1 /* Loop Vectorization 2 Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2011, 2012 3 Free Software Foundation, Inc. 4 Contributed by Dorit Naishlos <dorit@il.ibm.com> and 5 Ira Rosen <irar@il.ibm.com> 6 7 This file is part of GCC. 8 9 GCC is free software; you can redistribute it and/or modify it under 10 the terms of the GNU General Public License as published by the Free 11 Software Foundation; either version 3, or (at your option) any later 12 version. 13 14 GCC is distributed in the hope that it will be useful, but WITHOUT ANY 15 WARRANTY; without even the implied warranty of MERCHANTABILITY or 16 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 17 for more details. 18 19 You should have received a copy of the GNU General Public License 20 along with GCC; see the file COPYING3. If not see 21 <http://www.gnu.org/licenses/>. */ 22 23 #include "config.h" 24 #include "system.h" 25 #include "coretypes.h" 26 #include "tm.h" 27 #include "ggc.h" 28 #include "tree.h" 29 #include "basic-block.h" 30 #include "tree-pretty-print.h" 31 #include "gimple-pretty-print.h" 32 #include "tree-flow.h" 33 #include "tree-dump.h" 34 #include "cfgloop.h" 35 #include "cfglayout.h" 36 #include "expr.h" 37 #include "recog.h" 38 #include "optabs.h" 39 #include "params.h" 40 #include "diagnostic-core.h" 41 #include "tree-chrec.h" 42 #include "tree-scalar-evolution.h" 43 #include "tree-vectorizer.h" 44 #include "target.h" 45 46 /* Loop Vectorization Pass. 47 48 This pass tries to vectorize loops. 49 50 For example, the vectorizer transforms the following simple loop: 51 52 short a[N]; short b[N]; short c[N]; int i; 53 54 for (i=0; i<N; i++){ 55 a[i] = b[i] + c[i]; 56 } 57 58 as if it was manually vectorized by rewriting the source code into: 59 60 typedef int __attribute__((mode(V8HI))) v8hi; 61 short a[N]; short b[N]; short c[N]; int i; 62 v8hi *pa = (v8hi*)a, *pb = (v8hi*)b, *pc = (v8hi*)c; 63 v8hi va, vb, vc; 64 65 for (i=0; i<N/8; i++){ 66 vb = pb[i]; 67 vc = pc[i]; 68 va = vb + vc; 69 pa[i] = va; 70 } 71 72 The main entry to this pass is vectorize_loops(), in which 73 the vectorizer applies a set of analyses on a given set of loops, 74 followed by the actual vectorization transformation for the loops that 75 had successfully passed the analysis phase. 76 Throughout this pass we make a distinction between two types of 77 data: scalars (which are represented by SSA_NAMES), and memory references 78 ("data-refs"). These two types of data require different handling both 79 during analysis and transformation. The types of data-refs that the 80 vectorizer currently supports are ARRAY_REFS which base is an array DECL 81 (not a pointer), and INDIRECT_REFS through pointers; both array and pointer 82 accesses are required to have a simple (consecutive) access pattern. 83 84 Analysis phase: 85 =============== 86 The driver for the analysis phase is vect_analyze_loop(). 87 It applies a set of analyses, some of which rely on the scalar evolution 88 analyzer (scev) developed by Sebastian Pop. 89 90 During the analysis phase the vectorizer records some information 91 per stmt in a "stmt_vec_info" struct which is attached to each stmt in the 92 loop, as well as general information about the loop as a whole, which is 93 recorded in a "loop_vec_info" struct attached to each loop. 94 95 Transformation phase: 96 ===================== 97 The loop transformation phase scans all the stmts in the loop, and 98 creates a vector stmt (or a sequence of stmts) for each scalar stmt S in 99 the loop that needs to be vectorized. It inserts the vector code sequence 100 just before the scalar stmt S, and records a pointer to the vector code 101 in STMT_VINFO_VEC_STMT (stmt_info) (stmt_info is the stmt_vec_info struct 102 attached to S). This pointer will be used for the vectorization of following 103 stmts which use the def of stmt S. Stmt S is removed if it writes to memory; 104 otherwise, we rely on dead code elimination for removing it. 105 106 For example, say stmt S1 was vectorized into stmt VS1: 107 108 VS1: vb = px[i]; 109 S1: b = x[i]; STMT_VINFO_VEC_STMT (stmt_info (S1)) = VS1 110 S2: a = b; 111 112 To vectorize stmt S2, the vectorizer first finds the stmt that defines 113 the operand 'b' (S1), and gets the relevant vector def 'vb' from the 114 vector stmt VS1 pointed to by STMT_VINFO_VEC_STMT (stmt_info (S1)). The 115 resulting sequence would be: 116 117 VS1: vb = px[i]; 118 S1: b = x[i]; STMT_VINFO_VEC_STMT (stmt_info (S1)) = VS1 119 VS2: va = vb; 120 S2: a = b; STMT_VINFO_VEC_STMT (stmt_info (S2)) = VS2 121 122 Operands that are not SSA_NAMEs, are data-refs that appear in 123 load/store operations (like 'x[i]' in S1), and are handled differently. 124 125 Target modeling: 126 ================= 127 Currently the only target specific information that is used is the 128 size of the vector (in bytes) - "TARGET_VECTORIZE_UNITS_PER_SIMD_WORD". 129 Targets that can support different sizes of vectors, for now will need 130 to specify one value for "TARGET_VECTORIZE_UNITS_PER_SIMD_WORD". More 131 flexibility will be added in the future. 132 133 Since we only vectorize operations which vector form can be 134 expressed using existing tree codes, to verify that an operation is 135 supported, the vectorizer checks the relevant optab at the relevant 136 machine_mode (e.g, optab_handler (add_optab, V8HImode)). If 137 the value found is CODE_FOR_nothing, then there's no target support, and 138 we can't vectorize the stmt. 139 140 For additional information on this project see: 141 http://gcc.gnu.org/projects/tree-ssa/vectorization.html 142 */ 143 144 /* Function vect_determine_vectorization_factor 145 146 Determine the vectorization factor (VF). VF is the number of data elements 147 that are operated upon in parallel in a single iteration of the vectorized 148 loop. For example, when vectorizing a loop that operates on 4byte elements, 149 on a target with vector size (VS) 16byte, the VF is set to 4, since 4 150 elements can fit in a single vector register. 151 152 We currently support vectorization of loops in which all types operated upon 153 are of the same size. Therefore this function currently sets VF according to 154 the size of the types operated upon, and fails if there are multiple sizes 155 in the loop. 156 157 VF is also the factor by which the loop iterations are strip-mined, e.g.: 158 original loop: 159 for (i=0; i<N; i++){ 160 a[i] = b[i] + c[i]; 161 } 162 163 vectorized loop: 164 for (i=0; i<N; i+=VF){ 165 a[i:VF] = b[i:VF] + c[i:VF]; 166 } 167 */ 168 169 static bool 170 vect_determine_vectorization_factor (loop_vec_info loop_vinfo) 171 { 172 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); 173 basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); 174 int nbbs = loop->num_nodes; 175 gimple_stmt_iterator si; 176 unsigned int vectorization_factor = 0; 177 tree scalar_type; 178 gimple phi; 179 tree vectype; 180 unsigned int nunits; 181 stmt_vec_info stmt_info; 182 int i; 183 HOST_WIDE_INT dummy; 184 gimple stmt, pattern_stmt = NULL; 185 gimple_seq pattern_def_seq = NULL; 186 gimple_stmt_iterator pattern_def_si = gsi_start (NULL); 187 bool analyze_pattern_stmt = false; 188 189 if (vect_print_dump_info (REPORT_DETAILS)) 190 fprintf (vect_dump, "=== vect_determine_vectorization_factor ==="); 191 192 for (i = 0; i < nbbs; i++) 193 { 194 basic_block bb = bbs[i]; 195 196 for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) 197 { 198 phi = gsi_stmt (si); 199 stmt_info = vinfo_for_stmt (phi); 200 if (vect_print_dump_info (REPORT_DETAILS)) 201 { 202 fprintf (vect_dump, "==> examining phi: "); 203 print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM); 204 } 205 206 gcc_assert (stmt_info); 207 208 if (STMT_VINFO_RELEVANT_P (stmt_info)) 209 { 210 gcc_assert (!STMT_VINFO_VECTYPE (stmt_info)); 211 scalar_type = TREE_TYPE (PHI_RESULT (phi)); 212 213 if (vect_print_dump_info (REPORT_DETAILS)) 214 { 215 fprintf (vect_dump, "get vectype for scalar type: "); 216 print_generic_expr (vect_dump, scalar_type, TDF_SLIM); 217 } 218 219 vectype = get_vectype_for_scalar_type (scalar_type); 220 if (!vectype) 221 { 222 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) 223 { 224 fprintf (vect_dump, 225 "not vectorized: unsupported data-type "); 226 print_generic_expr (vect_dump, scalar_type, TDF_SLIM); 227 } 228 return false; 229 } 230 STMT_VINFO_VECTYPE (stmt_info) = vectype; 231 232 if (vect_print_dump_info (REPORT_DETAILS)) 233 { 234 fprintf (vect_dump, "vectype: "); 235 print_generic_expr (vect_dump, vectype, TDF_SLIM); 236 } 237 238 nunits = TYPE_VECTOR_SUBPARTS (vectype); 239 if (vect_print_dump_info (REPORT_DETAILS)) 240 fprintf (vect_dump, "nunits = %d", nunits); 241 242 if (!vectorization_factor 243 || (nunits > vectorization_factor)) 244 vectorization_factor = nunits; 245 } 246 } 247 248 for (si = gsi_start_bb (bb); !gsi_end_p (si) || analyze_pattern_stmt;) 249 { 250 tree vf_vectype; 251 252 if (analyze_pattern_stmt) 253 stmt = pattern_stmt; 254 else 255 stmt = gsi_stmt (si); 256 257 stmt_info = vinfo_for_stmt (stmt); 258 259 if (vect_print_dump_info (REPORT_DETAILS)) 260 { 261 fprintf (vect_dump, "==> examining statement: "); 262 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); 263 } 264 265 gcc_assert (stmt_info); 266 267 /* Skip stmts which do not need to be vectorized. */ 268 if (!STMT_VINFO_RELEVANT_P (stmt_info) 269 && !STMT_VINFO_LIVE_P (stmt_info)) 270 { 271 if (STMT_VINFO_IN_PATTERN_P (stmt_info) 272 && (pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info)) 273 && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) 274 || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) 275 { 276 stmt = pattern_stmt; 277 stmt_info = vinfo_for_stmt (pattern_stmt); 278 if (vect_print_dump_info (REPORT_DETAILS)) 279 { 280 fprintf (vect_dump, "==> examining pattern statement: "); 281 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); 282 } 283 } 284 else 285 { 286 if (vect_print_dump_info (REPORT_DETAILS)) 287 fprintf (vect_dump, "skip."); 288 gsi_next (&si); 289 continue; 290 } 291 } 292 else if (STMT_VINFO_IN_PATTERN_P (stmt_info) 293 && (pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info)) 294 && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) 295 || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) 296 analyze_pattern_stmt = true; 297 298 /* If a pattern statement has def stmts, analyze them too. */ 299 if (is_pattern_stmt_p (stmt_info)) 300 { 301 if (pattern_def_seq == NULL) 302 { 303 pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info); 304 pattern_def_si = gsi_start (pattern_def_seq); 305 } 306 else if (!gsi_end_p (pattern_def_si)) 307 gsi_next (&pattern_def_si); 308 if (pattern_def_seq != NULL) 309 { 310 gimple pattern_def_stmt = NULL; 311 stmt_vec_info pattern_def_stmt_info = NULL; 312 313 while (!gsi_end_p (pattern_def_si)) 314 { 315 pattern_def_stmt = gsi_stmt (pattern_def_si); 316 pattern_def_stmt_info 317 = vinfo_for_stmt (pattern_def_stmt); 318 if (STMT_VINFO_RELEVANT_P (pattern_def_stmt_info) 319 || STMT_VINFO_LIVE_P (pattern_def_stmt_info)) 320 break; 321 gsi_next (&pattern_def_si); 322 } 323 324 if (!gsi_end_p (pattern_def_si)) 325 { 326 if (vect_print_dump_info (REPORT_DETAILS)) 327 { 328 fprintf (vect_dump, 329 "==> examining pattern def stmt: "); 330 print_gimple_stmt (vect_dump, pattern_def_stmt, 0, 331 TDF_SLIM); 332 } 333 334 stmt = pattern_def_stmt; 335 stmt_info = pattern_def_stmt_info; 336 } 337 else 338 { 339 pattern_def_si = gsi_start (NULL); 340 analyze_pattern_stmt = false; 341 } 342 } 343 else 344 analyze_pattern_stmt = false; 345 } 346 347 if (gimple_get_lhs (stmt) == NULL_TREE) 348 { 349 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) 350 { 351 fprintf (vect_dump, "not vectorized: irregular stmt."); 352 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); 353 } 354 return false; 355 } 356 357 if (VECTOR_MODE_P (TYPE_MODE (gimple_expr_type (stmt)))) 358 { 359 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) 360 { 361 fprintf (vect_dump, "not vectorized: vector stmt in loop:"); 362 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); 363 } 364 return false; 365 } 366 367 if (STMT_VINFO_VECTYPE (stmt_info)) 368 { 369 /* The only case when a vectype had been already set is for stmts 370 that contain a dataref, or for "pattern-stmts" (stmts 371 generated by the vectorizer to represent/replace a certain 372 idiom). */ 373 gcc_assert (STMT_VINFO_DATA_REF (stmt_info) 374 || is_pattern_stmt_p (stmt_info) 375 || !gsi_end_p (pattern_def_si)); 376 vectype = STMT_VINFO_VECTYPE (stmt_info); 377 } 378 else 379 { 380 gcc_assert (!STMT_VINFO_DATA_REF (stmt_info)); 381 scalar_type = TREE_TYPE (gimple_get_lhs (stmt)); 382 if (vect_print_dump_info (REPORT_DETAILS)) 383 { 384 fprintf (vect_dump, "get vectype for scalar type: "); 385 print_generic_expr (vect_dump, scalar_type, TDF_SLIM); 386 } 387 vectype = get_vectype_for_scalar_type (scalar_type); 388 if (!vectype) 389 { 390 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) 391 { 392 fprintf (vect_dump, 393 "not vectorized: unsupported data-type "); 394 print_generic_expr (vect_dump, scalar_type, TDF_SLIM); 395 } 396 return false; 397 } 398 399 STMT_VINFO_VECTYPE (stmt_info) = vectype; 400 } 401 402 /* The vectorization factor is according to the smallest 403 scalar type (or the largest vector size, but we only 404 support one vector size per loop). */ 405 scalar_type = vect_get_smallest_scalar_type (stmt, &dummy, 406 &dummy); 407 if (vect_print_dump_info (REPORT_DETAILS)) 408 { 409 fprintf (vect_dump, "get vectype for scalar type: "); 410 print_generic_expr (vect_dump, scalar_type, TDF_SLIM); 411 } 412 vf_vectype = get_vectype_for_scalar_type (scalar_type); 413 if (!vf_vectype) 414 { 415 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) 416 { 417 fprintf (vect_dump, 418 "not vectorized: unsupported data-type "); 419 print_generic_expr (vect_dump, scalar_type, TDF_SLIM); 420 } 421 return false; 422 } 423 424 if ((GET_MODE_SIZE (TYPE_MODE (vectype)) 425 != GET_MODE_SIZE (TYPE_MODE (vf_vectype)))) 426 { 427 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) 428 { 429 fprintf (vect_dump, 430 "not vectorized: different sized vector " 431 "types in statement, "); 432 print_generic_expr (vect_dump, vectype, TDF_SLIM); 433 fprintf (vect_dump, " and "); 434 print_generic_expr (vect_dump, vf_vectype, TDF_SLIM); 435 } 436 return false; 437 } 438 439 if (vect_print_dump_info (REPORT_DETAILS)) 440 { 441 fprintf (vect_dump, "vectype: "); 442 print_generic_expr (vect_dump, vf_vectype, TDF_SLIM); 443 } 444 445 nunits = TYPE_VECTOR_SUBPARTS (vf_vectype); 446 if (vect_print_dump_info (REPORT_DETAILS)) 447 fprintf (vect_dump, "nunits = %d", nunits); 448 449 if (!vectorization_factor 450 || (nunits > vectorization_factor)) 451 vectorization_factor = nunits; 452 453 if (!analyze_pattern_stmt && gsi_end_p (pattern_def_si)) 454 { 455 pattern_def_seq = NULL; 456 gsi_next (&si); 457 } 458 } 459 } 460 461 /* TODO: Analyze cost. Decide if worth while to vectorize. */ 462 if (vect_print_dump_info (REPORT_DETAILS)) 463 fprintf (vect_dump, "vectorization factor = %d", vectorization_factor); 464 if (vectorization_factor <= 1) 465 { 466 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) 467 fprintf (vect_dump, "not vectorized: unsupported data-type"); 468 return false; 469 } 470 LOOP_VINFO_VECT_FACTOR (loop_vinfo) = vectorization_factor; 471 472 return true; 473 } 474 475 476 /* Function vect_is_simple_iv_evolution. 477 478 FORNOW: A simple evolution of an induction variables in the loop is 479 considered a polynomial evolution with constant step. */ 480 481 static bool 482 vect_is_simple_iv_evolution (unsigned loop_nb, tree access_fn, tree * init, 483 tree * step) 484 { 485 tree init_expr; 486 tree step_expr; 487 tree evolution_part = evolution_part_in_loop_num (access_fn, loop_nb); 488 489 /* When there is no evolution in this loop, the evolution function 490 is not "simple". */ 491 if (evolution_part == NULL_TREE) 492 return false; 493 494 /* When the evolution is a polynomial of degree >= 2 495 the evolution function is not "simple". */ 496 if (tree_is_chrec (evolution_part)) 497 return false; 498 499 step_expr = evolution_part; 500 init_expr = unshare_expr (initial_condition_in_loop_num (access_fn, loop_nb)); 501 502 if (vect_print_dump_info (REPORT_DETAILS)) 503 { 504 fprintf (vect_dump, "step: "); 505 print_generic_expr (vect_dump, step_expr, TDF_SLIM); 506 fprintf (vect_dump, ", init: "); 507 print_generic_expr (vect_dump, init_expr, TDF_SLIM); 508 } 509 510 *init = init_expr; 511 *step = step_expr; 512 513 if (TREE_CODE (step_expr) != INTEGER_CST) 514 { 515 if (vect_print_dump_info (REPORT_DETAILS)) 516 fprintf (vect_dump, "step unknown."); 517 return false; 518 } 519 520 return true; 521 } 522 523 /* Function vect_analyze_scalar_cycles_1. 524 525 Examine the cross iteration def-use cycles of scalar variables 526 in LOOP. LOOP_VINFO represents the loop that is now being 527 considered for vectorization (can be LOOP, or an outer-loop 528 enclosing LOOP). */ 529 530 static void 531 vect_analyze_scalar_cycles_1 (loop_vec_info loop_vinfo, struct loop *loop) 532 { 533 basic_block bb = loop->header; 534 tree dumy; 535 VEC(gimple,heap) *worklist = VEC_alloc (gimple, heap, 64); 536 gimple_stmt_iterator gsi; 537 bool double_reduc; 538 539 if (vect_print_dump_info (REPORT_DETAILS)) 540 fprintf (vect_dump, "=== vect_analyze_scalar_cycles ==="); 541 542 /* First - identify all inductions. Reduction detection assumes that all the 543 inductions have been identified, therefore, this order must not be 544 changed. */ 545 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) 546 { 547 gimple phi = gsi_stmt (gsi); 548 tree access_fn = NULL; 549 tree def = PHI_RESULT (phi); 550 stmt_vec_info stmt_vinfo = vinfo_for_stmt (phi); 551 552 if (vect_print_dump_info (REPORT_DETAILS)) 553 { 554 fprintf (vect_dump, "Analyze phi: "); 555 print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM); 556 } 557 558 /* Skip virtual phi's. The data dependences that are associated with 559 virtual defs/uses (i.e., memory accesses) are analyzed elsewhere. */ 560 if (!is_gimple_reg (SSA_NAME_VAR (def))) 561 continue; 562 563 STMT_VINFO_DEF_TYPE (stmt_vinfo) = vect_unknown_def_type; 564 565 /* Analyze the evolution function. */ 566 access_fn = analyze_scalar_evolution (loop, def); 567 if (access_fn) 568 { 569 STRIP_NOPS (access_fn); 570 if (vect_print_dump_info (REPORT_DETAILS)) 571 { 572 fprintf (vect_dump, "Access function of PHI: "); 573 print_generic_expr (vect_dump, access_fn, TDF_SLIM); 574 } 575 STMT_VINFO_LOOP_PHI_EVOLUTION_PART (stmt_vinfo) 576 = evolution_part_in_loop_num (access_fn, loop->num); 577 } 578 579 if (!access_fn 580 || !vect_is_simple_iv_evolution (loop->num, access_fn, &dumy, &dumy)) 581 { 582 VEC_safe_push (gimple, heap, worklist, phi); 583 continue; 584 } 585 586 gcc_assert (STMT_VINFO_LOOP_PHI_EVOLUTION_PART (stmt_vinfo) != NULL_TREE); 587 588 if (vect_print_dump_info (REPORT_DETAILS)) 589 fprintf (vect_dump, "Detected induction."); 590 STMT_VINFO_DEF_TYPE (stmt_vinfo) = vect_induction_def; 591 } 592 593 594 /* Second - identify all reductions and nested cycles. */ 595 while (VEC_length (gimple, worklist) > 0) 596 { 597 gimple phi = VEC_pop (gimple, worklist); 598 tree def = PHI_RESULT (phi); 599 stmt_vec_info stmt_vinfo = vinfo_for_stmt (phi); 600 gimple reduc_stmt; 601 bool nested_cycle; 602 603 if (vect_print_dump_info (REPORT_DETAILS)) 604 { 605 fprintf (vect_dump, "Analyze phi: "); 606 print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM); 607 } 608 609 gcc_assert (is_gimple_reg (SSA_NAME_VAR (def))); 610 gcc_assert (STMT_VINFO_DEF_TYPE (stmt_vinfo) == vect_unknown_def_type); 611 612 nested_cycle = (loop != LOOP_VINFO_LOOP (loop_vinfo)); 613 reduc_stmt = vect_force_simple_reduction (loop_vinfo, phi, !nested_cycle, 614 &double_reduc); 615 if (reduc_stmt) 616 { 617 if (double_reduc) 618 { 619 if (vect_print_dump_info (REPORT_DETAILS)) 620 fprintf (vect_dump, "Detected double reduction."); 621 622 STMT_VINFO_DEF_TYPE (stmt_vinfo) = vect_double_reduction_def; 623 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (reduc_stmt)) = 624 vect_double_reduction_def; 625 } 626 else 627 { 628 if (nested_cycle) 629 { 630 if (vect_print_dump_info (REPORT_DETAILS)) 631 fprintf (vect_dump, "Detected vectorizable nested cycle."); 632 633 STMT_VINFO_DEF_TYPE (stmt_vinfo) = vect_nested_cycle; 634 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (reduc_stmt)) = 635 vect_nested_cycle; 636 } 637 else 638 { 639 if (vect_print_dump_info (REPORT_DETAILS)) 640 fprintf (vect_dump, "Detected reduction."); 641 642 STMT_VINFO_DEF_TYPE (stmt_vinfo) = vect_reduction_def; 643 STMT_VINFO_DEF_TYPE (vinfo_for_stmt (reduc_stmt)) = 644 vect_reduction_def; 645 /* Store the reduction cycles for possible vectorization in 646 loop-aware SLP. */ 647 VEC_safe_push (gimple, heap, 648 LOOP_VINFO_REDUCTIONS (loop_vinfo), 649 reduc_stmt); 650 } 651 } 652 } 653 else 654 if (vect_print_dump_info (REPORT_DETAILS)) 655 fprintf (vect_dump, "Unknown def-use cycle pattern."); 656 } 657 658 VEC_free (gimple, heap, worklist); 659 } 660 661 662 /* Function vect_analyze_scalar_cycles. 663 664 Examine the cross iteration def-use cycles of scalar variables, by 665 analyzing the loop-header PHIs of scalar variables. Classify each 666 cycle as one of the following: invariant, induction, reduction, unknown. 667 We do that for the loop represented by LOOP_VINFO, and also to its 668 inner-loop, if exists. 669 Examples for scalar cycles: 670 671 Example1: reduction: 672 673 loop1: 674 for (i=0; i<N; i++) 675 sum += a[i]; 676 677 Example2: induction: 678 679 loop2: 680 for (i=0; i<N; i++) 681 a[i] = i; */ 682 683 static void 684 vect_analyze_scalar_cycles (loop_vec_info loop_vinfo) 685 { 686 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); 687 688 vect_analyze_scalar_cycles_1 (loop_vinfo, loop); 689 690 /* When vectorizing an outer-loop, the inner-loop is executed sequentially. 