1 /* Vectorizer 2 Copyright (C) 2003-2022 Free Software Foundation, Inc. 3 Contributed by Dorit Naishlos <dorit (at) il.ibm.com> 4 5 This file is part of GCC. 6 7 GCC is free software; you can redistribute it and/or modify it under 8 the terms of the GNU General Public License as published by the Free 9 Software Foundation; either version 3, or (at your option) any later 10 version. 11 12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY 13 WARRANTY; without even the implied warranty of MERCHANTABILITY or 14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 15 for more details. 16 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/>. */ 20 21 #ifndef GCC_TREE_VECTORIZER_H 22 #define GCC_TREE_VECTORIZER_H 23 24 typedef class _stmt_vec_info *stmt_vec_info; 25 typedef struct _slp_tree *slp_tree; 26 27 #include "tree-data-ref.h" 28 #include "tree-hash-traits.h" 29 #include "target.h" 30 #include "internal-fn.h" 31 #include "tree-ssa-operands.h" 32 #include "gimple-match.h" 33 34 /* Used for naming of new temporaries. */ 35 enum vect_var_kind { 36 vect_simple_var, 37 vect_pointer_var, 38 vect_scalar_var, 39 vect_mask_var 40 }; 41 42 /* Defines type of operation. */ 43 enum operation_type { 44 unary_op = 1, 45 binary_op, 46 ternary_op 47 }; 48 49 /* Define type of available alignment support. */ 50 enum dr_alignment_support { 51 dr_unaligned_unsupported, 52 dr_unaligned_supported, 53 dr_explicit_realign, 54 dr_explicit_realign_optimized, 55 dr_aligned 56 }; 57 58 /* Define type of def-use cross-iteration cycle. */ 59 enum vect_def_type { 60 vect_uninitialized_def = 0, 61 vect_constant_def = 1, 62 vect_external_def, 63 vect_internal_def, 64 vect_induction_def, 65 vect_reduction_def, 66 vect_double_reduction_def, 67 vect_nested_cycle, 68 vect_unknown_def_type 69 }; 70 71 /* Define type of reduction. */ 72 enum vect_reduction_type { 73 TREE_CODE_REDUCTION, 74 COND_REDUCTION, 75 INTEGER_INDUC_COND_REDUCTION, 76 CONST_COND_REDUCTION, 77 78 /* Retain a scalar phi and use a FOLD_EXTRACT_LAST within the loop 79 to implement: 80 81 for (int i = 0; i < VF; ++i) 82 res = cond[i] ? val[i] : res; */ 83 EXTRACT_LAST_REDUCTION, 84 85 /* Use a folding reduction within the loop to implement: 86 87 for (int i = 0; i < VF; ++i) 88 res = res OP val[i]; 89 90 (with no reassocation). */ 91 FOLD_LEFT_REDUCTION 92 }; 93 94 #define VECTORIZABLE_CYCLE_DEF(D) (((D) == vect_reduction_def) \ 95 || ((D) == vect_double_reduction_def) \ 96 || ((D) == vect_nested_cycle)) 97 98 /* Structure to encapsulate information about a group of like 99 instructions to be presented to the target cost model. */ 100 struct stmt_info_for_cost { 101 int count; 102 enum vect_cost_for_stmt kind; 103 enum vect_cost_model_location where; 104 stmt_vec_info stmt_info; 105 slp_tree node; 106 tree vectype; 107 int misalign; 108 }; 109 110 typedef vec<stmt_info_for_cost> stmt_vector_for_cost; 111 112 /* Maps base addresses to an innermost_loop_behavior and the stmt it was 113 derived from that gives the maximum known alignment for that base. */ 114 typedef hash_map<tree_operand_hash, 115 std::pair<stmt_vec_info, innermost_loop_behavior *> > 116 vec_base_alignments; 117 118 /* Represents elements [START, START + LENGTH) of cyclical array OPS* 119 (i.e. OPS repeated to give at least START + LENGTH elements) */ 120 struct vect_scalar_ops_slice 121 { 122 tree op (unsigned int i) const; 123 bool all_same_p () const; 124 125 vec<tree> *ops; 126 unsigned int start; 127 unsigned int length; 128 }; 129 130 /* Return element I of the slice. */ 131 inline tree 132 vect_scalar_ops_slice::op (unsigned int i) const 133 { 134 return (*ops)[(i + start) % ops->length ()]; 135 } 136 137 /* Hash traits for vect_scalar_ops_slice. */ 138 struct vect_scalar_ops_slice_hash : typed_noop_remove<vect_scalar_ops_slice> 139 { 140 typedef vect_scalar_ops_slice value_type; 141 typedef vect_scalar_ops_slice compare_type; 142 143 static const bool empty_zero_p = true; 144 145 static void mark_deleted (value_type &s) { s.length = ~0U; } 146 static void mark_empty (value_type &s) { s.length = 0; } 147 static bool is_deleted (const value_type &s) { return s.length == ~0U; } 148 static bool is_empty (const value_type &s) { return s.length == 0; } 149 static hashval_t hash (const value_type &); 150 static bool equal (const value_type &, const compare_type &); 151 }; 152 153 /************************************************************************ 154 SLP 155 ************************************************************************/ 156 typedef vec<std::pair<unsigned, unsigned> > lane_permutation_t; 157 typedef vec<unsigned> load_permutation_t; 158 159 /* A computation tree of an SLP instance. Each node corresponds to a group of 160 stmts to be packed in a SIMD stmt. */ 161 struct _slp_tree { 162 _slp_tree (); 163 ~_slp_tree (); 164 165 /* Nodes that contain def-stmts of this node statements operands. */ 166 vec<slp_tree> children; 167 168 /* A group of scalar stmts to be vectorized together. */ 169 vec<stmt_vec_info> stmts; 170 /* A group of scalar operands to be vectorized together. */ 171 vec<tree> ops; 172 /* The representative that should be used for analysis and 173 code generation. */ 174 stmt_vec_info representative; 175 176 /* Load permutation relative to the stores, NULL if there is no 177 permutation. */ 178 load_permutation_t load_permutation; 179 /* Lane permutation of the operands scalar lanes encoded as pairs 180 of { operand number, lane number }. The number of elements 181 denotes the number of output lanes. */ 182 lane_permutation_t lane_permutation; 183 184 tree vectype; 185 /* Vectorized stmt/s. */ 186 vec<gimple *> vec_stmts; 187 vec<tree> vec_defs; 188 /* Number of vector stmts that are created to replace the group of scalar 189 stmts. It is calculated during the transformation phase as the number of 190 scalar elements in one scalar iteration (GROUP_SIZE) multiplied by VF 191 divided by vector size. */ 192 unsigned int vec_stmts_size; 193 194 /* Reference count in the SLP graph. */ 195 unsigned int refcnt; 196 /* The maximum number of vector elements for the subtree rooted 197 at this node. */ 198 poly_uint64 max_nunits; 199 /* The DEF type of this node. */ 200 enum vect_def_type def_type; 201 /* The number of scalar lanes produced by this node. */ 202 unsigned int lanes; 203 /* The operation of this node. */ 204 enum tree_code code; 205 206 int vertex; 207 208 /* If not NULL this is a cached failed SLP discovery attempt with 209 the lanes that failed during SLP discovery as 'false'. This is 210 a copy of the matches array. */ 211 bool *failed; 212 213 /* Allocate from slp_tree_pool. */ 214 static void *operator new (size_t); 215 216 /* Return memory to slp_tree_pool. */ 217 static void operator delete (void *, size_t); 218 219 /* Linked list of nodes to release when we free the slp_tree_pool. */ 220 slp_tree next_node; 221 slp_tree prev_node; 222 }; 223 224 /* The enum describes the type of operations that an SLP instance 225 can perform. */ 226 227 enum slp_instance_kind { 228 slp_inst_kind_store, 229 slp_inst_kind_reduc_group, 230 slp_inst_kind_reduc_chain, 231 slp_inst_kind_bb_reduc, 232 slp_inst_kind_ctor 233 }; 234 235 /* SLP instance is a sequence of stmts in a loop that can be packed into 236 SIMD stmts. */ 237 typedef class _slp_instance { 238 public: 239 /* The root of SLP tree. */ 240 slp_tree root; 241 242 /* For vector constructors, the constructor stmt that the SLP tree is built 243 from, NULL otherwise. */ 244 vec<stmt_vec_info> root_stmts; 245 246 /* The unrolling factor required to vectorized this SLP instance. */ 247 poly_uint64 unrolling_factor; 248 249 /* The group of nodes that contain loads of this SLP instance. */ 250 vec<slp_tree> loads; 251 252 /* The SLP node containing the reduction PHIs. */ 253 slp_tree reduc_phis; 254 255 /* Vector cost of this entry to the SLP graph. */ 256 stmt_vector_for_cost cost_vec; 257 258 /* If this instance is the main entry of a subgraph the set of 259 entries into the same subgraph, including itself. */ 260 vec<_slp_instance *> subgraph_entries; 261 262 /* The type of operation the SLP instance is performing. */ 263 slp_instance_kind kind; 264 265 dump_user_location_t location () const; 266 } *slp_instance; 267 268 269 /* Access Functions. */ 270 #define SLP_INSTANCE_TREE(S) (S)->root 271 #define SLP_INSTANCE_UNROLLING_FACTOR(S) (S)->unrolling_factor 272 #define SLP_INSTANCE_LOADS(S) (S)->loads 273 #define SLP_INSTANCE_ROOT_STMTS(S) (S)->root_stmts 274 #define SLP_INSTANCE_KIND(S) (S)->kind 275 276 #define SLP_TREE_CHILDREN(S) (S)->children 277 #define SLP_TREE_SCALAR_STMTS(S) (S)->stmts 278 #define SLP_TREE_SCALAR_OPS(S) (S)->ops 279 #define SLP_TREE_REF_COUNT(S) (S)->refcnt 280 #define SLP_TREE_VEC_STMTS(S) (S)->vec_stmts 281 #define SLP_TREE_VEC_DEFS(S) (S)->vec_defs 282 #define SLP_TREE_NUMBER_OF_VEC_STMTS(S) (S)->vec_stmts_size 283 #define SLP_TREE_LOAD_PERMUTATION(S) (S)->load_permutation 284 #define SLP_TREE_LANE_PERMUTATION(S) (S)->lane_permutation 285 #define SLP_TREE_DEF_TYPE(S) (S)->def_type 286 #define SLP_TREE_VECTYPE(S) (S)->vectype 287 #define SLP_TREE_REPRESENTATIVE(S) (S)->representative 288 #define SLP_TREE_LANES(S) (S)->lanes 289 #define SLP_TREE_CODE(S) (S)->code 290 291 /* Key for map that records association between 292 scalar conditions and corresponding loop mask, and 293 is populated by vect_record_loop_mask. */ 294 295 struct scalar_cond_masked_key 296 { 297 scalar_cond_masked_key (tree t, unsigned ncopies_) 298 : ncopies (ncopies_) 299 { 300 get_cond_ops_from_tree (t); 301 } 302 303 void get_cond_ops_from_tree (tree); 304 305 unsigned ncopies; 306 bool inverted_p; 307 tree_code code; 308 tree op0; 309 tree op1; 310 }; 311 312 template<> 313 struct default_hash_traits<scalar_cond_masked_key> 314 { 315 typedef scalar_cond_masked_key compare_type; 316 typedef scalar_cond_masked_key value_type; 317 318 static inline hashval_t 319 hash (value_type v) 320 { 321 inchash::hash h; 322 h.add_int (v.code); 323 inchash::add_expr (v.op0, h, 0); 324 inchash::add_expr (v.op1, h, 0); 325 h.add_int (v.ncopies); 326 h.add_flag (v.inverted_p); 327 return h.end (); 328 } 329 330 static inline bool 331 equal (value_type existing, value_type candidate) 332 { 333 return (existing.ncopies == candidate.ncopies 334 && existing.code == candidate.code 335 && existing.inverted_p == candidate.inverted_p 336 && operand_equal_p (existing.op0, candidate.op0, 0) 337 && operand_equal_p (existing.op1, candidate.op1, 0)); 338 } 339 340 static const bool empty_zero_p = true; 341 342 static inline void 343 mark_empty (value_type &v) 344 { 345 v.ncopies = 0; 346 v.inverted_p = false; 347 } 348 349 static inline bool 350 is_empty (value_type v) 351 { 352 return v.ncopies == 0; 353 } 354 355 static inline void mark_deleted (value_type &) {} 356 357 static inline bool is_deleted (const value_type &) 358 { 359 return false; 360 } 361 362 static inline void remove (value_type &) {} 363 }; 364 365 typedef hash_set<scalar_cond_masked_key> scalar_cond_masked_set_type; 366 367 /* Key and map that records association between vector conditions and 368 corresponding loop mask, and is populated by prepare_vec_mask. */ 369 370 typedef pair_hash<tree_operand_hash, tree_operand_hash> tree_cond_mask_hash; 371 typedef hash_set<tree_cond_mask_hash> vec_cond_masked_set_type; 372 373 /* Describes two objects whose addresses must be unequal for the vectorized 374 loop to be valid. */ 375 typedef std::pair<tree, tree> vec_object_pair; 376 377 /* Records that vectorization is only possible if abs (EXPR) >= MIN_VALUE. 