1 /* Vectorizer 2 Copyright (C) 2003-2017 Free Software Foundation, Inc. 3 Contributed by Dorit Naishlos <dorit@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 #include "tree-data-ref.h" 25 #include "target.h" 26 27 /* Used for naming of new temporaries. */ 28 enum vect_var_kind { 29 vect_simple_var, 30 vect_pointer_var, 31 vect_scalar_var, 32 vect_mask_var 33 }; 34 35 /* Defines type of operation. */ 36 enum operation_type { 37 unary_op = 1, 38 binary_op, 39 ternary_op 40 }; 41 42 /* Define type of available alignment support. */ 43 enum dr_alignment_support { 44 dr_unaligned_unsupported, 45 dr_unaligned_supported, 46 dr_explicit_realign, 47 dr_explicit_realign_optimized, 48 dr_aligned 49 }; 50 51 /* Define type of def-use cross-iteration cycle. */ 52 enum vect_def_type { 53 vect_uninitialized_def = 0, 54 vect_constant_def = 1, 55 vect_external_def, 56 vect_internal_def, 57 vect_induction_def, 58 vect_reduction_def, 59 vect_double_reduction_def, 60 vect_nested_cycle, 61 vect_unknown_def_type 62 }; 63 64 /* Define type of reduction. */ 65 enum vect_reduction_type { 66 TREE_CODE_REDUCTION, 67 COND_REDUCTION, 68 INTEGER_INDUC_COND_REDUCTION, 69 CONST_COND_REDUCTION 70 }; 71 72 #define VECTORIZABLE_CYCLE_DEF(D) (((D) == vect_reduction_def) \ 73 || ((D) == vect_double_reduction_def) \ 74 || ((D) == vect_nested_cycle)) 75 76 /* Structure to encapsulate information about a group of like 77 instructions to be presented to the target cost model. */ 78 struct stmt_info_for_cost { 79 int count; 80 enum vect_cost_for_stmt kind; 81 gimple *stmt; 82 int misalign; 83 }; 84 85 typedef vec<stmt_info_for_cost> stmt_vector_for_cost; 86 87 /************************************************************************ 88 SLP 89 ************************************************************************/ 90 typedef struct _slp_tree *slp_tree; 91 92 /* A computation tree of an SLP instance. Each node corresponds to a group of 93 stmts to be packed in a SIMD stmt. */ 94 struct _slp_tree { 95 /* Nodes that contain def-stmts of this node statements operands. */ 96 vec<slp_tree> children; 97 /* A group of scalar stmts to be vectorized together. */ 98 vec<gimple *> stmts; 99 /* Load permutation relative to the stores, NULL if there is no 100 permutation. */ 101 vec<unsigned> load_permutation; 102 /* Vectorized stmt/s. */ 103 vec<gimple *> vec_stmts; 104 /* Number of vector stmts that are created to replace the group of scalar 105 stmts. It is calculated during the transformation phase as the number of 106 scalar elements in one scalar iteration (GROUP_SIZE) multiplied by VF 107 divided by vector size. */ 108 unsigned int vec_stmts_size; 109 /* Whether the scalar computations use two different operators. */ 110 bool two_operators; 111 /* The DEF type of this node. */ 112 enum vect_def_type def_type; 113 }; 114 115 116 /* SLP instance is a sequence of stmts in a loop that can be packed into 117 SIMD stmts. */ 118 typedef struct _slp_instance { 119 /* The root of SLP tree. */ 120 slp_tree root; 121 122 /* Size of groups of scalar stmts that will be replaced by SIMD stmt/s. */ 123 unsigned int group_size; 124 125 /* The unrolling factor required to vectorized this SLP instance. */ 126 unsigned int unrolling_factor; 127 128 /* The group of nodes that contain loads of this SLP instance. */ 129 vec<slp_tree> loads; 130 } *slp_instance; 131 132 133 /* Access Functions. */ 134 #define SLP_INSTANCE_TREE(S) (S)->root 135 #define SLP_INSTANCE_GROUP_SIZE(S) (S)->group_size 136 #define SLP_INSTANCE_UNROLLING_FACTOR(S) (S)->unrolling_factor 137 #define SLP_INSTANCE_LOADS(S) (S)->loads 138 139 #define SLP_TREE_CHILDREN(S) (S)->children 140 #define SLP_TREE_SCALAR_STMTS(S) (S)->stmts 141 #define SLP_TREE_VEC_STMTS(S) (S)->vec_stmts 142 #define SLP_TREE_NUMBER_OF_VEC_STMTS(S) (S)->vec_stmts_size 143 #define SLP_TREE_LOAD_PERMUTATION(S) (S)->load_permutation 144 #define SLP_TREE_TWO_OPERATORS(S) (S)->two_operators 145 #define SLP_TREE_DEF_TYPE(S) (S)->def_type 146 147 148 149 /* This struct is used to store the information of a data reference, 150 including the data ref itself and the segment length for aliasing 151 checks. This is used to merge alias checks. */ 152 153 struct dr_with_seg_len 154 { 155 dr_with_seg_len (data_reference_p d, tree len) 156 : dr (d), seg_len (len) {} 157 158 data_reference_p dr; 159 tree seg_len; 160 }; 161 162 /* This struct contains two dr_with_seg_len objects with aliasing data 163 refs. Two comparisons are generated from them. */ 164 165 struct dr_with_seg_len_pair_t 166 { 167 dr_with_seg_len_pair_t (const dr_with_seg_len& d1, 168 const dr_with_seg_len& d2) 169 : first (d1), second (d2) {} 170 171 dr_with_seg_len first; 172 dr_with_seg_len second; 173 }; 174 175 176 177 /* Vectorizer state common between loop and basic-block vectorization. */ 178 struct vec_info { 179 enum { bb, loop } kind; 180 181 /* All SLP instances. */ 182 vec<slp_instance> slp_instances; 183 184 /* All data references. */ 185 vec<data_reference_p> datarefs; 186 187 /* All data dependences. */ 188 vec<ddr_p> ddrs; 189 190 /* All interleaving chains of stores, represented by the first 191 stmt in the chain. */ 192 vec<gimple *> grouped_stores; 193 194 /* Cost data used by the target cost model. */ 195 void *target_cost_data; 196 }; 197 198 struct _loop_vec_info; 199 struct _bb_vec_info; 200 201 template<> 202 template<> 203 inline bool 204 is_a_helper <_loop_vec_info *>::test (vec_info *i) 