691 Reductions in such inner-loop therefore have different properties than 692 the reductions in the nest that gets vectorized: 693 1. When vectorized, they are executed in the same order as in the original 694 scalar loop, so we can't change the order of computation when 695 vectorizing them. 696 2. FIXME: Inner-loop reductions can be used in the inner-loop, so the 697 current checks are too strict. */ 698 699 if (loop->inner) 700 vect_analyze_scalar_cycles_1 (loop_vinfo, loop->inner); 701 } 702 703 /* Function vect_get_loop_niters. 704 705 Determine how many iterations the loop is executed. 706 If an expression that represents the number of iterations 707 can be constructed, place it in NUMBER_OF_ITERATIONS. 708 Return the loop exit condition. */ 709 710 static gimple 711 vect_get_loop_niters (struct loop *loop, tree *number_of_iterations) 712 { 713 tree niters; 714 715 if (vect_print_dump_info (REPORT_DETAILS)) 716 fprintf (vect_dump, "=== get_loop_niters ==="); 717 718 niters = number_of_exit_cond_executions (loop); 719 720 if (niters != NULL_TREE 721 && niters != chrec_dont_know) 722 { 723 *number_of_iterations = niters; 724 725 if (vect_print_dump_info (REPORT_DETAILS)) 726 { 727 fprintf (vect_dump, "==> get_loop_niters:" ); 728 print_generic_expr (vect_dump, *number_of_iterations, TDF_SLIM); 729 } 730 } 731 732 return get_loop_exit_condition (loop); 733 } 734 735 736 /* Function bb_in_loop_p 737 738 Used as predicate for dfs order traversal of the loop bbs. */ 739 740 static bool 741 bb_in_loop_p (const_basic_block bb, const void *data) 742 { 743 const struct loop *const loop = (const struct loop *)data; 744 if (flow_bb_inside_loop_p (loop, bb)) 745 return true; 746 return false; 747 } 748 749 750 /* Function new_loop_vec_info. 751 752 Create and initialize a new loop_vec_info struct for LOOP, as well as 753 stmt_vec_info structs for all the stmts in LOOP. */ 754 755 static loop_vec_info 756 new_loop_vec_info (struct loop *loop) 757 { 758 loop_vec_info res; 759 basic_block *bbs; 760 gimple_stmt_iterator si; 761 unsigned int i, nbbs; 762 763 res = (loop_vec_info) xcalloc (1, sizeof (struct _loop_vec_info)); 764 LOOP_VINFO_LOOP (res) = loop; 765 766 bbs = get_loop_body (loop); 767 768 /* Create/Update stmt_info for all stmts in the loop. */ 769 for (i = 0; i < loop->num_nodes; i++) 770 { 771 basic_block bb = bbs[i]; 772 773 /* BBs in a nested inner-loop will have been already processed (because 774 we will have called vect_analyze_loop_form for any nested inner-loop). 775 Therefore, for stmts in an inner-loop we just want to update the 776 STMT_VINFO_LOOP_VINFO field of their stmt_info to point to the new 777 loop_info of the outer-loop we are currently considering to vectorize 778 (instead of the loop_info of the inner-loop). 779 For stmts in other BBs we need to create a stmt_info from scratch. */ 780 if (bb->loop_father != loop) 781 { 782 /* Inner-loop bb. */ 783 gcc_assert (loop->inner && bb->loop_father == loop->inner); 784 for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) 785 { 786 gimple phi = gsi_stmt (si); 787 stmt_vec_info stmt_info = vinfo_for_stmt (phi); 788 loop_vec_info inner_loop_vinfo = 789 STMT_VINFO_LOOP_VINFO (stmt_info); 790 gcc_assert (loop->inner == LOOP_VINFO_LOOP (inner_loop_vinfo)); 791 STMT_VINFO_LOOP_VINFO (stmt_info) = res; 792 } 793 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) 794 { 795 gimple stmt = gsi_stmt (si); 796 stmt_vec_info stmt_info = vinfo_for_stmt (stmt); 797 loop_vec_info inner_loop_vinfo = 798 STMT_VINFO_LOOP_VINFO (stmt_info); 799 gcc_assert (loop->inner == LOOP_VINFO_LOOP (inner_loop_vinfo)); 800 STMT_VINFO_LOOP_VINFO (stmt_info) = res; 801 } 802 } 803 else 804 { 805 /* bb in current nest. */ 806 for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) 807 { 808 gimple phi = gsi_stmt (si); 809 gimple_set_uid (phi, 0); 810 set_vinfo_for_stmt (phi, new_stmt_vec_info (phi, res, NULL)); 811 } 812 813 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) 814 { 815 gimple stmt = gsi_stmt (si); 816 gimple_set_uid (stmt, 0); 817 set_vinfo_for_stmt (stmt, new_stmt_vec_info (stmt, res, NULL)); 818 } 819 } 820 } 821 822 /* CHECKME: We want to visit all BBs before their successors (except for 823 latch blocks, for which this assertion wouldn't hold). In the simple 824 case of the loop forms we allow, a dfs order of the BBs would the same 825 as reversed postorder traversal, so we are safe. */ 826 827 free (bbs); 828 bbs = XCNEWVEC (basic_block, loop->num_nodes); 829 nbbs = dfs_enumerate_from (loop->header, 0, bb_in_loop_p, 830 bbs, loop->num_nodes, loop); 831 gcc_assert (nbbs == loop->num_nodes); 832 833 LOOP_VINFO_BBS (res) = bbs; 834 LOOP_VINFO_NITERS (res) = NULL; 835 LOOP_VINFO_NITERS_UNCHANGED (res) = NULL; 836 LOOP_VINFO_COST_MODEL_MIN_ITERS (res) = 0; 837 LOOP_VINFO_VECTORIZABLE_P (res) = 0; 838 LOOP_PEELING_FOR_ALIGNMENT (res) = 0; 839 LOOP_VINFO_VECT_FACTOR (res) = 0; 840 LOOP_VINFO_LOOP_NEST (res) = VEC_alloc (loop_p, heap, 3); 841 LOOP_VINFO_DATAREFS (res) = VEC_alloc (data_reference_p, heap, 10); 842 LOOP_VINFO_DDRS (res) = VEC_alloc (ddr_p, heap, 10 * 10); 843 LOOP_VINFO_UNALIGNED_DR (res) = NULL; 844 LOOP_VINFO_MAY_MISALIGN_STMTS (res) = 845 VEC_alloc (gimple, heap, 846 PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIGNMENT_CHECKS)); 847 LOOP_VINFO_MAY_ALIAS_DDRS (res) = 848 VEC_alloc (ddr_p, heap, 849 PARAM_VALUE (PARAM_VECT_MAX_VERSION_FOR_ALIAS_CHECKS)); 850 LOOP_VINFO_STRIDED_STORES (res) = VEC_alloc (gimple, heap, 10); 851 LOOP_VINFO_REDUCTIONS (res) = VEC_alloc (gimple, heap, 10); 852 LOOP_VINFO_REDUCTION_CHAINS (res) = VEC_alloc (gimple, heap, 10); 853 LOOP_VINFO_SLP_INSTANCES (res) = VEC_alloc (slp_instance, heap, 10); 854 LOOP_VINFO_SLP_UNROLLING_FACTOR (res) = 1; 855 LOOP_VINFO_PEELING_HTAB (res) = NULL; 856 LOOP_VINFO_PEELING_FOR_GAPS (res) = false; 857 858 return res; 859 } 860 861 862 /* Function destroy_loop_vec_info. 863 864 Free LOOP_VINFO struct, as well as all the stmt_vec_info structs of all the 865 stmts in the loop. */ 866 867 void 868 destroy_loop_vec_info (loop_vec_info loop_vinfo, bool clean_stmts) 869 { 870 struct loop *loop; 871 basic_block *bbs; 872 int nbbs; 873 gimple_stmt_iterator si; 874 int j; 875 VEC (slp_instance, heap) *slp_instances; 876 slp_instance instance; 877 878 if (!loop_vinfo) 879 return; 880 881 loop = LOOP_VINFO_LOOP (loop_vinfo); 882 883 bbs = LOOP_VINFO_BBS (loop_vinfo); 884 nbbs = loop->num_nodes; 885 886 if (!clean_stmts) 887 { 888 free (LOOP_VINFO_BBS (loop_vinfo)); 889 free_data_refs (LOOP_VINFO_DATAREFS (loop_vinfo)); 890 free_dependence_relations (LOOP_VINFO_DDRS (loop_vinfo)); 891 VEC_free (loop_p, heap, LOOP_VINFO_LOOP_NEST (loop_vinfo)); 892 VEC_free (gimple, heap, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo)); 893 VEC_free (ddr_p, heap, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo)); 894 895 free (loop_vinfo); 896 loop->aux = NULL; 897 return; 898 } 899 900 for (j = 0; j < nbbs; j++) 901 { 902 basic_block bb = bbs[j]; 903 for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) 904 free_stmt_vec_info (gsi_stmt (si)); 905 906 for (si = gsi_start_bb (bb); !gsi_end_p (si); ) 907 { 908 gimple stmt = gsi_stmt (si); 909 /* Free stmt_vec_info. */ 910 free_stmt_vec_info (stmt); 911 gsi_next (&si); 912 } 913 } 914 915 free (LOOP_VINFO_BBS (loop_vinfo)); 916 free_data_refs (LOOP_VINFO_DATAREFS (loop_vinfo)); 917 free_dependence_relations (LOOP_VINFO_DDRS (loop_vinfo)); 918 VEC_free (loop_p, heap, LOOP_VINFO_LOOP_NEST (loop_vinfo)); 919 VEC_free (gimple, heap, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo)); 920 VEC_free (ddr_p, heap, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo)); 921 slp_instances = LOOP_VINFO_SLP_INSTANCES (loop_vinfo); 922 FOR_EACH_VEC_ELT (slp_instance, slp_instances, j, instance) 923 vect_free_slp_instance (instance); 924 925 VEC_free (slp_instance, heap, LOOP_VINFO_SLP_INSTANCES (loop_vinfo)); 926 VEC_free (gimple, heap, LOOP_VINFO_STRIDED_STORES (loop_vinfo)); 927 VEC_free (gimple, heap, LOOP_VINFO_REDUCTIONS (loop_vinfo)); 928 VEC_free (gimple, heap, LOOP_VINFO_REDUCTION_CHAINS (loop_vinfo)); 929 930 if (LOOP_VINFO_PEELING_HTAB (loop_vinfo)) 931 htab_delete (LOOP_VINFO_PEELING_HTAB (loop_vinfo)); 932 933 free (loop_vinfo); 934 loop->aux = NULL; 935 } 936 937 938 /* Function vect_analyze_loop_1. 939 940 Apply a set of analyses on LOOP, and create a loop_vec_info struct 941 for it. The different analyses will record information in the 942 loop_vec_info struct. This is a subset of the analyses applied in 943 vect_analyze_loop, to be applied on an inner-loop nested in the loop 944 that is now considered for (outer-loop) vectorization. */ 945 946 static loop_vec_info 947 vect_analyze_loop_1 (struct loop *loop) 948 { 949 loop_vec_info loop_vinfo; 950 951 if (vect_print_dump_info (REPORT_DETAILS)) 952 fprintf (vect_dump, "===== analyze_loop_nest_1 ====="); 953 954 /* Check the CFG characteristics of the loop (nesting, entry/exit, etc. */ 955 956 loop_vinfo = vect_analyze_loop_form (loop); 957 if (!loop_vinfo) 958 { 959 if (vect_print_dump_info (REPORT_DETAILS)) 960 fprintf (vect_dump, "bad inner-loop form."); 961 return NULL; 962 } 963 964 return loop_vinfo; 965 } 966 967 968 /* Function vect_analyze_loop_form. 969 970 Verify that certain CFG restrictions hold, including: 971 - the loop has a pre-header 972 - the loop has a single entry and exit 973 - the loop exit condition is simple enough, and the number of iterations 974 can be analyzed (a countable loop). */ 975 976 loop_vec_info 977 vect_analyze_loop_form (struct loop *loop) 978 { 979 loop_vec_info loop_vinfo; 980 gimple loop_cond; 981 tree number_of_iterations = NULL; 982 loop_vec_info inner_loop_vinfo = NULL; 983 984 if (vect_print_dump_info (REPORT_DETAILS)) 985 fprintf (vect_dump, "=== vect_analyze_loop_form ==="); 986 987 /* Different restrictions apply when we are considering an inner-most loop, 988 vs. an outer (nested) loop. 989 (FORNOW. May want to relax some of these restrictions in the future). */ 990 991 if (!loop->inner) 992 { 993 /* Inner-most loop. We currently require that the number of BBs is 994 exactly 2 (the header and latch). Vectorizable inner-most loops 995 look like this: 996 997 (pre-header) 998 | 999 header <--------+ 1000 | | | 1001 | +--> latch --+ 1002 | 1003 (exit-bb) */ 1004 1005 if (loop->num_nodes != 2) 1006 { 1007 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS)) 1008 fprintf (vect_dump, "not vectorized: control flow in loop."); 1009 return NULL; 1010 } 1011 1012 if (empty_block_p (loop->header)) 1013 { 1014 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS)) 1015 fprintf (vect_dump, "not vectorized: empty loop."); 1016 return NULL; 1017 } 1018 } 1019 else 1020 { 1021 struct loop *innerloop = loop->inner; 1022 edge entryedge; 1023 1024 /* Nested loop. We currently require that the loop is doubly-nested, 1025 contains a single inner loop, and the number of BBs is exactly 5. 1026 Vectorizable outer-loops look like this: 1027 1028 (pre-header) 1029 | 1030 header <---+ 1031 | | 1032 inner-loop | 1033 | | 1034 tail ------+ 1035 | 1036 (exit-bb) 1037 1038 The inner-loop has the properties expected of inner-most loops 1039 as described above. */ 1040 1041 if ((loop->inner)->inner || (loop->inner)->next) 1042 { 1043 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS)) 1044 fprintf (vect_dump, "not vectorized: multiple nested loops."); 1045 return NULL; 1046 } 1047 1048 /* Analyze the inner-loop. */ 1049 inner_loop_vinfo = vect_analyze_loop_1 (loop->inner); 1050 if (!inner_loop_vinfo) 1051 { 1052 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS)) 1053 fprintf (vect_dump, "not vectorized: Bad inner loop."); 1054 return NULL; 1055 } 1056 1057 if (!expr_invariant_in_loop_p (loop, 1058 LOOP_VINFO_NITERS (inner_loop_vinfo))) 1059 { 1060 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS)) 1061 fprintf (vect_dump, 1062 "not vectorized: inner-loop count not invariant."); 1063 destroy_loop_vec_info (inner_loop_vinfo, true); 1064 return NULL; 1065 } 1066 1067 if (loop->num_nodes != 5) 1068 { 1069 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS)) 1070 fprintf (vect_dump, "not vectorized: control flow in loop."); 1071 destroy_loop_vec_info (inner_loop_vinfo, true); 1072 return NULL; 1073 } 1074 1075 gcc_assert (EDGE_COUNT (innerloop->header->preds) == 2); 1076 entryedge = EDGE_PRED (innerloop->header, 0); 1077 if (EDGE_PRED (innerloop->header, 0)->src == innerloop->latch) 1078 entryedge = EDGE_PRED (innerloop->header, 1); 1079 1080 if (entryedge->src != loop->header 1081 || !single_exit (innerloop) 1082 || single_exit (innerloop)->dest != EDGE_PRED (loop->latch, 0)->src) 1083 { 1084 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS)) 1085 fprintf (vect_dump, "not vectorized: unsupported outerloop form."); 1086 destroy_loop_vec_info (inner_loop_vinfo, true); 1087 return NULL; 1088 } 1089 1090 if (vect_print_dump_info (REPORT_DETAILS)) 1091 fprintf (vect_dump, "Considering outer-loop vectorization."); 1092 } 1093 1094 if (!single_exit (loop) 1095 || EDGE_COUNT (loop->header->preds) != 2) 1096 { 1097 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS)) 1098 { 1099 if (!single_exit (loop)) 1100 fprintf (vect_dump, "not vectorized: multiple exits."); 1101 else if (EDGE_COUNT (loop->header->preds) != 2) 1102 fprintf (vect_dump, "not vectorized: too many incoming edges."); 1103 } 1104 if (inner_loop_vinfo) 1105 destroy_loop_vec_info (inner_loop_vinfo, true); 1106 return NULL; 1107 } 1108 1109 /* We assume that the loop exit condition is at the end of the loop. i.e, 1110 that the loop is represented as a do-while (with a proper if-guard 1111 before the loop if needed), where the loop header contains all the 1112 executable statements, and the latch is empty. */ 1113 if (!empty_block_p (loop->latch) 1114 || !gimple_seq_empty_p (phi_nodes (loop->latch))) 1115 { 1116 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS)) 1117 fprintf (vect_dump, "not vectorized: unexpected loop form."); 1118 if (inner_loop_vinfo) 1119 destroy_loop_vec_info (inner_loop_vinfo, true); 1120 return NULL; 1121 } 1122 1123 /* Make sure there exists a single-predecessor exit bb: */ 1124 if (!single_pred_p (single_exit (loop)->dest)) 1125 { 1126 edge e = single_exit (loop); 1127 if (!(e->flags & EDGE_ABNORMAL)) 1128 { 1129 split_loop_exit_edge (e); 1130 if (vect_print_dump_info (REPORT_DETAILS)) 1131 fprintf (vect_dump, "split exit edge."); 1132 } 1133 else 1134 { 1135 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS)) 1136 fprintf (vect_dump, "not vectorized: abnormal loop exit edge."); 1137 if (inner_loop_vinfo) 1138 destroy_loop_vec_info (inner_loop_vinfo, true); 1139 return NULL; 1140 } 1141 } 1142 1143 loop_cond = vect_get_loop_niters (loop, &number_of_iterations); 1144 if (!loop_cond) 1145 { 1146 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS)) 1147 fprintf (vect_dump, "not vectorized: complicated exit condition."); 1148 if (inner_loop_vinfo) 1149 destroy_loop_vec_info (inner_loop_vinfo, true); 1150 return NULL; 1151 } 1152 1153 if (!number_of_iterations) 1154 { 1155 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS)) 1156 fprintf (vect_dump, 1157 "not vectorized: number of iterations cannot be computed."); 1158 if (inner_loop_vinfo) 1159 destroy_loop_vec_info (inner_loop_vinfo, true); 1160 return NULL; 1161 } 1162 1163 if (chrec_contains_undetermined (number_of_iterations)) 1164 { 1165 if (vect_print_dump_info (REPORT_BAD_FORM_LOOPS)) 1166 fprintf (vect_dump, "Infinite number of iterations."); 1167 if (inner_loop_vinfo) 1168 destroy_loop_vec_info (inner_loop_vinfo, true); 1169 return NULL; 1170 } 1171 1172 if (!NITERS_KNOWN_P (number_of_iterations)) 1173 { 1174 if (vect_print_dump_info (REPORT_DETAILS)) 1175 { 1176 fprintf (vect_dump, "Symbolic number of iterations is "); 1177 print_generic_expr (vect_dump, number_of_iterations, TDF_DETAILS); 1178 } 1179 } 1180 else if (TREE_INT_CST_LOW (number_of_iterations) == 0) 1181 { 1182 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) 1183 fprintf (vect_dump, "not vectorized: number of iterations = 0."); 1184 if (inner_loop_vinfo) 1185 destroy_loop_vec_info (inner_loop_vinfo, false); 1186 return NULL; 1187 } 1188 1189 loop_vinfo = new_loop_vec_info (loop); 1190 LOOP_VINFO_NITERS (loop_vinfo) = number_of_iterations; 1191 LOOP_VINFO_NITERS_UNCHANGED (loop_vinfo) = number_of_iterations; 1192 1193 STMT_VINFO_TYPE (vinfo_for_stmt (loop_cond)) = loop_exit_ctrl_vec_info_type; 1194 1195 /* CHECKME: May want to keep it around it in the future. */ 1196 if (inner_loop_vinfo) 1197 destroy_loop_vec_info (inner_loop_vinfo, false); 1198 1199 gcc_assert (!loop->aux); 1200 loop->aux = loop_vinfo; 1201 return loop_vinfo; 1202 } 1203 1204 1205 /* Get cost by calling cost target builtin. */ 1206 1207 static inline int 1208 vect_get_cost (enum vect_cost_for_stmt type_of_cost) 1209 { 1210 tree dummy_type = NULL; 1211 int dummy = 0; 1212 1213 return targetm.vectorize.builtin_vectorization_cost (type_of_cost, 1214 dummy_type, dummy); 1215 } 1216 1217 1218 /* Function vect_analyze_loop_operations. 1219 1220 Scan the loop stmts and make sure they are all vectorizable. */ 1221 1222 static bool 1223 vect_analyze_loop_operations (loop_vec_info loop_vinfo, bool slp) 1224 { 1225 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); 1226 basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); 1227 int nbbs = loop->num_nodes; 1228 gimple_stmt_iterator si; 1229 unsigned int vectorization_factor = 0; 1230 int i; 1231 gimple phi; 1232 stmt_vec_info stmt_info; 1233 bool need_to_vectorize = false; 1234 int min_profitable_iters; 1235 int min_scalar_loop_bound; 1236 unsigned int th; 1237 bool only_slp_in_loop = true, ok; 1238 1239 if (vect_print_dump_info (REPORT_DETAILS)) 1240 fprintf (vect_dump, "=== vect_analyze_loop_operations ==="); 1241 1242 gcc_assert (LOOP_VINFO_VECT_FACTOR (loop_vinfo)); 1243 vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo); 1244 if (slp) 1245 { 1246 /* If all the stmts in the loop can be SLPed, we perform only SLP, and 1247 vectorization factor of the loop is the unrolling factor required by 1248 the SLP instances. If that unrolling factor is 1, we say, that we 1249 perform pure SLP on loop - cross iteration parallelism is not 1250 exploited. */ 1251 for (i = 0; i < nbbs; i++) 1252 { 1253 basic_block bb = bbs[i]; 1254 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) 1255 { 1256 gimple stmt = gsi_stmt (si); 1257 stmt_vec_info stmt_info = vinfo_for_stmt (stmt); 1258 gcc_assert (stmt_info); 1259 if ((STMT_VINFO_RELEVANT_P (stmt_info) 1260 || VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (stmt_info))) 1261 && !PURE_SLP_STMT (stmt_info)) 1262 /* STMT needs both SLP and loop-based vectorization. */ 1263 only_slp_in_loop = false; 1264 } 1265 } 1266 1267 if (only_slp_in_loop) 1268 vectorization_factor = LOOP_VINFO_SLP_UNROLLING_FACTOR (loop_vinfo); 1269 else 1270 vectorization_factor = least_common_multiple (vectorization_factor, 1271 LOOP_VINFO_SLP_UNROLLING_FACTOR (loop_vinfo)); 1272 1273 LOOP_VINFO_VECT_FACTOR (loop_vinfo) = vectorization_factor; 1274 if (vect_print_dump_info (REPORT_DETAILS)) 1275 fprintf (vect_dump, "Updating vectorization factor to %d ", 1276 vectorization_factor); 1277 } 1278 1279 for (i = 0; i < nbbs; i++) 1280 { 1281 basic_block bb = bbs[i]; 1282 1283 for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) 1284 { 1285 phi = gsi_stmt (si); 1286 ok = true; 1287 1288 stmt_info = vinfo_for_stmt (phi); 1289 if (vect_print_dump_info (REPORT_DETAILS)) 1290 { 1291 fprintf (vect_dump, "examining phi: "); 1292 print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM); 1293 } 1294 1295 /* Inner-loop loop-closed exit phi in outer-loop vectorization 1296 (i.e., a phi in the tail of the outer-loop). */ 1297 if (! is_loop_header_bb_p (bb)) 1298 { 1299 /* FORNOW: we currently don't support the case that these phis 1300 are not used in the outerloop (unless it is double reduction, 1301 i.e., this phi is vect_reduction_def), cause this case 1302 requires to actually do something here. */ 1303 if ((!STMT_VINFO_RELEVANT_P (stmt_info) 1304 || STMT_VINFO_LIVE_P (stmt_info)) 1305 && STMT_VINFO_DEF_TYPE (stmt_info) 1306 != vect_double_reduction_def) 1307 { 1308 if (vect_print_dump_info (REPORT_DETAILS)) 1309 fprintf (vect_dump, 1310 "Unsupported loop-closed phi in outer-loop."); 1311 return false; 1312 } 1313 1314 /* If PHI is used in the outer loop, we check that its operand 1315 is defined in the inner loop. */ 1316 if (STMT_VINFO_RELEVANT_P (stmt_info)) 1317 { 1318 tree phi_op; 1319 gimple op_def_stmt; 1320 1321 if (gimple_phi_num_args (phi) != 1) 1322 return false; 1323 1324 phi_op = PHI_ARG_DEF (phi, 0); 1325 if (TREE_CODE (phi_op) != SSA_NAME) 1326 return false; 1327 1328 op_def_stmt = SSA_NAME_DEF_STMT (phi_op); 1329 if (!op_def_stmt 1330 || !flow_bb_inside_loop_p (loop, gimple_bb (op_def_stmt)) 1331 || !vinfo_for_stmt (op_def_stmt)) 1332 return false; 1333 1334 if (STMT_VINFO_RELEVANT (vinfo_for_stmt (op_def_stmt)) 1335 != vect_used_in_outer 1336 && STMT_VINFO_RELEVANT (vinfo_for_stmt (op_def_stmt)) 1337 != vect_used_in_outer_by_reduction) 1338 return false; 1339 } 1340 1341 continue; 1342 } 1343 1344 gcc_assert (stmt_info); 1345 1346 if (STMT_VINFO_LIVE_P (stmt_info)) 1347 { 1348 /* FORNOW: not yet supported. */ 1349 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) 1350 fprintf (vect_dump, "not vectorized: value used after loop."); 1351 return false; 1352 } 1353 1354 if (STMT_VINFO_RELEVANT (stmt_info) == vect_used_in_scope 1355 && STMT_VINFO_DEF_TYPE (stmt_info) != vect_induction_def) 1356 { 1357 /* A scalar-dependence cycle that we don't support. */ 1358 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) 1359 fprintf (vect_dump, "not vectorized: scalar dependence cycle."); 1360 return false; 1361 } 1362 1363 if (STMT_VINFO_RELEVANT_P (stmt_info)) 1364 { 1365 need_to_vectorize = true; 1366 if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_induction_def) 1367 ok = vectorizable_induction (phi, NULL, NULL); 1368 } 1369 1370 if (!ok) 1371 { 1372 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) 1373 { 1374 fprintf (vect_dump, 1375 "not vectorized: relevant phi not supported: "); 1376 print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM); 1377 } 1378 return false; 1379 } 1380 } 1381 1382 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) 1383 { 1384 gimple stmt = gsi_stmt (si); 1385 if (!vect_analyze_stmt (stmt, &need_to_vectorize, NULL)) 1386 return false; 1387 } 1388 } /* bbs */ 1389 1390 /* All operations in the loop are either irrelevant (deal with loop 1391 control, or dead), or only used outside the loop and can be moved 1392 out of the loop (e.g. invariants, inductions). The loop can be 1393 optimized away by scalar optimizations. We're better off not 1394 touching this loop. */ 1395 if (!need_to_vectorize) 1396 { 1397 if (vect_print_dump_info (REPORT_DETAILS)) 1398 fprintf (vect_dump, 1399 "All the computation can be taken out of the loop."); 1400 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) 1401 fprintf (vect_dump, 1402 "not vectorized: redundant loop. no profit to vectorize."); 1403 return false; 1404 } 1405 1406 if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) 1407 && vect_print_dump_info (REPORT_DETAILS)) 1408 fprintf (vect_dump, 1409 "vectorization_factor = %d, niters = " HOST_WIDE_INT_PRINT_DEC, 1410 vectorization_factor, LOOP_VINFO_INT_NITERS (loop_vinfo)); 1411 1412 if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) 1413 && (LOOP_VINFO_INT_NITERS (loop_vinfo) < vectorization_factor)) 1414 { 1415 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) 1416 fprintf (vect_dump, "not vectorized: iteration count too small."); 1417 if (vect_print_dump_info (REPORT_DETAILS)) 1418 fprintf (vect_dump,"not vectorized: iteration count smaller than " 1419 "vectorization factor."); 1420 return false; 1421 } 1422 1423 /* Analyze cost. Decide if worth while to vectorize. */ 1424 1425 /* Once VF is set, SLP costs should be updated since the number of created 1426 vector stmts depends on VF. */ 1427 vect_update_slp_costs_according_to_vf (loop_vinfo); 1428 1429 min_profitable_iters = vect_estimate_min_profitable_iters (loop_vinfo); 1430 LOOP_VINFO_COST_MODEL_MIN_ITERS (loop_vinfo) = min_profitable_iters; 1431 1432 if (min_profitable_iters < 0) 1433 { 1434 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) 1435 fprintf (vect_dump, "not vectorized: vectorization not profitable."); 1436 if (vect_print_dump_info (REPORT_DETAILS)) 1437 fprintf (vect_dump, "not vectorized: vector version will never be " 1438 "profitable."); 1439 return false; 1440 } 1441 1442 min_scalar_loop_bound = ((PARAM_VALUE (PARAM_MIN_VECT_LOOP_BOUND) 1443 * vectorization_factor) - 1); 1444 1445 /* Use the cost model only if it is more conservative than user specified 1446 threshold. */ 1447 1448 th = (unsigned) min_scalar_loop_bound; 1449 if (min_profitable_iters 1450 && (!min_scalar_loop_bound 1451 || min_profitable_iters > min_scalar_loop_bound)) 1452 th = (unsigned) min_profitable_iters; 1453 1454 if (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) 1455 && LOOP_VINFO_INT_NITERS (loop_vinfo) <= th) 1456 { 1457 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) 1458 fprintf (vect_dump, "not vectorized: vectorization not " 1459 "profitable."); 1460 if (vect_print_dump_info (REPORT_DETAILS)) 1461 fprintf (vect_dump, "not vectorized: iteration count smaller than " 1462 "user specified loop bound parameter or minimum " 1463 "profitable iterations (whichever is more conservative)."); 1464 return false; 1465 } 1466 1467 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) 1468 || LOOP_VINFO_INT_NITERS (loop_vinfo) % vectorization_factor != 0 1469 || LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo)) 1470 { 1471 if (vect_print_dump_info (REPORT_DETAILS)) 1472 fprintf (vect_dump, "epilog loop required."); 1473 if (!vect_can_advance_ivs_p (loop_vinfo)) 1474 { 1475 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) 1476 fprintf (vect_dump, 1477 "not vectorized: can't create epilog loop 1."); 1478 return false; 1479 } 1480 if (!slpeel_can_duplicate_loop_p (loop, single_exit (loop))) 1481 { 1482 if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) 1483 fprintf (vect_dump, 1484 "not vectorized: can't create epilog loop 2."); 1485 return false; 1486 } 1487 } 1488 1489 return true; 1490 } 1491 1492 1493 /* Function vect_analyze_loop_2. 1494 1495 Apply a set of analyses on LOOP, and create a loop_vec_info struct 1496 for it. The different analyses will record information in the 1497 loop_vec_info struct. */ 1498 static bool 1499 vect_analyze_loop_2 (loop_vec_info loop_vinfo) 1500 { 1501 bool ok, slp = false; 1502 int max_vf = MAX_VECTORIZATION_FACTOR; 1503 int min_vf = 2; 1504 1505 /* Find all data references in the loop (which correspond to vdefs/vuses) 1506 and analyze their evolution in the loop. Also adjust the minimal 1507 vectorization factor according to the loads and stores. 1508 1509 FORNOW: Handle only simple, array references, which 1510 alignment can be forced, and aligned pointer-references. */ 1511 1512 ok = vect_analyze_data_refs (loop_vinfo, NULL, &min_vf); 1513 if (!ok) 1514 { 1515 if (vect_print_dump_info (REPORT_DETAILS)) 1516 fprintf (vect_dump, "bad data references."); 1517 return false; 1518 } 1519 1520 /* Classify all cross-iteration scalar data-flow cycles. 1521 Cross-iteration cycles caused by virtual phis are analyzed separately. */ 1522 1523 vect_analyze_scalar_cycles (loop_vinfo); 1524 1525 vect_pattern_recog (loop_vinfo); 1526 1527 /* Data-flow analysis to detect stmts that do not need to be vectorized. */ 1528 1529 ok = vect_mark_stmts_to_be_vectorized (loop_vinfo); 1530 if (!ok) 1531 { 1532 if (vect_print_dump_info (REPORT_DETAILS)) 1533 fprintf (vect_dump, "unexpected pattern."); 1534 return false; 1535 } 1536 1537 /* Analyze data dependences between the data-refs in the loop 1538 and adjust the maximum vectorization factor according to 1539 the dependences. 1540 FORNOW: fail at the first data dependence that we encounter. */ 1541 1542 ok = vect_analyze_data_ref_dependences (loop_vinfo, NULL, &max_vf); 1543 if (!ok 1544 || max_vf < min_vf) 1545 { 1546 if (vect_print_dump_info (REPORT_DETAILS)) 1547 fprintf (vect_dump, "bad data dependence."); 1548 return false; 1549 } 1550 1551 ok = vect_determine_vectorization_factor (loop_vinfo); 1552 if (!ok) 1553 { 1554 if (vect_print_dump_info (REPORT_DETAILS)) 1555 fprintf (vect_dump, "can't determine vectorization factor."); 1556 return false; 1557 } 1558 if (max_vf < LOOP_VINFO_VECT_FACTOR (loop_vinfo)) 1559 { 1560 if (vect_print_dump_info (REPORT_DETAILS)) 1561 fprintf (vect_dump, "bad data dependence."); 1562 return false; 1563 } 1564 1565 /* Analyze the alignment of the data-refs in the loop. 1566 Fail if a data reference is found that cannot be vectorized. */ 1567 1568 ok = vect_analyze_data_refs_alignment (loop_vinfo, NULL); 1569 if (!ok) 1570 { 1571 if (vect_print_dump_info (REPORT_DETAILS)) 1572 fprintf (vect_dump, "bad data alignment."); 1573 return false; 1574 } 1575 1576 /* Analyze the access patterns of the data-refs in the loop (consecutive, 1577 complex, etc.). FORNOW: Only handle consecutive access pattern. */ 1578 1579 ok = vect_analyze_data_ref_accesses (loop_vinfo, NULL); 1580 if (!ok) 1581 { 1582 if (vect_print_dump_info (REPORT_DETAILS)) 1583 fprintf (vect_dump, "bad data access."); 1584 return false; 1585 } 1586 1587 /* Prune the list of ddrs to be tested at run-time by versioning for alias. 1588 It is important to call pruning after vect_analyze_data_ref_accesses, 1589 since we use grouping information gathered by interleaving analysis. */ 1590 ok = vect_prune_runtime_alias_test_list (loop_vinfo); 1591 if (!ok) 1592 { 1593 if (vect_print_dump_info (REPORT_DETAILS)) 1594 fprintf (vect_dump, "too long list of versioning for alias " 1595 "run-time tests."); 1596 return false; 1597 } 1598 1599 /* This pass will decide on using loop versioning and/or loop peeling in 1600 order to enhance the alignment of data references in the loop. */ 1601 1602 ok = vect_enhance_data_refs_alignment (loop_vinfo); 1603 if (!ok) 1604 { 1605 if (vect_print_dump_info (REPORT_DETAILS)) 1606 fprintf (vect_dump, "bad data alignment."); 1607 return false; 1608 } 1609 1610 /* Check the SLP opportunities in the loop, analyze and build SLP trees. */ 1611 ok = vect_analyze_slp (loop_vinfo, NULL); 1612 if (ok) 1613 { 1614 /* Decide which possible SLP instances to SLP. */ 1615 slp = vect_make_slp_decision (loop_vinfo); 1616 1617 /* Find stmts that need to be both vectorized and SLPed. */ 1618 vect_detect_hybrid_slp (loop_vinfo); 1619 } 1620 else 1621 return false; 1622 1623 /* Scan all the operations in the loop and make sure they are 1624 vectorizable. */ 1625 1626 ok = vect_analyze_loop_operations (loop_vinfo, slp); 1627 if (!ok) 1628 { 1629 if (vect_print_dump_info (REPORT_DETAILS)) 1630 fprintf (vect_dump, "bad operation or unsupported loop bound."); 1631 return false; 1632 } 1633 1634 return true; 1635 } 1636 1637 /* Function vect_analyze_loop. 1638 1639 Apply a set of analyses on LOOP, and create a loop_vec_info struct 1640 for it. The different analyses will record information in the 1641 loop_vec_info struct. */ 1642 loop_vec_info 1643 vect_analyze_loop (struct loop *loop) 1644 { 1645 loop_vec_info loop_vinfo; 1646 unsigned int vector_sizes; 1647 1648 /* Autodetect first vector size we try. */ 1649 current_vector_size = 0; 1650 vector_sizes = targetm.vectorize.autovectorize_vector_sizes (); 1651 1652 if (vect_print_dump_info (REPORT_DETAILS)) 1653 fprintf (vect_dump, "===== analyze_loop_nest ====="); 1654 1655 if (loop_outer (loop) 1656 && loop_vec_info_for_loop (loop_outer (loop)) 1657 && LOOP_VINFO_VECTORIZABLE_P (loop_vec_info_for_loop (loop_outer (loop)))) 1658 { 1659 if (vect_print_dump_info (REPORT_DETAILS)) 1660 fprintf (vect_dump, "outer-loop already vectorized."); 1661 return NULL; 1662 } 1663 1664 while (1) 1665 { 1666 /* Check the CFG characteristics of the loop (nesting, entry/exit). */ 1667 loop_vinfo = vect_analyze_loop_form (loop); 1668 if (!loop_vinfo) 1669 { 1670 if (vect_print_dump_info (REPORT_DETAILS)) 1671 fprintf (vect_dump, "bad loop form."); 1672 return NULL; 1673 } 1674 1675 if (vect_analyze_loop_2 (loop_vinfo)) 1676 { 1677 LOOP_VINFO_VECTORIZABLE_P (loop_vinfo) = 1; 1678 1679 return loop_vinfo; 1680 } 1681 1682 destroy_loop_vec_info (loop_vinfo, true); 1683 1684 vector_sizes &= ~current_vector_size; 1685 if (vector_sizes == 0 1686 || current_vector_size == 0) 1687 return NULL; 1688 1689 /* Try the next biggest vector size. */ 1690 current_vector_size = 1 << floor_log2 (vector_sizes); 1691 if (vect_print_dump_info (REPORT_DETAILS)) 1692 fprintf (vect_dump, "***** Re-trying analysis with " 1693 "vector size %d\n", current_vector_size); 1694 } 1695 } 1696 1697 1698 /* Function reduction_code_for_scalar_code 1699 1700 Input: 1701 CODE - tree_code of a reduction operations. 1702 1703 Output: 1704 REDUC_CODE - the corresponding tree-code to be used to reduce the 1705 vector of partial results into a single scalar result (which 1706 will also reside in a vector) or ERROR_MARK if the operation is 1707 a supported reduction operation, but does not have such tree-code. 1708 1709 Return FALSE if CODE currently cannot be vectorized as reduction. */ 1710 1711 static bool 1712 reduction_code_for_scalar_code (enum tree_code code, 1713 enum tree_code *reduc_code) 1714 { 1715 switch (code) 1716 { 1717 case MAX_EXPR: 1718 *reduc_code = REDUC_MAX_EXPR; 1719 return true; 1720 1721 case MIN_EXPR: 1722 *reduc_code = REDUC_MIN_EXPR; 1723 return true; 1724 1725 case PLUS_EXPR: 1726 *reduc_code = REDUC_PLUS_EXPR; 1727 return true; 1728 1729 case MULT_EXPR: 1730 case MINUS_EXPR: 1731 case BIT_IOR_EXPR: 1732 case BIT_XOR_EXPR: 1733 case BIT_AND_EXPR: 1734 *reduc_code = ERROR_MARK; 1735 return true; 1736 1737 default: 1738 return false; 1739 } 1740 } 1741 1742 1743 /* Error reporting helper for vect_is_simple_reduction below. GIMPLE statement 1744 STMT is printed with a message MSG. */ 1745 1746 static void 1747 report_vect_op (gimple stmt, const char *msg) 1748 { 1749 fprintf (vect_dump, "%s", msg); 1750 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); 1751 } 1752 1753 1754 /* Detect SLP reduction of the form: 1755 1756 #a1 = phi <a5, a0> 1757 a2 = operation (a1) 1758 a3 = operation (a2) 1759 a4 = operation (a3) 1760 a5 = operation (a4) 1761 1762 #a = phi <a5> 1763 1764 PHI is the reduction phi node (#a1 = phi <a5, a0> above) 1765 FIRST_STMT is the first reduction stmt in the chain 1766 (a2 = operation (a1)). 1767 1768 Return TRUE if a reduction chain was detected. */ 1769 1770 static bool 1771 vect_is_slp_reduction (loop_vec_info loop_info, gimple phi, gimple first_stmt) 1772 { 1773 struct loop *loop = (gimple_bb (phi))->loop_father; 1774 struct loop *vect_loop = LOOP_VINFO_LOOP (loop_info); 1775 enum tree_code code; 1776 gimple current_stmt = NULL, loop_use_stmt = NULL, first, next_stmt; 1777 stmt_vec_info use_stmt_info, current_stmt_info; 1778 tree lhs; 1779 imm_use_iterator imm_iter; 1780 use_operand_p use_p; 1781 int nloop_uses, size = 0, n_out_of_loop_uses; 1782 bool found = false; 1783 1784 if (loop != vect_loop) 1785 return false; 1786 1787 lhs = PHI_RESULT (phi); 1788 code = gimple_assign_rhs_code (first_stmt); 1789 while (1) 1790 { 1791 nloop_uses = 0; 1792 n_out_of_loop_uses = 0; 1793 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, lhs) 1794 { 1795 gimple use_stmt = USE_STMT (use_p); 1796 if (is_gimple_debug (use_stmt)) 1797 continue; 1798 1799 use_stmt = USE_STMT (use_p); 1800 1801 /* Check if we got back to the reduction phi. */ 1802 if (use_stmt == phi) 1803 { 1804 loop_use_stmt = use_stmt; 1805 found = true; 1806 break; 1807 } 1808 1809 if (flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))) 1810 { 1811 if (vinfo_for_stmt (use_stmt) 1812 && !STMT_VINFO_IN_PATTERN_P (vinfo_for_stmt (use_stmt))) 1813 { 1814 loop_use_stmt = use_stmt; 1815 nloop_uses++; 1816 } 1817 } 1818 else 1819 n_out_of_loop_uses++; 1820 1821 /* There are can be either a single use in the loop or two uses in 1822 phi nodes. */ 1823 if (nloop_uses > 1 || (n_out_of_loop_uses && nloop_uses)) 1824 return false; 1825 } 1826 1827 if (found) 1828 break; 1829 1830 /* We reached a statement with no loop uses. */ 1831 if (nloop_uses == 0) 1832 return false; 1833 1834 /* This is a loop exit phi, and we haven't reached the reduction phi. */ 1835 if (gimple_code (loop_use_stmt) == GIMPLE_PHI) 1836 return false; 1837 1838 if (!is_gimple_assign (loop_use_stmt) 1839 || code != gimple_assign_rhs_code (loop_use_stmt) 1840 || !flow_bb_inside_loop_p (loop, gimple_bb (loop_use_stmt))) 1841 return false; 1842 1843 /* Insert USE_STMT into reduction chain. */ 1844 use_stmt_info = vinfo_for_stmt (loop_use_stmt); 1845 if (current_stmt) 1846 { 1847 current_stmt_info = vinfo_for_stmt (current_stmt); 1848 GROUP_NEXT_ELEMENT (current_stmt_info) = loop_use_stmt; 1849 GROUP_FIRST_ELEMENT (use_stmt_info) 1850 = GROUP_FIRST_ELEMENT (current_stmt_info); 1851 } 1852 else 1853 GROUP_FIRST_ELEMENT (use_stmt_info) = loop_use_stmt; 1854 1855 lhs = gimple_assign_lhs (loop_use_stmt); 1856 current_stmt = loop_use_stmt; 1857 size++; 1858 } 1859 1860 if (!found || loop_use_stmt != phi || size < 2) 1861 return false; 1862 1863 /* Swap the operands, if needed, to make the reduction operand be the second 1864 operand. */ 1865 lhs = PHI_RESULT (phi); 1866 next_stmt = GROUP_FIRST_ELEMENT (vinfo_for_stmt (current_stmt)); 1867 while (next_stmt) 1868 { 1869 if (gimple_assign_rhs2 (next_stmt) == lhs) 1870 { 1871 tree op = gimple_assign_rhs1 (next_stmt); 1872 gimple def_stmt = NULL; 1873 1874 if (TREE_CODE (op) == SSA_NAME) 1875 def_stmt = SSA_NAME_DEF_STMT (op); 1876 1877 /* Check that the other def is either defined in the loop 1878 ("vect_internal_def"), or it's an induction (defined by a 1879 loop-header phi-node). */ 1880 if (def_stmt 1881 && gimple_bb (def_stmt) 1882 && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt)) 1883 && (is_gimple_assign (def_stmt) 1884 || is_gimple_call (def_stmt) 1885 || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt)) 1886 == vect_induction_def 1887 || (gimple_code (def_stmt) == GIMPLE_PHI 1888 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt)) 1889 == vect_internal_def 1890 && !is_loop_header_bb_p (gimple_bb (def_stmt))))) 1891 { 1892 lhs = gimple_assign_lhs (next_stmt); 1893 next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); 1894 continue; 1895 } 1896 1897 return false; 1898 } 1899 else 1900 { 1901 tree op = gimple_assign_rhs2 (next_stmt); 1902 gimple def_stmt = NULL; 1903 1904 if (TREE_CODE (op) == SSA_NAME) 1905 def_stmt = SSA_NAME_DEF_STMT (op); 1906 1907 /* Check that the other def is either defined in the loop 1908 ("vect_internal_def"), or it's an induction (defined by a 1909 loop-header phi-node). */ 1910 if (def_stmt 1911 && gimple_bb (def_stmt) 1912 && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt)) 1913 && (is_gimple_assign (def_stmt) 1914 || is_gimple_call (def_stmt) 1915 || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt)) 1916 == vect_induction_def 1917 || (gimple_code (def_stmt) == GIMPLE_PHI 1918 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt)) 1919 == vect_internal_def 1920 && !is_loop_header_bb_p (gimple_bb (def_stmt))))) 1921 { 1922 if (vect_print_dump_info (REPORT_DETAILS)) 1923 { 1924 fprintf (vect_dump, "swapping oprnds: "); 1925 print_gimple_stmt (vect_dump, next_stmt, 0, TDF_SLIM); 1926 } 1927 1928 swap_tree_operands (next_stmt, 1929 gimple_assign_rhs1_ptr (next_stmt), 1930 gimple_assign_rhs2_ptr (next_stmt)); 1931 mark_symbols_for_renaming (next_stmt); 1932 } 1933 else 1934 return false; 1935 } 1936 1937 lhs = gimple_assign_lhs (next_stmt); 1938 next_stmt = GROUP_NEXT_ELEMENT (vinfo_for_stmt (next_stmt)); 1939 } 1940 1941 /* Save the chain for further analysis in SLP detection. */ 1942 first = GROUP_FIRST_ELEMENT (vinfo_for_stmt (current_stmt)); 1943 VEC_safe_push (gimple, heap, LOOP_VINFO_REDUCTION_CHAINS (loop_info), first); 1944 GROUP_SIZE (vinfo_for_stmt (first)) = size; 1945 1946 return true; 1947 } 1948 1949 1950 /* Function vect_is_simple_reduction_1 1951 1952 (1) Detect a cross-iteration def-use cycle that represents a simple 1953 reduction computation. We look for the following pattern: 1954 1955 loop_header: 1956 a1 = phi < a0, a2 > 1957 a3 = ... 1958 a2 = operation (a3, a1) 1959 1960 such that: 1961 1. operation is commutative and associative and it is safe to 1962 change the order of the computation (if CHECK_REDUCTION is true) 1963 2. no uses for a2 in the loop (a2 is used out of the loop) 1964 3. no uses of a1 in the loop besides the reduction operation 1965 4. no uses of a1 outside the loop. 1966 1967 Conditions 1,4 are tested here. 1968 Conditions 2,3 are tested in vect_mark_stmts_to_be_vectorized. 1969 1970 (2) Detect a cross-iteration def-use cycle in nested loops, i.e., 1971 nested cycles, if CHECK_REDUCTION is false. 1972 1973 (3) Detect cycles of phi nodes in outer-loop vectorization, i.e., double 1974 reductions: 1975 1976 a1 = phi < a0, a2 > 1977 inner loop (def of a3) 1978 a2 = phi < a3 > 1979 1980 If MODIFY is true it tries also to rework the code in-place to enable 1981 detection of more reduction patterns. For the time being we rewrite 1982 "res -= RHS" into "rhs += -RHS" when it seems worthwhile. 