378 UNSIGNED_P is true if we can assume that abs (EXPR) == EXPR. */ 379 class vec_lower_bound { 380 public: 381 vec_lower_bound () {} 382 vec_lower_bound (tree e, bool u, poly_uint64 m) 383 : expr (e), unsigned_p (u), min_value (m) {} 384 385 tree expr; 386 bool unsigned_p; 387 poly_uint64 min_value; 388 }; 389 390 /* Vectorizer state shared between different analyses like vector sizes 391 of the same CFG region. */ 392 class vec_info_shared { 393 public: 394 vec_info_shared(); 395 ~vec_info_shared(); 396 397 void save_datarefs(); 398 void check_datarefs(); 399 400 /* The number of scalar stmts. */ 401 unsigned n_stmts; 402 403 /* All data references. Freed by free_data_refs, so not an auto_vec. */ 404 vec<data_reference_p> datarefs; 405 vec<data_reference> datarefs_copy; 406 407 /* The loop nest in which the data dependences are computed. */ 408 auto_vec<loop_p> loop_nest; 409 410 /* All data dependences. Freed by free_dependence_relations, so not 411 an auto_vec. */ 412 vec<ddr_p> ddrs; 413 }; 414 415 /* Vectorizer state common between loop and basic-block vectorization. */ 416 class vec_info { 417 public: 418 typedef hash_set<int_hash<machine_mode, E_VOIDmode, E_BLKmode> > mode_set; 419 enum vec_kind { bb, loop }; 420 421 vec_info (vec_kind, vec_info_shared *); 422 ~vec_info (); 423 424 stmt_vec_info add_stmt (gimple *); 425 stmt_vec_info add_pattern_stmt (gimple *, stmt_vec_info); 426 stmt_vec_info lookup_stmt (gimple *); 427 stmt_vec_info lookup_def (tree); 428 stmt_vec_info lookup_single_use (tree); 429 class dr_vec_info *lookup_dr (data_reference *); 430 void move_dr (stmt_vec_info, stmt_vec_info); 431 void remove_stmt (stmt_vec_info); 432 void replace_stmt (gimple_stmt_iterator *, stmt_vec_info, gimple *); 433 void insert_on_entry (stmt_vec_info, gimple *); 434 void insert_seq_on_entry (stmt_vec_info, gimple_seq); 435 436 /* The type of vectorization. */ 437 vec_kind kind; 438 439 /* Shared vectorizer state. */ 440 vec_info_shared *shared; 441 442 /* The mapping of GIMPLE UID to stmt_vec_info. */ 443 vec<stmt_vec_info> stmt_vec_infos; 444 /* Whether the above mapping is complete. */ 445 bool stmt_vec_info_ro; 446 447 /* The SLP graph. */ 448 auto_vec<slp_instance> slp_instances; 449 450 /* Maps base addresses to an innermost_loop_behavior that gives the maximum 451 known alignment for that base. */ 452 vec_base_alignments base_alignments; 453 454 /* All interleaving chains of stores, represented by the first 455 stmt in the chain. */ 456 auto_vec<stmt_vec_info> grouped_stores; 457 458 /* The set of vector modes used in the vectorized region. */ 459 mode_set used_vector_modes; 460 461 /* The argument we should pass to related_vector_mode when looking up 462 the vector mode for a scalar mode, or VOIDmode if we haven't yet 463 made any decisions about which vector modes to use. */ 464 machine_mode vector_mode; 465 466 private: 467 stmt_vec_info new_stmt_vec_info (gimple *stmt); 468 void set_vinfo_for_stmt (gimple *, stmt_vec_info, bool = true); 469 void free_stmt_vec_infos (); 470 void free_stmt_vec_info (stmt_vec_info); 471 }; 472 473 class _loop_vec_info; 474 class _bb_vec_info; 475 476 template<> 477 template<> 478 inline bool 479 is_a_helper <_loop_vec_info *>::test (vec_info *i) 480 { 481 return i->kind == vec_info::loop; 482 } 483 484 template<> 485 template<> 486 inline bool 487 is_a_helper <_bb_vec_info *>::test (vec_info *i) 488 { 489 return i->kind == vec_info::bb; 490 } 491 492 /* In general, we can divide the vector statements in a vectorized loop 493 into related groups ("rgroups") and say that for each rgroup there is 494 some nS such that the rgroup operates on nS values from one scalar 495 iteration followed by nS values from the next. That is, if VF is the 496 vectorization factor of the loop, the rgroup operates on a sequence: 497 498 (1,1) (1,2) ... (1,nS) (2,1) ... (2,nS) ... (VF,1) ... (VF,nS) 499 500 where (i,j) represents a scalar value with index j in a scalar 501 iteration with index i. 502 503 [ We use the term "rgroup" to emphasise that this grouping isn't 504 necessarily the same as the grouping of statements used elsewhere. 505 For example, if we implement a group of scalar loads using gather 506 loads, we'll use a separate gather load for each scalar load, and 507 thus each gather load will belong to its own rgroup. ] 508 509 In general this sequence will occupy nV vectors concatenated 510 together. If these vectors have nL lanes each, the total number 511 of scalar values N is given by: 512 513 N = nS * VF = nV * nL 514 515 None of nS, VF, nV and nL are required to be a power of 2. nS and nV 516 are compile-time constants but VF and nL can be variable (if the target 517 supports variable-length vectors). 518 519 In classical vectorization, each iteration of the vector loop would 520 handle exactly VF iterations of the original scalar loop. However, 521 in vector loops that are able to operate on partial vectors, a 522 particular iteration of the vector loop might handle fewer than VF 523 iterations of the scalar loop. The vector lanes that correspond to 524 iterations of the scalar loop are said to be "active" and the other 525 lanes are said to be "inactive". 526 527 In such vector loops, many rgroups need to be controlled to ensure 528 that they have no effect for the inactive lanes. Conceptually, each 529 such rgroup needs a sequence of booleans in the same order as above, 530 but with each (i,j) replaced by a boolean that indicates whether 531 iteration i is active. This sequence occupies nV vector controls 532 that again have nL lanes each. Thus the control sequence as a whole 533 consists of VF independent booleans that are each repeated nS times. 534 535 Taking mask-based approach as a partially-populated vectors example. 536 We make the simplifying assumption that if a sequence of nV masks is 537 suitable for one (nS,nL) pair, we can reuse it for (nS/2,nL/2) by 538 VIEW_CONVERTing it. This holds for all current targets that support 539 fully-masked loops. For example, suppose the scalar loop is: 540 541 float *f; 542 double *d; 543 for (int i = 0; i < n; ++i) 544 { 545 f[i * 2 + 0] += 1.0f; 546 f[i * 2 + 1] += 2.0f; 547 d[i] += 3.0; 548 } 549 550 and suppose that vectors have 256 bits. The vectorized f accesses 551 will belong to one rgroup and the vectorized d access to another: 552 553 f rgroup: nS = 2, nV = 1, nL = 8 554 d rgroup: nS = 1, nV = 1, nL = 4 555 VF = 4 556 557 [ In this simple example the rgroups do correspond to the normal 558 SLP grouping scheme. ] 559 560 If only the first three lanes are active, the masks we need are: 561 562 f rgroup: 1 1 | 1 1 | 1 1 | 0 0 563 d rgroup: 1 | 1 | 1 | 0 564 565 Here we can use a mask calculated for f's rgroup for d's, but not 566 vice versa. 567 568 Thus for each value of nV, it is enough to provide nV masks, with the 569 mask being calculated based on the highest nL (or, equivalently, based 570 on the highest nS) required by any rgroup with that nV. We therefore 571 represent the entire collection of masks as a two-level table, with the 572 first level being indexed by nV - 1 (since nV == 0 doesn't exist) and 573 the second being indexed by the mask index 0 <= i < nV. */ 574 575 /* The controls (like masks or lengths) needed by rgroups with nV vectors, 576 according to the description above. */ 577 struct rgroup_controls { 578 /* The largest nS for all rgroups that use these controls. */ 579 unsigned int max_nscalars_per_iter; 580 581 /* For the largest nS recorded above, the loop controls divide each scalar 582 into FACTOR equal-sized pieces. This is useful if we need to split 583 element-based accesses into byte-based accesses. */ 584 unsigned int factor; 585 586 /* This is a vector type with MAX_NSCALARS_PER_ITER * VF / nV elements. 587 For mask-based controls, it is the type of the masks in CONTROLS. 588 For length-based controls, it can be any vector type that has the 589 specified number of elements; the type of the elements doesn't matter. */ 590 tree type; 591 592 /* A vector of nV controls, in iteration order. */ 593 vec<tree> controls; 594 595 /* In case of len_load and len_store with a bias there is only one 596 rgroup. This holds the adjusted loop length for the this rgroup. */ 597 tree bias_adjusted_ctrl; 598 }; 599 600 typedef auto_vec<rgroup_controls> vec_loop_masks; 601 602 typedef auto_vec<rgroup_controls> vec_loop_lens; 603 604 typedef auto_vec<std::pair<data_reference*, tree> > drs_init_vec; 605 606 /* Information about a reduction accumulator from the main loop that could 607 conceivably be reused as the input to a reduction in an epilogue loop. */ 608 struct vect_reusable_accumulator { 609 /* The final value of the accumulator, which forms the input to the 610 reduction operation. */ 611 tree reduc_input; 612 613 /* The stmt_vec_info that describes the reduction (i.e. the one for 614 which is_reduc_info is true). */ 615 stmt_vec_info reduc_info; 616 }; 617 618 /*-----------------------------------------------------------------*/ 619 /* Info on vectorized loops. */ 620 /*-----------------------------------------------------------------*/ 621 typedef class _loop_vec_info : public vec_info { 622 public: 623 _loop_vec_info (class loop *, vec_info_shared *); 624 ~_loop_vec_info (); 625 626 /* The loop to which this info struct refers to. */ 627 class loop *loop; 628 629 /* The loop basic blocks. */ 630 basic_block *bbs; 631 632 /* Number of latch executions. */ 633 tree num_itersm1; 634 /* Number of iterations. */ 635 tree num_iters; 636 /* Number of iterations of the original loop. */ 637 tree num_iters_unchanged; 638 /* Condition under which this loop is analyzed and versioned. */ 639 tree num_iters_assumptions; 640 641 /* The cost of the vector code. */ 642 class vector_costs *vector_costs; 643 644 /* The cost of the scalar code. */ 645 class vector_costs *scalar_costs; 646 647 /* Threshold of number of iterations below which vectorization will not be 648 performed. It is calculated from MIN_PROFITABLE_ITERS and 649 param_min_vect_loop_bound. */ 650 unsigned int th; 651 652 /* When applying loop versioning, the vector form should only be used 653 if the number of scalar iterations is >= this value, on top of all 654 the other requirements. Ignored when loop versioning is not being 655 used. */ 656 poly_uint64 versioning_threshold; 657 658 /* Unrolling factor */ 659 poly_uint64 vectorization_factor; 660 661 /* If this loop is an epilogue loop whose main loop can be skipped, 662 MAIN_LOOP_EDGE is the edge from the main loop to this loop's 663 preheader. SKIP_MAIN_LOOP_EDGE is then the edge that skips the 664 main loop and goes straight to this loop's preheader. 665 666 Both fields are null otherwise. */ 667 edge main_loop_edge; 668 edge skip_main_loop_edge; 669 670 /* If this loop is an epilogue loop that might be skipped after executing 671 the main loop, this edge is the one that skips the epilogue. */ 672 edge skip_this_loop_edge; 673 674 /* The vectorized form of a standard reduction replaces the original 675 scalar code's final result (a loop-closed SSA PHI) with the result 676 of a vector-to-scalar reduction operation. After vectorization, 677 this variable maps these vector-to-scalar results to information 678 about the reductions that generated them. */ 679 hash_map<tree, vect_reusable_accumulator> reusable_accumulators; 680 681 /* The number of times that the target suggested we unroll the vector loop 682 in order to promote more ILP. This value will be used to re-analyze the 683 loop for vectorization and if successful the value will be folded into 684 vectorization_factor (and therefore exactly divides 685 vectorization_factor). */ 686 unsigned int suggested_unroll_factor; 687 688 /* Maximum runtime vectorization factor, or MAX_VECTORIZATION_FACTOR 689 if there is no particular limit. */ 690 unsigned HOST_WIDE_INT max_vectorization_factor; 691 692 /* The masks that a fully-masked loop should use to avoid operating 693 on inactive scalars. */ 694 vec_loop_masks masks; 695 696 /* The lengths that a loop with length should use to avoid operating 697 on inactive scalars. */ 698 vec_loop_lens lens; 699 700 /* Set of scalar conditions that have loop mask applied. */ 701 scalar_cond_masked_set_type scalar_cond_masked_set; 702 703 /* Set of vector conditions that have loop mask applied. */ 704 vec_cond_masked_set_type vec_cond_masked_set; 705 706 /* If we are using a loop mask to align memory addresses, this variable 707 contains the number of vector elements that we should skip in the 708 first iteration of the vector loop (i.e. the number of leading 709 elements that should be false in the first mask). */ 710 tree mask_skip_niters; 711 712 /* The type that the loop control IV should be converted to before 713 testing which of the VF scalars are active and inactive. 