/pointer/struct access). A GIMPLE stmt is expected to have 566 at most one such data-ref. **/ 567 568 /* Information about the data-ref (access function, etc), 569 relative to the inner-most containing loop. */ 570 struct data_reference *data_ref_info; 571 572 /* Information about the data-ref relative to this loop 573 nest (the loop that is being considered for vectorization). */ 574 tree dr_base_address; 575 tree dr_init; 576 tree dr_offset; 577 tree dr_step; 578 tree dr_aligned_to; 579 580 /* For loop PHI nodes, the evolution part of it. This makes sure 581 this information is still available in vect_update_ivs_after_vectorizer 582 where we may not be able to re-analyze the PHI nodes evolution as 583 peeling for the prologue loop can make it unanalyzable. The evolution 584 part is still correct though. */ 585 tree loop_phi_evolution_part; 586 587 /* Used for various bookkeeping purposes, generally holding a pointer to 588 some other stmt S that is in some way "related" to this stmt. 589 Current use of this field is: 590 If this stmt is part of a pattern (i.e. the field 'in_pattern_p' is 591 true): S is the "pattern stmt" that represents (and replaces) the 592 sequence of stmts that constitutes the pattern. Similarly, the 593 related_stmt of the "pattern stmt" points back to this stmt (which is 594 the last stmt in the original sequence of stmts that constitutes the 595 pattern). */ 596 gimple related_stmt; 597 598 /* Used to keep a sequence of def stmts of a pattern stmt if such exists. */ 599 gimple_seq pattern_def_seq; 600 601 /* List of datarefs that are known to have the same alignment as the dataref 602 of this stmt. */ 603 vec<dr_p> same_align_refs; 604 605 /* Selected SIMD clone's function info. First vector element 606 is SIMD clone's function decl, followed by a pair of trees (base + step) 607 for linear arguments (pair of NULLs for other arguments). */ 608 vec<tree> simd_clone_info; 609 610 /* Classify the def of this stmt. */ 611 enum vect_def_type def_type; 612 613 /* Whether the stmt is SLPed, loop-based vectorized, or both. */ 614 enum slp_vect_type slp_type; 615 616 /* Interleaving and reduction chains info. */ 617 /* First element in the group. */ 618 gimple first_element; 619 /* Pointer to the next element in the group. */ 620 gimple next_element; 621 /* For data-refs, in case that two or more stmts share data-ref, this is the 622 pointer to the previously detected stmt with the same dr. */ 623 gimple same_dr_stmt; 624 /* The size of the group. */ 625 unsigned int size; 626 /* For stores, number of stores from this group seen. We vectorize the last 627 one. */ 628 unsigned int store_count; 629 /* For loads only, the gap from the previous load. For consecutive loads, GAP 630 is 1. */ 631 unsigned int gap; 632 633 /* The minimum negative dependence distance this stmt participates in 634 or zero if none. */ 635 unsigned int min_neg_dist; 636 637 /* Not all stmts in the loop need to be vectorized. e.g, the increment 638 of the loop induction variable and computation of array indexes. relevant 639 indicates whether the stmt needs to be vectorized. */ 640 enum vect_relevant relevant; 641 642 /* The bb_vec_info with respect to which STMT is vectorized. */ 643 bb_vec_info bb_vinfo; 644 645 /* Is this statement vectorizable or should it be skipped in (partial) 646 vectorization. */ 647 bool vectorizable; 648 649 /* For loads only, true if this is a gather load. */ 650 bool gather_p; 651 bool stride_load_p; 652 653 /* For both loads and stores. */ 654 bool