1983 */ 1984 1985 static gimple 1986 vect_is_simple_reduction_1 (loop_vec_info loop_info, gimple phi, 1987 bool check_reduction, bool *double_reduc, 1988 bool modify) 1989 { 1990 struct loop *loop = (gimple_bb (phi))->loop_father; 1991 struct loop *vect_loop = LOOP_VINFO_LOOP (loop_info); 1992 edge latch_e = loop_latch_edge (loop); 1993 tree loop_arg = PHI_ARG_DEF_FROM_EDGE (phi, latch_e); 1994 gimple def_stmt, def1 = NULL, def2 = NULL; 1995 enum tree_code orig_code, code; 1996 tree op1, op2, op3 = NULL_TREE, op4 = NULL_TREE; 1997 tree type; 1998 int nloop_uses; 1999 tree name; 2000 imm_use_iterator imm_iter; 2001 use_operand_p use_p; 2002 bool phi_def; 2003 2004 *double_reduc = false; 2005 2006 /* If CHECK_REDUCTION is true, we assume inner-most loop vectorization, 2007 otherwise, we assume outer loop vectorization. */ 2008 gcc_assert ((check_reduction && loop == vect_loop) 2009 || (!check_reduction && flow_loop_nested_p (vect_loop, loop))); 2010 2011 name = PHI_RESULT (phi); 2012 /* ??? If there are no uses of the PHI result the inner loop reduction 2013 won't be detected as possibly double-reduction by vectorizable_reduction 2014 because that tries to walk the PHI arg from the preheader edge which 2015 can be constant. See PR60382. */ 2016 if (has_zero_uses (name)) 2017 return NULL; 2018 nloop_uses = 0; 2019 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, name) 2020 { 2021 gimple use_stmt = USE_STMT (use_p); 2022 if (is_gimple_debug (use_stmt)) 2023 continue; 2024 2025 if (!flow_bb_inside_loop_p (loop, gimple_bb (use_stmt))) 2026 { 2027 if (vect_print_dump_info (REPORT_DETAILS)) 2028 fprintf (vect_dump, "intermediate value used outside loop."); 2029 2030 return NULL; 2031 } 2032 2033 if (vinfo_for_stmt (use_stmt) 2034 && !is_pattern_stmt_p (vinfo_for_stmt (use_stmt))) 2035 nloop_uses++; 2036 if (nloop_uses > 1) 2037 { 2038 if (vect_print_dump_info (REPORT_DETAILS)) 2039 fprintf (vect_dump, "reduction used in loop."); 2040 return NULL; 2041 } 2042 } 2043 2044 if (TREE_CODE (loop_arg) != SSA_NAME) 2045 { 2046 if (vect_print_dump_info (REPORT_DETAILS)) 2047 { 2048 fprintf (vect_dump, "reduction: not ssa_name: "); 2049 print_generic_expr (vect_dump, loop_arg, TDF_SLIM); 2050 } 2051 return NULL; 2052 } 2053 2054 def_stmt = SSA_NAME_DEF_STMT (loop_arg); 2055 if (!def_stmt) 2056 { 2057 if (vect_print_dump_info (REPORT_DETAILS)) 2058 fprintf (vect_dump, "reduction: no def_stmt."); 2059 return NULL; 2060 } 2061 2062 if (!is_gimple_assign (def_stmt) && gimple_code (def_stmt) != GIMPLE_PHI) 2063 { 2064 if (vect_print_dump_info (REPORT_DETAILS)) 2065 print_gimple_stmt (vect_dump, def_stmt, 0, TDF_SLIM); 2066 return NULL; 2067 } 2068 2069 if (is_gimple_assign (def_stmt)) 2070 { 2071 name = gimple_assign_lhs (def_stmt); 2072 phi_def = false; 2073 } 2074 else 2075 { 2076 name = PHI_RESULT (def_stmt); 2077 phi_def = true; 2078 } 2079 2080 nloop_uses = 0; 2081 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, name) 2082 { 2083 gimple use_stmt = USE_STMT (use_p); 2084 if (is_gimple_debug (use_stmt)) 2085 continue; 2086 if (flow_bb_inside_loop_p (loop, gimple_bb (use_stmt)) 2087 && vinfo_for_stmt (use_stmt) 2088 && !is_pattern_stmt_p (vinfo_for_stmt (use_stmt))) 2089 nloop_uses++; 2090 if (nloop_uses > 1) 2091 { 2092 if (vect_print_dump_info (REPORT_DETAILS)) 2093 fprintf (vect_dump, "reduction used in loop."); 2094 return NULL; 2095 } 2096 } 2097 2098 /* If DEF_STMT is a phi node itself, we expect it to have a single argument 2099 defined in the inner loop. */ 2100 if (phi_def) 2101 { 2102 op1 = PHI_ARG_DEF (def_stmt, 0); 2103 2104 if (gimple_phi_num_args (def_stmt) != 1 2105 || TREE_CODE (op1) != SSA_NAME) 2106 { 2107 if (vect_print_dump_info (REPORT_DETAILS)) 2108 fprintf (vect_dump, "unsupported phi node definition."); 2109 2110 return NULL; 2111 } 2112 2113 def1 = SSA_NAME_DEF_STMT (op1); 2114 if (flow_bb_inside_loop_p (loop, gimple_bb (def_stmt)) 2115 && loop->inner 2116 && flow_bb_inside_loop_p (loop->inner, gimple_bb (def1)) 2117 && is_gimple_assign (def1)) 2118 { 2119 if (vect_print_dump_info (REPORT_DETAILS)) 2120 report_vect_op (def_stmt, "detected double reduction: "); 2121 2122 *double_reduc = true; 2123 return def_stmt; 2124 } 2125 2126 return NULL; 2127 } 2128 2129 code = orig_code = gimple_assign_rhs_code (def_stmt); 2130 2131 /* We can handle "res -= x[i]", which is non-associative by 2132 simply rewriting this into "res += -x[i]". Avoid changing 2133 gimple instruction for the first simple tests and only do this 2134 if we're allowed to change code at all. */ 2135 if (code == MINUS_EXPR 2136 && modify 2137 && (op1 = gimple_assign_rhs1 (def_stmt)) 2138 && TREE_CODE (op1) == SSA_NAME 2139 && SSA_NAME_DEF_STMT (op1) == phi) 2140 code = PLUS_EXPR; 2141 2142 if (check_reduction 2143 && (!commutative_tree_code (code) || !associative_tree_code (code))) 2144 { 2145 if (vect_print_dump_info (REPORT_DETAILS)) 2146 report_vect_op (def_stmt, "reduction: not commutative/associative: "); 2147 return NULL; 2148 } 2149 2150 if (get_gimple_rhs_class (code) != GIMPLE_BINARY_RHS) 2151 { 2152 if (code != COND_EXPR) 2153 { 2154 if (vect_print_dump_info (REPORT_DETAILS)) 2155 report_vect_op (def_stmt, "reduction: not binary operation: "); 2156 2157 return NULL; 2158 } 2159 2160 op3 = gimple_assign_rhs1 (def_stmt); 2161 if (COMPARISON_CLASS_P (op3)) 2162 { 2163 op4 = TREE_OPERAND (op3, 1); 2164 op3 = TREE_OPERAND (op3, 0); 2165 } 2166 2167 op1 = gimple_assign_rhs2 (def_stmt); 2168 op2 = gimple_assign_rhs3 (def_stmt); 2169 2170 if (TREE_CODE (op1) != SSA_NAME && TREE_CODE (op2) != SSA_NAME) 2171 { 2172 if (vect_print_dump_info (REPORT_DETAILS)) 2173 report_vect_op (def_stmt, "reduction: uses not ssa_names: "); 2174 2175 return NULL; 2176 } 2177 } 2178 else 2179 { 2180 op1 = gimple_assign_rhs1 (def_stmt); 2181 op2 = gimple_assign_rhs2 (def_stmt); 2182 2183 if (TREE_CODE (op1) != SSA_NAME && TREE_CODE (op2) != SSA_NAME) 2184 { 2185 if (vect_print_dump_info (REPORT_DETAILS)) 2186 report_vect_op (def_stmt, "reduction: uses not ssa_names: "); 2187 2188 return NULL; 2189 } 2190 } 2191 2192 type = TREE_TYPE (gimple_assign_lhs (def_stmt)); 2193 if ((TREE_CODE (op1) == SSA_NAME 2194 && !types_compatible_p (type,TREE_TYPE (op1))) 2195 || (TREE_CODE (op2) == SSA_NAME 2196 && !types_compatible_p (type, TREE_TYPE (op2))) 2197 || (op3 && TREE_CODE (op3) == SSA_NAME 2198 && !types_compatible_p (type, TREE_TYPE (op3))) 2199 || (op4 && TREE_CODE (op4) == SSA_NAME 2200 && !types_compatible_p (type, TREE_TYPE (op4)))) 2201 { 2202 if (vect_print_dump_info (REPORT_DETAILS)) 2203 { 2204 fprintf (vect_dump, "reduction: multiple types: operation type: "); 2205 print_generic_expr (vect_dump, type, TDF_SLIM); 2206 fprintf (vect_dump, ", operands types: "); 2207 print_generic_expr (vect_dump, TREE_TYPE (op1), TDF_SLIM); 2208 fprintf (vect_dump, ","); 2209 print_generic_expr (vect_dump, TREE_TYPE (op2), TDF_SLIM); 2210 if (op3) 2211 { 2212 fprintf (vect_dump, ","); 2213 print_generic_expr (vect_dump, TREE_TYPE (op3), TDF_SLIM); 2214 } 2215 2216 if (op4) 2217 { 2218 fprintf (vect_dump, ","); 2219 print_generic_expr (vect_dump, TREE_TYPE (op4), TDF_SLIM); 2220 } 2221 } 2222 2223 return NULL; 2224 } 2225 2226 /* Check that it's ok to change the order of the computation. 2227 Generally, when vectorizing a reduction we change the order of the 2228 computation. This may change the behavior of the program in some 2229 cases, so we need to check that this is ok. One exception is when 2230 vectorizing an outer-loop: the inner-loop is executed sequentially, 2231 and therefore vectorizing reductions in the inner-loop during 2232 outer-loop vectorization is safe. */ 2233 2234 /* CHECKME: check for !flag_finite_math_only too? */ 2235 if (SCALAR_FLOAT_TYPE_P (type) && !flag_associative_math 2236 && check_reduction) 2237 { 2238 /* Changing the order of operations changes the semantics. */ 2239 if (vect_print_dump_info (REPORT_DETAILS)) 2240 report_vect_op (def_stmt, "reduction: unsafe fp math optimization: "); 2241 return NULL; 2242 } 2243 else if (INTEGRAL_TYPE_P (type) && TYPE_OVERFLOW_TRAPS (type) 2244 && check_reduction) 2245 { 2246 /* Changing the order of operations changes the semantics. */ 2247 if (vect_print_dump_info (REPORT_DETAILS)) 2248 report_vect_op (def_stmt, "reduction: unsafe int math optimization: "); 2249 return NULL; 2250 } 2251 else if (SAT_FIXED_POINT_TYPE_P (type) && check_reduction) 2252 { 2253 /* Changing the order of operations changes the semantics. */ 2254 if (vect_print_dump_info (REPORT_DETAILS)) 2255 report_vect_op (def_stmt, 2256 "reduction: unsafe fixed-point math optimization: "); 2257 return NULL; 2258 } 2259 2260 /* If we detected "res -= x[i]" earlier, rewrite it into 2261 "res += -x[i]" now. If this turns out to be useless reassoc 2262 will clean it up again. */ 2263 if (orig_code == MINUS_EXPR) 2264 { 2265 tree rhs = gimple_assign_rhs2 (def_stmt); 2266 tree var = TREE_CODE (rhs) == SSA_NAME 2267 ? SSA_NAME_VAR (rhs) 2268 : create_tmp_reg (TREE_TYPE (rhs), NULL); 2269 tree negrhs = make_ssa_name (var, NULL); 2270 gimple negate_stmt = gimple_build_assign_with_ops (NEGATE_EXPR, negrhs, 2271 rhs, NULL); 2272 gimple_stmt_iterator gsi = gsi_for_stmt (def_stmt); 2273 set_vinfo_for_stmt (negate_stmt, new_stmt_vec_info (negate_stmt, 2274 loop_info, NULL)); 2275 gsi_insert_before (&gsi, negate_stmt, GSI_NEW_STMT); 2276 gimple_assign_set_rhs2 (def_stmt, negrhs); 2277 gimple_assign_set_rhs_code (def_stmt, PLUS_EXPR); 2278 update_stmt (def_stmt); 2279 } 2280 2281 /* Reduction is safe. We're dealing with one of the following: 2282 1) integer arithmetic and no trapv 2283 2) floating point arithmetic, and special flags permit this optimization 2284 3) nested cycle (i.e., outer loop vectorization). */ 2285 if (TREE_CODE (op1) == SSA_NAME) 2286 def1 = SSA_NAME_DEF_STMT (op1); 2287 2288 if (TREE_CODE (op2) == SSA_NAME) 2289 def2 = SSA_NAME_DEF_STMT (op2); 2290 2291 if (code != COND_EXPR 2292 && ((!def1 || gimple_nop_p (def1)) && (!def2 || gimple_nop_p (def2)))) 2293 { 2294 if (vect_print_dump_info (REPORT_DETAILS)) 2295 report_vect_op (def_stmt, "reduction: no defs for operands: "); 2296 return NULL; 2297 } 2298 2299 /* Check that one def is the reduction def, defined by PHI, 2300 the other def is either defined in the loop ("vect_internal_def"), 2301 or it's an induction (defined by a loop-header phi-node). */ 2302 2303 if (def2 && def2 == phi 2304 && (code == COND_EXPR 2305 || !def1 || gimple_nop_p (def1) 2306 || (def1 && flow_bb_inside_loop_p (loop, gimple_bb (def1)) 2307 && (is_gimple_assign (def1) 2308 || is_gimple_call (def1) 2309 || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def1)) 2310 == vect_induction_def 2311 || (gimple_code (def1) == GIMPLE_PHI 2312 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def1)) 2313 == vect_internal_def 2314 && !is_loop_header_bb_p (gimple_bb (def1))))))) 2315 { 2316 if (vect_print_dump_info (REPORT_DETAILS)) 2317 report_vect_op (def_stmt, "detected reduction: "); 2318 return def_stmt; 2319 } 2320 2321 if (def1 && def1 == phi 2322 && (code == COND_EXPR 2323 || !def2 || gimple_nop_p (def2) 2324 || (def2 && flow_bb_inside_loop_p (loop, gimple_bb (def2)) 2325 && (is_gimple_assign (def2) 2326 || is_gimple_call (def2) 2327 || STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def2)) 2328 == vect_induction_def 2329 || (gimple_code (def2) == GIMPLE_PHI 2330 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def2)) 2331 == vect_internal_def 2332 && !is_loop_header_bb_p (gimple_bb (def2))))))) 2333 { 2334 if (check_reduction) 2335 { 2336 /* Swap operands (just for simplicity - so that the rest of the code 2337 can assume that the reduction variable is always the last (second) 2338 argument). */ 2339 if (vect_print_dump_info (REPORT_DETAILS)) 2340 report_vect_op (def_stmt, 2341 "detected reduction: need to swap operands: "); 2342 2343 swap_tree_operands (def_stmt, gimple_assign_rhs1_ptr (def_stmt), 2344 gimple_assign_rhs2_ptr (def_stmt)); 2345 } 2346 else 2347 { 2348 if (vect_print_dump_info (REPORT_DETAILS)) 2349 report_vect_op (def_stmt, "detected reduction: "); 2350 } 2351 2352 return def_stmt; 2353 } 2354 2355 /* Try to find SLP reduction chain. */ 2356 if (check_reduction && vect_is_slp_reduction (loop_info, phi, def_stmt)) 2357 { 2358 if (vect_print_dump_info (REPORT_DETAILS)) 2359 report_vect_op (def_stmt, "reduction: detected reduction chain: "); 2360 2361 return def_stmt; 2362 } 2363 2364 if (vect_print_dump_info (REPORT_DETAILS)) 2365 report_vect_op (def_stmt, "reduction: unknown pattern: "); 2366 2367 return NULL; 2368 } 2369 2370 /* Wrapper around vect_is_simple_reduction_1, that won't modify code 2371 in-place. Arguments as there. */ 2372 2373 static gimple 2374 vect_is_simple_reduction (loop_vec_info loop_info, gimple phi, 2375 bool check_reduction, bool *double_reduc) 2376 { 2377 return vect_is_simple_reduction_1 (loop_info, phi, check_reduction, 2378 double_reduc, false); 2379 } 2380 2381 /* Wrapper around vect_is_simple_reduction_1, which will modify code 2382 in-place if it enables detection of more reductions. Arguments 2383 as there. */ 2384 2385 gimple 2386 vect_force_simple_reduction (loop_vec_info loop_info, gimple phi, 2387 bool check_reduction, bool *double_reduc) 2388 { 2389 return vect_is_simple_reduction_1 (loop_info, phi, check_reduction, 2390 double_reduc, true); 2391 } 2392 2393 /* Calculate the cost of one scalar iteration of the loop. */ 2394 int 2395 vect_get_single_scalar_iteration_cost (loop_vec_info loop_vinfo) 2396 { 2397 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); 2398 basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); 2399 int nbbs = loop->num_nodes, factor, scalar_single_iter_cost = 0; 2400 int innerloop_iters, i, stmt_cost; 2401 2402 /* Count statements in scalar loop. Using this as scalar cost for a single 2403 iteration for now. 2404 2405 TODO: Add outer loop support. 2406 2407 TODO: Consider assigning different costs to different scalar 2408 statements. */ 2409 2410 /* FORNOW. */ 2411 innerloop_iters = 1; 2412 if (loop->inner) 2413 innerloop_iters = 50; /* FIXME */ 2414 2415 for (i = 0; i < nbbs; i++) 2416 { 2417 gimple_stmt_iterator si; 2418 basic_block bb = bbs[i]; 2419 2420 if (bb->loop_father == loop->inner) 2421 factor = innerloop_iters; 2422 else 2423 factor = 1; 2424 2425 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) 2426 { 2427 gimple stmt = gsi_stmt (si); 2428 stmt_vec_info stmt_info = vinfo_for_stmt (stmt); 2429 2430 if (!is_gimple_assign (stmt) && !is_gimple_call (stmt)) 2431 continue; 2432 2433 /* Skip stmts that are not vectorized inside the loop. */ 2434 if (stmt_info 2435 && !STMT_VINFO_RELEVANT_P (stmt_info) 2436 && (!STMT_VINFO_LIVE_P (stmt_info) 2437 || !VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (stmt_info))) 2438 && !STMT_VINFO_IN_PATTERN_P (stmt_info)) 2439 continue; 2440 2441 if (STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt))) 2442 { 2443 if (DR_IS_READ (STMT_VINFO_DATA_REF (vinfo_for_stmt (stmt)))) 2444 stmt_cost = vect_get_cost (scalar_load); 2445 else 2446 stmt_cost = vect_get_cost (scalar_store); 2447 } 2448 else 2449 stmt_cost = vect_get_cost (scalar_stmt); 2450 2451 scalar_single_iter_cost += stmt_cost * factor; 2452 } 2453 } 2454 return scalar_single_iter_cost; 2455 } 2456 2457 /* Calculate cost of peeling the loop PEEL_ITERS_PROLOGUE times. */ 2458 int 2459 vect_get_known_peeling_cost (loop_vec_info loop_vinfo, int peel_iters_prologue, 2460 int *peel_iters_epilogue, 2461 int scalar_single_iter_cost) 2462 { 2463 int peel_guard_costs = 0; 2464 int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); 2465 2466 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo)) 2467 { 2468 *peel_iters_epilogue = vf/2; 2469 if (vect_print_dump_info (REPORT_COST)) 2470 fprintf (vect_dump, "cost model: " 2471 "epilogue peel iters set to vf/2 because " 2472 "loop iterations are unknown ."); 2473 2474 /* If peeled iterations are known but number of scalar loop 2475 iterations are unknown, count a taken branch per peeled loop. */ 2476 peel_guard_costs = 2 * vect_get_cost (cond_branch_taken); 2477 } 2478 else 2479 { 2480 int niters = LOOP_VINFO_INT_NITERS (loop_vinfo); 2481 peel_iters_prologue = niters < peel_iters_prologue ? 2482 niters : peel_iters_prologue; 2483 *peel_iters_epilogue = (niters - peel_iters_prologue) % vf; 2484 /* If we need to peel for gaps, but no peeling is required, we have to 2485 peel VF iterations. */ 2486 if (LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo) && !*peel_iters_epilogue) 2487 *peel_iters_epilogue = vf; 2488 } 2489 2490 return (peel_iters_prologue * scalar_single_iter_cost) 2491 + (*peel_iters_epilogue * scalar_single_iter_cost) 2492 + peel_guard_costs; 2493 } 2494 2495 /* Function vect_estimate_min_profitable_iters 2496 2497 Return the number of iterations required for the vector version of the 2498 loop to be profitable relative to the cost of the scalar version of the 2499 loop. 2500 2501 TODO: Take profile info into account before making vectorization 2502 decisions, if available. */ 2503 2504 int 2505 vect_estimate_min_profitable_iters (loop_vec_info loop_vinfo) 2506 { 2507 int i; 2508 int min_profitable_iters; 2509 int peel_iters_prologue; 2510 int peel_iters_epilogue; 2511 int vec_inside_cost = 0; 2512 int vec_outside_cost = 0; 2513 int scalar_single_iter_cost = 0; 2514 int scalar_outside_cost = 0; 2515 int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); 2516 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); 2517 basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); 2518 int nbbs = loop->num_nodes; 2519 int npeel = LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo); 2520 int peel_guard_costs = 0; 2521 int innerloop_iters = 0, factor; 2522 VEC (slp_instance, heap) *slp_instances; 2523 slp_instance instance; 2524 2525 /* Cost model disabled. */ 2526 if (!flag_vect_cost_model) 2527 { 2528 if (vect_print_dump_info (REPORT_COST)) 2529 fprintf (vect_dump, "cost model disabled."); 2530 return 0; 2531 } 2532 2533 /* Requires loop versioning tests to handle misalignment. */ 2534 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo)) 2535 { 2536 /* FIXME: Make cost depend on complexity of individual check. */ 2537 vec_outside_cost += 2538 VEC_length (gimple, LOOP_VINFO_MAY_MISALIGN_STMTS (loop_vinfo)); 2539 if (vect_print_dump_info (REPORT_COST)) 2540 fprintf (vect_dump, "cost model: Adding cost of checks for loop " 2541 "versioning to treat misalignment.\n"); 2542 } 2543 2544 /* Requires loop versioning with alias checks. */ 2545 if (LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo)) 2546 { 2547 /* FIXME: Make cost depend on complexity of individual check. */ 2548 vec_outside_cost += 2549 VEC_length (ddr_p, LOOP_VINFO_MAY_ALIAS_DDRS (loop_vinfo)); 2550 if (vect_print_dump_info (REPORT_COST)) 2551 fprintf (vect_dump, "cost model: Adding cost of checks for loop " 2552 "versioning aliasing.\n"); 2553 } 2554 2555 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo) 2556 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo)) 2557 vec_outside_cost += vect_get_cost (cond_branch_taken); 2558 2559 /* Count statements in scalar loop. Using this as scalar cost for a single 2560 iteration for now. 2561 2562 TODO: Add outer loop support. 2563 2564 TODO: Consider assigning different costs to different scalar 2565 statements. */ 2566 2567 /* FORNOW. */ 2568 if (loop->inner) 2569 innerloop_iters = 50; /* FIXME */ 2570 2571 for (i = 0; i < nbbs; i++) 2572 { 2573 gimple_stmt_iterator si; 2574 basic_block bb = bbs[i]; 2575 2576 if (bb->loop_father == loop->inner) 2577 factor = innerloop_iters; 2578 else 2579 factor = 1; 2580 2581 for (si = gsi_start_bb (bb); !gsi_end_p (si); gsi_next (&si)) 2582 { 2583 gimple stmt = gsi_stmt (si); 2584 stmt_vec_info stmt_info = vinfo_for_stmt (stmt); 2585 2586 if (STMT_VINFO_IN_PATTERN_P (stmt_info)) 2587 { 2588 stmt = STMT_VINFO_RELATED_STMT (stmt_info); 2589 stmt_info = vinfo_for_stmt (stmt); 2590 } 2591 2592 /* Skip stmts that are not vectorized inside the loop. */ 2593 if (!STMT_VINFO_RELEVANT_P (stmt_info) 2594 && (!STMT_VINFO_LIVE_P (stmt_info) 2595 || !VECTORIZABLE_CYCLE_DEF (STMT_VINFO_DEF_TYPE (stmt_info)))) 2596 continue; 2597 2598 vec_inside_cost += STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info) * factor; 2599 /* FIXME: for stmts in the inner-loop in outer-loop vectorization, 2600 some of the "outside" costs are generated inside the outer-loop. */ 2601 vec_outside_cost += STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info); 2602 if (is_pattern_stmt_p (stmt_info) 2603 && STMT_VINFO_PATTERN_DEF_SEQ (stmt_info)) 2604 { 2605 gimple_stmt_iterator gsi; 2606 2607 for (gsi = gsi_start (STMT_VINFO_PATTERN_DEF_SEQ (stmt_info)); 2608 !gsi_end_p (gsi); gsi_next (&gsi)) 2609 { 2610 gimple pattern_def_stmt = gsi_stmt (gsi); 2611 stmt_vec_info pattern_def_stmt_info 2612 = vinfo_for_stmt (pattern_def_stmt); 2613 if (STMT_VINFO_RELEVANT_P (pattern_def_stmt_info) 2614 || STMT_VINFO_LIVE_P (pattern_def_stmt_info)) 2615 { 2616 vec_inside_cost 2617 += STMT_VINFO_INSIDE_OF_LOOP_COST 2618 (pattern_def_stmt_info) * factor; 2619 vec_outside_cost 2620 += STMT_VINFO_OUTSIDE_OF_LOOP_COST 2621 (pattern_def_stmt_info); 2622 } 2623 } 2624 } 2625 } 2626 } 2627 2628 scalar_single_iter_cost = vect_get_single_scalar_iteration_cost (loop_vinfo); 2629 2630 /* Add additional cost for the peeled instructions in prologue and epilogue 2631 loop. 2632 2633 FORNOW: If we don't know the value of peel_iters for prologue or epilogue 2634 at compile-time - we assume it's vf/2 (the worst would be vf-1). 2635 2636 TODO: Build an expression that represents peel_iters for prologue and 2637 epilogue to be used in a run-time test. */ 2638 2639 if (npeel < 0) 2640 { 2641 peel_iters_prologue = vf/2; 2642 if (vect_print_dump_info (REPORT_COST)) 2643 fprintf (vect_dump, "cost model: " 2644 "prologue peel iters set to vf/2."); 2645 2646 /* If peeling for alignment is unknown, loop bound of main loop becomes 2647 unknown. */ 2648 peel_iters_epilogue = vf/2; 2649 if (vect_print_dump_info (REPORT_COST)) 2650 fprintf (vect_dump, "cost model: " 2651 "epilogue peel iters set to vf/2 because " 2652 "peeling for alignment is unknown ."); 2653 2654 /* If peeled iterations are unknown, count a taken branch and a not taken 2655 branch per peeled loop. Even if scalar loop iterations are known, 2656 vector iterations are not known since peeled prologue iterations are 2657 not known. Hence guards remain the same. */ 2658 peel_guard_costs += 2 * (vect_get_cost (cond_branch_taken) 2659 + vect_get_cost (cond_branch_not_taken)); 2660 vec_outside_cost += (peel_iters_prologue * scalar_single_iter_cost) 2661 + (peel_iters_epilogue * scalar_single_iter_cost) 2662 + peel_guard_costs; 2663 } 2664 else 2665 { 2666 peel_iters_prologue = npeel; 2667 vec_outside_cost += vect_get_known_peeling_cost (loop_vinfo, 2668 peel_iters_prologue, &peel_iters_epilogue, 2669 scalar_single_iter_cost); 2670 } 2671 2672 /* FORNOW: The scalar outside cost is incremented in one of the 2673 following ways: 2674 2675 1. The vectorizer checks for alignment and aliasing and generates 2676 a condition that allows dynamic vectorization. A cost model 2677 check is ANDED with the versioning condition. Hence scalar code 2678 path now has the added cost of the versioning check. 2679 2680 if (cost > th & versioning_check) 2681 jmp to vector code 2682 2683 Hence run-time scalar is incremented by not-taken branch cost. 2684 2685 2. The vectorizer then checks if a prologue is required. If the 2686 cost model check was not done before during versioning, it has to 2687 be done before the prologue check. 