714 Only meaningful if LOOP_VINFO_USING_PARTIAL_VECTORS_P. */ 715 tree rgroup_compare_type; 716 717 /* For #pragma omp simd if (x) loops the x expression. If constant 0, 718 the loop should not be vectorized, if constant non-zero, simd_if_cond 719 shouldn't be set and loop vectorized normally, if SSA_NAME, the loop 720 should be versioned on that condition, using scalar loop if the condition 721 is false and vectorized loop otherwise. */ 722 tree simd_if_cond; 723 724 /* The type that the vector loop control IV should have when 725 LOOP_VINFO_USING_PARTIAL_VECTORS_P is true. */ 726 tree rgroup_iv_type; 727 728 /* Unknown DRs according to which loop was peeled. */ 729 class dr_vec_info *unaligned_dr; 730 731 /* peeling_for_alignment indicates whether peeling for alignment will take 732 place, and what the peeling factor should be: 733 peeling_for_alignment = X means: 734 If X=0: Peeling for alignment will not be applied. 735 If X>0: Peel first X iterations. 736 If X=-1: Generate a runtime test to calculate the number of iterations 737 to be peeled, using the dataref recorded in the field 738 unaligned_dr. */ 739 int peeling_for_alignment; 740 741 /* The mask used to check the alignment of pointers or arrays. */ 742 int ptr_mask; 743 744 /* Data Dependence Relations defining address ranges that are candidates 745 for a run-time aliasing check. */ 746 auto_vec<ddr_p> may_alias_ddrs; 747 748 /* Data Dependence Relations defining address ranges together with segment 749 lengths from which the run-time aliasing check is built. */ 750 auto_vec<dr_with_seg_len_pair_t> comp_alias_ddrs; 751 752 /* Check that the addresses of each pair of objects is unequal. */ 753 auto_vec<vec_object_pair> check_unequal_addrs; 754 755 /* List of values that are required to be nonzero. This is used to check 756 whether things like "x[i * n] += 1;" are safe and eventually gets added 757 to the checks for lower bounds below. */ 758 auto_vec<tree> check_nonzero; 759 760 /* List of values that need to be checked for a minimum value. */ 761 auto_vec<vec_lower_bound> lower_bounds; 762 763 /* Statements in the loop that have data references that are candidates for a 764 runtime (loop versioning) misalignment check. */ 765 auto_vec<stmt_vec_info> may_misalign_stmts; 766 767 /* Reduction cycles detected in the loop. Used in loop-aware SLP. */ 768 auto_vec<stmt_vec_info> reductions; 769 770 /* All reduction chains in the loop, represented by the first 771 stmt in the chain. */ 772 auto_vec<stmt_vec_info> reduction_chains; 773 774 /* Cost vector for a single scalar iteration. */ 775 auto_vec<stmt_info_for_cost> scalar_cost_vec; 776 777 /* Map of IV base/step expressions to inserted name in the preheader. */ 778 hash_map<tree_operand_hash, tree> *ivexpr_map; 779 780 /* Map of OpenMP "omp simd array" scan variables to corresponding 781 rhs of the store of the initializer. */ 782 hash_map<tree, tree> *scan_map; 783 784 /* The unrolling factor needed to SLP the loop. In case of that pure SLP is 785 applied to the loop, i.e., no unrolling is needed, this is 1. */ 786 poly_uint64 slp_unrolling_factor; 787 788 /* The factor used to over weight those statements in an inner loop 789 relative to the loop being vectorized. */ 790 unsigned int inner_loop_cost_factor; 791 792 /* Is the loop vectorizable? */ 793 bool vectorizable; 794 795 /* Records whether we still have the option of vectorizing this loop 796 using partially-populated vectors; in other words, whether it is 797 still possible for one iteration of the vector loop to handle 798 fewer than VF scalars. */ 799 bool can_use_partial_vectors_p; 800 801 /* True if we've decided to use partially-populated vectors, so that 802 the vector loop can handle fewer than VF scalars. */ 803 bool using_partial_vectors_p; 804 805 /* True if we've decided to use partially-populated vectors for the 806 epilogue of loop. */ 807 bool epil_using_partial_vectors_p; 808 809 /* The bias for len_load and len_store. For now, only 0 and -1 are 810 supported. -1 must be used when a backend does not support 811 len_load/len_store with a length of zero. */ 812 signed char partial_load_store_bias; 813 814 /* When we have grouped data accesses with gaps, we may introduce invalid 815 memory accesses. We peel the last iteration of the loop to prevent 816 this. */ 817 bool peeling_for_gaps; 818 819 /* When the number of iterations is not a multiple of the vector size 820 we need to peel off iterations at the end to form an epilogue loop. */ 821 bool peeling_for_niter; 822 823 /* True if there are no loop carried data dependencies in the loop. 824 If loop->safelen <= 1, then this is always true, either the loop 825 didn't have any loop carried data dependencies, or the loop is being 826 vectorized guarded with some runtime alias checks, or couldn't 827 be vectorized at all, but then this field shouldn't be used. 828 For loop->safelen >= 2, the user has asserted that there are no 829 backward dependencies, but there still could be loop carried forward 830 dependencies in such loops. This flag will be false if normal 831 vectorizer data dependency analysis would fail or require versioning 832 for alias, but because of loop->safelen >= 2 it has been vectorized 833 even without versioning for alias. E.g. in: 834 #pragma omp simd 835 for (int i = 0; i < m; i++) 836 a[i] = a[i + k] * c; 837 (or #pragma simd or #pragma ivdep) we can vectorize this and it will 838 DTRT even for k > 0 && k < m, but without safelen we would not 839 vectorize this, so this field would be false. */ 840 bool no_data_dependencies; 841 842 /* Mark loops having masked stores. */ 843 bool has_mask_store; 844 845 /* Queued scaling factor for the scalar loop. */ 846 profile_probability scalar_loop_scaling; 847 848 /* If if-conversion versioned this loop before conversion, this is the 849 loop version without if-conversion. */ 850 class loop *scalar_loop; 851 852 /* For loops being epilogues of already vectorized loops 853 this points to the original vectorized loop. Otherwise NULL. */ 854 _loop_vec_info *orig_loop_info; 855 856 /* Used to store loop_vec_infos of epilogues of this loop during 857 analysis. */ 858 vec<_loop_vec_info *> epilogue_vinfos; 859 860 } *loop_vec_info; 861 862 /* Access Functions. */ 863 #define LOOP_VINFO_LOOP(L) (L)->loop 864 #define LOOP_VINFO_BBS(L) (L)->bbs 865 #define LOOP_VINFO_NITERSM1(L) (L)->num_itersm1 866 #define LOOP_VINFO_NITERS(L) (L)->num_iters 867 /* Since LOOP_VINFO_NITERS and LOOP_VINFO_NITERSM1 can change after 868 prologue peeling retain total unchanged scalar loop iterations for 869 cost model. */ 870 #define LOOP_VINFO_NITERS_UNCHANGED(L) (L)->num_iters_unchanged 871 #define LOOP_VINFO_NITERS_ASSUMPTIONS(L) (L)->num_iters_assumptions 872 #define LOOP_VINFO_COST_MODEL_THRESHOLD(L) (L)->th 873 #define LOOP_VINFO_VERSIONING_THRESHOLD(L) (L)->versioning_threshold 874 #define LOOP_VINFO_VECTORIZABLE_P(L) (L)->vectorizable 875 #define LOOP_VINFO_CAN_USE_PARTIAL_VECTORS_P(L) (L)->can_use_partial_vectors_p 876 #define LOOP_VINFO_USING_PARTIAL_VECTORS_P(L) (L)->using_partial_vectors_p 877 #define LOOP_VINFO_EPIL_USING_PARTIAL_VECTORS_P(L) \ 878 (L)->epil_using_partial_vectors_p 879 #define LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS(L) (L)->partial_load_store_bias 880 #define LOOP_VINFO_VECT_FACTOR(L) (L)->vectorization_factor 881 #define LOOP_VINFO_MAX_VECT_FACTOR(L) (L)->max_vectorization_factor 882 #define LOOP_VINFO_MASKS(L) (L)->masks 883 #define LOOP_VINFO_LENS(L) (L)->lens 884 #define LOOP_VINFO_MASK_SKIP_NITERS(L) (L)->mask_skip_niters 885 #define LOOP_VINFO_RGROUP_COMPARE_TYPE(L) (L)->rgroup_compare_type 886 #define LOOP_VINFO_RGROUP_IV_TYPE(L) (L)->rgroup_iv_type 887 #define LOOP_VINFO_PTR_MASK(L) (L)->ptr_mask 888 #define LOOP_VINFO_N_STMTS(L) (L)->shared->n_stmts 889 #define LOOP_VINFO_LOOP_NEST(L) (L)->shared->loop_nest 890 #define LOOP_VINFO_DATAREFS(L) (L)->shared->datarefs 891 #define LOOP_VINFO_DDRS(L) (L)->shared->ddrs 892 #define LOOP_VINFO_INT_NITERS(L) (TREE_INT_CST_LOW ((L)->num_iters)) 893 #define LOOP_VINFO_PEELING_FOR_ALIGNMENT(L) (L)->peeling_for_alignment 894 #define LOOP_VINFO_UNALIGNED_DR(L) (L)->unaligned_dr 895 #define LOOP_VINFO_MAY_MISALIGN_STMTS(L) (L)->may_misalign_stmts 896 #define LOOP_VINFO_MAY_ALIAS_DDRS(L) (L)->may_alias_ddrs 897 #define LOOP_VINFO_COMP_ALIAS_DDRS(L) (L)->comp_alias_ddrs 898 #define LOOP_VINFO_CHECK_UNEQUAL_ADDRS(L) (L)->check_unequal_addrs 899 #define LOOP_VINFO_CHECK_NONZERO(L) (L)->check_nonzero 900 #define LOOP_VINFO_LOWER_BOUNDS(L) (L)->lower_bounds 901 #define LOOP_VINFO_GROUPED_STORES(L) (L)->grouped_stores 902 #define LOOP_VINFO_SLP_INSTANCES(L) (L)->slp_instances 903 #define LOOP_VINFO_SLP_UNROLLING_FACTOR(L) (L)->slp_unrolling_factor 904 #define LOOP_VINFO_REDUCTIONS(L) (L)->reductions 905 #define LOOP_VINFO_REDUCTION_CHAINS(L) (L)->reduction_chains 906 #define LOOP_VINFO_PEELING_FOR_GAPS(L) (L)->peeling_for_gaps 907 #define LOOP_VINFO_PEELING_FOR_NITER(L) (L)->peeling_for_niter 908 #define LOOP_VINFO_NO_DATA_DEPENDENCIES(L) (L)->no_data_dependencies 909 #define LOOP_VINFO_SCALAR_LOOP(L) (L)->scalar_loop 910 #define LOOP_VINFO_SCALAR_LOOP_SCALING(L) (L)->scalar_loop_scaling 911 #define LOOP_VINFO_HAS_MASK_STORE(L) (L)->has_mask_store 912 #define LOOP_VINFO_SCALAR_ITERATION_COST(L) (L)->scalar_cost_vec 913 #define LOOP_VINFO_ORIG_LOOP_INFO(L) (L)->orig_loop_info 914 #define LOOP_VINFO_SIMD_IF_COND(L) (L)->simd_if_cond 915 #define LOOP_VINFO_INNER_LOOP_COST_FACTOR(L) (L)->inner_loop_cost_factor 916 917 #define LOOP_VINFO_FULLY_MASKED_P(L) \ 918 (LOOP_VINFO_USING_PARTIAL_VECTORS_P (L) \ 919 && !LOOP_VINFO_MASKS (L).is_empty ()) 920 921 #define LOOP_VINFO_FULLY_WITH_LENGTH_P(L) \ 922 (LOOP_VINFO_USING_PARTIAL_VECTORS_P (L) \ 923 && !LOOP_VINFO_LENS (L).is_empty ()) 924 925 #define LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT(L) \ 926 ((L)->may_misalign_stmts.length () > 0) 927 #define LOOP_REQUIRES_VERSIONING_FOR_ALIAS(L) \ 928 ((L)->comp_alias_ddrs.length () > 0 \ 929 || (L)->check_unequal_addrs.length () > 0 \ 930 || (L)->lower_bounds.length () > 0) 931 #define LOOP_REQUIRES_VERSIONING_FOR_NITERS(L) \ 932 (LOOP_VINFO_NITERS_ASSUMPTIONS (L)) 933 #define LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND(L) \ 934 (LOOP_VINFO_SIMD_IF_COND (L)) 935 #define LOOP_REQUIRES_VERSIONING(L) \ 936 (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (L) \ 937 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (L) \ 938 || LOOP_REQUIRES_VERSIONING_FOR_NITERS (L) \ 939 || LOOP_REQUIRES_VERSIONING_FOR_SIMD_IF_COND (L)) 940 941 #define LOOP_VINFO_NITERS_KNOWN_P(L) \ 942 (tree_fits_shwi_p ((L)->num_iters) && tree_to_shwi ((L)->num_iters) > 0) 943 944 #define LOOP_VINFO_EPILOGUE_P(L) \ 945 (LOOP_VINFO_ORIG_LOOP_INFO (L) != NULL) 946 947 #define LOOP_VINFO_ORIG_MAX_VECT_FACTOR(L) \ 948 (LOOP_VINFO_MAX_VECT_FACTOR (LOOP_VINFO_ORIG_LOOP_INFO (L))) 949 950 /* Wrapper for loop_vec_info, for tracking success/failure, where a non-NULL 951 value signifies success, and a NULL value signifies failure, supporting 952 propagating an opt_problem * describing the failure back up the call 953 stack. */ 954 typedef opt_pointer_wrapper <loop_vec_info> opt_loop_vec_info; 955 956 static inline loop_vec_info 957 loop_vec_info_for_loop (class loop *loop) 958 { 959 return (loop_vec_info) loop->aux; 960 } 961 962 struct slp_root 963 { 964 slp_root (slp_instance_kind kind_, vec<stmt_vec_info> stmts_, 965 vec<stmt_vec_info> roots_) 966 : kind(kind_), stmts(stmts_), roots(roots_) {} 967 slp_instance_kind kind; 968 vec<stmt_vec_info> stmts; 969 vec<stmt_vec_info> roots; 970 }; 971 972 typedef class _bb_vec_info : public vec_info 973 { 974 public: 975 _bb_vec_info (vec<basic_block> bbs, vec_info_shared *); 976 ~_bb_vec_info (); 977 978 /* The region we are operating on. bbs[0] is the entry, excluding 979 its PHI nodes. In the future we might want to track an explicit 980 entry edge to cover bbs[0] PHI nodes and have a region entry 981 insert location. */ 982 vec<basic_block> bbs; 983 984 vec<slp_root> roots; 985 } *bb_vec_info; 986 987 #define BB_VINFO_BB(B) (B)->bb 988 #define BB_VINFO_GROUPED_STORES(B) (B)->grouped_stores 989 #define BB_VINFO_SLP_INSTANCES(B) (B)->slp_instances 990 #define BB_VINFO_DATAREFS(B) (B)->shared->datarefs 991 #define BB_VINFO_DDRS(B) (B)->shared->ddrs 992 993 /*-----------------------------------------------------------------*/ 994 /* Info on vectorized defs. */ 995 /*-----------------------------------------------------------------*/ 996 enum stmt_vec_info_type { 997 undef_vec_info_type = 0, 998 load_vec_info_type, 999 store_vec_info_type, 1000 shift_vec_info_type, 1001 op_vec_info_type, 1002 call_vec_info_type, 1003 call_simd_clone_vec_info_type, 1004 assignment_vec_info_type, 1005 condition_vec_info_type, 1006 comparison_vec_info_type, 1007 reduc_vec_info_type, 1008 induc_vec_info_type, 1009 type_promotion_vec_info_type, 1010 type_demotion_vec_info_type, 1011 type_conversion_vec_info_type, 1012 cycle_phi_info_type, 1013 lc_phi_info_type, 1014 phi_info_type, 1015 loop_exit_ctrl_vec_info_type 1016 }; 1017 1018 /* Indicates whether/how a variable is used in the scope of loop/basic 1019 block. */ 1020 enum vect_relevant { 1021 vect_unused_in_scope = 0, 1022 1023 /* The def is only used outside the loop. */ 1024 vect_used_only_live, 1025 /* The def is in the inner loop, and the use is in the outer loop, and the 1026 use is a reduction stmt. */ 1027 vect_used_in_outer_by_reduction, 1028 /* The def is in the inner loop, and the use is in the outer loop (and is 1029 not part of reduction). */ 1030 vect_used_in_outer, 1031 1032 /* defs that feed computations that end up (only) in a reduction. These 1033 defs may be used by non-reduction stmts, but eventually, any 1034 computations/values that are affected by these defs are used to compute 1035 a reduction (i.e. don't get stored to memory, for example). We use this 1036 to identify computations that we can change the order in which they are 1037 computed. */ 1038 vect_used_by_reduction, 1039 1040 vect_used_in_scope 1041 }; 1042 1043 /* The type of vectorization that can be applied to the stmt: regular loop-based 1044 vectorization; pure SLP - the stmt is a part of SLP instances and does not 1045 have uses outside SLP instances; or hybrid SLP and loop-based - the stmt is 1046 a part of SLP instance and also must be loop-based vectorized, since it has 1047 uses outside SLP sequences. 