simd_lane_access_p; 655} *stmt_vec_info; 656 657/* Access Functions. */ 658#define STMT_VINFO_TYPE(S) (S)->type 659#define STMT_VINFO_STMT(S) (S)->stmt 660#define STMT_VINFO_LOOP_VINFO(S) (S)->loop_vinfo 661#define STMT_VINFO_BB_VINFO(S) (S)->bb_vinfo 662#define STMT_VINFO_RELEVANT(S) (S)->relevant 663#define STMT_VINFO_LIVE_P(S) (S)->live 664#define STMT_VINFO_VECTYPE(S) (S)->vectype 665#define STMT_VINFO_VEC_STMT(S) (S)->vectorized_stmt 666#define STMT_VINFO_VECTORIZABLE(S) (S)->vectorizable 667#define STMT_VINFO_DATA_REF(S) (S)->data_ref_info 668#define STMT_VINFO_GATHER_P(S) (S)->gather_p 669#define STMT_VINFO_STRIDE_LOAD_P(S) (S)->stride_load_p 670#define STMT_VINFO_SIMD_LANE_ACCESS_P(S) (S)->simd_lane_access_p 671 672#define STMT_VINFO_DR_BASE_ADDRESS(S) (S)->dr_base_address 673#define STMT_VINFO_DR_INIT(S) (S)->dr_init 674#define STMT_VINFO_DR_OFFSET(S) (S)->dr_offset 675#define STMT_VINFO_DR_STEP(S) (S)->dr_step 676#define STMT_VINFO_DR_ALIGNED_TO(S) (S)->dr_aligned_to 677 678#define STMT_VINFO_IN_PATTERN_P(S) (S)->in_pattern_p 679#define STMT_VINFO_RELATED_STMT(S) (S)->related_stmt 680#define STMT_VINFO_PATTERN_DEF_SEQ(S) (S)->pattern_def_seq 681#define STMT_VINFO_SAME_ALIGN_REFS(S) (S)->same_align_refs 682#define STMT_VINFO_SIMD_CLONE_INFO(S) (S)->simd_clone_info 683#define STMT_VINFO_DEF_TYPE(S) (S)->def_type 684#define STMT_VINFO_GROUP_FIRST_ELEMENT(S) (S)->first_element 685#define STMT_VINFO_GROUP_NEXT_ELEMENT(S) (S)->next_element 686#define STMT_VINFO_GROUP_SIZE(S) (S)->size 687#define STMT_VINFO_GROUP_STORE_COUNT(S) (S)->store_count 688#define STMT_VINFO_GROUP_GAP(S) (S)->gap 689#define STMT_VINFO_GROUP_SAME_DR_STMT(S) (S)->same_dr_stmt 690#define STMT_VINFO_GROUPED_ACCESS(S) ((S)->first_element != NULL && (S)->data_ref_info) 691#define STMT_VINFO_LOOP_PHI_EVOLUTION_PART(S) (S)->loop_phi_evolution_part 692#define STMT_VINFO_MIN_NEG_DIST(S) (S)->min_neg_dist 693 694#define GROUP_FIRST_ELEMENT(S) (S)->first_element 695#define GROUP_NEXT_ELEMENT(S) (S)->next_element 696#define GROUP_SIZE(S) (S)->size 697#define GROUP_STORE_COUNT(S) (S)->store_count 698#define GROUP_GAP(S) (S)->gap 699#define GROUP_SAME_DR_STMT(S) (S)->same_dr_stmt 700 701#define STMT_VINFO_RELEVANT_P(S) ((S)->relevant != vect_unused_in_scope) 702 703#define HYBRID_SLP_STMT(S) ((S)->slp_type == hybrid) 704#define PURE_SLP_STMT(S) ((S)->slp_type == pure_slp) 705#define STMT_SLP_TYPE(S) (S)->slp_type 706 707struct dataref_aux { 708 int misalignment; 709 /* If true the alignment of base_decl needs to be increased. */ 710 bool base_misaligned; 711 /* If true we know the base is at least vector element alignment aligned. */ 712 bool base_element_aligned; 713 tree base_decl; 714}; 715 716#define DR_VECT_AUX(dr) ((dataref_aux *)(dr)->aux) 717 718#define VECT_MAX_COST 1000 719 720/* The maximum number of intermediate steps required in multi-step type 721 conversion. */ 722#define MAX_INTERM_CVT_STEPS 3 723 724/* The maximum vectorization factor supported by any target (V64QI). */ 725#define MAX_VECTORIZATION_FACTOR 64 726 727/* Avoid GTY(()) on stmt_vec_info. */ 728typedef void *vec_void_p; 729 730extern vec<vec_void_p> stmt_vec_info_vec; 731 732void init_stmt_vec_info_vec (void); 733void free_stmt_vec_info_vec (void); 734 735/* Return a stmt_vec_info corresponding to STMT. */ 736 737static inline stmt_vec_info 738vinfo_for_stmt (gimple stmt) 739{ 740 unsigned int uid = gimple_uid (stmt); 741 if (uid == 0) 742 return NULL; 743 744 return (stmt_vec_info) stmt_vec_info_vec[uid - 1]; 745} 746 747/* Set