2688 2689 if (cost <= th) 2690 prologue = scalar_iters 2691 if (prologue == 0) 2692 jmp to vector code 2693 else 2694 execute prologue 2695 if (prologue == num_iters) 2696 go to exit 2697 2698 Hence the run-time scalar cost is incremented by a taken branch, 2699 plus a not-taken branch, plus a taken branch cost. 2700 2701 3. The vectorizer then checks if an epilogue is required. If the 2702 cost model check was not done before during prologue check, it 2703 has to be done with the epilogue check. 2704 2705 if (prologue == 0) 2706 jmp to vector code 2707 else 2708 execute prologue 2709 if (prologue == num_iters) 2710 go to exit 2711 vector code: 2712 if ((cost <= th) | (scalar_iters-prologue-epilogue == 0)) 2713 jmp to epilogue 2714 2715 Hence the run-time scalar cost should be incremented by 2 taken 2716 branches. 2717 2718 TODO: The back end may reorder the BBS's differently and reverse 2719 conditions/branch directions. Change the estimates below to 2720 something more reasonable. */ 2721 2722 /* If the number of iterations is known and we do not do versioning, we can 2723 decide whether to vectorize at compile time. Hence the scalar version 2724 do not carry cost model guard costs. */ 2725 if (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) 2726 || LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo) 2727 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo)) 2728 { 2729 /* Cost model check occurs at versioning. */ 2730 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo) 2731 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo)) 2732 scalar_outside_cost += vect_get_cost (cond_branch_not_taken); 2733 else 2734 { 2735 /* Cost model check occurs at prologue generation. */ 2736 if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo) < 0) 2737 scalar_outside_cost += 2 * vect_get_cost (cond_branch_taken) 2738 + vect_get_cost (cond_branch_not_taken); 2739 /* Cost model check occurs at epilogue generation. */ 2740 else 2741 scalar_outside_cost += 2 * vect_get_cost (cond_branch_taken); 2742 } 2743 } 2744 2745 /* Add SLP costs. */ 2746 slp_instances = LOOP_VINFO_SLP_INSTANCES (loop_vinfo); 2747 FOR_EACH_VEC_ELT (slp_instance, slp_instances, i, instance) 2748 { 2749 vec_outside_cost += SLP_INSTANCE_OUTSIDE_OF_LOOP_COST (instance); 2750 vec_inside_cost += SLP_INSTANCE_INSIDE_OF_LOOP_COST (instance); 2751 } 2752 2753 /* Calculate number of iterations required to make the vector version 2754 profitable, relative to the loop bodies only. The following condition 2755 must hold true: 2756 SIC * niters + SOC > VIC * ((niters-PL_ITERS-EP_ITERS)/VF) + VOC 2757 where 2758 SIC = scalar iteration cost, VIC = vector iteration cost, 2759 VOC = vector outside cost, VF = vectorization factor, 2760 PL_ITERS = prologue iterations, EP_ITERS= epilogue iterations 2761 SOC = scalar outside cost for run time cost model check. */ 2762 2763 if ((scalar_single_iter_cost * vf) > vec_inside_cost) 2764 { 2765 if (vec_outside_cost <= 0) 2766 min_profitable_iters = 1; 2767 else 2768 { 2769 min_profitable_iters = ((vec_outside_cost - scalar_outside_cost) * vf 2770 - vec_inside_cost * peel_iters_prologue 2771 - vec_inside_cost * peel_iters_epilogue) 2772 / ((scalar_single_iter_cost * vf) 2773 - vec_inside_cost); 2774 2775 if ((scalar_single_iter_cost * vf * min_profitable_iters) 2776 <= ((vec_inside_cost * min_profitable_iters) 2777 + ((vec_outside_cost - scalar_outside_cost) * vf))) 2778 min_profitable_iters++; 2779 } 2780 } 2781 /* vector version will never be profitable. */ 2782 else 2783 { 2784 if (vect_print_dump_info (REPORT_COST)) 2785 fprintf (vect_dump, "cost model: the vector iteration cost = %d " 2786 "divided by the scalar iteration cost = %d " 2787 "is greater or equal to the vectorization factor = %d.", 2788 vec_inside_cost, scalar_single_iter_cost, vf); 2789 return -1; 2790 } 2791 2792 if (vect_print_dump_info (REPORT_COST)) 2793 { 2794 fprintf (vect_dump, "Cost model analysis: \n"); 2795 fprintf (vect_dump, " Vector inside of loop cost: %d\n", 2796 vec_inside_cost); 2797 fprintf (vect_dump, " Vector outside of loop cost: %d\n", 2798 vec_outside_cost); 2799 fprintf (vect_dump, " Scalar iteration cost: %d\n", 2800 scalar_single_iter_cost); 2801 fprintf (vect_dump, " Scalar outside cost: %d\n", scalar_outside_cost); 2802 fprintf (vect_dump, " prologue iterations: %d\n", 2803 peel_iters_prologue); 2804 fprintf (vect_dump, " epilogue iterations: %d\n", 2805 peel_iters_epilogue); 2806 fprintf (vect_dump, " Calculated minimum iters for profitability: %d\n", 2807 min_profitable_iters); 2808 } 2809 2810 min_profitable_iters = 2811 min_profitable_iters < vf ? vf : min_profitable_iters; 2812 2813 /* Because the condition we create is: 2814 if (niters <= min_profitable_iters) 2815 then skip the vectorized loop. */ 2816 min_profitable_iters--; 2817 2818 if (vect_print_dump_info (REPORT_COST)) 2819 fprintf (vect_dump, " Profitability threshold = %d\n", 2820 min_profitable_iters); 2821 2822 return min_profitable_iters; 2823 } 2824 2825 2826 /* TODO: Close dependency between vect_model_*_cost and vectorizable_* 2827 functions. Design better to avoid maintenance issues. */ 2828 2829 /* Function vect_model_reduction_cost. 2830 2831 Models cost for a reduction operation, including the vector ops 2832 generated within the strip-mine loop, the initial definition before 2833 the loop, and the epilogue code that must be generated. */ 2834 2835 static bool 2836 vect_model_reduction_cost (stmt_vec_info stmt_info, enum tree_code reduc_code, 2837 int ncopies) 2838 { 2839 int outer_cost = 0; 2840 enum tree_code code; 2841 optab optab; 2842 tree vectype; 2843 gimple stmt, orig_stmt; 2844 tree reduction_op; 2845 enum machine_mode mode; 2846 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); 2847 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); 2848 2849 2850 /* Cost of reduction op inside loop. */ 2851 STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info) 2852 += ncopies * vect_get_cost (vector_stmt); 2853 2854 stmt = STMT_VINFO_STMT (stmt_info); 2855 2856 switch (get_gimple_rhs_class (gimple_assign_rhs_code (stmt))) 2857 { 2858 case GIMPLE_SINGLE_RHS: 2859 gcc_assert (TREE_OPERAND_LENGTH (gimple_assign_rhs1 (stmt)) == ternary_op); 2860 reduction_op = TREE_OPERAND (gimple_assign_rhs1 (stmt), 2); 2861 break; 2862 case GIMPLE_UNARY_RHS: 2863 reduction_op = gimple_assign_rhs1 (stmt); 2864 break; 2865 case GIMPLE_BINARY_RHS: 2866 reduction_op = gimple_assign_rhs2 (stmt); 2867 break; 2868 case GIMPLE_TERNARY_RHS: 2869 reduction_op = gimple_assign_rhs3 (stmt); 2870 break; 2871 default: 2872 gcc_unreachable (); 2873 } 2874 2875 vectype = get_vectype_for_scalar_type (TREE_TYPE (reduction_op)); 2876 if (!vectype) 2877 { 2878 if (vect_print_dump_info (REPORT_COST)) 2879 { 2880 fprintf (vect_dump, "unsupported data-type "); 2881 print_generic_expr (vect_dump, TREE_TYPE (reduction_op), TDF_SLIM); 2882 } 2883 return false; 2884 } 2885 2886 mode = TYPE_MODE (vectype); 2887 orig_stmt = STMT_VINFO_RELATED_STMT (stmt_info); 2888 2889 if (!orig_stmt) 2890 orig_stmt = STMT_VINFO_STMT (stmt_info); 2891 2892 code = gimple_assign_rhs_code (orig_stmt); 2893 2894 /* Add in cost for initial definition. */ 2895 outer_cost += vect_get_cost (scalar_to_vec); 2896 2897 /* Determine cost of epilogue code. 2898 2899 We have a reduction operator that will reduce the vector in one statement. 2900 Also requires scalar extract. */ 2901 2902 if (!nested_in_vect_loop_p (loop, orig_stmt)) 2903 { 2904 if (reduc_code != ERROR_MARK) 2905 outer_cost += vect_get_cost (vector_stmt) 2906 + vect_get_cost (vec_to_scalar); 2907 else 2908 { 2909 int vec_size_in_bits = tree_low_cst (TYPE_SIZE (vectype), 1); 2910 tree bitsize = 2911 TYPE_SIZE (TREE_TYPE (gimple_assign_lhs (orig_stmt))); 2912 int element_bitsize = tree_low_cst (bitsize, 1); 2913 int nelements = vec_size_in_bits / element_bitsize; 2914 2915 optab = optab_for_tree_code (code, vectype, optab_default); 2916 2917 /* We have a whole vector shift available. */ 2918 if (VECTOR_MODE_P (mode) 2919 && optab_handler (optab, mode) != CODE_FOR_nothing 2920 && optab_handler (vec_shr_optab, mode) != CODE_FOR_nothing) 2921 /* Final reduction via vector shifts and the reduction operator. Also 2922 requires scalar extract. */ 2923 outer_cost += ((exact_log2(nelements) * 2) 2924 * vect_get_cost (vector_stmt) 2925 + vect_get_cost (vec_to_scalar)); 2926 else 2927 /* Use extracts and reduction op for final reduction. For N elements, 2928 we have N extracts and N-1 reduction ops. */ 2929 outer_cost += ((nelements + nelements - 1) 2930 * vect_get_cost (vector_stmt)); 2931 } 2932 } 2933 2934 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info) = outer_cost; 2935 2936 if (vect_print_dump_info (REPORT_COST)) 2937 fprintf (vect_dump, "vect_model_reduction_cost: inside_cost = %d, " 2938 "outside_cost = %d .", STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info), 2939 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info)); 2940 2941 return true; 2942 } 2943 2944 2945 /* Function vect_model_induction_cost. 2946 2947 Models cost for induction operations. */ 2948 2949 static void 2950 vect_model_induction_cost (stmt_vec_info stmt_info, int ncopies) 2951 { 2952 /* loop cost for vec_loop. */ 2953 STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info) 2954 = ncopies * vect_get_cost (vector_stmt); 2955 /* prologue cost for vec_init and vec_step. */ 2956 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info) 2957 = 2 * vect_get_cost (scalar_to_vec); 2958 2959 if (vect_print_dump_info (REPORT_COST)) 2960 fprintf (vect_dump, "vect_model_induction_cost: inside_cost = %d, " 2961 "outside_cost = %d .", STMT_VINFO_INSIDE_OF_LOOP_COST (stmt_info), 2962 STMT_VINFO_OUTSIDE_OF_LOOP_COST (stmt_info)); 2963 } 2964 2965 2966 /* Function get_initial_def_for_induction 2967 2968 Input: 2969 STMT - a stmt that performs an induction operation in the loop. 2970 IV_PHI - the initial value of the induction variable 2971 2972 Output: 2973 Return a vector variable, initialized with the first VF values of 2974 the induction variable. E.g., for an iv with IV_PHI='X' and 2975 evolution S, for a vector of 4 units, we want to return: 2976 [X, X + S, X + 2*S, X + 3*S]. */ 2977 2978 static tree 2979 get_initial_def_for_induction (gimple iv_phi) 2980 { 2981 stmt_vec_info stmt_vinfo = vinfo_for_stmt (iv_phi); 2982 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); 2983 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); 2984 tree scalar_type; 2985 tree vectype; 2986 int nunits; 2987 edge pe = loop_preheader_edge (loop); 2988 struct loop *iv_loop; 2989 basic_block new_bb; 2990 tree vec, vec_init, vec_step, t; 2991 tree access_fn; 2992 tree new_var; 2993 tree new_name; 2994 gimple init_stmt, induction_phi, new_stmt; 2995 tree induc_def, vec_def, vec_dest; 2996 tree init_expr, step_expr; 2997 int vf = LOOP_VINFO_VECT_FACTOR (loop_vinfo); 2998 int i; 2999 bool ok; 3000 int ncopies; 3001 tree expr; 3002 stmt_vec_info phi_info = vinfo_for_stmt (iv_phi); 3003 bool nested_in_vect_loop = false; 3004 gimple_seq stmts = NULL; 3005 imm_use_iterator imm_iter; 3006 use_operand_p use_p; 3007 gimple exit_phi; 3008 edge latch_e; 3009 tree loop_arg; 3010 gimple_stmt_iterator si; 3011 basic_block bb = gimple_bb (iv_phi); 3012 tree stepvectype; 3013 tree resvectype; 3014 3015 /* Is phi in an inner-loop, while vectorizing an enclosing outer-loop? */ 3016 if (nested_in_vect_loop_p (loop, iv_phi)) 3017 { 3018 nested_in_vect_loop = true; 3019 iv_loop = loop->inner; 3020 } 3021 else 3022 iv_loop = loop; 3023 gcc_assert (iv_loop == (gimple_bb (iv_phi))->loop_father); 3024 3025 latch_e = loop_latch_edge (iv_loop); 3026 loop_arg = PHI_ARG_DEF_FROM_EDGE (iv_phi, latch_e); 3027 3028 access_fn = analyze_scalar_evolution (iv_loop, PHI_RESULT (iv_phi)); 3029 gcc_assert (access_fn); 3030 STRIP_NOPS (access_fn); 3031 ok = vect_is_simple_iv_evolution (iv_loop->num, access_fn, 3032 &init_expr, &step_expr); 3033 gcc_assert (ok); 3034 pe = loop_preheader_edge (iv_loop); 3035 3036 scalar_type = TREE_TYPE (init_expr); 3037 vectype = get_vectype_for_scalar_type (scalar_type); 3038 resvectype = get_vectype_for_scalar_type (TREE_TYPE (PHI_RESULT (iv_phi))); 3039 gcc_assert (vectype); 3040 nunits = TYPE_VECTOR_SUBPARTS (vectype); 3041 ncopies = vf / nunits; 3042 3043 gcc_assert (phi_info); 3044 gcc_assert (ncopies >= 1); 3045 3046 /* Find the first insertion point in the BB. */ 3047 si = gsi_after_labels (bb); 3048 3049 /* Create the vector that holds the initial_value of the induction. */ 3050 if (nested_in_vect_loop) 3051 { 3052 /* iv_loop is nested in the loop to be vectorized. init_expr had already 3053 been created during vectorization of previous stmts. We obtain it 3054 from the STMT_VINFO_VEC_STMT of the defining stmt. */ 3055 tree iv_def = PHI_ARG_DEF_FROM_EDGE (iv_phi, 3056 loop_preheader_edge (iv_loop)); 3057 vec_init = vect_get_vec_def_for_operand (iv_def, iv_phi, NULL); 3058 } 3059 else 3060 { 3061 /* iv_loop is the loop to be vectorized. Create: 3062 vec_init = [X, X+S, X+2*S, X+3*S] (S = step_expr, X = init_expr) */ 3063 new_var = vect_get_new_vect_var (scalar_type, vect_scalar_var, "var_"); 3064 add_referenced_var (new_var); 3065 3066 new_name = force_gimple_operand (init_expr, &stmts, false, new_var); 3067 if (stmts) 3068 { 3069 new_bb = gsi_insert_seq_on_edge_immediate (pe, stmts); 3070 gcc_assert (!new_bb); 3071 } 3072 3073 t = NULL_TREE; 3074 t = tree_cons (NULL_TREE, new_name, t); 3075 for (i = 1; i < nunits; i++) 3076 { 3077 /* Create: new_name_i = new_name + step_expr */ 3078 enum tree_code code = POINTER_TYPE_P (scalar_type) 3079 ? POINTER_PLUS_EXPR : PLUS_EXPR; 3080 init_stmt = gimple_build_assign_with_ops (code, new_var, 3081 new_name, step_expr); 3082 new_name = make_ssa_name (new_var, init_stmt); 3083 gimple_assign_set_lhs (init_stmt, new_name); 3084 3085 new_bb = gsi_insert_on_edge_immediate (pe, init_stmt); 3086 gcc_assert (!new_bb); 3087 3088 if (vect_print_dump_info (REPORT_DETAILS)) 3089 { 3090 fprintf (vect_dump, "created new init_stmt: "); 3091 print_gimple_stmt (vect_dump, init_stmt, 0, TDF_SLIM); 3092 } 3093 t = tree_cons (NULL_TREE, new_name, t); 3094 } 3095 /* Create a vector from [new_name_0, new_name_1, ..., new_name_nunits-1] */ 3096 vec = build_constructor_from_list (vectype, nreverse (t)); 3097 vec_init = vect_init_vector (iv_phi, vec, vectype, NULL); 3098 } 3099 3100 3101 /* Create the vector that holds the step of the induction. */ 3102 if (nested_in_vect_loop) 3103 /* iv_loop is nested in the loop to be vectorized. Generate: 3104 vec_step = [S, S, S, S] */ 3105 new_name = step_expr; 3106 else 3107 { 3108 /* iv_loop is the loop to be vectorized. Generate: 3109 vec_step = [VF*S, VF*S, VF*S, VF*S] */ 3110 expr = build_int_cst (TREE_TYPE (step_expr), vf); 3111 new_name = fold_build2 (MULT_EXPR, TREE_TYPE (step_expr), 3112 expr, step_expr); 3113 } 3114 3115 t = unshare_expr (new_name); 3116 gcc_assert (CONSTANT_CLASS_P (new_name)); 3117 stepvectype = get_vectype_for_scalar_type (TREE_TYPE (new_name)); 3118 gcc_assert (stepvectype); 3119 vec = build_vector_from_val (stepvectype, t); 3120 vec_step = vect_init_vector (iv_phi, vec, stepvectype, NULL); 3121 3122 3123 /* Create the following def-use cycle: 3124 loop prolog: 3125 vec_init = ... 3126 vec_step = ... 3127 loop: 3128 vec_iv = PHI <vec_init, vec_loop> 3129 ... 3130 STMT 3131 ... 3132 vec_loop = vec_iv + vec_step; */ 3133 3134 /* Create the induction-phi that defines the induction-operand. */ 3135 vec_dest = vect_get_new_vect_var (vectype, vect_simple_var, "vec_iv_"); 3136 add_referenced_var (vec_dest); 3137 induction_phi = create_phi_node (vec_dest, iv_loop->header); 3138 set_vinfo_for_stmt (induction_phi, 3139 new_stmt_vec_info (induction_phi, loop_vinfo, NULL)); 3140 induc_def = PHI_RESULT (induction_phi); 3141 3142 /* Create the iv update inside the loop */ 3143 new_stmt = gimple_build_assign_with_ops (PLUS_EXPR, vec_dest, 3144 induc_def, vec_step); 3145 vec_def = make_ssa_name (vec_dest, new_stmt); 3146 gimple_assign_set_lhs (new_stmt, vec_def); 3147 gsi_insert_before (&si, new_stmt, GSI_SAME_STMT); 3148 set_vinfo_for_stmt (new_stmt, new_stmt_vec_info (new_stmt, loop_vinfo, 3149 NULL)); 3150 3151 /* Set the arguments of the phi node: */ 3152 add_phi_arg (induction_phi, vec_init, pe, UNKNOWN_LOCATION); 3153 add_phi_arg (induction_phi, vec_def, loop_latch_edge (iv_loop), 3154 UNKNOWN_LOCATION); 3155 3156 3157 /* In case that vectorization factor (VF) is bigger than the number 3158 of elements that we can fit in a vectype (nunits), we have to generate 3159 more than one vector stmt - i.e - we need to "unroll" the 3160 vector stmt by a factor VF/nunits. For more details see documentation 3161 in vectorizable_operation. */ 3162 3163 if (ncopies > 1) 3164 { 3165 stmt_vec_info prev_stmt_vinfo; 3166 /* FORNOW. This restriction should be relaxed. */ 3167 gcc_assert (!nested_in_vect_loop); 3168 3169 /* Create the vector that holds the step of the induction. */ 3170 expr = build_int_cst (TREE_TYPE (step_expr), nunits); 3171 new_name = fold_build2 (MULT_EXPR, TREE_TYPE (step_expr), 3172 expr, step_expr); 3173 t = unshare_expr (new_name); 3174 gcc_assert (CONSTANT_CLASS_P (new_name)); 3175 vec = build_vector_from_val (stepvectype, t); 3176 vec_step = vect_init_vector (iv_phi, vec, stepvectype, NULL); 3177 3178 vec_def = induc_def; 3179 prev_stmt_vinfo = vinfo_for_stmt (induction_phi); 3180 for (i = 1; i < ncopies; i++) 3181 { 3182 /* vec_i = vec_prev + vec_step */ 3183 new_stmt = gimple_build_assign_with_ops (PLUS_EXPR, vec_dest, 3184 vec_def, vec_step); 3185 vec_def = make_ssa_name (vec_dest, new_stmt); 3186 gimple_assign_set_lhs (new_stmt, vec_def); 3187 3188 gsi_insert_before (&si, new_stmt, GSI_SAME_STMT); 3189 if (!useless_type_conversion_p (resvectype, vectype)) 3190 { 3191 new_stmt = gimple_build_assign_with_ops 3192 (VIEW_CONVERT_EXPR, 3193 vect_get_new_vect_var (resvectype, vect_simple_var, 3194 "vec_iv_"), 3195 build1 (VIEW_CONVERT_EXPR, resvectype, 3196 gimple_assign_lhs (new_stmt)), NULL_TREE); 3197 gimple_assign_set_lhs (new_stmt, 3198 make_ssa_name 3199 (gimple_assign_lhs (new_stmt), new_stmt)); 3200 gsi_insert_before (&si, new_stmt, GSI_SAME_STMT); 3201 } 3202 set_vinfo_for_stmt (new_stmt, 3203 new_stmt_vec_info (new_stmt, loop_vinfo, NULL)); 3204 STMT_VINFO_RELATED_STMT (prev_stmt_vinfo) = new_stmt; 3205 prev_stmt_vinfo = vinfo_for_stmt (new_stmt); 3206 } 3207 } 3208 3209 if (nested_in_vect_loop) 3210 { 3211 /* Find the loop-closed exit-phi of the induction, and record 3212 the final vector of induction results: */ 3213 exit_phi = NULL; 3214 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, loop_arg) 3215 { 3216 if (!flow_bb_inside_loop_p (iv_loop, gimple_bb (USE_STMT (use_p)))) 3217 { 3218 exit_phi = USE_STMT (use_p); 3219 break; 3220 } 3221 } 3222 if (exit_phi) 3223 { 3224 stmt_vec_info stmt_vinfo = vinfo_for_stmt (exit_phi); 3225 /* FORNOW. Currently not supporting the case that an inner-loop induction 3226 is not used in the outer-loop (i.e. only outside the outer-loop). */ 3227 gcc_assert (STMT_VINFO_RELEVANT_P (stmt_vinfo) 3228 && !STMT_VINFO_LIVE_P (stmt_vinfo)); 3229 3230 STMT_VINFO_VEC_STMT (stmt_vinfo) = new_stmt; 3231 if (vect_print_dump_info (REPORT_DETAILS)) 3232 { 3233 fprintf (vect_dump, "vector of inductions after inner-loop:"); 3234 print_gimple_stmt (vect_dump, new_stmt, 0, TDF_SLIM); 3235 } 3236 } 3237 } 3238 3239 3240 if (vect_print_dump_info (REPORT_DETAILS)) 3241 { 3242 fprintf (vect_dump, "transform induction: created def-use cycle: "); 3243 print_gimple_stmt (vect_dump, induction_phi, 0, TDF_SLIM); 3244 fprintf (vect_dump, "\n"); 3245 print_gimple_stmt (vect_dump, SSA_NAME_DEF_STMT (vec_def), 0, TDF_SLIM); 3246 } 3247 3248 STMT_VINFO_VEC_STMT (phi_info) = induction_phi; 3249 if (!useless_type_conversion_p (resvectype, vectype)) 3250 { 3251 new_stmt = gimple_build_assign_with_ops 3252 (VIEW_CONVERT_EXPR, 3253 vect_get_new_vect_var (resvectype, vect_simple_var, "vec_iv_"), 3254 build1 (VIEW_CONVERT_EXPR, resvectype, induc_def), NULL_TREE); 3255 induc_def = make_ssa_name (gimple_assign_lhs (new_stmt), new_stmt); 3256 gimple_assign_set_lhs (new_stmt, induc_def); 3257 si = gsi_start_bb (bb); 3258 gsi_insert_before (&si, new_stmt, GSI_SAME_STMT); 3259 set_vinfo_for_stmt (new_stmt, 3260 new_stmt_vec_info (new_stmt, loop_vinfo, NULL)); 3261 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (new_stmt)) 3262 = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (induction_phi)); 3263 } 3264 3265 return induc_def; 3266 } 3267 3268 3269 /* Function get_initial_def_for_reduction 3270 3271 Input: 3272 STMT - a stmt that performs a reduction operation in the loop. 3273 INIT_VAL - the initial value of the reduction variable 3274 3275 Output: 3276 ADJUSTMENT_DEF - a tree that holds a value to be added to the final result 3277 of the reduction (used for adjusting the epilog - see below). 3278 Return a vector variable, initialized according to the operation that STMT 3279 performs. This vector will be used as the initial value of the 3280 vector of partial results. 3281 3282 Option1 (adjust in epilog): Initialize the vector as follows: 3283 add/bit or/xor: [0,0,...,0,0] 3284 mult/bit and: [1,1,...,1,1] 3285 min/max/cond_expr: [init_val,init_val,..,init_val,init_val] 3286 and when necessary (e.g. add/mult case) let the caller know 3287 that it needs to adjust the result by init_val. 3288 3289 Option2: Initialize the vector as follows: 3290 add/bit or/xor: [init_val,0,0,...,0] 3291 mult/bit and: [init_val,1,1,...,1] 3292 min/max/cond_expr: [init_val,init_val,...,init_val] 3293 and no adjustments are needed. 3294 3295 For example, for the following code: 3296 3297 s = init_val; 3298 for (i=0;i<n;i++) 3299 s = s + a[i]; 3300 3301 STMT is 's = s + a[i]', and the reduction variable is 's'. 