1048 1049 In the loop context the meanings of pure and hybrid SLP are slightly 1050 different. By saying that pure SLP is applied to the loop, we mean that we 1051 exploit only intra-iteration parallelism in the loop; i.e., the loop can be 1052 vectorized without doing any conceptual unrolling, cause we don't pack 1053 together stmts from different iterations, only within a single iteration. 1054 Loop hybrid SLP means that we exploit both intra-iteration and 1055 inter-iteration parallelism (e.g., number of elements in the vector is 4 1056 and the slp-group-size is 2, in which case we don't have enough parallelism 1057 within an iteration, so we obtain the rest of the parallelism from subsequent 1058 iterations by unrolling the loop by 2). */ 1059 enum slp_vect_type { 1060 loop_vect = 0, 1061 pure_slp, 1062 hybrid 1063 }; 1064 1065 /* Says whether a statement is a load, a store of a vectorized statement 1066 result, or a store of an invariant value. */ 1067 enum vec_load_store_type { 1068 VLS_LOAD, 1069 VLS_STORE, 1070 VLS_STORE_INVARIANT 1071 }; 1072 1073 /* Describes how we're going to vectorize an individual load or store, 1074 or a group of loads or stores. */ 1075 enum vect_memory_access_type { 1076 /* An access to an invariant address. This is used only for loads. */ 1077 VMAT_INVARIANT, 1078 1079 /* A simple contiguous access. */ 1080 VMAT_CONTIGUOUS, 1081 1082 /* A contiguous access that goes down in memory rather than up, 1083 with no additional permutation. This is used only for stores 1084 of invariants. */ 1085 VMAT_CONTIGUOUS_DOWN, 1086 1087 /* A simple contiguous access in which the elements need to be permuted 1088 after loading or before storing. Only used for loop vectorization; 1089 SLP uses separate permutes. */ 1090 VMAT_CONTIGUOUS_PERMUTE, 1091 1092 /* A simple contiguous access in which the elements need to be reversed 1093 after loading or before storing. */ 1094 VMAT_CONTIGUOUS_REVERSE, 1095 1096 /* An access that uses IFN_LOAD_LANES or IFN_STORE_LANES. */ 1097 VMAT_LOAD_STORE_LANES, 1098 1099 /* An access in which each scalar element is loaded or stored 1100 individually. */ 1101 VMAT_ELEMENTWISE, 1102 1103 /* A hybrid of VMAT_CONTIGUOUS and VMAT_ELEMENTWISE, used for grouped 1104 SLP accesses. Each unrolled iteration uses a contiguous load 1105 or store for the whole group, but the groups from separate iterations 1106 are combined in the same way as for VMAT_ELEMENTWISE. */ 1107 VMAT_STRIDED_SLP, 1108 1109 /* The access uses gather loads or scatter stores. */ 1110 VMAT_GATHER_SCATTER 1111 }; 1112 1113 class dr_vec_info { 1114 public: 1115 /* The data reference itself. */ 1116 data_reference *dr; 1117 /* The statement that contains the data reference. */ 1118 stmt_vec_info stmt; 1119 /* The analysis group this DR belongs to when doing BB vectorization. 1120 DRs of the same group belong to the same conditional execution context. */ 1121 unsigned group; 1122 /* The misalignment in bytes of the reference, or -1 if not known. */ 1123 int misalignment; 1124 /* The byte alignment that we'd ideally like the reference to have, 1125 and the value that misalignment is measured against. */ 1126 poly_uint64 target_alignment; 1127 /* If true the alignment of base_decl needs to be increased. */ 1128 bool base_misaligned; 1129 tree base_decl; 1130 1131 /* Stores current vectorized loop's offset. To be added to the DR's 1132 offset to calculate current offset of data reference. */ 1133 tree offset; 1134 }; 1135 1136 typedef struct data_reference *dr_p; 1137 1138 class _stmt_vec_info { 1139 public: 1140 1141 enum stmt_vec_info_type type; 1142 1143 /* Indicates whether this stmts is part of a computation whose result is 1144 used outside the loop. */ 1145 bool live; 1146 1147 /* Stmt is part of some pattern (computation idiom) */ 1148 bool in_pattern_p; 1149 1150 /* True if the statement was created during pattern recognition as 1151 part of the replacement for RELATED_STMT. This implies that the 1152 statement isn't part of any basic block, although for convenience 1153 its gimple_bb is the same as for RELATED_STMT. */ 1154 bool pattern_stmt_p; 1155 1156 /* Is this statement vectorizable or should it be skipped in (partial) 1157 vectorization. */ 1158 bool vectorizable; 1159 1160 /* The stmt to which this info struct refers to. */ 1161 gimple *stmt; 1162 1163 /* The vector type to be used for the LHS of this statement. */ 1164 tree vectype; 1165 1166 /* The vectorized stmts. */ 1167 vec<gimple *> vec_stmts; 1168 1169 /* The following is relevant only for stmts that contain a non-scalar 1170 data-ref (array/pointer/struct access). A GIMPLE stmt is expected to have 1171 at most one such data-ref. */ 1172 1173 dr_vec_info dr_aux; 1174 1175 /* Information about the data-ref relative to this loop 1176 nest (the loop that is being considered for vectorization). */ 1177 innermost_loop_behavior dr_wrt_vec_loop; 1178 1179 /* For loop PHI nodes, the base and evolution part of it. This makes sure 1180 this information is still available in vect_update_ivs_after_vectorizer 1181 where we may not be able to re-analyze the PHI nodes evolution as 1182 peeling for the prologue loop can make it unanalyzable. The evolution 1183 part is still correct after peeling, but the base may have changed from 1184 the version here. */ 1185 tree loop_phi_evolution_base_unchanged; 1186 tree loop_phi_evolution_part; 1187 1188 /* Used for various bookkeeping purposes, generally holding a pointer to 1189 some other stmt S that is in some way "related" to this stmt. 1190 Current use of this field is: 1191 If this stmt is part of a pattern (i.e. the field 'in_pattern_p' is 1192 true): S is the "pattern stmt" that represents (and replaces) the 1193 sequence of stmts that constitutes the pattern. Similarly, the 1194 related_stmt of the "pattern stmt" points back to this stmt (which is 1195 the last stmt in the original sequence of stmts that constitutes the 1196 pattern). */ 1197 stmt_vec_info related_stmt; 1198 1199 /* Used to keep a sequence of def stmts of a pattern stmt if such exists. 1200 The sequence is attached to the original statement rather than the 1201 pattern statement. */ 1202 gimple_seq pattern_def_seq; 1203 1204 /* Selected SIMD clone's function info. First vector element 1205 is SIMD clone's function decl, followed by a pair of trees (base + step) 1206 for linear arguments (pair of NULLs for other arguments). */ 1207 vec<tree> simd_clone_info; 1208 1209 /* Classify the def of this stmt. */ 1210 enum vect_def_type def_type; 1211 1212 /* Whether the stmt is SLPed, loop-based vectorized, or both. */ 1213 enum slp_vect_type slp_type; 1214 1215 /* Interleaving and reduction chains info. */ 1216 /* First element in the group. */ 1217 stmt_vec_info first_element; 1218 /* Pointer to the next element in the group. */ 1219 stmt_vec_info next_element; 1220 /* The size of the group. */ 1221 unsigned int size; 1222 /* For stores, number of stores from this group seen. We vectorize the last 1223 one. */ 1224 unsigned int store_count; 1225 /* For loads only, the gap from the previous load. For consecutive loads, GAP 1226 is 1. */ 1227 unsigned int gap; 1228 1229 /* The minimum negative dependence distance this stmt participates in 1230 or zero if none. */ 1231 unsigned int min_neg_dist; 1232 1233 /* Not all stmts in the loop need to be vectorized. e.g, the increment 1234 of the loop induction variable and computation of array indexes. relevant 1235 indicates whether the stmt needs to be vectorized. */ 1236 enum vect_relevant relevant; 1237 1238 /* For loads if this is a gather, for stores if this is a scatter. */ 1239 bool gather_scatter_p; 1240 1241 /* True if this is an access with loop-invariant stride. */ 1242 bool strided_p; 1243 1244 /* For both loads and stores. */ 1245 unsigned simd_lane_access_p : 3; 1246 1247 /* Classifies how the load or store is going to be implemented 1248 for loop vectorization. */ 1249 vect_memory_access_type memory_access_type; 1250 1251 /* For INTEGER_INDUC_COND_REDUCTION, the initial value to be used. */ 1252 tree induc_cond_initial_val; 1253 1254 /* If not NULL the value to be added to compute final reduction value. */ 1255 tree reduc_epilogue_adjustment; 1256 1257 /* On a reduction PHI the reduction type as detected by 1258 vect_is_simple_reduction and vectorizable_reduction. */ 1259 enum vect_reduction_type reduc_type; 1260 1261 /* The original reduction code, to be used in the epilogue. */ 1262 code_helper reduc_code; 1263 /* An internal function we should use in the epilogue. */ 1264 internal_fn reduc_fn; 1265 1266 /* On a stmt participating in the reduction the index of the operand 1267 on the reduction SSA cycle. */ 1268 int reduc_idx; 1269 1270 /* On a reduction PHI the def returned by vect_force_simple_reduction. 1271 On the def returned by vect_force_simple_reduction the 1272 corresponding PHI. */ 1273 stmt_vec_info reduc_def; 1274 1275 /* The vector input type relevant for reduction vectorization. */ 1276 tree reduc_vectype_in; 1277 1278 /* The vector type for performing the actual reduction. */ 1279 tree reduc_vectype; 1280 1281 /* If IS_REDUC_INFO is true and if the vector code is performing 1282 N scalar reductions in parallel, this variable gives the initial 1283 scalar values of those N reductions. */ 1284 vec<tree> reduc_initial_values; 1285 1286 /* If IS_REDUC_INFO is true and if the vector code is performing 1287 N scalar reductions in parallel, this variable gives the vectorized code's 1288 final (scalar) result for each of those N reductions. In other words, 1289 REDUC_SCALAR_RESULTS[I] replaces the original scalar code's loop-closed 1290 SSA PHI for reduction number I. */ 1291 vec<tree> reduc_scalar_results; 1292 1293 /* Only meaningful if IS_REDUC_INFO. If non-null, the reduction is 1294 being performed by an epilogue loop and we have decided to reuse 1295 this accumulator from the main loop. */ 1296 vect_reusable_accumulator *reused_accumulator; 1297 1298 /* Whether we force a single cycle PHI during reduction vectorization. */ 1299 bool force_single_cycle; 1300 1301 /* Whether on this stmt reduction meta is recorded. */ 1302 bool is_reduc_info; 1303 1304 /* If nonzero, the lhs of the statement could be truncated to this 1305 many bits without affecting any users of the result. */ 1306 unsigned int min_output_precision; 1307 1308 /* If nonzero, all non-boolean input operands have the same precision, 1309 and they could each be truncated to this many bits without changing 1310 the result. */ 1311 unsigned int min_input_precision; 1312 1313 /* If OPERATION_BITS is nonzero, the statement could be performed on 1314 an integer with the sign and number of bits given by OPERATION_SIGN 1315 and OPERATION_BITS without changing the result. */ 1316 unsigned int operation_precision; 1317 signop operation_sign; 1318 1319 /* If the statement produces a boolean result, this value describes 1320 how we should choose the associated vector type. The possible 1321 values are: 1322 1323 - an integer precision N if we should use the vector mask type 1324 associated with N-bit integers. This is only used if all relevant 1325 input booleans also want the vector mask type for N-bit integers, 1326 or if we can convert them into that form by pattern-matching. 