vectorizer information INFO for STMT. */ 748 749static inline void 750set_vinfo_for_stmt (gimple stmt, stmt_vec_info info) 751{ 752 unsigned int uid = gimple_uid (stmt); 753 if (uid == 0) 754 { 755 gcc_checking_assert (info); 756 uid = stmt_vec_info_vec.length () + 1; 757 gimple_set_uid (stmt, uid); 758 stmt_vec_info_vec.safe_push ((vec_void_p) info); 759 } 760 else 761 stmt_vec_info_vec[uid - 1] = (vec_void_p) info; 762} 763 764/* Return the earlier statement between STMT1 and STMT2. */ 765 766static inline gimple 767get_earlier_stmt (gimple stmt1, gimple stmt2) 768{ 769 unsigned int uid1, uid2; 770 771 if (stmt1 == NULL) 772 return stmt2; 773 774 if (stmt2 == NULL) 775 return stmt1; 776 777 uid1 = gimple_uid (stmt1); 778 uid2 = gimple_uid (stmt2); 779 780 if (uid1 == 0 || uid2 == 0) 781 return NULL; 782 783 gcc_checking_assert (uid1 <= stmt_vec_info_vec.length () 784 && uid2 <= stmt_vec_info_vec.length ()); 785 786 if (uid1 < uid2) 787 return stmt1; 788 else 789 return stmt2; 790} 791 792/* Return the later statement between STMT1 and STMT2. */ 793 794static inline gimple 795get_later_stmt (gimple stmt1, gimple stmt2) 796{ 797 unsigned int uid1, uid2; 798 799 if (stmt1 == NULL) 800 return stmt2; 801 802 if (stmt2 == NULL) 803 return stmt1; 804 805 uid1 = gimple_uid (stmt1); 806 uid2 = gimple_uid (stmt2); 807 808 if (uid1 == 0 || uid2 == 0) 809 return NULL; 810 811 gcc_assert (uid1 <= stmt_vec_info_vec.length ()); 812 gcc_assert (uid2 <= stmt_vec_info_vec.length ()); 813 814 if (uid1 > uid2) 815 return stmt1; 816 else 817 return stmt2; 818} 819 820/* Return TRUE if a statement represented by STMT_INFO is a part of a 821 pattern. */ 822 823static inline bool 824is_pattern_stmt_p (stmt_vec_info stmt_info) 825{ 826 gimple related_stmt; 827 stmt_vec_info related_stmt_info; 828 829 related_stmt = STMT_VINFO_RELATED_STMT (stmt_info); 830 if (related_stmt 831 && (related_stmt_info = vinfo_for_stmt (related_stmt)) 832 && STMT_VINFO_IN_PATTERN_P (related_stmt_info)) 833 return true; 834 835 return false; 836} 837 838/* Return true if BB is a loop header. */ 839 840static inline bool 841is_loop_header_bb_p (basic_block bb) 842{ 843 if (bb == (bb->loop_father)->header) 844 return true; 845 gcc_checking_assert (EDGE_COUNT (bb->preds) == 1); 846 return false; 847} 848 849/* Return pow2 (X). */ 850 851static inline int 852vect_pow2 (int x) 853{ 854 int i, res = 1; 855 856 for (i = 0; i < x; i++) 857 res *= 2; 858 859 return res; 860} 861 862/* Alias targetm.vectorize.builtin_vectorization_cost. */ 863 864static inline int 865builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost, 866 tree vectype, int misalign) 867{ 868 return targetm.vectorize.builtin_vectorization_cost (type_of_cost, 869 vectype, misalign); 870} 871 872/* Get cost by calling cost target builtin. */ 873 874static inline 875int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost) 876{ 877 return builtin_vectorization_cost (type_of_cost, NULL, 0); 878} 879 880/* Alias targetm.vectorize.init_cost. */ 881 882static inline void * 883init_cost (struct loop *loop_info) 884{ 885 return targetm.vectorize.init_cost (loop_info); 886} 887 888/* Alias targetm.vectorize.add_stmt_cost. */ 889 890static inline unsigned 891add_stmt_cost (void *data, int count, enum vect_cost_for_stmt kind, 892 stmt_vec_info stmt_info, int misalign, 893 enum vect_cost_model_location where) 894{ 895 return targetm.vectorize.add_stmt_cost (data, count, kind, 896 stmt_info, misalign, where); 897} 898 899/* Alias targetm.vectorize.finish_cost. */ 