3302 For a vector of 4 units, we want to return either [0,0,0,init_val], 3303 or [0,0,0,0] and let the caller know that it needs to adjust 3304 the result at the end by 'init_val'. 3305 3306 FORNOW, we are using the 'adjust in epilog' scheme, because this way the 3307 initialization vector is simpler (same element in all entries), if 3308 ADJUSTMENT_DEF is not NULL, and Option2 otherwise. 3309 3310 A cost model should help decide between these two schemes. */ 3311 3312 tree 3313 get_initial_def_for_reduction (gimple stmt, tree init_val, 3314 tree *adjustment_def) 3315 { 3316 stmt_vec_info stmt_vinfo = vinfo_for_stmt (stmt); 3317 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_vinfo); 3318 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); 3319 tree scalar_type = TREE_TYPE (init_val); 3320 tree vectype = get_vectype_for_scalar_type (scalar_type); 3321 int nunits; 3322 enum tree_code code = gimple_assign_rhs_code (stmt); 3323 tree def_for_init; 3324 tree init_def; 3325 tree t = NULL_TREE; 3326 int i; 3327 bool nested_in_vect_loop = false; 3328 tree init_value; 3329 REAL_VALUE_TYPE real_init_val = dconst0; 3330 int int_init_val = 0; 3331 gimple def_stmt = NULL; 3332 3333 gcc_assert (vectype); 3334 nunits = TYPE_VECTOR_SUBPARTS (vectype); 3335 3336 gcc_assert (POINTER_TYPE_P (scalar_type) || INTEGRAL_TYPE_P (scalar_type) 3337 || SCALAR_FLOAT_TYPE_P (scalar_type)); 3338 3339 if (nested_in_vect_loop_p (loop, stmt)) 3340 nested_in_vect_loop = true; 3341 else 3342 gcc_assert (loop == (gimple_bb (stmt))->loop_father); 3343 3344 /* In case of double reduction we only create a vector variable to be put 3345 in the reduction phi node. The actual statement creation is done in 3346 vect_create_epilog_for_reduction. */ 3347 if (adjustment_def && nested_in_vect_loop 3348 && TREE_CODE (init_val) == SSA_NAME 3349 && (def_stmt = SSA_NAME_DEF_STMT (init_val)) 3350 && gimple_code (def_stmt) == GIMPLE_PHI 3351 && flow_bb_inside_loop_p (loop, gimple_bb (def_stmt)) 3352 && vinfo_for_stmt (def_stmt) 3353 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_stmt)) 3354 == vect_double_reduction_def) 3355 { 3356 *adjustment_def = NULL; 3357 return vect_create_destination_var (init_val, vectype); 3358 } 3359 3360 if (TREE_CONSTANT (init_val)) 3361 { 3362 if (SCALAR_FLOAT_TYPE_P (scalar_type)) 3363 init_value = build_real (scalar_type, TREE_REAL_CST (init_val)); 3364 else 3365 init_value = build_int_cst (scalar_type, TREE_INT_CST_LOW (init_val)); 3366 } 3367 else 3368 init_value = init_val; 3369 3370 switch (code) 3371 { 3372 case WIDEN_SUM_EXPR: 3373 case DOT_PROD_EXPR: 3374 case PLUS_EXPR: 3375 case MINUS_EXPR: 3376 case BIT_IOR_EXPR: 3377 case BIT_XOR_EXPR: 3378 case MULT_EXPR: 3379 case BIT_AND_EXPR: 3380 /* ADJUSMENT_DEF is NULL when called from 3381 vect_create_epilog_for_reduction to vectorize double reduction. */ 3382 if (adjustment_def) 3383 { 3384 if (nested_in_vect_loop) 3385 *adjustment_def = vect_get_vec_def_for_operand (init_val, stmt, 3386 NULL); 3387 else 3388 *adjustment_def = init_val; 3389 } 3390 3391 if (code == MULT_EXPR) 3392 { 3393 real_init_val = dconst1; 3394 int_init_val = 1; 3395 } 3396 3397 if (code == BIT_AND_EXPR) 3398 int_init_val = -1; 3399 3400 if (SCALAR_FLOAT_TYPE_P (scalar_type)) 3401 def_for_init = build_real (scalar_type, real_init_val); 3402 else 3403 def_for_init = build_int_cst (scalar_type, int_init_val); 3404 3405 /* Create a vector of '0' or '1' except the first element. */ 3406 for (i = nunits - 2; i >= 0; --i) 3407 t = tree_cons (NULL_TREE, def_for_init, t); 3408 3409 /* Option1: the first element is '0' or '1' as well. */ 3410 if (adjustment_def) 3411 { 3412 t = tree_cons (NULL_TREE, def_for_init, t); 3413 init_def = build_vector (vectype, t); 3414 break; 3415 } 3416 3417 /* Option2: the first element is INIT_VAL. */ 3418 t = tree_cons (NULL_TREE, init_value, t); 3419 if (TREE_CONSTANT (init_val)) 3420 init_def = build_vector (vectype, t); 3421 else 3422 init_def = build_constructor_from_list (vectype, t); 3423 3424 break; 3425 3426 case MIN_EXPR: 3427 case MAX_EXPR: 3428 case COND_EXPR: 3429 if (adjustment_def) 3430 { 3431 *adjustment_def = NULL_TREE; 3432 init_def = vect_get_vec_def_for_operand (init_val, stmt, NULL); 3433 break; 3434 } 3435 3436 init_def = build_vector_from_val (vectype, init_value); 3437 break; 3438 3439 default: 3440 gcc_unreachable (); 3441 } 3442 3443 return init_def; 3444 } 3445 3446 3447 /* Function vect_create_epilog_for_reduction 3448 3449 Create code at the loop-epilog to finalize the result of a reduction 3450 computation. 3451 3452 VECT_DEFS is list of vector of partial results, i.e., the lhs's of vector 3453 reduction statements. 3454 STMT is the scalar reduction stmt that is being vectorized. 3455 NCOPIES is > 1 in case the vectorization factor (VF) is bigger than the 3456 number of elements that we can fit in a vectype (nunits). In this case 3457 we have to generate more than one vector stmt - i.e - we need to "unroll" 3458 the vector stmt by a factor VF/nunits. For more details see documentation 3459 in vectorizable_operation. 3460 REDUC_CODE is the tree-code for the epilog reduction. 3461 REDUCTION_PHIS is a list of the phi-nodes that carry the reduction 3462 computation. 3463 REDUC_INDEX is the index of the operand in the right hand side of the 3464 statement that is defined by REDUCTION_PHI. 3465 DOUBLE_REDUC is TRUE if double reduction phi nodes should be handled. 3466 SLP_NODE is an SLP node containing a group of reduction statements. The 3467 first one in this group is STMT. 3468 3469 This function: 3470 1. Creates the reduction def-use cycles: sets the arguments for 3471 REDUCTION_PHIS: 3472 The loop-entry argument is the vectorized initial-value of the reduction. 3473 The loop-latch argument is taken from VECT_DEFS - the vector of partial 3474 sums. 3475 2. "Reduces" each vector of partial results VECT_DEFS into a single result, 3476 by applying the operation specified by REDUC_CODE if available, or by 3477 other means (whole-vector shifts or a scalar loop). 3478 The function also creates a new phi node at the loop exit to preserve 3479 loop-closed form, as illustrated below. 3480 3481 The flow at the entry to this function: 3482 3483 loop: 3484 vec_def = phi <null, null> # REDUCTION_PHI 3485 VECT_DEF = vector_stmt # vectorized form of STMT 3486 s_loop = scalar_stmt # (scalar) STMT 3487 loop_exit: 3488 s_out0 = phi <s_loop> # (scalar) EXIT_PHI 3489 use <s_out0> 3490 use <s_out0> 3491 3492 The above is transformed by this function into: 3493 3494 loop: 3495 vec_def = phi <vec_init, VECT_DEF> # REDUCTION_PHI 3496 VECT_DEF = vector_stmt # vectorized form of STMT 3497 s_loop = scalar_stmt # (scalar) STMT 3498 loop_exit: 3499 s_out0 = phi <s_loop> # (scalar) EXIT_PHI 3500 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI 3501 v_out2 = reduce <v_out1> 3502 s_out3 = extract_field <v_out2, 0> 3503 s_out4 = adjust_result <s_out3> 3504 use <s_out4> 3505 use <s_out4> 3506 */ 3507 3508 static void 3509 vect_create_epilog_for_reduction (VEC (tree, heap) *vect_defs, gimple stmt, 3510 int ncopies, enum tree_code reduc_code, 3511 VEC (gimple, heap) *reduction_phis, 3512 int reduc_index, bool double_reduc, 3513 slp_tree slp_node) 3514 { 3515 stmt_vec_info stmt_info = vinfo_for_stmt (stmt); 3516 stmt_vec_info prev_phi_info; 3517 tree vectype; 3518 enum machine_mode mode; 3519 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); 3520 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo), *outer_loop = NULL; 3521 basic_block exit_bb; 3522 tree scalar_dest; 3523 tree scalar_type; 3524 gimple new_phi = NULL, phi; 3525 gimple_stmt_iterator exit_gsi; 3526 tree vec_dest; 3527 tree new_temp = NULL_TREE, new_dest, new_name, new_scalar_dest; 3528 gimple epilog_stmt = NULL; 3529 enum tree_code code = gimple_assign_rhs_code (stmt); 3530 gimple exit_phi; 3531 tree bitsize, bitpos; 3532 tree adjustment_def = NULL; 3533 tree vec_initial_def = NULL; 3534 tree reduction_op, expr, def; 3535 tree orig_name, scalar_result; 3536 imm_use_iterator imm_iter, phi_imm_iter; 3537 use_operand_p use_p, phi_use_p; 3538 bool extract_scalar_result = false; 3539 gimple use_stmt, orig_stmt, reduction_phi = NULL; 3540 bool nested_in_vect_loop = false; 3541 VEC (gimple, heap) *new_phis = NULL; 3542 VEC (gimple, heap) *inner_phis = NULL; 3543 enum vect_def_type dt = vect_unknown_def_type; 3544 int j, i; 3545 VEC (tree, heap) *scalar_results = NULL; 3546 unsigned int group_size = 1, k, ratio; 3547 VEC (tree, heap) *vec_initial_defs = NULL; 3548 VEC (gimple, heap) *phis; 3549 bool slp_reduc = false; 3550 tree new_phi_result; 3551 gimple inner_phi = NULL; 3552 3553 if (slp_node) 3554 group_size = VEC_length (gimple, SLP_TREE_SCALAR_STMTS (slp_node)); 3555 3556 if (nested_in_vect_loop_p (loop, stmt)) 3557 { 3558 outer_loop = loop; 3559 loop = loop->inner; 3560 nested_in_vect_loop = true; 3561 gcc_assert (!slp_node); 3562 } 3563 3564 switch (get_gimple_rhs_class (gimple_assign_rhs_code (stmt))) 3565 { 3566 case GIMPLE_SINGLE_RHS: 3567 gcc_assert (TREE_OPERAND_LENGTH (gimple_assign_rhs1 (stmt)) 3568 == ternary_op); 3569 reduction_op = TREE_OPERAND (gimple_assign_rhs1 (stmt), reduc_index); 3570 break; 3571 case GIMPLE_UNARY_RHS: 3572 reduction_op = gimple_assign_rhs1 (stmt); 3573 break; 3574 case GIMPLE_BINARY_RHS: 3575 reduction_op = reduc_index ? 3576 gimple_assign_rhs2 (stmt) : gimple_assign_rhs1 (stmt); 3577 break; 3578 case GIMPLE_TERNARY_RHS: 3579 reduction_op = gimple_op (stmt, reduc_index + 1); 3580 break; 3581 default: 3582 gcc_unreachable (); 3583 } 3584 3585 vectype = get_vectype_for_scalar_type (TREE_TYPE (reduction_op)); 3586 gcc_assert (vectype); 3587 mode = TYPE_MODE (vectype); 3588 3589 /* 1. Create the reduction def-use cycle: 3590 Set the arguments of REDUCTION_PHIS, i.e., transform 3591 3592 loop: 3593 vec_def = phi <null, null> # REDUCTION_PHI 3594 VECT_DEF = vector_stmt # vectorized form of STMT 3595 ... 3596 3597 into: 3598 3599 loop: 3600 vec_def = phi <vec_init, VECT_DEF> # REDUCTION_PHI 3601 VECT_DEF = vector_stmt # vectorized form of STMT 3602 ... 3603 3604 (in case of SLP, do it for all the phis). */ 3605 3606 /* Get the loop-entry arguments. */ 3607 if (slp_node) 3608 vect_get_vec_defs (reduction_op, NULL_TREE, stmt, &vec_initial_defs, 3609 NULL, slp_node, reduc_index); 3610 else 3611 { 3612 vec_initial_defs = VEC_alloc (tree, heap, 1); 3613 /* For the case of reduction, vect_get_vec_def_for_operand returns 3614 the scalar def before the loop, that defines the initial value 3615 of the reduction variable. */ 3616 vec_initial_def = vect_get_vec_def_for_operand (reduction_op, stmt, 3617 &adjustment_def); 3618 VEC_quick_push (tree, vec_initial_defs, vec_initial_def); 3619 } 3620 3621 /* Set phi nodes arguments. */ 3622 FOR_EACH_VEC_ELT (gimple, reduction_phis, i, phi) 3623 { 3624 tree vec_init_def = VEC_index (tree, vec_initial_defs, i); 3625 tree def = VEC_index (tree, vect_defs, i); 3626 for (j = 0; j < ncopies; j++) 3627 { 3628 /* Set the loop-entry arg of the reduction-phi. */ 3629 add_phi_arg (phi, vec_init_def, loop_preheader_edge (loop), 3630 UNKNOWN_LOCATION); 3631 3632 /* Set the loop-latch arg for the reduction-phi. */ 3633 if (j > 0) 3634 def = vect_get_vec_def_for_stmt_copy (vect_unknown_def_type, def); 3635 3636 add_phi_arg (phi, def, loop_latch_edge (loop), UNKNOWN_LOCATION); 3637 3638 if (vect_print_dump_info (REPORT_DETAILS)) 3639 { 3640 fprintf (vect_dump, "transform reduction: created def-use" 3641 " cycle: "); 3642 print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM); 3643 fprintf (vect_dump, "\n"); 3644 print_gimple_stmt (vect_dump, SSA_NAME_DEF_STMT (def), 0, 3645 TDF_SLIM); 3646 } 3647 3648 phi = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (phi)); 3649 } 3650 } 3651 3652 VEC_free (tree, heap, vec_initial_defs); 3653 3654 /* 2. Create epilog code. 3655 The reduction epilog code operates across the elements of the vector 3656 of partial results computed by the vectorized loop. 3657 The reduction epilog code consists of: 3658 3659 step 1: compute the scalar result in a vector (v_out2) 3660 step 2: extract the scalar result (s_out3) from the vector (v_out2) 3661 step 3: adjust the scalar result (s_out3) if needed. 3662 3663 Step 1 can be accomplished using one the following three schemes: 3664 (scheme 1) using reduc_code, if available. 3665 (scheme 2) using whole-vector shifts, if available. 3666 (scheme 3) using a scalar loop. In this case steps 1+2 above are 3667 combined. 3668 3669 The overall epilog code looks like this: 3670 3671 s_out0 = phi <s_loop> # original EXIT_PHI 3672 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI 3673 v_out2 = reduce <v_out1> # step 1 3674 s_out3 = extract_field <v_out2, 0> # step 2 3675 s_out4 = adjust_result <s_out3> # step 3 3676 3677 (step 3 is optional, and steps 1 and 2 may be combined). 3678 Lastly, the uses of s_out0 are replaced by s_out4. */ 3679 3680 3681 /* 2.1 Create new loop-exit-phis to preserve loop-closed form: 3682 v_out1 = phi <VECT_DEF> 3683 Store them in NEW_PHIS. */ 3684 3685 exit_bb = single_exit (loop)->dest; 3686 prev_phi_info = NULL; 3687 new_phis = VEC_alloc (gimple, heap, VEC_length (tree, vect_defs)); 3688 FOR_EACH_VEC_ELT (tree, vect_defs, i, def) 3689 { 3690 for (j = 0; j < ncopies; j++) 3691 { 3692 phi = create_phi_node (SSA_NAME_VAR (def), exit_bb); 3693 set_vinfo_for_stmt (phi, new_stmt_vec_info (phi, loop_vinfo, NULL)); 3694 if (j == 0) 3695 VEC_quick_push (gimple, new_phis, phi); 3696 else 3697 { 3698 def = vect_get_vec_def_for_stmt_copy (dt, def); 3699 STMT_VINFO_RELATED_STMT (prev_phi_info) = phi; 3700 } 3701 3702 SET_PHI_ARG_DEF (phi, single_exit (loop)->dest_idx, def); 3703 prev_phi_info = vinfo_for_stmt (phi); 3704 } 3705 } 3706 3707 /* The epilogue is created for the outer-loop, i.e., for the loop being 3708 vectorized. Create exit phis for the outer loop. */ 3709 if (double_reduc) 3710 { 3711 loop = outer_loop; 3712 exit_bb = single_exit (loop)->dest; 3713 inner_phis = VEC_alloc (gimple, heap, VEC_length (tree, vect_defs)); 3714 FOR_EACH_VEC_ELT (gimple, new_phis, i, phi) 3715 { 3716 gimple outer_phi = create_phi_node (SSA_NAME_VAR (PHI_RESULT (phi)), 3717 exit_bb); 3718 SET_PHI_ARG_DEF (outer_phi, single_exit (loop)->dest_idx, 3719 PHI_RESULT (phi)); 3720 set_vinfo_for_stmt (outer_phi, new_stmt_vec_info (outer_phi, 3721 loop_vinfo, NULL)); 3722 VEC_quick_push (gimple, inner_phis, phi); 3723 VEC_replace (gimple, new_phis, i, outer_phi); 3724 prev_phi_info = vinfo_for_stmt (outer_phi); 3725 while (STMT_VINFO_RELATED_STMT (vinfo_for_stmt (phi))) 3726 { 3727 phi = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (phi)); 3728 outer_phi = create_phi_node (SSA_NAME_VAR (PHI_RESULT (phi)), 3729 exit_bb); 3730 SET_PHI_ARG_DEF (outer_phi, single_exit (loop)->dest_idx, 3731 PHI_RESULT (phi)); 3732 set_vinfo_for_stmt (outer_phi, new_stmt_vec_info (outer_phi, 3733 loop_vinfo, NULL)); 3734 STMT_VINFO_RELATED_STMT (prev_phi_info) = outer_phi; 3735 prev_phi_info = vinfo_for_stmt (outer_phi); 3736 } 3737 } 3738 } 3739 3740 exit_gsi = gsi_after_labels (exit_bb); 3741 3742 /* 2.2 Get the relevant tree-code to use in the epilog for schemes 2,3 3743 (i.e. when reduc_code is not available) and in the final adjustment 3744 code (if needed). Also get the original scalar reduction variable as 3745 defined in the loop. In case STMT is a "pattern-stmt" (i.e. - it 3746 represents a reduction pattern), the tree-code and scalar-def are 3747 taken from the original stmt that the pattern-stmt (STMT) replaces. 3748 Otherwise (it is a regular reduction) - the tree-code and scalar-def 3749 are taken from STMT. */ 3750 3751 orig_stmt = STMT_VINFO_RELATED_STMT (stmt_info); 3752 if (!orig_stmt) 3753 { 3754 /* Regular reduction */ 3755 orig_stmt = stmt; 3756 } 3757 else 3758 { 3759 /* Reduction pattern */ 3760 stmt_vec_info stmt_vinfo = vinfo_for_stmt (orig_stmt); 3761 gcc_assert (STMT_VINFO_IN_PATTERN_P (stmt_vinfo)); 3762 gcc_assert (STMT_VINFO_RELATED_STMT (stmt_vinfo) == stmt); 3763 } 3764 3765 code = gimple_assign_rhs_code (orig_stmt); 3766 /* For MINUS_EXPR the initial vector is [init_val,0,...,0], therefore, 3767 partial results are added and not subtracted. */ 3768 if (code == MINUS_EXPR) 3769 code = PLUS_EXPR; 3770 3771 scalar_dest = gimple_assign_lhs (orig_stmt); 3772 scalar_type = TREE_TYPE (scalar_dest); 3773 scalar_results = VEC_alloc (tree, heap, group_size); 3774 new_scalar_dest = vect_create_destination_var (scalar_dest, NULL); 3775 bitsize = TYPE_SIZE (scalar_type); 3776 3777 /* In case this is a reduction in an inner-loop while vectorizing an outer 3778 loop - we don't need to extract a single scalar result at the end of the 3779 inner-loop (unless it is double reduction, i.e., the use of reduction is 3780 outside the outer-loop). The final vector of partial results will be used 3781 in the vectorized outer-loop, or reduced to a scalar result at the end of 3782 the outer-loop. */ 3783 if (nested_in_vect_loop && !double_reduc) 3784 goto vect_finalize_reduction; 3785 3786 /* SLP reduction without reduction chain, e.g., 3787 # a1 = phi <a2, a0> 3788 # b1 = phi <b2, b0> 3789 a2 = operation (a1) 3790 b2 = operation (b1) */ 3791 slp_reduc = (slp_node && !GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))); 3792 3793 /* In case of reduction chain, e.g., 3794 # a1 = phi <a3, a0> 3795 a2 = operation (a1) 3796 a3 = operation (a2), 3797 3798 we may end up with more than one vector result. Here we reduce them to 3799 one vector. */ 3800 if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))) 3801 { 3802 tree first_vect = PHI_RESULT (VEC_index (gimple, new_phis, 0)); 3803 tree tmp; 3804 gimple new_vec_stmt = NULL; 3805 3806 vec_dest = vect_create_destination_var (scalar_dest, vectype); 3807 for (k = 1; k < VEC_length (gimple, new_phis); k++) 3808 { 3809 gimple next_phi = VEC_index (gimple, new_phis, k); 3810 tree second_vect = PHI_RESULT (next_phi); 3811 3812 tmp = build2 (code, vectype, first_vect, second_vect); 3813 new_vec_stmt = gimple_build_assign (vec_dest, tmp); 3814 first_vect = make_ssa_name (vec_dest, new_vec_stmt); 3815 gimple_assign_set_lhs (new_vec_stmt, first_vect); 3816 gsi_insert_before (&exit_gsi, new_vec_stmt, GSI_SAME_STMT); 3817 } 3818 3819 new_phi_result = first_vect; 3820 if (new_vec_stmt) 3821 { 3822 VEC_truncate (gimple, new_phis, 0); 3823 VEC_safe_push (gimple, heap, new_phis, new_vec_stmt); 3824 } 3825 } 3826 else 3827 new_phi_result = PHI_RESULT (VEC_index (gimple, new_phis, 0)); 3828 3829 /* 2.3 Create the reduction code, using one of the three schemes described 3830 above. In SLP we simply need to extract all the elements from the 3831 vector (without reducing them), so we use scalar shifts. */ 3832 if (reduc_code != ERROR_MARK && !slp_reduc) 3833 { 3834 tree tmp; 3835 3836 /*** Case 1: Create: 3837 v_out2 = reduc_expr <v_out1> */ 3838 3839 if (vect_print_dump_info (REPORT_DETAILS)) 3840 fprintf (vect_dump, "Reduce using direct vector reduction."); 3841 3842 vec_dest = vect_create_destination_var (scalar_dest, vectype); 3843 tmp = build1 (reduc_code, vectype, new_phi_result); 3844 epilog_stmt = gimple_build_assign (vec_dest, tmp); 3845 new_temp = make_ssa_name (vec_dest, epilog_stmt); 3846 gimple_assign_set_lhs (epilog_stmt, new_temp); 3847 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); 3848 3849 extract_scalar_result = true; 3850 } 3851 else 3852 { 3853 enum tree_code shift_code = ERROR_MARK; 3854 bool have_whole_vector_shift = true; 3855 int bit_offset; 3856 int element_bitsize = tree_low_cst (bitsize, 1); 3857 int vec_size_in_bits = tree_low_cst (TYPE_SIZE (vectype), 1); 3858 tree vec_temp; 3859 3860 if (optab_handler (vec_shr_optab, mode) != CODE_FOR_nothing) 3861 shift_code = VEC_RSHIFT_EXPR; 3862 else 3863 have_whole_vector_shift = false; 3864 3865 /* Regardless of whether we have a whole vector shift, if we're 3866 emulating the operation via tree-vect-generic, we don't want 3867 to use it. Only the first round of the reduction is likely 3868 to still be profitable via emulation. */ 3869 /* ??? It might be better to emit a reduction tree code here, so that 3870 tree-vect-generic can expand the first round via bit tricks. */ 3871 if (!VECTOR_MODE_P (mode)) 3872 have_whole_vector_shift = false; 3873 else 3874 { 3875 optab optab = optab_for_tree_code (code, vectype, optab_default); 3876 if (optab_handler (optab, mode) == CODE_FOR_nothing) 3877 have_whole_vector_shift = false; 3878 } 3879 3880 if (have_whole_vector_shift && !slp_reduc) 3881 { 3882 /*** Case 2: Create: 3883 for (offset = VS/2; offset >= element_size; offset/=2) 3884 { 3885 Create: va' = vec_shift <va, offset> 3886 Create: va = vop <va, va'> 3887 } */ 3888 3889 if (vect_print_dump_info (REPORT_DETAILS)) 3890 fprintf (vect_dump, "Reduce using vector shifts"); 3891 3892 vec_dest = vect_create_destination_var (scalar_dest, vectype); 3893 new_temp = new_phi_result; 3894 for (bit_offset = vec_size_in_bits/2; 3895 bit_offset >= element_bitsize; 3896 bit_offset /= 2) 3897 { 3898 tree bitpos = size_int (bit_offset); 3899 3900 epilog_stmt = gimple_build_assign_with_ops (shift_code, 3901 vec_dest, new_temp, bitpos); 3902 new_name = make_ssa_name (vec_dest, epilog_stmt); 3903 gimple_assign_set_lhs (epilog_stmt, new_name); 3904 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); 3905 3906 epilog_stmt = gimple_build_assign_with_ops (code, vec_dest, 3907 new_name, new_temp); 3908 new_temp = make_ssa_name (vec_dest, epilog_stmt); 3909 gimple_assign_set_lhs (epilog_stmt, new_temp); 3910 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); 3911 } 3912 3913 extract_scalar_result = true; 3914 } 3915 else 3916 { 3917 tree rhs; 3918 3919 /*** Case 3: Create: 3920 s = extract_field <v_out2, 0> 3921 for (offset = element_size; 3922 offset < vector_size; 3923 