1327 1328 - ~0U if we considered choosing a vector mask type but decided 1329 to treat the boolean as a normal integer type instead. 1330 1331 - 0 otherwise. This means either that the operation isn't one that 1332 could have a vector mask type (and so should have a normal vector 1333 type instead) or that we simply haven't made a choice either way. */ 1334 unsigned int mask_precision; 1335 1336 /* True if this is only suitable for SLP vectorization. */ 1337 bool slp_vect_only_p; 1338 1339 /* True if this is a pattern that can only be handled by SLP 1340 vectorization. */ 1341 bool slp_vect_pattern_only_p; 1342 }; 1343 1344 /* Information about a gather/scatter call. */ 1345 struct gather_scatter_info { 1346 /* The internal function to use for the gather/scatter operation, 1347 or IFN_LAST if a built-in function should be used instead. */ 1348 internal_fn ifn; 1349 1350 /* The FUNCTION_DECL for the built-in gather/scatter function, 1351 or null if an internal function should be used instead. */ 1352 tree decl; 1353 1354 /* The loop-invariant base value. */ 1355 tree base; 1356 1357 /* The original scalar offset, which is a non-loop-invariant SSA_NAME. */ 1358 tree offset; 1359 1360 /* Each offset element should be multiplied by this amount before 1361 being added to the base. */ 1362 int scale; 1363 1364 /* The definition type for the vectorized offset. */ 1365 enum vect_def_type offset_dt; 1366 1367 /* The type of the vectorized offset. */ 1368 tree offset_vectype; 1369 1370 /* The type of the scalar elements after loading or before storing. */ 1371 tree element_type; 1372 1373 /* The type of the scalar elements being loaded or stored. */ 1374 tree memory_type; 1375 }; 1376 1377 /* Access Functions. */ 1378 #define STMT_VINFO_TYPE(S) (S)->type 1379 #define STMT_VINFO_STMT(S) (S)->stmt 1380 #define STMT_VINFO_RELEVANT(S) (S)->relevant 1381 #define STMT_VINFO_LIVE_P(S) (S)->live 1382 #define STMT_VINFO_VECTYPE(S) (S)->vectype 1383 #define STMT_VINFO_VEC_STMTS(S) (S)->vec_stmts 1384 #define STMT_VINFO_VECTORIZABLE(S) (S)->vectorizable 1385 #define STMT_VINFO_DATA_REF(S) ((S)->dr_aux.dr + 0) 1386 #define STMT_VINFO_GATHER_SCATTER_P(S) (S)->gather_scatter_p 1387 #define STMT_VINFO_STRIDED_P(S) (S)->strided_p 1388 #define STMT_VINFO_MEMORY_ACCESS_TYPE(S) (S)->memory_access_type 1389 #define STMT_VINFO_SIMD_LANE_ACCESS_P(S) (S)->simd_lane_access_p 1390 #define STMT_VINFO_VEC_INDUC_COND_INITIAL_VAL(S) (S)->induc_cond_initial_val 1391 #define STMT_VINFO_REDUC_EPILOGUE_ADJUSTMENT(S) (S)->reduc_epilogue_adjustment 1392 #define STMT_VINFO_REDUC_IDX(S) (S)->reduc_idx 1393 #define STMT_VINFO_FORCE_SINGLE_CYCLE(S) (S)->force_single_cycle 1394 1395 #define STMT_VINFO_DR_WRT_VEC_LOOP(S) (S)->dr_wrt_vec_loop 1396 #define STMT_VINFO_DR_BASE_ADDRESS(S) (S)->dr_wrt_vec_loop.base_address 1397 #define STMT_VINFO_DR_INIT(S) (S)->dr_wrt_vec_loop.init 1398 #define STMT_VINFO_DR_OFFSET(S) (S)->dr_wrt_vec_loop.offset 1399 #define STMT_VINFO_DR_STEP(S) (S)->dr_wrt_vec_loop.step 1400 #define STMT_VINFO_DR_BASE_ALIGNMENT(S) (S)->dr_wrt_vec_loop.base_alignment 1401 #define STMT_VINFO_DR_BASE_MISALIGNMENT(S) \ 1402 (S)->dr_wrt_vec_loop.base_misalignment 1403 #define STMT_VINFO_DR_OFFSET_ALIGNMENT(S) \ 1404 (S)->dr_wrt_vec_loop.offset_alignment 1405 #define STMT_VINFO_DR_STEP_ALIGNMENT(S) \ 1406 (S)->dr_wrt_vec_loop.step_alignment 1407 1408 #define STMT_VINFO_DR_INFO(S) \ 1409 (gcc_checking_assert ((S)->dr_aux.stmt == (S)), &(S)->dr_aux) 1410 1411 #define STMT_VINFO_IN_PATTERN_P(S) (S)->in_pattern_p 1412 #define STMT_VINFO_RELATED_STMT(S) (S)->related_stmt 1413 #define STMT_VINFO_PATTERN_DEF_SEQ(S) (S)->pattern_def_seq 1414 #define STMT_VINFO_SIMD_CLONE_INFO(S) (S)->simd_clone_info 1415 #define STMT_VINFO_DEF_TYPE(S) (S)->def_type 1416 #define STMT_VINFO_GROUPED_ACCESS(S) \ 1417 ((S)->dr_aux.dr && DR_GROUP_FIRST_ELEMENT(S)) 1418 #define STMT_VINFO_LOOP_PHI_EVOLUTION_BASE_UNCHANGED(S) (S)->loop_phi_evolution_base_unchanged 1419 #define STMT_VINFO_LOOP_PHI_EVOLUTION_PART(S) (S)->loop_phi_evolution_part 1420 #define STMT_VINFO_MIN_NEG_DIST(S) (S)->min_neg_dist 1421 #define STMT_VINFO_REDUC_TYPE(S) (S)->reduc_type 1422 #define STMT_VINFO_REDUC_CODE(S) (S)->reduc_code 1423 #define STMT_VINFO_REDUC_FN(S) (S)->reduc_fn 1424 #define STMT_VINFO_REDUC_DEF(S) (S)->reduc_def 1425 #define STMT_VINFO_REDUC_VECTYPE(S) (S)->reduc_vectype 1426 #define STMT_VINFO_REDUC_VECTYPE_IN(S) (S)->reduc_vectype_in 1427 #define STMT_VINFO_SLP_VECT_ONLY(S) (S)->slp_vect_only_p 1428 #define STMT_VINFO_SLP_VECT_ONLY_PATTERN(S) (S)->slp_vect_pattern_only_p 1429 1430 #define DR_GROUP_FIRST_ELEMENT(S) \ 1431 (gcc_checking_assert ((S)->dr_aux.dr), (S)->first_element) 1432 #define DR_GROUP_NEXT_ELEMENT(S) \ 1433 (gcc_checking_assert ((S)->dr_aux.dr), (S)->next_element) 1434 #define DR_GROUP_SIZE(S) \ 1435 (gcc_checking_assert ((S)->dr_aux.dr), (S)->size) 1436 #define DR_GROUP_STORE_COUNT(S) \ 1437 (gcc_checking_assert ((S)->dr_aux.dr), (S)->store_count) 1438 #define DR_GROUP_GAP(S) \ 1439 (gcc_checking_assert ((S)->dr_aux.dr), (S)->gap) 1440 1441 #define REDUC_GROUP_FIRST_ELEMENT(S) \ 1442 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->first_element) 1443 #define REDUC_GROUP_NEXT_ELEMENT(S) \ 1444 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->next_element) 1445 #define REDUC_GROUP_SIZE(S) \ 1446 (gcc_checking_assert (!(S)->dr_aux.dr), (S)->size) 1447 1448 #define STMT_VINFO_RELEVANT_P(S) ((S)->relevant != vect_unused_in_scope) 1449 1450 #define HYBRID_SLP_STMT(S) ((S)->slp_type == hybrid) 1451 #define PURE_SLP_STMT(S) ((S)->slp_type == pure_slp) 1452 #define STMT_SLP_TYPE(S) (S)->slp_type 1453 1454 /* Contains the scalar or vector costs for a vec_info. */ 1455 class vector_costs 1456 { 1457 public: 1458 vector_costs (vec_info *, bool); 1459 virtual ~vector_costs () {} 1460 1461 /* Update the costs in response to adding COUNT copies of a statement. 1462 1463 - WHERE specifies whether the cost occurs in the loop prologue, 1464 the loop body, or the loop epilogue. 1465 - KIND is the kind of statement, which is always meaningful. 1466 - STMT_INFO or NODE, if nonnull, describe the statement that will be 1467 vectorized. 1468 - VECTYPE, if nonnull, is the vector type that the vectorized 1469 statement will operate on. Note that this should be used in 1470 preference to STMT_VINFO_VECTYPE (STMT_INFO) since the latter 1471 is not correct for SLP. 1472 - for unaligned_load and unaligned_store statements, MISALIGN is 1473 the byte misalignment of the load or store relative to the target's 1474 preferred alignment for VECTYPE, or DR_MISALIGNMENT_UNKNOWN 1475 if the misalignment is not known. 1476 1477 Return the calculated cost as well as recording it. The return 1478 value is used for dumping purposes. */ 1479 virtual unsigned int add_stmt_cost (int count, vect_cost_for_stmt kind, 1480 stmt_vec_info stmt_info, 1481 slp_tree node, 1482 tree vectype, int misalign, 1483 vect_cost_model_location where); 1484 1485 /* Finish calculating the cost of the code. The results can be 1486 read back using the functions below. 1487 1488 If the costs describe vector code, SCALAR_COSTS gives the costs 1489 of the corresponding scalar code, otherwise it is null. */ 1490 virtual void finish_cost (const vector_costs *scalar_costs); 1491 1492 /* The costs in THIS and OTHER both describe ways of vectorizing 1493 a main loop. Return true if the costs described by THIS are 1494 cheaper than the costs described by OTHER. Return false if any 1495 of the following are true: 1496 1497 - THIS and OTHER are of equal cost 1498 - OTHER is better than THIS 1499 - we can't be sure about the relative costs of THIS and OTHER. */ 1500 virtual bool better_main_loop_than_p (const vector_costs *other) const; 1501 1502 /* Likewise, but the costs in THIS and OTHER both describe ways of 1503 vectorizing an epilogue loop of MAIN_LOOP. */ 1504 virtual bool better_epilogue_loop_than_p (const vector_costs *other, 1505 loop_vec_info main_loop) const; 1506 1507 unsigned int prologue_cost () const; 1508 unsigned int body_cost () const; 1509 unsigned int epilogue_cost () const; 1510 unsigned int outside_cost () const; 1511 unsigned int total_cost () const; 1512 unsigned int suggested_unroll_factor () const; 1513 1514 protected: 1515 unsigned int record_stmt_cost (stmt_vec_info, vect_cost_model_location, 1516 unsigned int); 1517 unsigned int adjust_cost_for_freq (stmt_vec_info, vect_cost_model_location, 1518 unsigned int); 1519 int compare_inside_loop_cost (const vector_costs *) const; 1520 int compare_outside_loop_cost (const vector_costs *) const; 1521 1522 /* The region of code that we're considering vectorizing. */ 1523 vec_info *m_vinfo; 1524 1525 /* True if we're costing the scalar code, false if we're costing 1526 the vector code. */ 1527 bool m_costing_for_scalar; 1528 1529 /* The costs of the three regions, indexed by vect_cost_model_location. */ 1530 unsigned int m_costs[3]; 1531 1532 /* The suggested unrolling factor determined at finish_cost. */ 1533 unsigned int m_suggested_unroll_factor; 1534 1535 /* True if finish_cost has been called. */ 1536 bool m_finished; 1537 }; 1538 1539 /* Create costs for VINFO. COSTING_FOR_SCALAR is true if the costs 1540 are for scalar code, false if they are for vector code. */ 1541 1542 inline 1543 vector_costs::vector_costs (vec_info *vinfo, bool costing_for_scalar) 1544 : m_vinfo (vinfo), 1545 m_costing_for_scalar (costing_for_scalar), 1546 m_costs (), 1547 m_suggested_unroll_factor(1), 1548 m_finished (false) 1549 { 1550 } 1551 1552 /* Return the cost of the prologue code (in abstract units). */ 1553 1554 inline unsigned int 1555 vector_costs::prologue_cost () const 1556 { 1557 gcc_checking_assert (m_finished); 1558 return m_costs[vect_prologue]; 1559 } 1560 1561 /* Return the cost of the body code (in abstract units). */ 1562 1563 inline unsigned int 1564 vector_costs::body_cost () const 1565 { 1566 gcc_checking_assert (m_finished); 1567 return m_costs[vect_body]; 1568 } 1569 1570 /* Return the cost of the epilogue code (in abstract units). */ 1571 1572 inline unsigned int 1573 vector_costs::epilogue_cost () const 1574 { 1575 gcc_checking_assert (m_finished); 1576 return m_costs[vect_epilogue]; 1577 } 1578 1579 /* Return the cost of the prologue and epilogue code (in abstract units). */ 1580 1581 inline unsigned int 1582 vector_costs::outside_cost () const 1583 { 1584 return prologue_cost () + epilogue_cost (); 1585 } 1586 1587 /* Return the cost of the prologue, body and epilogue code 1588 (in abstract units). */ 1589 1590 inline unsigned int 1591 vector_costs::total_cost () const 1592 { 1593 return body_cost () + outside_cost (); 1594 } 1595 1596 /* Return the suggested unroll factor. */ 1597 1598 inline unsigned int 1599 vector_costs::suggested_unroll_factor () const 1600 { 1601 gcc_checking_assert (m_finished); 1602 return m_suggested_unroll_factor; 1603 } 1604 1605 #define VECT_MAX_COST 1000 1606 1607 /* The maximum number of intermediate steps required in multi-step type 1608 conversion. */ 1609 #define MAX_INTERM_CVT_STEPS 3 1610 1611 #define MAX_VECTORIZATION_FACTOR INT_MAX 1612 1613 /* Nonzero if TYPE represents a (scalar) boolean type or type 1614 in the middle-end compatible with it (unsigned precision 1 integral 1615 types). Used to determine which types should be vectorized as 1616 VECTOR_BOOLEAN_TYPE_P. */ 1617 1618 #define VECT_SCALAR_BOOLEAN_TYPE_P(TYPE) \ 1619 (TREE_CODE (TYPE) == BOOLEAN_TYPE \ 1620 || ((TREE_CODE (TYPE) == INTEGER_TYPE \ 1621 || TREE_CODE (TYPE) == ENUMERAL_TYPE) \ 1622 && TYPE_PRECISION (TYPE) == 1 \ 1623 && TYPE_UNSIGNED (TYPE))) 1624 1625 static inline bool 1626 nested_in_vect_loop_p (class loop *loop, stmt_vec_info stmt_info) 1627 { 1628 return (loop->inner 1629 && (loop->inner == (gimple_bb (stmt_info->stmt))->loop_father)); 1630 } 1631 1632 /* PHI is either a scalar reduction phi or a scalar induction phi. 1633 Return the initial value of the variable on entry to the containing 1634 loop. */ 1635 1636 static inline tree 1637 vect_phi_initial_value (gphi *phi) 1638 { 1639 basic_block bb = gimple_bb (phi); 1640 edge pe = loop_preheader_edge (bb->loop_father); 1641 gcc_assert (pe->dest == bb); 1642 return PHI_ARG_DEF_FROM_EDGE (phi, pe); 1643 } 1644 1645 /* Return true if STMT_INFO should produce a vector mask type rather than 1646 a normal nonmask type. */ 1647 1648 static inline bool 1649 vect_use_mask_type_p (stmt_vec_info stmt_info) 1650 { 1651 return stmt_info->mask_precision && stmt_info->mask_precision != ~0U; 1652 } 1653 1654 /* Return TRUE if a statement represented by STMT_INFO is a part of a 1655 pattern. */ 1656 1657 static inline bool 1658 is_pattern_stmt_p (stmt_vec_info stmt_info) 1659 { 1660 return stmt_info->pattern_stmt_p; 1661 } 1662 1663 /* If STMT_INFO is a pattern statement, return the statement that it 1664 replaces, otherwise return STMT_INFO itself. */ 1665 1666 inline stmt_vec_info 1667 vect_orig_stmt (stmt_vec_info stmt_info) 1668 { 1669 if (is_pattern_stmt_p (stmt_info)) 1670 return STMT_VINFO_RELATED_STMT (stmt_info); 1671 return stmt_info; 1672 } 1673 1674 /* Return the later statement between STMT1_INFO and STMT2_INFO. */ 1675 1676 static inline stmt_vec_info 1677 get_later_stmt (stmt_vec_info stmt1_info, stmt_vec_info stmt2_info) 1678 { 1679 if (gimple_uid (vect_orig_stmt (stmt1_info)->stmt) 1680 > gimple_uid (vect_orig_stmt (stmt2_info)->stmt)) 1681 return stmt1_info; 1682 else 1683 return stmt2_info; 1684 } 1685 1686 /* If STMT_INFO has been replaced by a pattern statement, return the 1687 replacement statement, otherwise return STMT_INFO itself. */ 1688 1689 inline stmt_vec_info 1690 vect_stmt_to_vectorize (stmt_vec_info stmt_info) 1691 { 1692 if (STMT_VINFO_IN_PATTERN_P (stmt_info)) 1693 return STMT_VINFO_RELATED_STMT (stmt_info); 1694 return stmt_info; 1695 } 1696 1697 /* Return true if BB is a loop header. */ 1698 1699 static inline bool 1700 is_loop_header_bb_p (basic_block bb) 1701 { 1702 if (bb == (bb->loop_father)->header) 1703 return true; 1704 gcc_checking_assert (EDGE_COUNT (bb->preds) == 1); 1705 return false; 1706 } 1707 1708 /* Return pow2 (X). */ 1709 1710 static inline int 1711 vect_pow2 (int x) 1712 { 1713 int i, res = 1; 1714 1715 for (i = 0; i < x; i++) 1716 res *= 2; 1717 1718 return res; 1719 } 1720 1721 /* Alias targetm.vectorize.builtin_vectorization_cost. */ 1722 1723 static inline int 1724 builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost, 1725 tree vectype, int misalign) 1726 { 1727 return targetm.vectorize.builtin_vectorization_cost (type_of_cost, 1728 vectype, misalign); 1729 } 1730 1731 /* Get cost by calling cost target builtin. */ 1732 1733 static inline 1734 int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost) 1735 { 1736 return builtin_vectorization_cost (type_of_cost, NULL, 0); 1737 } 1738 1739 /* Alias targetm.vectorize.init_cost. */ 1740 1741 static inline vector_costs * 1742 init_cost (vec_info *vinfo, bool costing_for_scalar) 1743 { 1744 return targetm.vectorize.create_costs (vinfo, costing_for_scalar); 1745 } 1746 1747 extern void dump_stmt_cost (FILE *, int, enum vect_cost_for_stmt, 1748 stmt_vec_info, slp_tree, tree, int, unsigned, 1749 enum vect_cost_model_location); 1750 1751 /* Alias targetm.vectorize.add_stmt_cost. */ 1752 1753 static inline unsigned 1754 add_stmt_cost (vector_costs *costs, int count, 1755 enum vect_cost_for_stmt kind, 1756 stmt_vec_info stmt_info, slp_tree node, 1757 tree vectype, int misalign, 1758 enum vect_cost_model_location where) 1759 { 1760 unsigned cost = costs->add_stmt_cost (count, kind, stmt_info, node, vectype, 1761 misalign, where); 1762 if (dump_file && (dump_flags & TDF_DETAILS)) 1763 dump_stmt_cost (dump_file, count, kind, stmt_info, node, vectype, misalign, 1764 cost, where); 1765 return cost; 1766 } 1767 1768 static inline unsigned 1769 add_stmt_cost (vector_costs *costs, int count, enum vect_cost_for_stmt kind, 1770 enum vect_cost_model_location where) 1771 { 1772 gcc_assert (kind == cond_branch_taken || kind == cond_branch_not_taken 1773 || kind == scalar_stmt); 1774 return add_stmt_cost (costs, count, kind, NULL, NULL, NULL_TREE, 0, where); 1775 } 1776 1777 /* Alias targetm.vectorize.add_stmt_cost. */ 1778 1779 static inline unsigned 1780 add_stmt_cost (vector_costs *costs, stmt_info_for_cost *i) 1781 { 1782 return add_stmt_cost (costs, i->count, i->kind, i->stmt_info, i->node, 1783 i->vectype, i->misalign, i->where); 1784 } 1785 1786 /* Alias targetm.vectorize.finish_cost. */ 1787 1788 static inline void 1789 finish_cost (vector_costs *costs, const vector_costs *scalar_costs, 1790 unsigned *prologue_cost, unsigned *body_cost, 1791 unsigned *epilogue_cost, unsigned *suggested_unroll_factor = NULL) 1792 { 1793 costs->finish_cost (scalar_costs); 1794 *prologue_cost = costs->prologue_cost (); 1795 *body_cost = costs->body_cost (); 1796 *epilogue_cost = costs->epilogue_cost (); 1797 if (suggested_unroll_factor) 1798 *suggested_unroll_factor = costs->suggested_unroll_factor (); 1799 } 1800 1801 inline void 1802 add_stmt_costs (vector_costs *costs, stmt_vector_for_cost *cost_vec) 1803 { 1804 stmt_info_for_cost *cost; 1805 unsigned i; 1806 FOR_EACH_VEC_ELT (*cost_vec, i, cost) 1807 add_stmt_cost (costs, cost->count, cost->kind, cost->stmt_info, 1808 cost->node, cost->vectype, cost->misalign, cost->where); 1809 } 1810 1811 /*-----------------------------------------------------------------*/ 1812 /* Info on data references alignment. */ 1813 /*-----------------------------------------------------------------*/ 1814 #define DR_MISALIGNMENT_UNKNOWN (-1) 1815 #define DR_MISALIGNMENT_UNINITIALIZED (-2) 1816 1817 inline void 1818 set_dr_misalignment (dr_vec_info *dr_info, int val) 1819 { 1820 dr_info->misalignment = val; 1821 } 1822 1823 extern int dr_misalignment (dr_vec_info *dr_info, tree vectype, 1824 poly_int64 offset = 0); 1825 1826 #define SET_DR_MISALIGNMENT(DR, VAL) set_dr_misalignment (DR, VAL) 1827 1828 /* Only defined once DR_MISALIGNMENT is defined. */ 1829 static inline const poly_uint64 1830 dr_target_alignment (dr_vec_info *dr_info) 1831 { 1832 if (STMT_VINFO_GROUPED_ACCESS (dr_info->stmt)) 1833 dr_info = STMT_VINFO_DR_INFO (DR_GROUP_FIRST_ELEMENT (dr_info->stmt)); 1834 return dr_info->target_alignment; 1835 } 1836 #define DR_TARGET_ALIGNMENT(DR) dr_target_alignment (DR) 1837 1838 static inline void 1839 set_dr_target_alignment (dr_vec_info *dr_info, poly_uint64 val) 1840 { 1841 dr_info->target_alignment = val; 1842 } 1843 #define SET_DR_TARGET_ALIGNMENT(DR, VAL) set_dr_target_alignment (DR, VAL) 1844 1845 /* Return true if data access DR_INFO is aligned to the targets 1846 preferred alignment for VECTYPE (which may be less than a full vector). */ 1847 1848 static inline bool 1849 aligned_access_p (dr_vec_info *dr_info, tree vectype) 1850 { 1851 return (dr_misalignment (dr_info, vectype) == 0); 1852 } 1853 1854 /* Return TRUE if the (mis-)alignment of the data access is known with 1855 respect to the targets preferred alignment for VECTYPE, and FALSE 1856 otherwise. */ 1857 1858 static inline bool 1859 known_alignment_for_access_p (dr_vec_info *dr_info, tree vectype) 1860 { 1861 return (dr_misalignment (dr_info, vectype) != DR_MISALIGNMENT_UNKNOWN); 1862 } 1863 1864 /* Return the minimum alignment in bytes that the vectorized version 1865 of DR_INFO is guaranteed to have. */ 1866 1867 static inline unsigned int 1868 vect_known_alignment_in_bytes (dr_vec_info *dr_info, tree vectype) 1869 { 1870 int misalignment = dr_misalignment (dr_info, vectype); 1871 if (misalignment == DR_MISALIGNMENT_UNKNOWN) 1872 return TYPE_ALIGN_UNIT (TREE_TYPE (DR_REF (dr_info->dr))); 1873 else if (misalignment == 0) 1874 return known_alignment (DR_TARGET_ALIGNMENT (dr_info)); 1875 return misalignment & -misalignment; 1876 } 1877 1878 /* Return the behavior of DR_INFO with respect to the vectorization context 1879 (which for outer loop vectorization might not be the behavior recorded 1880 in DR_INFO itself). */ 1881 1882 static inline innermost_loop_behavior * 1883 vect_dr_behavior (vec_info *vinfo, dr_vec_info *dr_info) 1884 { 1885 stmt_vec_info stmt_info = dr_info->stmt; 1886 loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo); 1887 if (loop_vinfo == NULL 1888 || !nested_in_vect_loop_p (LOOP_VINFO_LOOP (loop_vinfo), stmt_info)) 1889 return &DR_INNERMOST (dr_info->dr); 1890 else 1891 return &STMT_VINFO_DR_WRT_VEC_LOOP (stmt_info); 1892 } 1893 1894 /* Return the offset calculated by adding the offset of this DR_INFO to the 1895 corresponding data_reference's offset. If CHECK_OUTER then use 1896 vect_dr_behavior to select the appropriate data_reference to use. */ 1897 1898 inline tree 1899 get_dr_vinfo_offset (vec_info *vinfo, 1900 dr_vec_info *dr_info, bool check_outer = false) 1901 { 1902 innermost_loop_behavior *base; 1903 if (check_outer) 1904 base = vect_dr_behavior (vinfo, dr_info); 1905 else 1906 base = &dr_info->dr->innermost; 1907 1908 tree offset = base->offset; 1909 1910 if (!dr_info->offset) 1911 return offset; 1912 1913 offset = fold_convert (sizetype, offset); 1914 return fold_build2 (PLUS_EXPR, TREE_TYPE (dr_info->offset), offset, 1915 dr_info->offset); 1916 } 1917 1918 1919 /* Return the vect cost model for LOOP. */ 1920 static inline enum vect_cost_model 1921 loop_cost_model (loop_p loop) 1922 { 1923 if (loop != NULL 1924 && loop->force_vectorize 1925 && flag_simd_cost_model != VECT_COST_MODEL_DEFAULT) 1926 return flag_simd_cost_model; 1927 return flag_vect_cost_model; 1928 } 1929 1930 /* Return true if the vect cost model is unlimited. */ 1931 static inline bool 1932 unlimited_cost_model (loop_p loop) 1933 { 1934 return loop_cost_model (loop) == VECT_COST_MODEL_UNLIMITED; 1935 } 1936 1937 /* Return true if the loop described by LOOP_VINFO is fully-masked and 1938 if the first iteration should use a partial mask in order to achieve 1939 alignment. */ 1940 1941 static inline bool 1942 vect_use_loop_mask_for_alignment_p (loop_vec_info loop_vinfo) 1943 { 1944 return (LOOP_VINFO_FULLY_MASKED_P (loop_vinfo) 1945 && LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo)); 1946 } 1947 1948 /* Return the number of vectors of type VECTYPE that are needed to get 1949 NUNITS elements. NUNITS should be based on the vectorization factor, 1950 so it is always a known multiple of the number of elements in VECTYPE. */ 1951 1952 static inline unsigned int 1953 vect_get_num_vectors (poly_uint64 nunits, tree vectype) 1954 { 1955 return exact_div (nunits, TYPE_VECTOR_SUBPARTS (vectype)).to_constant (); 1956 } 1957 1958 /* Return the number of copies needed for loop vectorization when 1959 a statement operates on vectors of type VECTYPE. This is the 1960 vectorization factor divided by the number of elements in 1961 VECTYPE and is always known at compile time. */ 1962 1963 static inline unsigned int 1964 vect_get_num_copies (loop_vec_info loop_vinfo, tree vectype) 1965 { 1966 return vect_get_num_vectors (LOOP_VINFO_VECT_FACTOR (loop_vinfo), vectype); 1967 } 1968 1969 /* Update maximum unit count *MAX_NUNITS so that it accounts for 1970 NUNITS. *MAX_NUNITS can be 1 if we haven't yet recorded anything. */ 1971 1972 static inline void 1973 vect_update_max_nunits (poly_uint64 *max_nunits, poly_uint64 nunits) 1974 { 1975 /* All unit counts have the form vec_info::vector_size * X for some 1976 rational X, so two unit sizes must have a common multiple. 1977 Everything is a multiple of the initial value of 1. */ 1978 *max_nunits = force_common_multiple (*max_nunits, nunits); 1979 } 1980 1981 /* Update maximum unit count *MAX_NUNITS so that it accounts for 1982 the number of units in vector type VECTYPE. *MAX_NUNITS can be 1 1983 if we haven't yet recorded any vector types. */ 1984 1985 static inline void 1986 vect_update_max_nunits (poly_uint64 *max_nunits, tree vectype) 1987 { 1988 vect_update_max_nunits (max_nunits, TYPE_VECTOR_SUBPARTS (vectype)); 1989 } 1990 1991 /* Return the vectorization factor that should be used for costing 1992 purposes while vectorizing the loop described by LOOP_VINFO. 