900 901static inline void 902finish_cost (void *data, unsigned *prologue_cost, 903 unsigned *body_cost, unsigned *epilogue_cost) 904{ 905 targetm.vectorize.finish_cost (data, prologue_cost, body_cost, epilogue_cost); 906} 907 908/* Alias targetm.vectorize.destroy_cost_data. */ 909 910static inline void 911destroy_cost_data (void *data) 912{ 913 targetm.vectorize.destroy_cost_data (data); 914} 915 916/*-----------------------------------------------------------------*/ 917/* Info on data references alignment. */ 918/*-----------------------------------------------------------------*/ 919inline void 920set_dr_misalignment (struct data_reference *dr, int val) 921{ 922 dataref_aux *data_aux = DR_VECT_AUX (dr); 923 924 if (!data_aux) 925 { 926 data_aux = XCNEW (dataref_aux); 927 dr->aux = data_aux; 928 } 929 930 data_aux->misalignment = val; 931} 932 933inline int 934dr_misalignment (struct data_reference *dr) 935{ 936 return DR_VECT_AUX (dr)->misalignment; 937} 938 939/* Reflects actual alignment of first access in the vectorized loop, 940 taking into account peeling/versioning if applied. */ 941#define DR_MISALIGNMENT(DR) dr_misalignment (DR) 942#define SET_DR_MISALIGNMENT(DR, VAL) set_dr_misalignment (DR, VAL) 943 944/* Return TRUE if the data access is aligned, and FALSE otherwise. */ 945 946static inline bool 947aligned_access_p (struct data_reference *data_ref_info) 948{ 949 return (DR_MISALIGNMENT (data_ref_info) == 0); 950} 951 952/* Return TRUE if the alignment of the data access is known, and FALSE 953 otherwise. */ 954 955static inline bool 956known_alignment_for_access_p (struct data_reference *data_ref_info) 957{ 958 return (DR_MISALIGNMENT (data_ref_info) != -1); 959} 960 961 962/* Return true if the vect cost model is unlimited. */ 963static inline bool 964unlimited_cost_model (loop_p loop) 965{ 966 if (loop != NULL && loop->force_vectorize 967 && flag_simd_cost_model != VECT_COST_MODEL_DEFAULT) 968 return flag_simd_cost_model == VECT_COST_MODEL_UNLIMITED; 969 return (flag_vect_cost_model == VECT_COST_MODEL_UNLIMITED); 970} 971 972/* Source location */ 973extern source_location vect_location; 974 975/*-----------------------------------------------------------------*/ 976/* Function prototypes. */ 977/*-----------------------------------------------------------------*/ 978 979/* Simple loop peeling and versioning utilities for vectorizer's purposes - 980 in tree-vect-loop-manip.c. */ 981extern void slpeel_make_loop_iterate_ntimes (struct loop *, tree); 982extern bool slpeel_can_duplicate_loop_p (const struct loop *, const_edge); 983struct loop *slpeel_tree_duplicate_loop_to_edge_cfg (struct loop *, 984 struct loop *, edge); 985extern void vect_loop_versioning (loop_vec_info, unsigned int, bool); 986extern void vect_do_peeling_for_loop_bound (loop_vec_info, tree, tree, 987 unsigned int, bool); 988extern void vect_do_peeling_for_alignment (loop_vec_info, tree, 989 unsigned int, bool); 990extern source_location find_loop_location (struct loop *); 991extern bool vect_can_advance_ivs_p (loop_vec_info); 992 993/* In tree-vect-stmts.c. */ 994extern unsigned int current_vector_size; 995extern tree get_vectype_for_scalar_type (tree); 996extern tree get_same_sized_vectype (tree, tree); 997extern bool vect_is_simple_use (tree, gimple, loop_vec_info, 998 bb_vec_info, gimple *, 999 tree *, enum vect_def_type *); 1000extern bool vect_is_simple_use_1 (tree, gimple, loop_vec_info, 1001 bb_vec_info, gimple *, 1002 tree *, enum vect_def_type *, tree *); 1003extern bool