offset += element_size;) 3924 { 3925 Create: s' = extract_field <v_out2, offset> 3926 Create: s = op <s, s'> // For non SLP cases 3927 } */ 3928 3929 if (vect_print_dump_info (REPORT_DETAILS)) 3930 fprintf (vect_dump, "Reduce using scalar code. "); 3931 3932 vec_size_in_bits = tree_low_cst (TYPE_SIZE (vectype), 1); 3933 FOR_EACH_VEC_ELT (gimple, new_phis, i, new_phi) 3934 { 3935 if (gimple_code (new_phi) == GIMPLE_PHI) 3936 vec_temp = PHI_RESULT (new_phi); 3937 else 3938 vec_temp = gimple_assign_lhs (new_phi); 3939 rhs = build3 (BIT_FIELD_REF, scalar_type, vec_temp, bitsize, 3940 bitsize_zero_node); 3941 epilog_stmt = gimple_build_assign (new_scalar_dest, rhs); 3942 new_temp = make_ssa_name (new_scalar_dest, epilog_stmt); 3943 gimple_assign_set_lhs (epilog_stmt, new_temp); 3944 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); 3945 3946 /* In SLP we don't need to apply reduction operation, so we just 3947 collect s' values in SCALAR_RESULTS. */ 3948 if (slp_reduc) 3949 VEC_safe_push (tree, heap, scalar_results, new_temp); 3950 3951 for (bit_offset = element_bitsize; 3952 bit_offset < vec_size_in_bits; 3953 bit_offset += element_bitsize) 3954 { 3955 tree bitpos = bitsize_int (bit_offset); 3956 tree rhs = build3 (BIT_FIELD_REF, scalar_type, vec_temp, 3957 bitsize, bitpos); 3958 3959 epilog_stmt = gimple_build_assign (new_scalar_dest, rhs); 3960 new_name = make_ssa_name (new_scalar_dest, epilog_stmt); 3961 gimple_assign_set_lhs (epilog_stmt, new_name); 3962 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); 3963 3964 if (slp_reduc) 3965 { 3966 /* In SLP we don't need to apply reduction operation, so 3967 we just collect s' values in SCALAR_RESULTS. */ 3968 new_temp = new_name; 3969 VEC_safe_push (tree, heap, scalar_results, new_name); 3970 } 3971 else 3972 { 3973 epilog_stmt = gimple_build_assign_with_ops (code, 3974 new_scalar_dest, new_name, new_temp); 3975 new_temp = make_ssa_name (new_scalar_dest, epilog_stmt); 3976 gimple_assign_set_lhs (epilog_stmt, new_temp); 3977 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); 3978 } 3979 } 3980 } 3981 3982 /* The only case where we need to reduce scalar results in SLP, is 3983 unrolling. If the size of SCALAR_RESULTS is greater than 3984 GROUP_SIZE, we reduce them combining elements modulo 3985 GROUP_SIZE. */ 3986 if (slp_reduc) 3987 { 3988 tree res, first_res, new_res; 3989 gimple new_stmt; 3990 3991 /* Reduce multiple scalar results in case of SLP unrolling. */ 3992 for (j = group_size; VEC_iterate (tree, scalar_results, j, res); 3993 j++) 3994 { 3995 first_res = VEC_index (tree, scalar_results, j % group_size); 3996 new_stmt = gimple_build_assign_with_ops (code, 3997 new_scalar_dest, first_res, res); 3998 new_res = make_ssa_name (new_scalar_dest, new_stmt); 3999 gimple_assign_set_lhs (new_stmt, new_res); 4000 gsi_insert_before (&exit_gsi, new_stmt, GSI_SAME_STMT); 4001 VEC_replace (tree, scalar_results, j % group_size, new_res); 4002 } 4003 } 4004 else 4005 /* Not SLP - we have one scalar to keep in SCALAR_RESULTS. */ 4006 VEC_safe_push (tree, heap, scalar_results, new_temp); 4007 4008 extract_scalar_result = false; 4009 } 4010 } 4011 4012 /* 2.4 Extract the final scalar result. Create: 4013 s_out3 = extract_field <v_out2, bitpos> */ 4014 4015 if (extract_scalar_result) 4016 { 4017 tree rhs; 4018 4019 if (vect_print_dump_info (REPORT_DETAILS)) 4020 fprintf (vect_dump, "extract scalar result"); 4021 4022 if (BYTES_BIG_ENDIAN) 4023 bitpos = size_binop (MULT_EXPR, 4024 bitsize_int (TYPE_VECTOR_SUBPARTS (vectype) - 1), 4025 TYPE_SIZE (scalar_type)); 4026 else 4027 bitpos = bitsize_zero_node; 4028 4029 rhs = build3 (BIT_FIELD_REF, scalar_type, new_temp, bitsize, bitpos); 4030 epilog_stmt = gimple_build_assign (new_scalar_dest, rhs); 4031 new_temp = make_ssa_name (new_scalar_dest, epilog_stmt); 4032 gimple_assign_set_lhs (epilog_stmt, new_temp); 4033 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); 4034 VEC_safe_push (tree, heap, scalar_results, new_temp); 4035 } 4036 4037 vect_finalize_reduction: 4038 4039 if (double_reduc) 4040 loop = loop->inner; 4041 4042 /* 2.5 Adjust the final result by the initial value of the reduction 4043 variable. (When such adjustment is not needed, then 4044 'adjustment_def' is zero). For example, if code is PLUS we create: 4045 new_temp = loop_exit_def + adjustment_def */ 4046 4047 if (adjustment_def) 4048 { 4049 gcc_assert (!slp_reduc); 4050 if (nested_in_vect_loop) 4051 { 4052 new_phi = VEC_index (gimple, new_phis, 0); 4053 gcc_assert (TREE_CODE (TREE_TYPE (adjustment_def)) == VECTOR_TYPE); 4054 expr = build2 (code, vectype, PHI_RESULT (new_phi), adjustment_def); 4055 new_dest = vect_create_destination_var (scalar_dest, vectype); 4056 } 4057 else 4058 { 4059 new_temp = VEC_index (tree, scalar_results, 0); 4060 gcc_assert (TREE_CODE (TREE_TYPE (adjustment_def)) != VECTOR_TYPE); 4061 expr = build2 (code, scalar_type, new_temp, adjustment_def); 4062 new_dest = vect_create_destination_var (scalar_dest, scalar_type); 4063 } 4064 4065 epilog_stmt = gimple_build_assign (new_dest, expr); 4066 new_temp = make_ssa_name (new_dest, epilog_stmt); 4067 gimple_assign_set_lhs (epilog_stmt, new_temp); 4068 SSA_NAME_DEF_STMT (new_temp) = epilog_stmt; 4069 gsi_insert_before (&exit_gsi, epilog_stmt, GSI_SAME_STMT); 4070 if (nested_in_vect_loop) 4071 { 4072 set_vinfo_for_stmt (epilog_stmt, 4073 new_stmt_vec_info (epilog_stmt, loop_vinfo, 4074 NULL)); 4075 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (epilog_stmt)) = 4076 STMT_VINFO_RELATED_STMT (vinfo_for_stmt (new_phi)); 4077 4078 if (!double_reduc) 4079 VEC_quick_push (tree, scalar_results, new_temp); 4080 else 4081 VEC_replace (tree, scalar_results, 0, new_temp); 4082 } 4083 else 4084 VEC_replace (tree, scalar_results, 0, new_temp); 4085 4086 VEC_replace (gimple, new_phis, 0, epilog_stmt); 4087 } 4088 4089 /* 2.6 Handle the loop-exit phis. Replace the uses of scalar loop-exit 4090 phis with new adjusted scalar results, i.e., replace use <s_out0> 4091 with use <s_out4>. 4092 4093 Transform: 4094 loop_exit: 4095 s_out0 = phi <s_loop> # (scalar) EXIT_PHI 4096 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI 4097 v_out2 = reduce <v_out1> 4098 s_out3 = extract_field <v_out2, 0> 4099 s_out4 = adjust_result <s_out3> 4100 use <s_out0> 4101 use <s_out0> 4102 4103 into: 4104 4105 loop_exit: 4106 s_out0 = phi <s_loop> # (scalar) EXIT_PHI 4107 v_out1 = phi <VECT_DEF> # NEW_EXIT_PHI 4108 v_out2 = reduce <v_out1> 4109 s_out3 = extract_field <v_out2, 0> 4110 s_out4 = adjust_result <s_out3> 4111 use <s_out4> 4112 use <s_out4> */ 4113 4114 4115 /* In SLP reduction chain we reduce vector results into one vector if 4116 necessary, hence we set here GROUP_SIZE to 1. SCALAR_DEST is the LHS of 4117 the last stmt in the reduction chain, since we are looking for the loop 4118 exit phi node. */ 4119 if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))) 4120 { 4121 scalar_dest = gimple_assign_lhs (VEC_index (gimple, 4122 SLP_TREE_SCALAR_STMTS (slp_node), 4123 group_size - 1)); 4124 group_size = 1; 4125 } 4126 4127 /* In SLP we may have several statements in NEW_PHIS and REDUCTION_PHIS (in 4128 case that GROUP_SIZE is greater than vectorization factor). Therefore, we 4129 need to match SCALAR_RESULTS with corresponding statements. The first 4130 (GROUP_SIZE / number of new vector stmts) scalar results correspond to 4131 the first vector stmt, etc. 4132 (RATIO is equal to (GROUP_SIZE / number of new vector stmts)). */ 4133 if (group_size > VEC_length (gimple, new_phis)) 4134 { 4135 ratio = group_size / VEC_length (gimple, new_phis); 4136 gcc_assert (!(group_size % VEC_length (gimple, new_phis))); 4137 } 4138 else 4139 ratio = 1; 4140 4141 for (k = 0; k < group_size; k++) 4142 { 4143 if (k % ratio == 0) 4144 { 4145 epilog_stmt = VEC_index (gimple, new_phis, k / ratio); 4146 reduction_phi = VEC_index (gimple, reduction_phis, k / ratio); 4147 if (double_reduc) 4148 inner_phi = VEC_index (gimple, inner_phis, k / ratio); 4149 } 4150 4151 if (slp_reduc) 4152 { 4153 gimple current_stmt = VEC_index (gimple, 4154 SLP_TREE_SCALAR_STMTS (slp_node), k); 4155 4156 orig_stmt = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (current_stmt)); 4157 /* SLP statements can't participate in patterns. */ 4158 gcc_assert (!orig_stmt); 4159 scalar_dest = gimple_assign_lhs (current_stmt); 4160 } 4161 4162 phis = VEC_alloc (gimple, heap, 3); 4163 /* Find the loop-closed-use at the loop exit of the original scalar 4164 result. (The reduction result is expected to have two immediate uses - 4165 one at the latch block, and one at the loop exit). */ 4166 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) 4167 if (!flow_bb_inside_loop_p (loop, gimple_bb (USE_STMT (use_p))) 4168 && !is_gimple_debug (USE_STMT (use_p))) 4169 VEC_safe_push (gimple, heap, phis, USE_STMT (use_p)); 4170 4171 /* We expect to have found an exit_phi because of loop-closed-ssa 4172 form. */ 4173 gcc_assert (!VEC_empty (gimple, phis)); 4174 4175 FOR_EACH_VEC_ELT (gimple, phis, i, exit_phi) 4176 { 4177 if (outer_loop) 4178 { 4179 stmt_vec_info exit_phi_vinfo = vinfo_for_stmt (exit_phi); 4180 gimple vect_phi; 4181 4182 /* FORNOW. Currently not supporting the case that an inner-loop 4183 reduction is not used in the outer-loop (but only outside the 4184 outer-loop), unless it is double reduction. */ 4185 gcc_assert ((STMT_VINFO_RELEVANT_P (exit_phi_vinfo) 4186 && !STMT_VINFO_LIVE_P (exit_phi_vinfo)) 4187 || double_reduc); 4188 4189 STMT_VINFO_VEC_STMT (exit_phi_vinfo) = epilog_stmt; 4190 if (!double_reduc 4191 || STMT_VINFO_DEF_TYPE (exit_phi_vinfo) 4192 != vect_double_reduction_def) 4193 continue; 4194 4195 /* Handle double reduction: 4196 4197 stmt1: s1 = phi <s0, s2> - double reduction phi (outer loop) 4198 stmt2: s3 = phi <s1, s4> - (regular) reduc phi (inner loop) 4199 stmt3: s4 = use (s3) - (regular) reduc stmt (inner loop) 4200 stmt4: s2 = phi <s4> - double reduction stmt (outer loop) 4201 4202 At that point the regular reduction (stmt2 and stmt3) is 4203 already vectorized, as well as the exit phi node, stmt4. 4204 Here we vectorize the phi node of double reduction, stmt1, and 4205 update all relevant statements. */ 4206 4207 /* Go through all the uses of s2 to find double reduction phi 4208 node, i.e., stmt1 above. */ 4209 orig_name = PHI_RESULT (exit_phi); 4210 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, orig_name) 4211 { 4212 stmt_vec_info use_stmt_vinfo = vinfo_for_stmt (use_stmt); 4213 stmt_vec_info new_phi_vinfo; 4214 tree vect_phi_init, preheader_arg, vect_phi_res, init_def; 4215 basic_block bb = gimple_bb (use_stmt); 4216 gimple use; 4217 4218 /* Check that USE_STMT is really double reduction phi 4219 node. */ 4220 if (gimple_code (use_stmt) != GIMPLE_PHI 4221 || gimple_phi_num_args (use_stmt) != 2 4222 || !use_stmt_vinfo 4223 || STMT_VINFO_DEF_TYPE (use_stmt_vinfo) 4224 != vect_double_reduction_def 4225 || bb->loop_father != outer_loop) 4226 continue; 4227 4228 /* Create vector phi node for double reduction: 4229 vs1 = phi <vs0, vs2> 4230 vs1 was created previously in this function by a call to 4231 vect_get_vec_def_for_operand and is stored in 4232 vec_initial_def; 4233 vs2 is defined by INNER_PHI, the vectorized EXIT_PHI; 4234 vs0 is created here. */ 4235 4236 /* Create vector phi node. */ 4237 vect_phi = create_phi_node (vec_initial_def, bb); 4238 new_phi_vinfo = new_stmt_vec_info (vect_phi, 4239 loop_vec_info_for_loop (outer_loop), NULL); 4240 set_vinfo_for_stmt (vect_phi, new_phi_vinfo); 4241 4242 /* Create vs0 - initial def of the double reduction phi. */ 4243 preheader_arg = PHI_ARG_DEF_FROM_EDGE (use_stmt, 4244 loop_preheader_edge (outer_loop)); 4245 init_def = get_initial_def_for_reduction (stmt, 4246 preheader_arg, NULL); 4247 vect_phi_init = vect_init_vector (use_stmt, init_def, 4248 vectype, NULL); 4249 4250 /* Update phi node arguments with vs0 and vs2. */ 4251 add_phi_arg (vect_phi, vect_phi_init, 4252 loop_preheader_edge (outer_loop), 4253 UNKNOWN_LOCATION); 4254 add_phi_arg (vect_phi, PHI_RESULT (inner_phi), 4255 loop_latch_edge (outer_loop), UNKNOWN_LOCATION); 4256 if (vect_print_dump_info (REPORT_DETAILS)) 4257 { 4258 fprintf (vect_dump, "created double reduction phi " 4259 "node: "); 4260 print_gimple_stmt (vect_dump, vect_phi, 0, TDF_SLIM); 4261 } 4262 4263 vect_phi_res = PHI_RESULT (vect_phi); 4264 4265 /* Replace the use, i.e., set the correct vs1 in the regular 4266 reduction phi node. FORNOW, NCOPIES is always 1, so the 4267 loop is redundant. */ 4268 use = reduction_phi; 4269 for (j = 0; j < ncopies; j++) 4270 { 4271 edge pr_edge = loop_preheader_edge (loop); 4272 SET_PHI_ARG_DEF (use, pr_edge->dest_idx, vect_phi_res); 4273 use = STMT_VINFO_RELATED_STMT (vinfo_for_stmt (use)); 4274 } 4275 } 4276 } 4277 } 4278 4279 VEC_free (gimple, heap, phis); 4280 if (nested_in_vect_loop) 4281 { 4282 if (double_reduc) 4283 loop = outer_loop; 4284 else 4285 continue; 4286 } 4287 4288 phis = VEC_alloc (gimple, heap, 3); 4289 /* Find the loop-closed-use at the loop exit of the original scalar 4290 result. (The reduction result is expected to have two immediate uses, 4291 one at the latch block, and one at the loop exit). For double 4292 reductions we are looking for exit phis of the outer loop. */ 4293 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, scalar_dest) 4294 { 4295 if (!flow_bb_inside_loop_p (loop, gimple_bb (USE_STMT (use_p)))) 4296 { 4297 if (!is_gimple_debug (USE_STMT (use_p))) 4298 VEC_safe_push (gimple, heap, phis, USE_STMT (use_p)); 4299 } 4300 else 4301 { 4302 if (double_reduc && gimple_code (USE_STMT (use_p)) == GIMPLE_PHI) 4303 { 4304 tree phi_res = PHI_RESULT (USE_STMT (use_p)); 4305 4306 FOR_EACH_IMM_USE_FAST (phi_use_p, phi_imm_iter, phi_res) 4307 { 4308 if (!flow_bb_inside_loop_p (loop, 4309 gimple_bb (USE_STMT (phi_use_p))) 4310 && !is_gimple_debug (USE_STMT (phi_use_p))) 4311 VEC_safe_push (gimple, heap, phis, 4312 USE_STMT (phi_use_p)); 4313 } 4314 } 4315 } 4316 } 4317 4318 FOR_EACH_VEC_ELT (gimple, phis, i, exit_phi) 4319 { 4320 /* Replace the uses: */ 4321 orig_name = PHI_RESULT (exit_phi); 4322 scalar_result = VEC_index (tree, scalar_results, k); 4323 FOR_EACH_IMM_USE_STMT (use_stmt, imm_iter, orig_name) 4324 FOR_EACH_IMM_USE_ON_STMT (use_p, imm_iter) 4325 SET_USE (use_p, scalar_result); 4326 } 4327 4328 VEC_free (gimple, heap, phis); 4329 } 4330 4331 VEC_free (tree, heap, scalar_results); 4332 VEC_free (gimple, heap, new_phis); 4333 } 4334 4335 4336 /* Function vectorizable_reduction. 4337 4338 Check if STMT performs a reduction operation that can be vectorized. 4339 If VEC_STMT is also passed, vectorize the STMT: create a vectorized 4340 stmt to replace it, put it in VEC_STMT, and insert it at GSI. 4341 Return FALSE if not a vectorizable STMT, TRUE otherwise. 4342 4343 This function also handles reduction idioms (patterns) that have been 4344 recognized in advance during vect_pattern_recog. In this case, STMT may be 4345 of this form: 4346 X = pattern_expr (arg0, arg1, ..., X) 4347 and it's STMT_VINFO_RELATED_STMT points to the last stmt in the original 4348 sequence that had been detected and replaced by the pattern-stmt (STMT). 4349 4350 In some cases of reduction patterns, the type of the reduction variable X is 4351 different than the type of the other arguments of STMT. 4352 In such cases, the vectype that is used when transforming STMT into a vector 4353 stmt is different than the vectype that is used to determine the 4354 vectorization factor, because it consists of a different number of elements 4355 than the actual number of elements that are being operated upon in parallel. 4356 4357 For example, consider an accumulation of shorts into an int accumulator. 4358 On some targets it's possible to vectorize this pattern operating on 8 4359 shorts at a time (hence, the vectype for purposes of determining the 4360 vectorization factor should be V8HI); on the other hand, the vectype that 4361 is used to create the vector form is actually V4SI (the type of the result). 4362 4363 Upon entry to this function, STMT_VINFO_VECTYPE records the vectype that 4364 indicates what is the actual level of parallelism (V8HI in the example), so 4365 that the right vectorization factor would be derived. This vectype 4366 corresponds to the type of arguments to the reduction stmt, and should *NOT* 4367 be used to create the vectorized stmt. The right vectype for the vectorized 4368 stmt is obtained from the type of the result X: 4369 get_vectype_for_scalar_type (TREE_TYPE (X)) 4370 4371 This means that, contrary to "regular" reductions (or "regular" stmts in 4372 general), the following equation: 4373 STMT_VINFO_VECTYPE == get_vectype_for_scalar_type (TREE_TYPE (X)) 4374 does *NOT* necessarily hold for reduction patterns. */ 4375 4376 bool 4377 vectorizable_reduction (gimple stmt, gimple_stmt_iterator *gsi, 4378 gimple *vec_stmt, slp_tree slp_node) 4379 { 4380 tree vec_dest; 4381 tree scalar_dest; 4382 tree loop_vec_def0 = NULL_TREE, loop_vec_def1 = NULL_TREE; 4383 stmt_vec_info stmt_info = vinfo_for_stmt (stmt); 4384 tree vectype_out = STMT_VINFO_VECTYPE (stmt_info); 4385 tree vectype_in = NULL_TREE; 4386 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); 4387 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); 4388 enum tree_code code, orig_code, epilog_reduc_code; 4389 enum machine_mode vec_mode; 4390 int op_type; 4391 optab optab, reduc_optab; 4392 tree new_temp = NULL_TREE; 4393 tree def; 4394 gimple def_stmt; 4395 enum vect_def_type dt; 4396 gimple new_phi = NULL; 4397 tree scalar_type; 4398 bool is_simple_use; 4399 gimple orig_stmt; 4400 stmt_vec_info orig_stmt_info; 4401 tree expr = NULL_TREE; 4402 int i; 4403 int ncopies; 4404 int epilog_copies; 4405 stmt_vec_info prev_stmt_info, prev_phi_info; 4406 bool single_defuse_cycle = false; 4407 tree reduc_def = NULL_TREE; 4408 gimple new_stmt = NULL; 4409 int j; 4410 tree ops[3]; 4411 bool nested_cycle = false, found_nested_cycle_def = false; 4412 gimple reduc_def_stmt = NULL; 4413 /* The default is that the reduction variable is the last in statement. */ 4414 int reduc_index = 2; 4415 bool double_reduc = false, dummy; 4416 basic_block def_bb; 4417 struct loop * def_stmt_loop, *outer_loop = NULL; 4418 tree def_arg; 4419 gimple def_arg_stmt; 4420 VEC (tree, heap) *vec_oprnds0 = NULL, *vec_oprnds1 = NULL, *vect_defs = NULL; 4421 VEC (gimple, heap) *phis = NULL; 4422 int vec_num; 4423 tree def0, def1, tem, op0, op1 = NULL_TREE; 4424 4425 /* In case of reduction chain we switch to the first stmt in the chain, but 4426 we don't update STMT_INFO, since only the last stmt is marked as reduction 4427 and has reduction properties. */ 4428 if (GROUP_FIRST_ELEMENT (vinfo_for_stmt (stmt))) 4429 stmt = GROUP_FIRST_ELEMENT (stmt_info); 4430 4431 if (nested_in_vect_loop_p (loop, stmt)) 4432 { 4433 outer_loop = loop; 4434 loop = loop->inner; 4435 nested_cycle = true; 4436 } 4437 4438 /* 1. Is vectorizable reduction? */ 4439 /* Not supportable if the reduction variable is used in the loop, unless 4440 it's a reduction chain. */ 4441 if (STMT_VINFO_RELEVANT (stmt_info) > vect_used_in_outer 4442 && !GROUP_FIRST_ELEMENT (stmt_info)) 4443 return false; 4444 4445 /* Reductions that are not used even in an enclosing outer-loop, 4446 are expected to be "live" (used out of the loop). */ 4447 if (STMT_VINFO_RELEVANT (stmt_info) == vect_unused_in_scope 4448 && !STMT_VINFO_LIVE_P (stmt_info)) 4449 return false; 4450 4451 /* Make sure it was already recognized as a reduction computation. */ 4452 if (STMT_VINFO_DEF_TYPE (stmt_info) != vect_reduction_def 4453 && STMT_VINFO_DEF_TYPE (stmt_info) != vect_nested_cycle) 4454 return false; 4455 4456 /* 2. Has this been recognized as a reduction pattern? 4457 4458 Check if STMT represents a pattern that has been recognized 4459 in earlier analysis stages. For stmts that represent a pattern, 4460 the STMT_VINFO_RELATED_STMT field records the last stmt in 4461 the original sequence that constitutes the pattern. */ 4462 4463 orig_stmt = STMT_VINFO_RELATED_STMT (stmt_info); 4464 if (orig_stmt) 4465 { 4466 orig_stmt_info = vinfo_for_stmt (orig_stmt); 4467 gcc_assert (STMT_VINFO_IN_PATTERN_P (orig_stmt_info)); 4468 gcc_assert (!STMT_VINFO_IN_PATTERN_P (stmt_info)); 4469 } 4470 4471 /* 3. Check the operands of the operation. The first operands are defined 4472 inside the loop body. The last operand is the reduction variable, 4473 which is defined by the loop-header-phi. */ 4474 4475 gcc_assert (is_gimple_assign (stmt)); 4476 4477 /* Flatten RHS. */ 4478 switch (get_gimple_rhs_class (gimple_assign_rhs_code (stmt))) 4479 { 4480 case GIMPLE_SINGLE_RHS: 4481 op_type = TREE_OPERAND_LENGTH (gimple_assign_rhs1 (stmt)); 4482 if (op_type == ternary_op) 4483 { 4484 tree rhs = gimple_assign_rhs1 (stmt); 4485 ops[0] = TREE_OPERAND (rhs, 0); 4486 ops[1] = TREE_OPERAND (rhs, 1); 4487 ops[2] = TREE_OPERAND (rhs, 2); 4488 code = TREE_CODE (rhs); 4489 } 4490 else 4491 return false; 4492 break; 4493 4494 case GIMPLE_BINARY_RHS: 4495 code = gimple_assign_rhs_code (stmt); 4496 op_type = TREE_CODE_LENGTH (code); 4497 gcc_assert (op_type == binary_op); 4498 ops[0] = gimple_assign_rhs1 (stmt); 4499 ops[1] = gimple_assign_rhs2 (stmt); 4500 break; 4501 4502 case GIMPLE_TERNARY_RHS: 4503 code = gimple_assign_rhs_code (stmt); 4504 op_type = TREE_CODE_LENGTH (code); 4505 gcc_assert (op_type == ternary_op); 4506 ops[0] = gimple_assign_rhs1 (stmt); 4507 ops[1] = gimple_assign_rhs2 (stmt); 4508 ops[2] = gimple_assign_rhs3 (stmt); 4509 break; 4510 4511 case GIMPLE_UNARY_RHS: 4512 return false; 4513 4514 default: 4515 gcc_unreachable (); 4516 } 4517 4518 if (code == COND_EXPR && slp_node) 4519 return false; 4520 4521 scalar_dest = gimple_assign_lhs (stmt); 4522 scalar_type = TREE_TYPE (scalar_dest); 4523 if (!POINTER_TYPE_P (scalar_type) && !INTEGRAL_TYPE_P (scalar_type) 4524 && !SCALAR_FLOAT_TYPE_P (scalar_type)) 4525 return false; 4526 4527 /* Do not try to vectorize bit-precision reductions. */ 4528 if ((TYPE_PRECISION (scalar_type) 4529 != GET_MODE_PRECISION (TYPE_MODE (scalar_type)))) 4530 return false; 4531 4532 /* All uses but the last are expected to be defined in the loop. 4533 The last use is the reduction variable. In case of nested cycle this 4534 assumption is not true: we use reduc_index to record the index of the 4535 reduction variable. */ 4536 for (i = 0; i < op_type - 1; i++) 4537 { 4538 /* The condition of COND_EXPR is checked in vectorizable_condition(). */ 4539 if (i == 0 && code == COND_EXPR) 4540 continue; 4541 4542 is_simple_use = vect_is_simple_use_1 (ops[i], stmt, loop_vinfo, NULL, 4543 &def_stmt, &def, &dt, &tem); 4544 if (!vectype_in) 4545 vectype_in = tem; 4546 gcc_assert (is_simple_use); 4547 4548 if (dt != vect_internal_def 4549 && dt != vect_external_def 4550 && dt != vect_constant_def 4551 && dt != vect_induction_def 4552 && !