205 { 206 return i->kind == vec_info::loop; 207 } 208 209 template<> 210 template<> 211 inline bool 212 is_a_helper <_bb_vec_info *>::test (vec_info *i) 213 { 214 return i->kind == vec_info::bb; 215 } 216 217 218 /*-----------------------------------------------------------------*/ 219 /* Info on vectorized loops. */ 220 /*-----------------------------------------------------------------*/ 221 typedef struct _loop_vec_info : public vec_info { 222 223 /* The loop to which this info struct refers to. */ 224 struct loop *loop; 225 226 /* The loop basic blocks. */ 227 basic_block *bbs; 228 229 /* Number of latch executions. */ 230 tree num_itersm1; 231 /* Number of iterations. */ 232 tree num_iters; 233 /* Number of iterations of the original loop. */ 234 tree num_iters_unchanged; 235 /* Condition under which this loop is analyzed and versioned. */ 236 tree num_iters_assumptions; 237 238 /* Threshold of number of iterations below which vectorzation will not be 239 performed. It is calculated from MIN_PROFITABLE_ITERS and 240 PARAM_MIN_VECT_LOOP_BOUND. */ 241 unsigned int th; 242 243 /* Unrolling factor */ 244 int vectorization_factor; 245 246 /* Unknown DRs according to which loop was peeled. */ 247 struct data_reference *unaligned_dr; 248 249 /* peeling_for_alignment indicates whether peeling for alignment will take 250 place, and what the peeling factor should be: 251 peeling_for_alignment = X means: 252 If X=0: Peeling for alignment will not be applied. 253 If X>0: Peel first X iterations. 254 If X=-1: Generate a runtime test to calculate the number of iterations 255 to be peeled, using the dataref recorded in the field 256 unaligned_dr. */ 257 int peeling_for_alignment; 258 259 /* The mask used to check the alignment of pointers or arrays. */ 260 int ptr_mask; 261 262 /* The loop nest in which the data dependences are computed. */ 263 vec<loop_p> loop_nest; 264 265 /* Data Dependence Relations defining address ranges that are candidates 266 for a run-time aliasing check. */ 267 vec<ddr_p> may_alias_ddrs; 268 269 /* Data Dependence Relations defining address ranges together with segment 270 lengths from which the run-time aliasing check is built. */ 271 vec<dr_with_seg_len_pair_t> comp_alias_ddrs; 272 273 /* Statements in the loop that have data references that are candidates for a 274 runtime (loop versioning) misalignment check. */ 275 vec<gimple *> may_misalign_stmts; 276 277 /* Reduction cycles detected in the loop. Used in loop-aware SLP. */ 278 vec<gimple *> reductions; 279 280 /* All reduction chains in the loop, represented by the first 281 stmt in the chain. */ 282 vec<gimple *> reduction_chains; 283 284 /* Cost vector for a single scalar iteration. */ 285 vec<stmt_info_for_cost> scalar_cost_vec; 286 287 /* The unrolling factor needed to SLP the loop. In case of that pure SLP is 288 applied to the loop, i.e., no unrolling is needed, this is 1. */ 289 unsigned slp_unrolling_factor; 290 291 /* Cost of a single scalar iteration. */ 292 int single_scalar_iteration_cost; 293 294 /* Is the loop vectorizable? */ 295 bool vectorizable; 296 297 /* When we have grouped data accesses with gaps, we may introduce invalid 298 memory accesses. We peel the last iteration of the loop to prevent 299 this. */ 300 bool peeling_for_gaps; 301 302 /* When the number of iterations is not a multiple of the vector size 303 we need to peel off iterations at the end to form an epilogue loop. */ 304 bool peeling_for_niter; 305 306 /* Reductions are canonicalized so that the last operand is the reduction 307 operand. If this places a constant into RHS1, this decanonicalizes 308 GIMPLE for other phases, so we must track when this has occurred and 309 fix it up. */ 310 bool operands_swapped; 311 312 /* True if there are no loop carried data dependencies in the loop. 313 If loop->safelen <= 1, then this is always true, either the loop 314 didn't have any loop carried data dependencies, or the loop is being 315 vectorized guarded with some runtime alias checks, or couldn't 316 be vectorized at all, but then this field shouldn't be used. 317 For loop->safelen >= 2, the user has asserted that there are no 318 backward dependencies, but there still could be loop carried forward 319 dependencies in such loops. This flag will be false if normal 320 vectorizer data dependency analysis would fail or require versioning 321 for alias, but because of loop->safelen >= 2 it has been vectorized 322 even without versioning for alias. E.g. in: 323 #pragma omp simd 324 for (int i = 0; i < m; i++) 325 a[i] = a[i + k] * c; 326 (or #pragma simd or #pragma ivdep) we can vectorize this and it will 327 DTRT even for k > 0 && k < m, but without safelen we would not 328 vectorize this, so this field would be false. */ 329 bool no_data_dependencies; 330 331 /* Mark loops having masked stores. */ 332 bool has_mask_store; 333 334 /* If if-conversion versioned this loop before conversion, this is the 335 loop version without if-conversion. */ 336 struct loop *scalar_loop; 337 338 /* For loops being epilogues of already vectorized loops 339 this points to the original vectorized loop. Otherwise NULL. */ 340 _loop_vec_info *orig_loop_info; 341 342 } *loop_vec_info; 343 344 /* Access Functions. */ 345 #define LOOP_VINFO_LOOP(L) (L)->loop 346 #define LOOP_VINFO_BBS(L) (L)->bbs 347 #define LOOP_VINFO_NITERSM1(L) (L)->num_itersm1 348 #define LOOP_VINFO_NITERS(L) (L)->num_iters 349 /* Since LOOP_VINFO_NITERS