supportable_widening_operation (enum tree_code, gimple, tree, tree, 1004 enum tree_code *, enum tree_code *, 1005 int *, vec<tree> *); 1006extern bool supportable_narrowing_operation (enum tree_code, tree, tree, 1007 enum tree_code *, 1008 int *, vec<tree> *); 1009extern stmt_vec_info new_stmt_vec_info (gimple stmt, loop_vec_info, 1010 bb_vec_info); 1011extern void free_stmt_vec_info (gimple stmt); 1012extern tree vectorizable_function (gcall *, tree, tree); 1013extern void vect_model_simple_cost (stmt_vec_info, int, enum vect_def_type *, 1014 stmt_vector_for_cost *, 1015 stmt_vector_for_cost *); 1016extern void vect_model_store_cost (stmt_vec_info, int, bool, 1017 enum vect_def_type, slp_tree, 1018 stmt_vector_for_cost *, 1019 stmt_vector_for_cost *); 1020extern void vect_model_load_cost (stmt_vec_info, int, bool, slp_tree, 1021 stmt_vector_for_cost *, 1022 stmt_vector_for_cost *); 1023extern unsigned record_stmt_cost (stmt_vector_for_cost *, int, 1024 enum vect_cost_for_stmt, stmt_vec_info, 1025 int, enum vect_cost_model_location); 1026extern void vect_finish_stmt_generation (gimple, gimple, 1027 gimple_stmt_iterator *); 1028extern bool vect_mark_stmts_to_be_vectorized (loop_vec_info); 1029extern tree vect_get_vec_def_for_operand (tree, gimple, tree *); 1030extern tree vect_init_vector (gimple, tree, tree, 1031 gimple_stmt_iterator *); 1032extern tree vect_get_vec_def_for_stmt_copy (enum vect_def_type, tree); 1033extern bool vect_transform_stmt (gimple, gimple_stmt_iterator *, 1034 bool *, slp_tree, slp_instance); 1035extern void vect_remove_stores (gimple); 1036extern bool vect_analyze_stmt (gimple, bool *, slp_tree); 1037extern bool vectorizable_condition (gimple, gimple_stmt_iterator *, gimple *, 1038 tree, int, slp_tree); 1039extern void vect_get_load_cost (struct data_reference *, int, bool, 1040 unsigned int *, unsigned int *, 1041 stmt_vector_for_cost *, 1042 stmt_vector_for_cost *, bool); 1043extern void vect_get_store_cost (struct data_reference *, int, 1044 unsigned int *, stmt_vector_for_cost *); 1045extern bool vect_supportable_shift (enum tree_code, tree); 1046extern void vect_get_vec_defs (tree, tree, gimple, vec<tree> *, 1047 vec<tree> *, slp_tree, int); 1048extern tree vect_gen_perm_mask_any (tree, const unsigned char *); 1049extern tree vect_gen_perm_mask_checked (tree, const unsigned char *); 1050 1051/* In tree-vect-data-refs.c. */ 1052extern bool vect_can_force_dr_alignment_p (const_tree, unsigned int); 1053extern enum dr_alignment_support vect_supportable_dr_alignment 1054 (struct data_reference *, bool); 1055extern tree vect_get_smallest_scalar_type (gimple, HOST_WIDE_INT *, 1056 HOST_WIDE_INT *); 1057extern bool vect_analyze_data_ref_dependences (loop_vec_info, int *); 1058extern bool vect_slp_analyze_data_ref_dependences (bb_vec_info); 1059extern bool vect_enhance_data_refs_alignment (loop_vec_info); 1060extern bool vect_analyze_data_refs_alignment (loop_vec_info, bb_vec_info); 1061extern bool vect_verify_datarefs_alignment (loop_vec_info, bb_vec_info); 1062extern bool vect_analyze_data_ref_accesses (loop_vec_info, bb_vec_info); 1063extern bool vect_prune_runtime_alias_test_list (loop_vec_info); 1064extern tree vect_check_gather (gimple, loop_vec_info, tree *, tree *, 1065 int *); 1066extern bool vect_analyze_data_refs (loop_vec_info, bb_vec_info, int *, 1067 unsigned *); 1068extern tree vect_create_data_ref_ptr (gimple, tree, struct loop *, tree, 1069 tree *, gimple_stmt_iterator *, 1070 gimple *, bool, bool *, 1071 tree = NULL_TREE); 1072extern tree