1993 Pick a reasonable estimate if the vectorization factor isn't 1994 known at compile time. */ 1995 1996 static inline unsigned int 1997 vect_vf_for_cost (loop_vec_info loop_vinfo) 1998 { 1999 return estimated_poly_value (LOOP_VINFO_VECT_FACTOR (loop_vinfo)); 2000 } 2001 2002 /* Estimate the number of elements in VEC_TYPE for costing purposes. 2003 Pick a reasonable estimate if the exact number isn't known at 2004 compile time. */ 2005 2006 static inline unsigned int 2007 vect_nunits_for_cost (tree vec_type) 2008 { 2009 return estimated_poly_value (TYPE_VECTOR_SUBPARTS (vec_type)); 2010 } 2011 2012 /* Return the maximum possible vectorization factor for LOOP_VINFO. */ 2013 2014 static inline unsigned HOST_WIDE_INT 2015 vect_max_vf (loop_vec_info loop_vinfo) 2016 { 2017 unsigned HOST_WIDE_INT vf; 2018 if (LOOP_VINFO_VECT_FACTOR (loop_vinfo).is_constant (&vf)) 2019 return vf; 2020 return MAX_VECTORIZATION_FACTOR; 2021 } 2022 2023 /* Return the size of the value accessed by unvectorized data reference 2024 DR_INFO. This is only valid once STMT_VINFO_VECTYPE has been calculated 2025 for the associated gimple statement, since that guarantees that DR_INFO 2026 accesses either a scalar or a scalar equivalent. ("Scalar equivalent" 2027 here includes things like V1SI, which can be vectorized in the same way 2028 as a plain SI.) */ 2029 2030 inline unsigned int 2031 vect_get_scalar_dr_size (dr_vec_info *dr_info) 2032 { 2033 return tree_to_uhwi (TYPE_SIZE_UNIT (TREE_TYPE (DR_REF (dr_info->dr)))); 2034 } 2035 2036 /* Return true if LOOP_VINFO requires a runtime check for whether the 2037 vector loop is profitable. */ 2038 2039 inline bool 2040 vect_apply_runtime_profitability_check_p (loop_vec_info loop_vinfo) 2041 { 2042 unsigned int th = LOOP_VINFO_COST_MODEL_THRESHOLD (loop_vinfo); 2043 return (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) 2044 && th >= vect_vf_for_cost (loop_vinfo)); 2045 } 2046 2047 /* Source location + hotness information. */ 2048 extern dump_user_location_t vect_location; 2049 2050 /* A macro for calling: 2051 dump_begin_scope (MSG, vect_location); 2052 via an RAII object, thus printing "=== MSG ===\n" to the dumpfile etc, 2053 and then calling 2054 dump_end_scope (); 2055 once the object goes out of scope, thus capturing the nesting of 2056 the scopes. 2057 2058 These scopes affect dump messages within them: dump messages at the 2059 top level implicitly default to MSG_PRIORITY_USER_FACING, whereas those 2060 in a nested scope implicitly default to MSG_PRIORITY_INTERNALS. */ 2061 2062 #define DUMP_VECT_SCOPE(MSG) \ 2063 AUTO_DUMP_SCOPE (MSG, vect_location) 2064 2065 /* A sentinel class for ensuring that the "vect_location" global gets 2066 reset at the end of a scope. 2067 2068 The "vect_location" global is used during dumping and contains a 2069 location_t, which could contain references to a tree block via the 2070 ad-hoc data. This data is used for tracking inlining information, 2071 but it's not a GC root; it's simply assumed that such locations never 2072 get accessed if the blocks are optimized away. 2073 2074 Hence we need to ensure that such locations are purged at the end 2075 of any operations using them (e.g. via this class). */ 2076 2077 class auto_purge_vect_location 2078 { 2079 public: 2080 ~auto_purge_vect_location (); 2081 }; 2082 2083 /*-----------------------------------------------------------------*/ 2084 /* Function prototypes. */ 2085 /*-----------------------------------------------------------------*/ 2086 2087 /* Simple loop peeling and versioning utilities for vectorizer's purposes - 2088 in tree-vect-loop-manip.cc. */ 2089 extern void vect_set_loop_condition (class loop *, loop_vec_info, 2090 tree, tree, tree, bool); 2091 extern bool slpeel_can_duplicate_loop_p (const class loop *, const_edge); 2092 class loop *slpeel_tree_duplicate_loop_to_edge_cfg (class loop *, 2093 class loop *, edge); 2094 class loop *vect_loop_versioning (loop_vec_info, gimple *); 2095 extern class loop *vect_do_peeling (loop_vec_info, tree, tree, 2096 tree *, tree *, tree *, int, bool, bool, 2097 tree *); 2098 extern tree vect_get_main_loop_result (loop_vec_info, tree, tree); 2099 extern void vect_prepare_for_masked_peels (loop_vec_info); 2100 extern dump_user_location_t find_loop_location (class loop *); 2101 extern bool vect_can_advance_ivs_p (loop_vec_info); 2102 extern void vect_update_inits_of_drs (loop_vec_info, tree, tree_code); 2103 2104 /* In tree-vect-stmts.cc. */ 2105 extern tree get_related_vectype_for_scalar_type (machine_mode, tree, 2106 poly_uint64 = 0); 2107 extern tree get_vectype_for_scalar_type (vec_info *, tree, unsigned int = 0); 2108 extern tree get_vectype_for_scalar_type (vec_info *, tree, slp_tree); 2109 extern tree get_mask_type_for_scalar_type (vec_info *, tree, unsigned int = 0); 2110 extern tree get_same_sized_vectype (tree, tree); 2111 extern bool vect_chooses_same_modes_p (vec_info *, machine_mode); 2112 extern bool vect_get_loop_mask_type (loop_vec_info); 2113 extern bool vect_is_simple_use (tree, vec_info *, enum vect_def_type *, 2114 stmt_vec_info * = NULL, gimple ** = NULL); 2115 extern bool vect_is_simple_use (tree, vec_info *, enum vect_def_type *, 2116 tree *, stmt_vec_info * = NULL, 2117 gimple ** = NULL); 2118 extern bool vect_is_simple_use (vec_info *, stmt_vec_info, slp_tree, 2119 unsigned, tree *, slp_tree *, 2120 enum vect_def_type *, 2121 tree *, stmt_vec_info * = NULL); 2122 extern bool vect_maybe_update_slp_op_vectype (slp_tree, tree); 2123 extern bool supportable_widening_operation (vec_info *, 2124 enum tree_code, stmt_vec_info, 2125 tree, tree, enum tree_code *, 2126 enum tree_code *, int *, 2127 vec<tree> *); 2128 extern bool supportable_narrowing_operation (enum tree_code, tree, tree, 2129 enum tree_code *, int *, 2130 vec<tree> *); 2131 2132 extern unsigned record_stmt_cost (stmt_vector_for_cost *, int, 2133 enum vect_cost_for_stmt, stmt_vec_info, 2134 tree, int, enum vect_cost_model_location); 2135 extern unsigned record_stmt_cost (stmt_vector_for_cost *, int, 2136 enum vect_cost_for_stmt, slp_tree, 2137 tree, int, enum vect_cost_model_location); 2138 extern unsigned record_stmt_cost (stmt_vector_for_cost *, int, 2139 enum vect_cost_for_stmt, 2140 enum vect_cost_model_location); 2141 2142 /* Overload of record_stmt_cost with VECTYPE derived from STMT_INFO. */ 2143 2144 static inline unsigned 2145 record_stmt_cost (stmt_vector_for_cost *body_cost_vec, int count, 2146 enum vect_cost_for_stmt kind, stmt_vec_info stmt_info, 2147 int misalign, enum vect_cost_model_location where) 2148 { 2149 return record_stmt_cost (body_cost_vec, count, kind, stmt_info, 2150 STMT_VINFO_VECTYPE (stmt_info), misalign, where); 2151 } 2152 2153 extern void vect_finish_replace_stmt (vec_info *, stmt_vec_info, gimple *); 2154 extern void vect_finish_stmt_generation (vec_info *, stmt_vec_info, gimple *, 2155 gimple_stmt_iterator *); 2156 extern opt_result vect_mark_stmts_to_be_vectorized (loop_vec_info, bool *); 2157 extern tree vect_get_store_rhs (stmt_vec_info); 2158 void vect_get_vec_defs_for_operand (vec_info *vinfo, stmt_vec_info, unsigned, 2159 tree op, vec<tree> *, tree = NULL); 2160 void vect_get_vec_defs (vec_info *, stmt_vec_info, slp_tree, unsigned, 2161 tree, vec<tree> *, 2162 tree = NULL, vec<tree> * = NULL, 2163 tree = NULL, vec<tree> * = NULL, 2164 tree = NULL, vec<tree> * = NULL); 2165 void vect_get_vec_defs (vec_info *, stmt_vec_info, slp_tree, unsigned, 2166 tree, vec<tree> *, tree, 2167 tree = NULL, vec<tree> * = NULL, tree = NULL, 2168 tree = NULL, vec<tree> * = NULL, tree = NULL, 2169 tree = NULL, vec<tree> * = NULL, tree = NULL); 2170 extern tree vect_init_vector (vec_info *, stmt_vec_info, tree, tree, 2171 gimple_stmt_iterator *); 2172 extern tree vect_get_slp_vect_def (slp_tree, unsigned); 2173 extern bool vect_transform_stmt (vec_info *, stmt_vec_info, 2174 gimple_stmt_iterator *, 2175 slp_tree, slp_instance); 2176 extern void vect_remove_stores (vec_info *, stmt_vec_info); 2177 extern bool vect_nop_conversion_p (stmt_vec_info); 2178 extern opt_result vect_analyze_stmt (vec_info *, stmt_vec_info, bool *, 2179 slp_tree, 2180 slp_instance, stmt_vector_for_cost *); 2181 extern void vect_get_load_cost (vec_info *, stmt_vec_info, int, 2182 dr_alignment_support, int, bool, 2183 unsigned int *, unsigned int *, 2184 stmt_vector_for_cost *, 2185 stmt_vector_for_cost *, bool); 2186 extern void vect_get_store_cost (vec_info *, stmt_vec_info, int, 2187 dr_alignment_support, int, 2188 unsigned int *, stmt_vector_for_cost *); 2189 extern bool vect_supportable_shift (vec_info *, enum tree_code, tree); 2190 extern tree vect_gen_perm_mask_any (tree, const vec_perm_indices &); 2191 extern tree vect_gen_perm_mask_checked (tree, const vec_perm_indices &); 2192 extern void optimize_mask_stores (class loop*); 2193 extern tree vect_gen_while (gimple_seq *, tree, tree, tree, 2194 const char * = nullptr); 2195 extern tree vect_gen_while_not (gimple_seq *, tree, tree, tree); 2196 extern opt_result vect_get_vector_types_for_stmt (vec_info *, 2197 stmt_vec_info, tree *, 2198 tree *, unsigned int = 0); 2199 extern opt_tree vect_get_mask_type_for_stmt (stmt_vec_info, unsigned int = 0); 2200 2201 /* In tree-vect-data-refs.cc. */ 2202 extern bool vect_can_force_dr_alignment_p (const_tree, poly_uint64); 2203 extern enum dr_alignment_support vect_supportable_dr_alignment 2204 (vec_info *, dr_vec_info *, tree, int); 2205 extern tree vect_get_smallest_scalar_type (stmt_vec_info, tree); 2206 extern opt_result vect_analyze_data_ref_dependences (loop_vec_info, unsigned int *); 2207 extern bool vect_slp_analyze_instance_dependence (vec_info *, slp_instance); 2208 extern opt_result vect_enhance_data_refs_alignment (loop_vec_info); 2209 extern opt_result vect_analyze_data_refs_alignment (loop_vec_info); 2210 extern bool vect_slp_analyze_instance_alignment (vec_info *, slp_instance); 2211 extern opt_result vect_analyze_data_ref_accesses (vec_info *, vec<int> *); 2212 extern opt_result vect_prune_runtime_alias_test_list (loop_vec_info); 2213 extern bool vect_gather_scatter_fn_p (vec_info *, bool, bool, tree, tree, 2214 tree, int, internal_fn *, tree *); 2215 extern bool vect_check_gather_scatter (stmt_vec_info, loop_vec_info, 2216 gather_scatter_info *); 2217 extern opt_result vect_find_stmt_data_reference (loop_p, gimple *, 2218 vec<data_reference_p> *, 2219 vec<int> *, int); 2220 extern opt_result vect_analyze_data_refs (vec_info *, poly_uint64 *, bool *); 2221 extern void vect_record_base_alignments (vec_info *); 2222 extern tree vect_create_data_ref_ptr (vec_info *, 2223 stmt_vec_info, tree, class loop *, tree, 2224 tree *, gimple_stmt_iterator *, 2225 gimple **, bool, 2226 tree = NULL_TREE); 2227 extern tree bump_vector_ptr (vec_info *, tree, gimple *, gimple_stmt_iterator *, 2228 stmt_vec_info, tree); 2229 extern void vect_copy_ref_info (tree, tree); 2230 extern tree vect_create_destination_var (tree, tree); 2231 extern bool vect_grouped_store_supported (tree, unsigned HOST_WIDE_INT); 2232 extern bool vect_store_lanes_supported (tree, unsigned HOST_WIDE_INT, bool); 2233 extern bool vect_grouped_load_supported (tree, bool, unsigned HOST_WIDE_INT); 2234 extern bool vect_load_lanes_supported (tree, unsigned HOST_WIDE_INT, bool); 2235 extern void vect_permute_store_chain (vec_info *, vec<tree> &, 2236 unsigned int, stmt_vec_info, 2237 gimple_stmt_iterator *, vec<tree> *); 2238 extern tree vect_setup_realignment (vec_info *, 2239 stmt_vec_info, gimple_stmt_iterator *, 2240 tree *, enum dr_alignment_support, tree, 2241 class loop **); 2242 extern void vect_transform_grouped_load (vec_info *, stmt_vec_info, vec<tree>, 2243 int, gimple_stmt_iterator *); 2244 extern void vect_record_grouped_load_vectors (vec_info *, 2245 stmt_vec_info, vec<tree>); 2246 extern tree vect_get_new_vect_var (tree, enum vect_var_kind, const char *); 2247 extern tree vect_get_new_ssa_name (tree, enum vect_var_kind, 2248 const char * = NULL); 2249 extern tree vect_create_addr_base_for_vector_ref (vec_info *, 2250 stmt_vec_info, gimple_seq *, 2251 tree); 2252 2253 /* In tree-vect-loop.cc. */ 2254 extern tree neutral_op_for_reduction (tree, code_helper, tree); 2255 extern widest_int vect_iv_limit_for_partial_vectors (loop_vec_info loop_vinfo); 2256 bool vect_rgroup_iv_might_wrap_p (loop_vec_info, rgroup_controls *); 2257 /* Used in tree-vect-loop-manip.cc */ 2258 extern opt_result vect_determine_partial_vectors_and_peeling (loop_vec_info, 2259 bool); 2260 /* Used in gimple-loop-interchange.c and tree-parloops.cc. */ 2261 extern bool check_reduction_path (dump_user_location_t, loop_p, gphi *, tree, 2262 enum tree_code); 2263 extern bool needs_fold_left_reduction_p (tree, code_helper); 2264 /* Drive for loop analysis stage. */ 2265 extern opt_loop_vec_info vect_analyze_loop (class loop *, vec_info_shared *); 2266 extern tree vect_build_loop_niters (loop_vec_info, bool * = NULL); 2267 extern void vect_gen_vector_loop_niters (loop_vec_info, tree, tree *, 2268 tree *, bool); 2269 extern tree vect_halve_mask_nunits (tree, machine_mode); 2270 extern tree vect_double_mask_nunits (tree, machine_mode); 2271 extern void vect_record_loop_mask (loop_vec_info, vec_loop_masks *, 2272 unsigned int, tree, tree); 2273 extern tree vect_get_loop_mask (gimple_stmt_iterator *, vec_loop_masks *, 2274 unsigned int, tree, unsigned int); 2275 extern void vect_record_loop_len (loop_vec_info, vec_loop_lens *, unsigned int, 2276 tree, unsigned int); 2277 extern tree vect_get_loop_len (loop_vec_info, vec_loop_lens *, unsigned int, 2278 unsigned int); 2279 extern gimple_seq vect_gen_len (tree, tree, tree, tree); 2280 extern stmt_vec_info info_for_reduction (vec_info *, stmt_vec_info); 2281 extern bool reduction_fn_for_scalar_code (code_helper, internal_fn *); 2282 2283 /* Drive for loop transformation stage. */ 2284 extern class loop *vect_transform_loop (loop_vec_info, gimple *); 2285 struct vect_loop_form_info 2286 { 2287 tree number_of_iterations; 2288 tree number_of_iterationsm1; 2289 tree assumptions; 2290 gcond *loop_cond; 2291 gcond *inner_loop_cond; 2292 }; 2293 extern opt_result vect_analyze_loop_form (class loop *, vect_loop_form_info *); 2294 extern loop_vec_info vect_create_loop_vinfo (class loop *, vec_info_shared *, 2295 const vect_loop_form_info *, 2296 loop_vec_info = nullptr); 2297 extern bool vectorizable_live_operation (vec_info *, 2298 stmt_vec_info, gimple_stmt_iterator *, 2299 slp_tree, slp_instance, int, 2300 bool, stmt_vector_for_cost *); 2301 extern bool vectorizable_reduction (loop_vec_info, stmt_vec_info, 2302 slp_tree, slp_instance, 2303 stmt_vector_for_cost *); 2304 extern bool vectorizable_induction (loop_vec_info, stmt_vec_info, 2305 gimple **, slp_tree, 2306 stmt_vector_for_cost *); 2307 extern bool vect_transform_reduction (loop_vec_info, stmt_vec_info, 2308 gimple_stmt_iterator *, 2309 gimple **, slp_tree); 2310 extern bool vect_transform_cycle_phi (loop_vec_info, stmt_vec_info, 2311 gimple **, 2312 slp_tree, slp_instance); 2313 extern bool vectorizable_lc_phi (loop_vec_info, stmt_vec_info, 2314 gimple **, slp_tree); 2315 extern bool vectorizable_phi (vec_info *, stmt_vec_info, gimple **, slp_tree, 2316 stmt_vector_for_cost *); 2317 extern bool vect_emulated_vector_p (tree); 2318 extern bool vect_can_vectorize_without_simd_p (tree_code); 2319 extern bool vect_can_vectorize_without_simd_p (code_helper); 2320 extern int vect_get_known_peeling_cost (loop_vec_info, int, int *, 2321 stmt_vector_for_cost *, 2322 stmt_vector_for_cost *, 2323 stmt_vector_for_cost *); 2324 extern tree cse_and_gimplify_to_preheader (loop_vec_info, tree); 2325 2326 /* In tree-vect-slp.cc. */ 2327 extern void vect_slp_init (void); 2328 extern void vect_slp_fini (void); 2329 extern void vect_free_slp_instance (slp_instance); 2330 extern bool vect_transform_slp_perm_load (vec_info *, slp_tree, const vec<tree> &, 2331 gimple_stmt_iterator *, poly_uint64, 2332 bool, unsigned *, 2333 unsigned * = nullptr, bool = false); 2334 extern bool vect_slp_analyze_operations (vec_info *); 2335 extern void vect_schedule_slp (vec_info *, const vec<slp_instance> &); 2336 extern opt_result vect_analyze_slp (vec_info *, unsigned); 2337 extern bool vect_make_slp_decision (loop_vec_info); 2338 extern void vect_detect_hybrid_slp (loop_vec_info); 2339 extern void vect_optimize_slp (vec_info *); 2340 extern void vect_gather_slp_loads (vec_info *); 2341 extern void vect_get_slp_defs (slp_tree, vec<tree> *); 2342 extern void vect_get_slp_defs (vec_info *, slp_tree, vec<vec<tree> > *, 2343 unsigned n = -1U); 2344 extern bool vect_slp_if_converted_bb (basic_block bb, loop_p orig_loop); 2345 extern bool vect_slp_function (function *); 2346 extern stmt_vec_info vect_find_last_scalar_stmt_in_slp (slp_tree); 2347 extern stmt_vec_info vect_find_first_scalar_stmt_in_slp (slp_tree); 2348 extern bool is_simple_and_all_uses_invariant (stmt_vec_info, loop_vec_info); 2349 extern bool can_duplicate_and_interleave_p (vec_info *, unsigned int, tree, 2350 unsigned int * = NULL, 2351 tree * = NULL, tree * = NULL); 2352 extern void duplicate_and_interleave (vec_info *, gimple_seq *, tree, 2353 const vec<tree> &, unsigned int, vec<tree> &); 2354 extern int vect_get_place_in_interleaving_chain (stmt_vec_info, stmt_vec_info); 2355 extern slp_tree vect_create_new_slp_node (unsigned, tree_code); 2356 extern void vect_free_slp_tree (slp_tree); 2357 extern bool compatible_calls_p (gcall *, gcall *); 2358 2359 /* In tree-vect-patterns.cc. */ 2360 extern void 2361 vect_mark_pattern_stmts (vec_info *, stmt_vec_info, gimple *, tree); 2362 2363 /* Pattern recognition functions. 2364 Additional pattern recognition functions can (and will) be added 2365 in the future. */ 2366 void vect_pattern_recog (vec_info *); 2367 2368 /* In tree-vectorizer.cc. */ 2369 unsigned vectorize_loops (void); 2370 void vect_free_loop_info_assumptions (class loop *); 2371 gimple *vect_loop_vectorized_call (class loop *, gcond **cond = NULL); 2372 bool vect_stmt_dominates_stmt_p (gimple *, gimple *); 2373 2374 /* SLP Pattern matcher types, tree-vect-slp-patterns.cc. */ 2375 2376 /* Forward declaration of possible two operands operation that can be matched 2377 by the complex numbers pattern matchers. */ 2378 enum _complex_operation : unsigned; 2379 2380 /* All possible load permute values that could result from the partial data-flow 2381 analysis. */ 2382 typedef enum _complex_perm_kinds { 2383 PERM_UNKNOWN, 2384 PERM_EVENODD, 2385 PERM_ODDEVEN, 2386 PERM_ODDODD, 2387 PERM_EVENEVEN, 2388 /* Can be combined with any other PERM values. */ 2389 PERM_TOP 2390 } complex_perm_kinds_t; 2391 2392 /* Cache from nodes to the load permutation they represent. */ 2393 typedef hash_map <slp_tree, complex_perm_kinds_t> 2394 slp_tree_to_load_perm_map_t; 2395 2396 /* Cache from nodes pair to being compatible or not. */ 2397 typedef pair_hash <nofree_ptr_hash <_slp_tree>, 2398 nofree_ptr_hash <_slp_tree>> slp_node_hash; 2399 typedef hash_map <slp_node_hash, bool> slp_compat_nodes_map_t; 2400 2401 2402 /* Vector pattern matcher base class. All SLP pattern matchers must inherit 2403 from this type. */ 2404 2405 class vect_pattern 2406 { 2407 protected: 2408 /* The number of arguments that the IFN requires. */ 2409 unsigned m_num_args; 2410 2411 /* The internal function that will be used when a pattern is created. */ 2412 internal_fn m_ifn; 2413 2414 /* The current node being inspected. */ 2415 slp_tree *m_node; 2416 2417 /* The list of operands to be the children for the node produced when the 2418 internal function is created. */ 2419 vec<slp_tree> m_ops; 2420 2421 /* Default constructor where NODE is the root of the tree to inspect. */ 2422 vect_pattern (slp_tree *node, vec<slp_tree> *m_ops, internal_fn ifn) 2423 { 2424 this->m_ifn = ifn; 2425 this->m_node = node; 2426 this->m_ops.create (0); 2427 if (m_ops) 2428 this->m_ops.safe_splice (*m_ops); 2429 } 2430 2431 public: 2432 2433 /* Create a new instance of the pattern matcher class of the given type. */ 2434 static vect_pattern* recognize (slp_tree_to_load_perm_map_t *, 2435 slp_compat_nodes_map_t *, slp_tree *); 2436 2437 /* Build the pattern from the data collected so far. */ 2438 virtual void build (vec_info *) = 0; 2439 2440 /* Default destructor. */ 2441 virtual ~vect_pattern () 2442 { 2443 this->m_ops.release (); 2444 } 2445 }; 2446 2447 /* Function pointer to create a new pattern matcher from a generic type. */ 2448 typedef vect_pattern* (*vect_pattern_decl_t) (slp_tree_to_load_perm_map_t *, 2449 slp_compat_nodes_map_t *, 2450 slp_tree *); 2451 2452 /* List of supported pattern matchers. */ 2453 extern vect_pattern_decl_t slp_patterns[]; 2454 2455 /* Number of supported pattern matchers. */ 2456 extern size_t num__slp_patterns; 2457 2458 /* ---------------------------------------------------------------------- 2459 Target support routines 2460 ----------------------------------------------------------------------- 2461 The following routines are provided to simplify costing decisions in 2462 target code. Please add more as needed. */ 2463 2464 /* Return true if an operaton of kind KIND for STMT_INFO represents 2465 the extraction of an element from a vector in preparation for 2466 storing the element to memory. */ 2467 inline bool 2468 vect_is_store_elt_extraction (vect_cost_for_stmt kind, stmt_vec_info stmt_info) 2469 { 2470 return (kind == vec_to_scalar 2471 && STMT_VINFO_DATA_REF (stmt_info) 2472 && DR_IS_WRITE (STMT_VINFO_DATA_REF (stmt_info))); 2473 } 2474 2475 /* Return true if STMT_INFO represents part of a reduction. */ 2476 inline bool 2477 vect_is_reduction (stmt_vec_info stmt_info) 2478 { 2479 return STMT_VINFO_REDUC_IDX (stmt_info) >= 0; 2480 } 2481 2482 /* If STMT_INFO describes a reduction, return the vect_reduction_type 2483 of the reduction it describes, otherwise return -1. */ 2484 inline int 2485 vect_reduc_type (vec_info *vinfo, stmt_vec_info stmt_info) 2486 { 2487 if (loop_vec_info loop_vinfo = dyn_cast<loop_vec_info> (vinfo)) 2488 if (STMT_VINFO_REDUC_DEF (stmt_info)) 2489 { 2490 stmt_vec_info reduc_info = info_for_reduction (loop_vinfo, stmt_info); 2491 return int (STMT_VINFO_REDUC_TYPE (reduc_info)); 2492 } 2493 return -1; 2494 } 2495 2496 /* If STMT_INFO is a COND_EXPR that includes an embedded comparison, return the 2497 scalar type of the values being compared. Return null otherwise. */ 2498 inline tree 2499 vect_embedded_comparison_type (stmt_vec_info stmt_info) 2500 { 2501 if (auto *assign = dyn_cast<gassign *> (stmt_info->stmt)) 2502 if (gimple_assign_rhs_code (assign) == COND_EXPR) 2503 { 2504 tree cond = gimple_assign_rhs1 (assign); 2505 if (COMPARISON_CLASS_P (cond)) 2506 return TREE_TYPE (TREE_OPERAND (cond, 0)); 2507 } 2508 return NULL_TREE; 2509 } 2510 2511 /* If STMT_INFO is a comparison or contains an embedded comparison, return the 2512 scalar type of the values being compared. Return null otherwise. */ 2513 inline tree 2514 vect_comparison_type (stmt_vec_info stmt_info) 2515 { 2516 if (auto *assign = dyn_cast<gassign *> (stmt_info->stmt)) 2517 if (TREE_CODE_CLASS (gimple_assign_rhs_code (assign)) == tcc_comparison) 2518 return TREE_TYPE (gimple_assign_rhs1 (assign)); 2519 return vect_embedded_comparison_type (stmt_info); 2520 } 2521 2522 /* Return true if STMT_INFO extends the result of a load. */ 2523 inline bool 2524 vect_is_extending_load (class vec_info *vinfo, stmt_vec_info stmt_info) 2525 { 2526 /* Although this is quite large for an inline function, this part 2527 at least should be inline. */ 2528 gassign *assign = dyn_cast <gassign *> (stmt_info->stmt); 2529 if (!assign || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (assign))) 2530 return false; 2531 2532 tree rhs = gimple_assign_rhs1 (stmt_info->stmt); 2533 tree lhs_type = TREE_TYPE (gimple_assign_lhs (assign)); 2534 tree rhs_type = TREE_TYPE (rhs); 2535 if (!INTEGRAL_TYPE_P (lhs_type) 2536 || !INTEGRAL_TYPE_P (rhs_type) 2537 || TYPE_PRECISION (lhs_type) <= TYPE_PRECISION (rhs_type)) 2538 return false; 2539 2540 stmt_vec_info def_stmt_info = vinfo->lookup_def (rhs); 2541 return (def_stmt_info 2542 && STMT_VINFO_DATA_REF (def_stmt_info) 2543 && DR_IS_READ (STMT_VINFO_DATA_REF (def_stmt_info))); 2544 } 2545 2546 /* Return true if STMT_INFO is an integer truncation. */ 2547 inline bool 2548 vect_is_integer_truncation (stmt_vec_info stmt_info) 2549 { 2550 gassign *assign = dyn_cast <gassign *> (stmt_info->stmt); 2551 if (!assign || !CONVERT_EXPR_CODE_P (gimple_assign_rhs_code (assign))) 2552 return false; 2553 2554 tree lhs_type = TREE_TYPE (gimple_assign_lhs (assign)); 2555 tree rhs_type = TREE_TYPE (gimple_assign_rhs1 (assign)); 2556 return (INTEGRAL_TYPE_P (lhs_type) 2557 && INTEGRAL_TYPE_P (rhs_type) 2558 && TYPE_PRECISION (lhs_type) < TYPE_PRECISION (rhs_type)); 2559 } 2560 2561 #endif /* GCC_TREE_VECTORIZER_H */ 2562