and LOOP_VINFO_NITERSM1 can change after 350 prologue peeling retain total unchanged scalar loop iterations for 351 cost model. */ 352 #define LOOP_VINFO_NITERS_UNCHANGED(L) (L)->num_iters_unchanged 353 #define LOOP_VINFO_NITERS_ASSUMPTIONS(L) (L)->num_iters_assumptions 354 #define LOOP_VINFO_COST_MODEL_THRESHOLD(L) (L)->th 355 #define LOOP_VINFO_VECTORIZABLE_P(L) (L)->vectorizable 356 #define LOOP_VINFO_VECT_FACTOR(L) (L)->vectorization_factor 357 #define LOOP_VINFO_PTR_MASK(L) (L)->ptr_mask 358 #define LOOP_VINFO_LOOP_NEST(L) (L)->loop_nest 359 #define LOOP_VINFO_DATAREFS(L) (L)->datarefs 360 #define LOOP_VINFO_DDRS(L) (L)->ddrs 361 #define LOOP_VINFO_INT_NITERS(L) (TREE_INT_CST_LOW ((L)->num_iters)) 362 #define LOOP_VINFO_PEELING_FOR_ALIGNMENT(L) (L)->peeling_for_alignment 363 #define LOOP_VINFO_UNALIGNED_DR(L) (L)->unaligned_dr 364 #define LOOP_VINFO_MAY_MISALIGN_STMTS(L) (L)->may_misalign_stmts 365 #define LOOP_VINFO_MAY_ALIAS_DDRS(L) (L)->may_alias_ddrs 366 #define LOOP_VINFO_COMP_ALIAS_DDRS(L) (L)->comp_alias_ddrs 367 #define LOOP_VINFO_GROUPED_STORES(L) (L)->grouped_stores 368 #define LOOP_VINFO_SLP_INSTANCES(L) (L)->slp_instances 369 #define LOOP_VINFO_SLP_UNROLLING_FACTOR(L) (L)->slp_unrolling_factor 370 #define LOOP_VINFO_REDUCTIONS(L) (L)->reductions 371 #define LOOP_VINFO_REDUCTION_CHAINS(L) (L)->reduction_chains 372 #define LOOP_VINFO_TARGET_COST_DATA(L) (L)->target_cost_data 373 #define LOOP_VINFO_PEELING_FOR_GAPS(L) (L)->peeling_for_gaps 374 #define LOOP_VINFO_OPERANDS_SWAPPED(L) (L)->operands_swapped 375 #define LOOP_VINFO_PEELING_FOR_NITER(L) (L)->peeling_for_niter 376 #define LOOP_VINFO_NO_DATA_DEPENDENCIES(L) (L)->no_data_dependencies 377 #define LOOP_VINFO_SCALAR_LOOP(L) (L)->scalar_loop 378 #define LOOP_VINFO_HAS_MASK_STORE(L) (L)->has_mask_store 379 #define LOOP_VINFO_SCALAR_ITERATION_COST(L) (L)->scalar_cost_vec 380 #define LOOP_VINFO_SINGLE_SCALAR_ITERATION_COST(L) (L)->single_scalar_iteration_cost 381 #define LOOP_VINFO_ORIG_LOOP_INFO(L) (L)->orig_loop_info 382 383 #define LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT(L) \ 384 ((L)->may_misalign_stmts.length () > 0) 385 #define LOOP_REQUIRES_VERSIONING_FOR_ALIAS(L) \ 386 ((L)->may_alias_ddrs.length () > 0) 387 #define LOOP_REQUIRES_VERSIONING_FOR_NITERS(L) \ 388 (LOOP_VINFO_NITERS_ASSUMPTIONS (L)) 389 #define LOOP_REQUIRES_VERSIONING(L) \ 390 (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (L) \ 391 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (L) \ 392 || LOOP_REQUIRES_VERSIONING_FOR_NITERS (L)) 393 394 #define LOOP_VINFO_NITERS_KNOWN_P(L) \ 395 (tree_fits_shwi_p ((L)->num_iters) && tree_to_shwi ((L)->num_iters) > 0) 396 397 #define LOOP_VINFO_EPILOGUE_P(L) \ 398 (LOOP_VINFO_ORIG_LOOP_INFO (L) != NULL) 399 400 #define LOOP_VINFO_ORIG_VECT_FACTOR(L) \ 401 (LOOP_VINFO_VECT_FACTOR (LOOP_VINFO_ORIG_LOOP_INFO (L))) 402 403 static inline loop_vec_info 404 loop_vec_info_for_loop (struct loop *loop) 405 { 406 return (loop_vec_info) loop->aux; 407 } 408 409 static inline bool 410 nested_in_vect_loop_p (struct loop *loop, gimple *stmt) 411 { 412 return (loop->inner 413 && (loop->inner == (gimple_bb (stmt))->loop_father)); 414 } 415 416 typedef struct _bb_vec_info : public vec_info 417 { 418 basic_block bb; 419 gimple_stmt_iterator region_begin; 420 gimple_stmt_iterator region_end; 421 } *bb_vec_info; 422 423 #define BB_VINFO_BB(B) (B)->bb 424 #define BB_VINFO_GROUPED_STORES(B) (B)->grouped_stores 425 #define BB_VINFO_SLP_INSTANCES(B) (B)->slp_instances 426 #define BB_VINFO_DATAREFS(B) (B)->datarefs 427 #define BB_VINFO_DDRS(B) (B)->ddrs 428 #define BB_VINFO_TARGET_COST_DATA(B) (B)->target_cost_data 429 430 static inline bb_vec_info 431 vec_info_for_bb (basic_block bb) 432 { 433 return (bb_vec_info) bb->aux; 434 } 435 436 /*-----------------------------------------------------------------*/ 437 /* Info on vectorized defs. */ 438 /*-----------------------------------------------------------------*/ 439 enum stmt_vec_info_type { 440 undef_vec_info_type = 0, 441 load_vec_info_type, 442 store_vec_info_type, 443 shift_vec_info_type, 444 op_vec_info_type, 445 call_vec_info_type, 446 call_simd_clone_vec_info_type, 447 assignment_vec_info_type, 448 condition_vec_info_type, 449 comparison_vec_info_type, 450 reduc_vec_info_type, 451 induc_vec_info_type, 452 type_promotion_vec_info_type, 453 type_demotion_vec_info_type, 454 type_conversion_vec_info_type, 455 loop_exit_ctrl_vec_info_type 456 }; 457 458 /* Indicates whether/how a variable is used in the scope of loop/basic 459 block. */ 460 enum vect_relevant { 461 vect_unused_in_scope = 0, 462 463 /* The def is only used outside the loop. */ 464 vect_used_only_live, 465 /* The def is in the inner loop, and the use is in the outer loop, and the 466 use is a reduction stmt. */ 467 vect_used_in_outer_by_reduction, 468 /* The def is in the inner loop, and the use is in the outer loop (and is 469 not part of reduction). */ 470 vect_used_in_outer, 471 472 /* defs that feed computations that end up (only) in a reduction. These 473 defs may be used by non-reduction stmts, but eventually, any 474 computations/values that are affected by these defs are used to compute 475 a reduction (i.e. don't get stored to memory, for example). We use this 476 to identify computations that we can change the order in which they are 477 computed. */ 478 vect_used_by_reduction, 479 480 vect_used_in_scope 481 }; 482 483 /* The type of vectorization that can be applied to the stmt: regular loop-based 484 vectorization; pure SLP - the stmt is a part of SLP instances and does not 485 have uses outside SLP instances; or hybrid SLP and loop-based - the stmt is 486 a part of SLP instance and also must be loop-based vectorized, since it has 487 uses outside SLP sequences. 