bump_vector_ptr (tree, gimple, gimple_stmt_iterator *, gimple, tree); 1073extern tree vect_create_destination_var (tree, tree); 1074extern bool vect_grouped_store_supported (tree, unsigned HOST_WIDE_INT); 1075extern bool vect_store_lanes_supported (tree, unsigned HOST_WIDE_INT); 1076extern bool vect_grouped_load_supported (tree, unsigned HOST_WIDE_INT); 1077extern bool vect_load_lanes_supported (tree, unsigned HOST_WIDE_INT); 1078extern void vect_permute_store_chain (vec<tree> ,unsigned int, gimple, 1079 gimple_stmt_iterator *, vec<tree> *); 1080extern tree vect_setup_realignment (gimple, gimple_stmt_iterator *, tree *, 1081 enum dr_alignment_support, tree, 1082 struct loop **); 1083extern void vect_transform_grouped_load (gimple, vec<tree> , int, 1084 gimple_stmt_iterator *); 1085extern void vect_record_grouped_load_vectors (gimple, vec<tree> ); 1086extern tree vect_get_new_vect_var (tree, enum vect_var_kind, const char *); 1087extern tree vect_create_addr_base_for_vector_ref (gimple, gimple_seq *, 1088 tree, struct loop *, 1089 tree = NULL_TREE); 1090 1091/* In tree-vect-loop.c. */ 1092/* FORNOW: Used in tree-parloops.c. */ 1093extern void destroy_loop_vec_info (loop_vec_info, bool); 1094extern gimple vect_force_simple_reduction (loop_vec_info, gimple, bool, bool *); 1095/* Drive for loop analysis stage. */ 1096extern loop_vec_info vect_analyze_loop (struct loop *); 1097/* Drive for loop transformation stage. */ 1098extern void vect_transform_loop (loop_vec_info); 1099extern loop_vec_info vect_analyze_loop_form (struct loop *); 1100extern bool vectorizable_live_operation (gimple, gimple_stmt_iterator *, 1101 gimple *); 1102extern bool vectorizable_reduction (gimple, gimple_stmt_iterator *, gimple *, 1103 slp_tree); 1104extern bool vectorizable_induction (gimple, gimple_stmt_iterator *, gimple *); 1105extern tree get_initial_def_for_reduction (gimple, tree, tree *); 1106extern int vect_min_worthwhile_factor (enum tree_code); 1107extern int vect_get_known_peeling_cost (loop_vec_info, int, int *, 1108 stmt_vector_for_cost *, 1109 stmt_vector_for_cost *, 1110 stmt_vector_for_cost *); 1111extern int vect_get_single_scalar_iteration_cost (loop_vec_info, 1112 stmt_vector_for_cost *); 1113 1114/* In tree-vect-slp.c. */ 1115extern void vect_free_slp_instance (slp_instance); 1116extern bool vect_transform_slp_perm_load (slp_tree, vec<tree> , 1117 gimple_stmt_iterator *, int, 1118 slp_instance, bool); 1119extern bool vect_schedule_slp (loop_vec_info, bb_vec_info); 1120extern void vect_update_slp_costs_according_to_vf (loop_vec_info); 1121extern bool vect_analyze_slp (loop_vec_info, bb_vec_info, unsigned); 1122extern bool vect_make_slp_decision (loop_vec_info); 1123extern void vect_detect_hybrid_slp (loop_vec_info); 1124extern void vect_get_slp_defs (vec<tree> , slp_tree, 1125 vec<vec<tree> > *, int); 1126 1127extern source_location find_bb_location (basic_block); 1128extern bb_vec_info vect_slp_analyze_bb (basic_block); 1129extern void vect_slp_transform_bb (basic_block); 1130 1131/* In tree-vect-patterns.c. */ 1132/* Pattern recognition functions. 1133 Additional pattern recognition functions can (and will) be added 1134 in the future. */ 1135typedef gimple (* vect_recog_func_ptr) (vec<gimple> *, tree *, tree *); 1136#define NUM_PATTERNS 12 1137void vect_pattern_recog (loop_vec_info, bb_vec_info); 1138 1139/* In tree-vectorizer.c. */ 1140unsigned vectorize_loops (void); 1141void vect_destroy_datarefs (loop_vec_info, bb_vec_info); 1142 1143#endif /* GCC_TREE_VECTORIZER_H */ 1144