488 489 In the loop context the meanings of pure and hybrid SLP are slightly 490 different. By saying that pure SLP is applied to the loop, we mean that we 491 exploit only intra-iteration parallelism in the loop; i.e., the loop can be 492 vectorized without doing any conceptual unrolling, cause we don't pack 493 together stmts from different iterations, only within a single iteration. 494 Loop hybrid SLP means that we exploit both intra-iteration and 495 inter-iteration parallelism (e.g., number of elements in the vector is 4 496 and the slp-group-size is 2, in which case we don't have enough parallelism 497 within an iteration, so we obtain the rest of the parallelism from subsequent 498 iterations by unrolling the loop by 2). */ 499 enum slp_vect_type { 500 loop_vect = 0, 501 pure_slp, 502 hybrid 503 }; 504 505 /* Describes how we're going to vectorize an individual load or store, 506 or a group of loads or stores. */ 507 enum vect_memory_access_type { 508 /* An access to an invariant address. This is used only for loads. */ 509 VMAT_INVARIANT, 510 511 /* A simple contiguous access. */ 512 VMAT_CONTIGUOUS, 513 514 /* A contiguous access that goes down in memory rather than up, 515 with no additional permutation. This is used only for stores 516 of invariants. */ 517 VMAT_CONTIGUOUS_DOWN, 518 519 /* A simple contiguous access in which the elements need to be permuted 520 after loading or before storing. Only used for loop vectorization; 521 SLP uses separate permutes. */ 522 VMAT_CONTIGUOUS_PERMUTE, 523 524 /* A simple contiguous access in which the elements need to be reversed 525 after loading or before storing. */ 526 VMAT_CONTIGUOUS_REVERSE, 527 528 /* An access that uses IFN_LOAD_LANES or IFN_STORE_LANES. */ 529 VMAT_LOAD_STORE_LANES, 530 531 /* An access in which each scalar element is loaded or stored 532 individually. */ 533 VMAT_ELEMENTWISE, 534 535 /* A hybrid of VMAT_CONTIGUOUS and VMAT_ELEMENTWISE, used for grouped 536 SLP accesses. Each unrolled iteration uses a contiguous load 537 or store for the whole group, but the groups from separate iterations 538 are combined in the same way as for VMAT_ELEMENTWISE. */ 539 VMAT_STRIDED_SLP, 540 541 /* The access uses gather loads or scatter stores. */ 542 VMAT_GATHER_SCATTER 543 }; 544 545 typedef struct data_reference *dr_p; 546 547 typedef struct _stmt_vec_info { 548 549 enum stmt_vec_info_type type; 550 551 /* Indicates whether this stmts is part of a computation whose result is 552 used outside the loop. */ 553 bool live; 554 555 /* Stmt is part of some pattern (computation idiom) */ 556 bool in_pattern_p; 557 558 /* Is this statement vectorizable or should it be skipped in (partial) 559 vectorization. */ 560 bool vectorizable; 561 562 /* The stmt to which this info struct refers to. */ 563 gimple *stmt; 564 565 /* The vec_info with respect to which STMT is vectorized. */ 566 vec_info *vinfo; 567 568 /* The vector type to be used for the LHS of this statement. */ 569 tree vectype; 570 571 /* The vectorized version of the stmt. */ 572 gimple *vectorized_stmt; 573 574 575 /** The following is relevant only for stmts that contain a non-scalar 576 data-ref (array/pointer/struct access). A GIMPLE stmt is expected to have 577 at most one such data-ref. **/ 578 579 /* Information about the data-ref (access function, etc), 580 relative to the inner-most containing loop. */ 581 struct data_reference *data_ref_info; 582 583 /* Information about the data-ref relative to this loop 584 nest (the loop that is being considered for vectorization). */ 585 tree dr_base_address; 586 tree dr_init; 587 tree dr_offset; 588 tree dr_step; 589 tree dr_aligned_to; 590 591 /* For loop PHI nodes, the base and evolution part of it. This makes sure 592 this information is still available in vect_update_ivs_after_vectorizer 593 where we may not be able to re-analyze the PHI nodes evolution as 594 peeling for the prologue loop can make it unanalyzable. The evolution 595 part is still correct after peeling, but the base may have changed from 596 the version here. */ 597 tree loop_phi_evolution_base_unchanged; 598 tree loop_phi_evolution_part; 599 600 /* Used for various bookkeeping purposes, generally holding a pointer to 601 some other stmt S that is in some way "related" to this stmt. 602 Current use of this field is: 603 If this stmt is part of a pattern (i.e. the field 'in_pattern_p' is 604 true): S is the "pattern stmt" that represents (and replaces) the 605 sequence of stmts that constitutes the pattern. Similarly, the 606 related_stmt of the "pattern stmt" points back to this stmt (which is 607 the last stmt in the original sequence of stmts that constitutes the 608 pattern). */ 609 gimple *related_stmt; 610 611 /* Used to keep a sequence of def stmts of a pattern stmt if such exists. */ 612 gimple_seq pattern_def_seq; 613 614 /* List of datarefs that are known to have the same alignment as the dataref 615 of this stmt. */ 616 vec<dr_p> same_align_refs; 617 618 /* Selected SIMD clone's function info. First vector element 619 is SIMD clone's function decl, followed by a pair of trees (base + step) 620 for linear arguments (pair of NULLs for other arguments). */ 621 vec<tree> simd_clone_info; 622 623 /* Classify the def of this stmt. */ 624 enum vect_def_type def_type; 625 626 /* Whether the stmt is SLPed, loop-based vectorized, or both. */ 627 enum slp_vect_type slp_type; 628 629 /* Interleaving and reduction chains info. */ 630 /* First element in the group. */ 631 gimple *first_element; 632 /* Pointer to the next element in the group. */ 633 gimple *next_element; 634 /* For data-refs, in case that two or more stmts share data-ref, this is the 635 pointer to the previously detected stmt with the same dr. */ 636 gimple *same_dr_stmt; 637 /* The size of the group. */ 638 unsigned int size; 639 /* For stores, number of stores from this group seen. We vectorize the last 640 one. */ 641 unsigned int store_count; 642 /* For loads only, the gap from the previous load. For consecutive loads, GAP 643 is 1. */ 644 unsigned int gap; 645 646 /* The minimum negative dependence distance this stmt participates in 647 or zero if none. */ 648 unsigned int min_neg_dist; 649 650 /* Not all stmts in the loop need to be vectorized. e.g, the increment 651 of the loop induction variable and computation of array indexes. relevant 652 indicates whether the stmt needs to be vectorized. */ 653 enum vect_relevant relevant; 654 655 /* For loads if this is a gather, for stores if this is a scatter. */ 656 bool gather_scatter_p; 657 658 /* True if this is an access with loop-invariant stride. */ 659 bool strided_p; 660 661 /* For both loads and stores. */ 662 bool simd_lane_access_p; 663 664 /* Classifies how the load or store is going to be implemented 665 for loop vectorization. */ 666 vect_memory_access_type memory_access_type; 667 668 /* For reduction loops, this is the type of reduction. */ 669 enum vect_reduction_type v_reduc_type; 670 671 /* For CONST_COND_REDUCTION, record the reduc code. */ 672 enum tree_code const_cond_reduc_code; 673 674 /* The number of scalar stmt references from active SLP instances. */ 675 unsigned int num_slp_uses; 676 } *stmt_vec_info; 677 678 /* Information about a gather/scatter call. */ 679 struct gather_scatter_info { 680 /* The FUNCTION_DECL for the built-in gather/scatter function. */ 681 tree decl; 682 683 /* The loop-invariant base value. */ 684 tree base; 685 686 /* The original scalar offset, which is a non-loop-invariant SSA_NAME. */ 687 tree offset; 688 689 /* Each offset element should be multiplied by this amount before 690 being added to the base. */ 691 int scale; 692 693 /* The definition type for the vectorized offset. */ 694 enum vect_def_type offset_dt; 695 696 /* The type of the vectorized offset. */ 697 tree offset_vectype; 698 }; 699 700 /* Access Functions. */ 701 #define STMT_VINFO_TYPE(S) (S)->type 702 #define STMT_VINFO_STMT(S) (S)->stmt 703 inline loop_vec_info 704 STMT_VINFO_LOOP_VINFO (stmt_vec_info stmt_vinfo) 705 { 706 if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (stmt_vinfo->vinfo)) 707 return loop_vinfo; 708 return NULL; 709 } 710 inline bb_vec_info 711 STMT_VINFO_BB_VINFO (stmt_vec_info stmt_vinfo) 712 { 713 if (bb_vec_info bb_vinfo = dyn_cast <bb_vec_info> (stmt_vinfo->vinfo)) 714 return bb_vinfo; 715 return NULL; 716 } 717 #define STMT_VINFO_RELEVANT(S) (S)->relevant 718 #define STMT_VINFO_LIVE_P(S) (S)->live 719 #define STMT_VINFO_VECTYPE(S) (S)->vectype 720 #define STMT_VINFO_VEC_STMT(S) (S)->vectorized_stmt 721 #define STMT_VINFO_VECTORIZABLE(S) (S)->vectorizable 722 #define STMT_VINFO_DATA_REF(S) (S)->data_ref_info 723 #define STMT_VINFO_GATHER_SCATTER_P(S) (S)->gather_scatter_p 724 #define STMT_VINFO_STRIDED_P(S) (S)->strided_p 725 #define STMT_VINFO_MEMORY_ACCESS_TYPE(S) (S)->memory_access_type 726 #define STMT_VINFO_SIMD_LANE_ACCESS_P(S) (S)->simd_lane_access_p 727 #define STMT_VINFO_VEC_REDUCTION_TYPE(S) (S)->v_reduc_type 728 #define STMT_VINFO_VEC_CONST_COND_REDUC_CODE(S) (S)->const_cond_reduc_code 729 730 #define STMT_VINFO_DR_BASE_ADDRESS(S) (S)->dr_base_address 731 #define STMT_VINFO_DR_INIT(S) (S)->dr_init 732 #define STMT_VINFO_DR_OFFSET(S) (S)->dr_offset 733 #define STMT_VINFO_DR_STEP(S) (S)->dr_step 734 #define STMT_VINFO_DR_ALIGNED_TO(S) (S)->dr_aligned_to 735 736 #define STMT_VINFO_IN_PATTERN_P(S) (S)->in_pattern_p 737 #define STMT_VINFO_RELATED_STMT(S) (S)->related_stmt 738 #define STMT_VINFO_PATTERN_DEF_SEQ(S) (S)->pattern_def_seq 739 #define STMT_VINFO_SAME_ALIGN_REFS(S) (S)->same_align_refs 740 #define STMT_VINFO_SIMD_CLONE_INFO(S) (S)->simd_clone_info 741 #define STMT_VINFO_DEF_TYPE(S) (S)->def_type 742 #define STMT_VINFO_GROUP_FIRST_ELEMENT(S) (S)->first_element 743 #define STMT_VINFO_GROUP_NEXT_ELEMENT(S) (S)->next_element 744 #define STMT_VINFO_GROUP_SIZE(S) (S)->size 745 #define STMT_VINFO_GROUP_STORE_COUNT(S) (S)->store_count 746 #define STMT_VINFO_GROUP_GAP(S) (S)->gap 747 #define STMT_VINFO_GROUP_SAME_DR_STMT(S) (S)->same_dr_stmt 748 #define STMT_VINFO_GROUPED_ACCESS(S) ((S)->first_element != NULL && (S)->data_ref_info) 749 #define STMT_VINFO_LOOP_PHI_EVOLUTION_BASE_UNCHANGED(S) (S)->loop_phi_evolution_base_unchanged 750 #define STMT_VINFO_LOOP_PHI_EVOLUTION_PART(S) (S)->loop_phi_evolution_part 751 #define STMT_VINFO_MIN_NEG_DIST(S) (S)->min_neg_dist 752 #define STMT_VINFO_NUM_SLP_USES(S) (S)->num_slp_uses 753 754 #define GROUP_FIRST_ELEMENT(S) (S)->first_element 755 #define GROUP_NEXT_ELEMENT(S) (S)->next_element 756 #define GROUP_SIZE(S) (S)->size 757 #define GROUP_STORE_COUNT(S) (S)->store_count 758 #define GROUP_GAP(S) (S)->gap 759 #define GROUP_SAME_DR_STMT(S) (S)->same_dr_stmt 760 761 #define STMT_VINFO_RELEVANT_P(S) ((S)->relevant != vect_unused_in_scope) 762 763 #define HYBRID_SLP_STMT(S) ((S)->slp_type == hybrid) 764 #define PURE_SLP_STMT(S) ((S)->slp_type == pure_slp) 765 #define STMT_SLP_TYPE(S) (S)->slp_type 766 767 struct dataref_aux { 768 int misalignment; 769 /* If true the alignment of base_decl needs to be increased. */ 770 bool base_misaligned; 771 /* If true we know the base is at least vector element alignment aligned. */ 772 bool base_element_aligned; 773 tree base_decl; 774 }; 775 776 #define DR_VECT_AUX(dr) ((dataref_aux *)(dr)->aux) 777 778 #define VECT_MAX_COST 1000 779 780 /* The maximum number of intermediate steps required in multi-step type 781 conversion. */ 782 #define MAX_INTERM_CVT_STEPS 3 783 784 /* The maximum vectorization factor supported by any target (V64QI). */ 785 #define MAX_VECTORIZATION_FACTOR 64 786 787 /* Nonzero if TYPE represents a (scalar) boolean type or type 788 in the middle-end compatible with it (unsigned precision 1 integral 789 types). Used to determine which types should be vectorized as 790 VECTOR_BOOLEAN_TYPE_P. */ 791 792 #define VECT_SCALAR_BOOLEAN_TYPE_P(TYPE) \ 793 (TREE_CODE (TYPE) == BOOLEAN_TYPE \ 794 || ((TREE_CODE (TYPE) == INTEGER_TYPE \ 795 || TREE_CODE (TYPE) == ENUMERAL_TYPE) \ 796 && TYPE_PRECISION (TYPE) == 1 \ 797 && TYPE_UNSIGNED (TYPE))) 798 799 extern vec<stmt_vec_info> stmt_vec_info_vec; 800 801 void init_stmt_vec_info_vec (void); 802 void free_stmt_vec_info_vec (void); 803 804 /* Return a stmt_vec_info corresponding to STMT. */ 805 806 static inline stmt_vec_info 807 vinfo_for_stmt (gimple *stmt) 808 { 809 unsigned int uid = gimple_uid (stmt); 810 if (uid == 0) 811 return NULL; 812 813 return stmt_vec_info_vec[uid - 1]; 814 } 815 816 /* Set vectorizer information INFO for STMT. */ 817 818 static inline void 819 set_vinfo_for_stmt (gimple *stmt, stmt_vec_info info) 820 { 821 unsigned int uid = gimple_uid (stmt); 822 if (uid == 0) 823 { 824 gcc_checking_assert (info); 825 uid = stmt_vec_info_vec.length () + 1; 826 gimple_set_uid (stmt, uid); 827 stmt_vec_info_vec.safe_push (info); 828 } 829 else 830 { 831 gcc_checking_assert (info == NULL); 832 stmt_vec_info_vec[uid - 1] = info; 833 } 834 } 835 836 /* Return TRUE if a statement represented by STMT_INFO is a part of a 837 pattern. */ 838 839 static inline bool 840 is_pattern_stmt_p (stmt_vec_info stmt_info) 841 { 842 gimple *related_stmt; 843 stmt_vec_info related_stmt_info; 844 845 related_stmt = STMT_VINFO_RELATED_STMT (stmt_info); 846 if (related_stmt 847 && (related_stmt_info = vinfo_for_stmt (related_stmt)) 848 && STMT_VINFO_IN_PATTERN_P (related_stmt_info)) 849 return true; 850 851 return false; 852 } 853 854 /* Return the later statement between STMT1 and STMT2. */ 855 856 static inline gimple * 857 get_later_stmt (gimple *stmt1, gimple *stmt2) 858 { 859 unsigned int uid1, uid2; 860 861 if (stmt1 == NULL) 862 return stmt2; 863 864 if (stmt2 == NULL) 865 return stmt1; 866 867 stmt_vec_info stmt_info1 = vinfo_for_stmt (stmt1); 868 stmt_vec_info stmt_info2 = vinfo_for_stmt (stmt2); 869 uid1 = gimple_uid (is_pattern_stmt_p (stmt_info1) 870 ? STMT_VINFO_RELATED_STMT (stmt_info1) : stmt1); 871 uid2 = gimple_uid (is_pattern_stmt_p (stmt_info2) 872 ? STMT_VINFO_RELATED_STMT (stmt_info2) : stmt2); 873 874 if (uid1 == 0 || uid2 == 0) 875 return NULL; 876 877 gcc_assert (uid1 <= stmt_vec_info_vec.length ()); 878 gcc_assert (uid2 <= stmt_vec_info_vec.length ()); 879 880 if (uid1 > uid2) 881 return stmt1; 882 else 883 return stmt2; 884 } 885 886 /* Return true if BB is a loop header. */ 887 888 static inline bool 889 is_loop_header_bb_p (basic_block bb) 890 { 891 if (bb == (bb->loop_father)->header) 892 return true; 893 gcc_checking_assert (EDGE_COUNT (bb->preds) == 1); 894 return false; 895 } 896 897 /* Return pow2 (X). */ 898 899 static inline int 900 vect_pow2 (int x) 901 { 902 int i, res = 1; 903 904 for (i = 0; i < x; i++) 905 res *= 2; 906 907 return res; 908 } 909 910 /* Alias targetm.vectorize.builtin_vectorization_cost. */ 911 912 static inline int 913 builtin_vectorization_cost (enum vect_cost_for_stmt type_of_cost, 914 tree vectype, int misalign) 915 { 916 return targetm.vectorize.builtin_vectorization_cost (type_of_cost, 917 vectype, misalign); 918 } 919 920 /* Get cost by calling cost target builtin. */ 921 922 static inline 923 int vect_get_stmt_cost (enum vect_cost_for_stmt type_of_cost) 924 { 925 return builtin_vectorization_cost (type_of_cost, NULL, 0); 926 } 927 928 /* Alias targetm.vectorize.init_cost. */ 929 930 static inline void * 931 init_cost (struct loop *loop_info) 932 { 933 return targetm.vectorize.init_cost (loop_info); 934 } 935 936 /* Alias targetm.vectorize.add_stmt_cost. */ 937 938 static inline unsigned 939 add_stmt_cost (void *data, int count, enum vect_cost_for_stmt kind, 940 stmt_vec_info stmt_info, int misalign, 941 enum vect_cost_model_location where) 942 { 943 return targetm.vectorize.add_stmt_cost (data, count, kind, 944 stmt_info, misalign, where); 945 } 946 947 /* Alias targetm.vectorize.finish_cost. */ 948 949 static inline void 950 finish_cost (void *data, unsigned *prologue_cost, 951 unsigned *body_cost, unsigned *epilogue_cost) 952 { 953 targetm.vectorize.finish_cost (data, prologue_cost, body_cost, epilogue_cost); 954 } 955 956 /* Alias targetm.vectorize.destroy_cost_data. */ 957 958 static inline void 959 destroy_cost_data (void *data) 960 { 961 targetm.vectorize.destroy_cost_data (data); 962 } 963 964 /*-----------------------------------------------------------------*/ 965 /* Info on data references alignment. */ 966 /*-----------------------------------------------------------------*/ 967 inline void 968 set_dr_misalignment (struct data_reference *dr, int val) 969 { 970 dataref_aux *data_aux = DR_VECT_AUX (dr); 971 972 if (!data_aux) 973 { 974 data_aux = XCNEW (dataref_aux); 975 dr->aux = data_aux; 976 } 977 978 data_aux->misalignment = val; 979 } 980 981 inline int 982 dr_misalignment (struct data_reference *dr) 983 { 984 return DR_VECT_AUX (dr)->misalignment; 985 } 986 987 /* Reflects actual alignment of first access in the vectorized loop, 988 taking into account peeling/versioning if applied. */ 989 #define DR_MISALIGNMENT(DR) dr_misalignment (DR) 990 #define SET_DR_MISALIGNMENT(DR, VAL) set_dr_misalignment (DR, VAL) 991 992 /* Return TRUE if the data access is aligned, and FALSE otherwise. */ 993 994 static inline bool 995 aligned_access_p (struct data_reference *data_ref_info) 996 { 997 return (DR_MISALIGNMENT (data_ref_info) == 0); 998 } 999 1000 /* Return TRUE if the alignment of the data access is known, and FALSE 1001 otherwise. */ 1002 1003 static inline bool 1004 known_alignment_for_access_p (struct data_reference *data_ref_info) 1005 { 1006 return (DR_MISALIGNMENT (data_ref_info) != -1); 1007 } 1008 1009 1010 /* Return true if the vect cost model is unlimited. */ 1011 static inline bool 1012 unlimited_cost_model (loop_p loop) 1013 { 1014 if (loop != NULL && loop->force_vectorize 1015 && flag_simd_cost_model != VECT_COST_MODEL_DEFAULT) 1016 return flag_simd_cost_model == VECT_COST_MODEL_UNLIMITED; 1017 return (flag_vect_cost_model == VECT_COST_MODEL_UNLIMITED); 1018 } 1019 1020 /* Source location */ 1021 extern source_location vect_location; 1022 1023 /*-----------------------------------------------------------------*/ 1024 /* Function prototypes. */ 1025 /*-----------------------------------------------------------------*/ 1026 1027 /* Simple loop peeling and versioning utilities for vectorizer's purposes - 1028 in tree-vect-loop-manip.c. */ 1029 extern void slpeel_make_loop_iterate_ntimes (struct loop *, tree); 1030 extern bool slpeel_can_duplicate_loop_p (const struct loop *, const_edge); 1031 struct loop *slpeel_tree_duplicate_loop_to_edge_cfg (struct loop *, 1032 struct loop *, edge); 1033 extern void vect_loop_versioning (loop_vec_info, unsigned int, bool); 1034 extern struct loop *vect_do_peeling (loop_vec_info, tree, tree, 1035 tree *, int, bool, bool); 1036 extern source_location find_loop_location (struct loop *); 1037 extern bool vect_can_advance_ivs_p (loop_vec_info); 1038 1039 /* In tree-vect-stmts.c. */ 1040 extern unsigned int current_vector_size; 1041 extern tree get_vectype_for_scalar_type (tree); 1042 extern tree get_mask_type_for_scalar_type (tree); 1043 extern tree get_same_sized_vectype (tree, tree); 1044 extern bool vect_is_simple_use (tree, vec_info *, gimple **, 1045 enum vect_def_type *); 1046 extern bool vect_is_simple_use (tree, vec_info *, gimple **, 1047 enum vect_def_type *, tree *); 1048 extern bool supportable_widening_operation (enum tree_code, gimple *, tree, 1049 tree, enum tree_code *, 1050 enum tree_code *, int *, 1051 vec<tree> *); 1052 extern bool supportable_narrowing_operation (enum tree_code, tree, tree, 1053 enum tree_code *, 1054 int *, vec<tree> *); 1055 extern stmt_vec_info new_stmt_vec_info (gimple *stmt, vec_info *); 1056 extern void free_stmt_vec_info (gimple *stmt); 1057 extern void vect_model_simple_cost (stmt_vec_info, int, enum vect_def_type *, 1058 stmt_vector_for_cost *, 1059 stmt_vector_for_cost *); 1060 extern void vect_model_store_cost (stmt_vec_info, int, vect_memory_access_type, 1061 enum vect_def_type, slp_tree, 1062 stmt_vector_for_cost *, 1063 stmt_vector_for_cost *); 1064 extern void vect_model_load_cost (stmt_vec_info, int, vect_memory_access_type, 1065 slp_tree, stmt_vector_for_cost *, 1066 stmt_vector_for_cost *); 1067 extern unsigned record_stmt_cost (stmt_vector_for_cost *, int, 1068 enum vect_cost_for_stmt, stmt_vec_info, 1069 int, enum vect_cost_model_location); 1070 extern void vect_finish_stmt_generation (gimple *, gimple *, 1071 gimple_stmt_iterator *); 1072 extern bool vect_mark_stmts_to_be_vectorized (loop_vec_info); 1073 extern tree vect_get_vec_def_for_operand_1 (gimple *, enum vect_def_type); 1074 extern tree vect_get_vec_def_for_operand (tree, gimple *, tree = NULL); 1075 extern tree vect_init_vector (gimple *, tree, tree, 1076 gimple_stmt_iterator *); 1077 extern tree vect_get_vec_def_for_stmt_copy (enum vect_def_type, tree); 1078 extern bool vect_transform_stmt (gimple *, gimple_stmt_iterator *, 1079 bool *, slp_tree, slp_instance); 1080 extern void vect_remove_stores (gimple *); 1081 extern bool vect_analyze_stmt (gimple *, bool *, slp_tree); 1082 extern bool vectorizable_condition (gimple *, gimple_stmt_iterator *, 1083 gimple **, tree, int, slp_tree); 1084 extern void vect_get_load_cost (struct data_reference *, int, bool, 1085 unsigned int *, unsigned int *, 1086 stmt_vector_for_cost *, 1087 stmt_vector_for_cost *, bool); 1088 extern void vect_get_store_cost (struct data_reference *, int, 1089 unsigned int *, stmt_vector_for_cost *); 1090 extern bool vect_supportable_shift (enum tree_code, tree); 1091 extern void vect_get_vec_defs (tree, tree, gimple *, vec<tree> *, 1092 vec<tree> *, slp_tree, int); 1093 extern tree vect_gen_perm_mask_any (tree, const unsigned char *); 1094 extern tree vect_gen_perm_mask_checked (tree, const unsigned char *); 1095 extern void optimize_mask_stores (struct loop*); 1096 1097 /* In tree-vect-data-refs.c. */ 1098 extern bool vect_can_force_dr_alignment_p (const_tree, unsigned int); 1099 extern enum dr_alignment_support vect_supportable_dr_alignment 1100 (struct data_reference *, bool); 1101 extern tree vect_get_smallest_scalar_type (gimple *, HOST_WIDE_INT *, 1102 HOST_WIDE_INT *); 1103 extern bool vect_analyze_data_ref_dependences (loop_vec_info, int *); 1104 extern bool vect_slp_analyze_instance_dependence (slp_instance); 1105 extern bool vect_enhance_data_refs_alignment (loop_vec_info); 1106 extern bool vect_analyze_data_refs_alignment (loop_vec_info); 1107 extern bool vect_verify_datarefs_alignment (loop_vec_info); 1108 extern bool vect_slp_analyze_and_verify_instance_alignment (slp_instance); 1109 extern bool vect_analyze_data_ref_accesses (vec_info *); 1110 extern bool vect_prune_runtime_alias_test_list (loop_vec_info); 1111 extern bool vect_check_gather_scatter (gimple *, loop_vec_info, 1112 gather_scatter_info *); 1113 extern bool vect_analyze_data_refs (vec_info *, int *); 1114 extern tree vect_create_data_ref_ptr (gimple *, tree, struct loop *, tree, 1115 tree *, gimple_stmt_iterator *, 1116 gimple **, bool, bool *, 1117 tree = NULL_TREE); 1118 extern tree bump_vector_ptr (tree, gimple *, gimple_stmt_iterator *, gimple *, 1119 tree); 1120 extern tree vect_create_destination_var (tree, tree); 1121 extern bool vect_grouped_store_supported (tree, unsigned HOST_WIDE_INT); 1122 extern bool vect_store_lanes_supported (tree, unsigned HOST_WIDE_INT); 1123 extern bool vect_grouped_load_supported (tree, bool, unsigned HOST_WIDE_INT); 1124 extern bool vect_load_lanes_supported (tree, unsigned HOST_WIDE_INT); 1125 extern void vect_permute_store_chain (vec<tree> ,unsigned int, gimple *, 1126 gimple_stmt_iterator *, vec<tree> *); 1127 extern tree vect_setup_realignment (gimple *, gimple_stmt_iterator *, tree *, 1128 enum dr_alignment_support, tree, 1129 struct loop **); 1130 extern void vect_transform_grouped_load (gimple *, vec<tree> , int, 1131 gimple_stmt_iterator *); 1132 extern void vect_record_grouped_load_vectors (gimple *, vec<tree> ); 1133 extern tree vect_get_new_vect_var (tree, enum vect_var_kind, const char *); 1134 extern tree vect_get_new_ssa_name (tree, enum vect_var_kind, 1135 const char * = NULL); 1136 extern tree vect_create_addr_base_for_vector_ref (gimple *, gimple_seq *, 1137 tree, struct loop *, 1138 tree = NULL_TREE); 1139 1140 /* In tree-vect-loop.c. */ 1141 /* FORNOW: Used in tree-parloops.c. */ 1142 extern void destroy_loop_vec_info (loop_vec_info, bool); 1143 extern gimple *vect_force_simple_reduction (loop_vec_info, gimple *, bool, 1144 bool *, bool); 1145 /* Drive for loop analysis stage. */ 1146 extern loop_vec_info vect_analyze_loop (struct loop *, loop_vec_info); 1147 extern tree vect_build_loop_niters (loop_vec_info); 1148 extern void vect_gen_vector_loop_niters (loop_vec_info, tree, tree *, bool); 1149 /* Drive for loop transformation stage. */ 1150 extern struct loop *vect_transform_loop (loop_vec_info); 1151 extern loop_vec_info vect_analyze_loop_form (struct loop *); 1152 extern bool vectorizable_live_operation (gimple *, gimple_stmt_iterator *, 1153 slp_tree, int, gimple **); 1154 extern bool vectorizable_reduction (gimple *, gimple_stmt_iterator *, 1155 gimple **, slp_tree); 1156 extern bool vectorizable_induction (gimple *, gimple_stmt_iterator *, gimple **); 1157 extern tree get_initial_def_for_reduction (gimple *, tree, tree *); 1158 extern int vect_min_worthwhile_factor (enum tree_code); 1159 extern int vect_get_known_peeling_cost (loop_vec_info, int, int *, 1160 stmt_vector_for_cost *, 1161 stmt_vector_for_cost *, 1162 stmt_vector_for_cost *); 1163 1164 /* In tree-vect-slp.c. */ 1165 extern void vect_free_slp_instance (slp_instance); 1166 extern bool vect_transform_slp_perm_load (slp_tree, vec<tree> , 1167 gimple_stmt_iterator *, int, 1168 slp_instance, bool, unsigned *); 1169 extern bool vect_slp_analyze_operations (vec<slp_instance> slp_instances, 1170 void *); 1171 extern bool vect_schedule_slp (vec_info *); 1172 extern bool vect_analyze_slp (vec_info *, unsigned); 1173 extern bool vect_make_slp_decision (loop_vec_info); 1174 extern void vect_detect_hybrid_slp (loop_vec_info); 1175 extern void vect_get_slp_defs (vec<tree> , slp_tree, 1176 vec<vec<tree> > *, int); 1177 extern bool vect_slp_bb (basic_block); 1178 extern gimple *vect_find_last_scalar_stmt_in_slp (slp_tree); 1179 extern bool is_simple_and_all_uses_invariant (gimple *, loop_vec_info); 1180 1181 /* In tree-vect-patterns.c. */ 1182 /* Pattern recognition functions. 1183 Additional pattern recognition functions can (and will) be added 1184 in the future. */ 1185 typedef gimple *(* vect_recog_func_ptr) (vec<gimple *> *, tree *, tree *); 1186 #define NUM_PATTERNS 14 1187 void vect_pattern_recog (vec_info *); 1188 1189 /* In tree-vectorizer.c. */ 1190 unsigned vectorize_loops (void); 1191 void vect_destroy_datarefs (vec_info *); 1192 bool vect_stmt_in_region_p (vec_info *, gimple *); 1193 void vect_free_loop_info_assumptions (struct loop *); 1194 1195 #endif /* GCC_TREE_VECTORIZER_H */ 1196