(dt == vect_nested_cycle && nested_cycle)) 4553 return false; 4554 4555 if (dt == vect_nested_cycle) 4556 { 4557 found_nested_cycle_def = true; 4558 reduc_def_stmt = def_stmt; 4559 reduc_index = i; 4560 } 4561 } 4562 4563 is_simple_use = vect_is_simple_use_1 (ops[i], stmt, loop_vinfo, NULL, 4564 &def_stmt, &def, &dt, &tem); 4565 if (!vectype_in) 4566 vectype_in = tem; 4567 gcc_assert (is_simple_use); 4568 if (!(dt == vect_reduction_def 4569 || dt == vect_nested_cycle 4570 || ((dt == vect_internal_def || dt == vect_external_def 4571 || dt == vect_constant_def || dt == vect_induction_def) 4572 && nested_cycle && found_nested_cycle_def))) 4573 { 4574 /* For pattern recognized stmts, orig_stmt might be a reduction, 4575 but some helper statements for the pattern might not, or 4576 might be COND_EXPRs with reduction uses in the condition. */ 4577 gcc_assert (orig_stmt); 4578 return false; 4579 } 4580 if (!found_nested_cycle_def) 4581 reduc_def_stmt = def_stmt; 4582 4583 gcc_assert (gimple_code (reduc_def_stmt) == GIMPLE_PHI); 4584 if (orig_stmt) 4585 gcc_assert (orig_stmt == vect_is_simple_reduction (loop_vinfo, 4586 reduc_def_stmt, 4587 !nested_cycle, 4588 &dummy)); 4589 else 4590 { 4591 gimple tmp = vect_is_simple_reduction (loop_vinfo, reduc_def_stmt, 4592 !nested_cycle, &dummy); 4593 /* We changed STMT to be the first stmt in reduction chain, hence we 4594 check that in this case the first element in the chain is STMT. */ 4595 gcc_assert (stmt == tmp 4596 || GROUP_FIRST_ELEMENT (vinfo_for_stmt (tmp)) == stmt); 4597 } 4598 4599 if (STMT_VINFO_LIVE_P (vinfo_for_stmt (reduc_def_stmt))) 4600 return false; 4601 4602 if (slp_node || PURE_SLP_STMT (stmt_info)) 4603 ncopies = 1; 4604 else 4605 ncopies = (LOOP_VINFO_VECT_FACTOR (loop_vinfo) 4606 / TYPE_VECTOR_SUBPARTS (vectype_in)); 4607 4608 gcc_assert (ncopies >= 1); 4609 4610 vec_mode = TYPE_MODE (vectype_in); 4611 4612 if (code == COND_EXPR) 4613 { 4614 if (!vectorizable_condition (stmt, gsi, NULL, ops[reduc_index], 0, NULL)) 4615 { 4616 if (vect_print_dump_info (REPORT_DETAILS)) 4617 fprintf (vect_dump, "unsupported condition in reduction"); 4618 4619 return false; 4620 } 4621 } 4622 else 4623 { 4624 /* 4. Supportable by target? */ 4625 4626 /* 4.1. check support for the operation in the loop */ 4627 optab = optab_for_tree_code (code, vectype_in, optab_default); 4628 if (!optab) 4629 { 4630 if (vect_print_dump_info (REPORT_DETAILS)) 4631 fprintf (vect_dump, "no optab."); 4632 4633 return false; 4634 } 4635 4636 if (optab_handler (optab, vec_mode) == CODE_FOR_nothing) 4637 { 4638 if (vect_print_dump_info (REPORT_DETAILS)) 4639 fprintf (vect_dump, "op not supported by target."); 4640 4641 if (GET_MODE_SIZE (vec_mode) != UNITS_PER_WORD 4642 || LOOP_VINFO_VECT_FACTOR (loop_vinfo) 4643 < vect_min_worthwhile_factor (code)) 4644 return false; 4645 4646 if (vect_print_dump_info (REPORT_DETAILS)) 4647 fprintf (vect_dump, "proceeding using word mode."); 4648 } 4649 4650 /* Worthwhile without SIMD support? */ 4651 if (!VECTOR_MODE_P (TYPE_MODE (vectype_in)) 4652 && LOOP_VINFO_VECT_FACTOR (loop_vinfo) 4653 < vect_min_worthwhile_factor (code)) 4654 { 4655 if (vect_print_dump_info (REPORT_DETAILS)) 4656 fprintf (vect_dump, "not worthwhile without SIMD support."); 4657 4658 return false; 4659 } 4660 } 4661 4662 /* 4.2. Check support for the epilog operation. 4663 4664 If STMT represents a reduction pattern, then the type of the 4665 reduction variable may be different than the type of the rest 4666 of the arguments. For example, consider the case of accumulation 4667 of shorts into an int accumulator; The original code: 4668 S1: int_a = (int) short_a; 4669 orig_stmt-> S2: int_acc = plus <int_a ,int_acc>; 4670 4671 was replaced with: 4672 STMT: int_acc = widen_sum <short_a, int_acc> 4673 4674 This means that: 4675 1. The tree-code that is used to create the vector operation in the 4676 epilog code (that reduces the partial results) is not the 4677 tree-code of STMT, but is rather the tree-code of the original 4678 stmt from the pattern that STMT is replacing. I.e, in the example 4679 above we want to use 'widen_sum' in the loop, but 'plus' in the 4680 epilog. 4681 2. The type (mode) we use to check available target support 4682 for the vector operation to be created in the *epilog*, is 4683 determined by the type of the reduction variable (in the example 4684 above we'd check this: optab_handler (plus_optab, vect_int_mode])). 4685 However the type (mode) we use to check available target support 4686 for the vector operation to be created *inside the loop*, is 4687 determined by the type of the other arguments to STMT (in the 4688 example we'd check this: optab_handler (widen_sum_optab, 4689 vect_short_mode)). 4690 4691 This is contrary to "regular" reductions, in which the types of all 4692 the arguments are the same as the type of the reduction variable. 4693 For "regular" reductions we can therefore use the same vector type 4694 (and also the same tree-code) when generating the epilog code and 4695 when generating the code inside the loop. */ 4696 4697 if (orig_stmt) 4698 { 4699 /* This is a reduction pattern: get the vectype from the type of the 4700 reduction variable, and get the tree-code from orig_stmt. */ 4701 orig_code = gimple_assign_rhs_code (orig_stmt); 4702 gcc_assert (vectype_out); 4703 vec_mode = TYPE_MODE (vectype_out); 4704 } 4705 else 4706 { 4707 /* Regular reduction: use the same vectype and tree-code as used for 4708 the vector code inside the loop can be used for the epilog code. */ 4709 orig_code = code; 4710 } 4711 4712 if (nested_cycle) 4713 { 4714 def_bb = gimple_bb (reduc_def_stmt); 4715 def_stmt_loop = def_bb->loop_father; 4716 def_arg = PHI_ARG_DEF_FROM_EDGE (reduc_def_stmt, 4717 loop_preheader_edge (def_stmt_loop)); 4718 if (TREE_CODE (def_arg) == SSA_NAME 4719 && (def_arg_stmt = SSA_NAME_DEF_STMT (def_arg)) 4720 && gimple_code (def_arg_stmt) == GIMPLE_PHI 4721 && flow_bb_inside_loop_p (outer_loop, gimple_bb (def_arg_stmt)) 4722 && vinfo_for_stmt (def_arg_stmt) 4723 && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_arg_stmt)) 4724 == vect_double_reduction_def) 4725 double_reduc = true; 4726 } 4727 4728 epilog_reduc_code = ERROR_MARK; 4729 if (reduction_code_for_scalar_code (orig_code, &epilog_reduc_code)) 4730 { 4731 reduc_optab = optab_for_tree_code (epilog_reduc_code, vectype_out, 4732 optab_default); 4733 if (!reduc_optab) 4734 { 4735 if (vect_print_dump_info (REPORT_DETAILS)) 4736 fprintf (vect_dump, "no optab for reduction."); 4737 4738 epilog_reduc_code = ERROR_MARK; 4739 } 4740 4741 if (reduc_optab 4742 && optab_handler (reduc_optab, vec_mode) == CODE_FOR_nothing) 4743 { 4744 if (vect_print_dump_info (REPORT_DETAILS)) 4745 fprintf (vect_dump, "reduc op not supported by target."); 4746 4747 epilog_reduc_code = ERROR_MARK; 4748 } 4749 } 4750 else 4751 { 4752 if (!nested_cycle || double_reduc) 4753 { 4754 if (vect_print_dump_info (REPORT_DETAILS)) 4755 fprintf (vect_dump, "no reduc code for scalar code."); 4756 4757 return false; 4758 } 4759 } 4760 4761 if (double_reduc && ncopies > 1) 4762 { 4763 if (vect_print_dump_info (REPORT_DETAILS)) 4764 fprintf (vect_dump, "multiple types in double reduction"); 4765 4766 return false; 4767 } 4768 4769 /* In case of widenning multiplication by a constant, we update the type 4770 of the constant to be the type of the other operand. We check that the 4771 constant fits the type in the pattern recognition pass. */ 4772 if (code == DOT_PROD_EXPR 4773 && !types_compatible_p (TREE_TYPE (ops[0]), TREE_TYPE (ops[1]))) 4774 { 4775 if (TREE_CODE (ops[0]) == INTEGER_CST) 4776 ops[0] = fold_convert (TREE_TYPE (ops[1]), ops[0]); 4777 else if (TREE_CODE (ops[1]) == INTEGER_CST) 4778 ops[1] = fold_convert (TREE_TYPE (ops[0]), ops[1]); 4779 else 4780 { 4781 if (vect_print_dump_info (REPORT_DETAILS)) 4782 fprintf (vect_dump, "invalid types in dot-prod"); 4783 4784 return false; 4785 } 4786 } 4787 4788 if (!vec_stmt) /* transformation not required. */ 4789 { 4790 if (!vect_model_reduction_cost (stmt_info, epilog_reduc_code, ncopies)) 4791 return false; 4792 STMT_VINFO_TYPE (stmt_info) = reduc_vec_info_type; 4793 return true; 4794 } 4795 4796 /** Transform. **/ 4797 4798 if (vect_print_dump_info (REPORT_DETAILS)) 4799 fprintf (vect_dump, "transform reduction."); 4800 4801 /* FORNOW: Multiple types are not supported for condition. */ 4802 if (code == COND_EXPR) 4803 gcc_assert (ncopies == 1); 4804 4805 /* Create the destination vector */ 4806 vec_dest = vect_create_destination_var (scalar_dest, vectype_out); 4807 4808 /* In case the vectorization factor (VF) is bigger than the number 4809 of elements that we can fit in a vectype (nunits), we have to generate 4810 more than one vector stmt - i.e - we need to "unroll" the 4811 vector stmt by a factor VF/nunits. For more details see documentation 4812 in vectorizable_operation. */ 4813 4814 /* If the reduction is used in an outer loop we need to generate 4815 VF intermediate results, like so (e.g. for ncopies=2): 4816 r0 = phi (init, r0) 4817 r1 = phi (init, r1) 4818 r0 = x0 + r0; 4819 r1 = x1 + r1; 4820 (i.e. we generate VF results in 2 registers). 4821 In this case we have a separate def-use cycle for each copy, and therefore 4822 for each copy we get the vector def for the reduction variable from the 4823 respective phi node created for this copy. 4824 4825 Otherwise (the reduction is unused in the loop nest), we can combine 4826 together intermediate results, like so (e.g. for ncopies=2): 4827 r = phi (init, r) 4828 r = x0 + r; 4829 r = x1 + r; 4830 (i.e. we generate VF/2 results in a single register). 4831 In this case for each copy we get the vector def for the reduction variable 4832 from the vectorized reduction operation generated in the previous iteration. 4833 */ 4834 4835 if (STMT_VINFO_RELEVANT (stmt_info) == vect_unused_in_scope) 4836 { 4837 single_defuse_cycle = true; 4838 epilog_copies = 1; 4839 } 4840 else 4841 epilog_copies = ncopies; 4842 4843 prev_stmt_info = NULL; 4844 prev_phi_info = NULL; 4845 if (slp_node) 4846 { 4847 vec_num = SLP_TREE_NUMBER_OF_VEC_STMTS (slp_node); 4848 gcc_assert (TYPE_VECTOR_SUBPARTS (vectype_out) 4849 == TYPE_VECTOR_SUBPARTS (vectype_in)); 4850 } 4851 else 4852 { 4853 vec_num = 1; 4854 vec_oprnds0 = VEC_alloc (tree, heap, 1); 4855 if (op_type == ternary_op) 4856 vec_oprnds1 = VEC_alloc (tree, heap, 1); 4857 } 4858 4859 phis = VEC_alloc (gimple, heap, vec_num); 4860 vect_defs = VEC_alloc (tree, heap, vec_num); 4861 if (!slp_node) 4862 VEC_quick_push (tree, vect_defs, NULL_TREE); 4863 4864 for (j = 0; j < ncopies; j++) 4865 { 4866 if (j == 0 || !single_defuse_cycle) 4867 { 4868 for (i = 0; i < vec_num; i++) 4869 { 4870 /* Create the reduction-phi that defines the reduction 4871 operand. */ 4872 new_phi = create_phi_node (vec_dest, loop->header); 4873 set_vinfo_for_stmt (new_phi, 4874 new_stmt_vec_info (new_phi, loop_vinfo, 4875 NULL)); 4876 if (j == 0 || slp_node) 4877 VEC_quick_push (gimple, phis, new_phi); 4878 } 4879 } 4880 4881 if (code == COND_EXPR) 4882 { 4883 gcc_assert (!slp_node); 4884 vectorizable_condition (stmt, gsi, vec_stmt, 4885 PHI_RESULT (VEC_index (gimple, phis, 0)), 4886 reduc_index, NULL); 4887 /* Multiple types are not supported for condition. */ 4888 break; 4889 } 4890 4891 /* Handle uses. */ 4892 if (j == 0) 4893 { 4894 op0 = ops[!reduc_index]; 4895 if (op_type == ternary_op) 4896 { 4897 if (reduc_index == 0) 4898 op1 = ops[2]; 4899 else 4900 op1 = ops[1]; 4901 } 4902 4903 if (slp_node) 4904 vect_get_vec_defs (op0, op1, stmt, &vec_oprnds0, &vec_oprnds1, 4905 slp_node, -1); 4906 else 4907 { 4908 loop_vec_def0 = vect_get_vec_def_for_operand (ops[!reduc_index], 4909 stmt, NULL); 4910 VEC_quick_push (tree, vec_oprnds0, loop_vec_def0); 4911 if (op_type == ternary_op) 4912 { 4913 loop_vec_def1 = vect_get_vec_def_for_operand (op1, stmt, 4914 NULL); 4915 VEC_quick_push (tree, vec_oprnds1, loop_vec_def1); 4916 } 4917 } 4918 } 4919 else 4920 { 4921 if (!slp_node) 4922 { 4923 enum vect_def_type dt; 4924 gimple dummy_stmt; 4925 tree dummy; 4926 4927 vect_is_simple_use (ops[!reduc_index], stmt, loop_vinfo, NULL, 4928 &dummy_stmt, &dummy, &dt); 4929 loop_vec_def0 = vect_get_vec_def_for_stmt_copy (dt, 4930 loop_vec_def0); 4931 VEC_replace (tree, vec_oprnds0, 0, loop_vec_def0); 4932 if (op_type == ternary_op) 4933 { 4934 vect_is_simple_use (op1, stmt, loop_vinfo, NULL, &dummy_stmt, 4935 &dummy, &dt); 4936 loop_vec_def1 = vect_get_vec_def_for_stmt_copy (dt, 4937 loop_vec_def1); 4938 VEC_replace (tree, vec_oprnds1, 0, loop_vec_def1); 4939 } 4940 } 4941 4942 if (single_defuse_cycle) 4943 reduc_def = gimple_assign_lhs (new_stmt); 4944 4945 STMT_VINFO_RELATED_STMT (prev_phi_info) = new_phi; 4946 } 4947 4948 FOR_EACH_VEC_ELT (tree, vec_oprnds0, i, def0) 4949 { 4950 if (slp_node) 4951 reduc_def = PHI_RESULT (VEC_index (gimple, phis, i)); 4952 else 4953 { 4954 if (!single_defuse_cycle || j == 0) 4955 reduc_def = PHI_RESULT (new_phi); 4956 } 4957 4958 def1 = ((op_type == ternary_op) 4959 ? VEC_index (tree, vec_oprnds1, i) : NULL); 4960 if (op_type == binary_op) 4961 { 4962 if (reduc_index == 0) 4963 expr = build2 (code, vectype_out, reduc_def, def0); 4964 else 4965 expr = build2 (code, vectype_out, def0, reduc_def); 4966 } 4967 else 4968 { 4969 if (reduc_index == 0) 4970 expr = build3 (code, vectype_out, reduc_def, def0, def1); 4971 else 4972 { 4973 if (reduc_index == 1) 4974 expr = build3 (code, vectype_out, def0, reduc_def, def1); 4975 else 4976 expr = build3 (code, vectype_out, def0, def1, reduc_def); 4977 } 4978 } 4979 4980 new_stmt = gimple_build_assign (vec_dest, expr); 4981 new_temp = make_ssa_name (vec_dest, new_stmt); 4982 gimple_assign_set_lhs (new_stmt, new_temp); 4983 vect_finish_stmt_generation (stmt, new_stmt, gsi); 4984 4985 if (slp_node) 4986 { 4987 VEC_quick_push (gimple, SLP_TREE_VEC_STMTS (slp_node), new_stmt); 4988 VEC_quick_push (tree, vect_defs, new_temp); 4989 } 4990 else 4991 VEC_replace (tree, vect_defs, 0, new_temp); 4992 } 4993 4994 if (slp_node) 4995 continue; 4996 4997 if (j == 0) 4998 STMT_VINFO_VEC_STMT (stmt_info) = *vec_stmt = new_stmt; 4999 else 5000 STMT_VINFO_RELATED_STMT (prev_stmt_info) = new_stmt; 5001 5002 prev_stmt_info = vinfo_for_stmt (new_stmt); 5003 prev_phi_info = vinfo_for_stmt (new_phi); 5004 } 5005 5006 /* Finalize the reduction-phi (set its arguments) and create the 5007 epilog reduction code. */ 5008 if ((!single_defuse_cycle || code == COND_EXPR) && !slp_node) 5009 { 5010 new_temp = gimple_assign_lhs (*vec_stmt); 5011 VEC_replace (tree, vect_defs, 0, new_temp); 5012 } 5013 5014 vect_create_epilog_for_reduction (vect_defs, stmt, epilog_copies, 5015 epilog_reduc_code, phis, reduc_index, 5016 double_reduc, slp_node); 5017 5018 VEC_free (gimple, heap, phis); 5019 VEC_free (tree, heap, vec_oprnds0); 5020 if (vec_oprnds1) 5021 VEC_free (tree, heap, vec_oprnds1); 5022 5023 return true; 5024 } 5025 5026 /* Function vect_min_worthwhile_factor. 5027 5028 For a loop where we could vectorize the operation indicated by CODE, 5029 return the minimum vectorization factor that makes it worthwhile 5030 to use generic vectors. */ 5031 int 5032 vect_min_worthwhile_factor (enum tree_code code) 5033 { 5034 switch (code) 5035 { 5036 case PLUS_EXPR: 5037 case MINUS_EXPR: 5038 case NEGATE_EXPR: 5039 return 4; 5040 5041 case BIT_AND_EXPR: 5042 case BIT_IOR_EXPR: 5043 case BIT_XOR_EXPR: 5044 case BIT_NOT_EXPR: 5045 return 2; 5046 5047 default: 5048 return INT_MAX; 5049 } 5050 } 5051 5052 5053 /* Function vectorizable_induction 5054 5055 Check if PHI performs an induction computation that can be vectorized. 5056 If VEC_STMT is also passed, vectorize the induction PHI: create a vectorized 5057 phi to replace it, put it in VEC_STMT, and add it to the same basic block. 5058 Return FALSE if not a vectorizable STMT, TRUE otherwise. */ 5059 5060 bool 5061 vectorizable_induction (gimple phi, gimple_stmt_iterator *gsi ATTRIBUTE_UNUSED, 5062 gimple *vec_stmt) 5063 { 5064 stmt_vec_info stmt_info = vinfo_for_stmt (phi); 5065 tree vectype = STMT_VINFO_VECTYPE (stmt_info); 5066 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); 5067 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); 5068 int nunits = TYPE_VECTOR_SUBPARTS (vectype); 5069 int ncopies = LOOP_VINFO_VECT_FACTOR (loop_vinfo) / nunits; 5070 tree vec_def; 5071 5072 gcc_assert (ncopies >= 1); 5073 /* FORNOW. These restrictions should be relaxed. */ 5074 if (nested_in_vect_loop_p (loop, phi)) 5075 { 5076 imm_use_iterator imm_iter; 5077 use_operand_p use_p; 5078 gimple exit_phi; 5079 edge latch_e; 5080 tree loop_arg; 5081 5082 if (ncopies > 1) 5083 { 5084 if (vect_print_dump_info (REPORT_DETAILS)) 5085 fprintf (vect_dump, "multiple types in nested loop."); 5086 return false; 5087 } 5088 5089 exit_phi = NULL; 5090 latch_e = loop_latch_edge (loop->inner); 5091 loop_arg = PHI_ARG_DEF_FROM_EDGE (phi, latch_e); 5092 FOR_EACH_IMM_USE_FAST (use_p, imm_iter, loop_arg) 5093 { 5094 if (!flow_bb_inside_loop_p (loop->inner, 5095 gimple_bb (USE_STMT (use_p)))) 5096 { 5097 exit_phi = USE_STMT (use_p); 5098 break; 5099 } 5100 } 5101 if (exit_phi) 5102 { 5103 stmt_vec_info exit_phi_vinfo = vinfo_for_stmt (exit_phi); 5104 if (!(STMT_VINFO_RELEVANT_P (exit_phi_vinfo) 5105 && !STMT_VINFO_LIVE_P (exit_phi_vinfo))) 5106 { 5107 if (vect_print_dump_info (REPORT_DETAILS)) 5108 fprintf (vect_dump, "inner-loop induction only used outside " 5109 "of the outer vectorized loop."); 5110 return false; 5111 } 5112 } 5113 } 5114 5115 if (!STMT_VINFO_RELEVANT_P (stmt_info)) 5116 return false; 5117 5118 /* FORNOW: SLP not supported. */ 5119 if (STMT_SLP_TYPE (stmt_info)) 5120 return false; 5121 5122 gcc_assert (STMT_VINFO_DEF_TYPE (stmt_info) == vect_induction_def); 5123 5124 if (gimple_code (phi) != GIMPLE_PHI) 5125 return false; 5126 5127 if (!vec_stmt) /* transformation not required. */ 5128 { 5129 STMT_VINFO_TYPE (stmt_info) = induc_vec_info_type; 5130 if (vect_print_dump_info (REPORT_DETAILS)) 5131 fprintf (vect_dump, "=== vectorizable_induction ==="); 5132 vect_model_induction_cost (stmt_info, ncopies); 5133 return true; 5134 } 5135 5136 /** Transform. **/ 5137 5138 if (vect_print_dump_info (REPORT_DETAILS)) 5139 fprintf (vect_dump, "transform induction phi."); 5140 5141 vec_def = get_initial_def_for_induction (phi); 5142 *vec_stmt = SSA_NAME_DEF_STMT (vec_def); 5143 return true; 5144 } 5145 5146 /* Function vectorizable_live_operation. 5147 5148 STMT computes a value that is used outside the loop. Check if 5149 it can be supported. */ 5150 5151 bool 5152 vectorizable_live_operation (gimple stmt, 5153 gimple_stmt_iterator *gsi ATTRIBUTE_UNUSED, 5154 gimple *vec_stmt ATTRIBUTE_UNUSED) 5155 { 5156 stmt_vec_info stmt_info = vinfo_for_stmt (stmt); 5157 loop_vec_info loop_vinfo = STMT_VINFO_LOOP_VINFO (stmt_info); 5158 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); 5159 int i; 5160 int op_type; 5161 tree op; 5162 tree def; 5163 gimple def_stmt; 5164 enum vect_def_type dt; 5165 enum tree_code code; 5166 enum gimple_rhs_class rhs_class; 5167 5168 gcc_assert (STMT_VINFO_LIVE_P (stmt_info)); 5169 5170 if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def) 5171 return false; 5172 5173 if (!is_gimple_assign (stmt)) 5174 return false; 5175 5176 if (TREE_CODE (gimple_assign_lhs (stmt)) != SSA_NAME) 5177 return false; 5178 5179 /* FORNOW. CHECKME. */ 5180 if (nested_in_vect_loop_p (loop, stmt)) 5181 return false; 5182 5183 code = gimple_assign_rhs_code (stmt); 5184 op_type = TREE_CODE_LENGTH (code); 5185 rhs_class = get_gimple_rhs_class (code); 5186 gcc_assert (rhs_class != GIMPLE_UNARY_RHS || op_type == unary_op); 5187 gcc_assert (rhs_class != GIMPLE_BINARY_RHS || op_type == binary_op); 5188 5189 /* FORNOW: support only if all uses are invariant. This means 5190 that the scalar operations can remain in place, unvectorized. 5191 The original last scalar value that they compute will be used. */ 5192 5193 for (i = 0; i < op_type; i++) 5194 { 5195 if (rhs_class == GIMPLE_SINGLE_RHS) 5196 op = TREE_OPERAND (gimple_op (stmt, 1), i); 5197 else 5198 op = gimple_op (stmt, i + 1); 5199 if (op 5200 && !vect_is_simple_use (op, stmt, loop_vinfo, NULL, &def_stmt, &def, 5201 &dt)) 5202 { 5203 if (vect_print_dump_info (REPORT_DETAILS)) 5204 fprintf (vect_dump, "use not simple."); 5205 return false; 5206 } 5207 5208 if (dt != vect_external_def && dt != vect_constant_def) 5209 return false; 5210 } 5211 5212 /* No transformation is required for the cases we currently support. */ 5213 return true; 5214 } 5215 5216 /* Kill any debug uses outside LOOP of SSA names defined in STMT. */ 5217 5218 static void 5219 vect_loop_kill_debug_uses (struct loop *loop, gimple stmt) 5220 { 5221 ssa_op_iter op_iter; 5222 imm_use_iterator imm_iter; 5223 def_operand_p def_p; 5224 gimple ustmt; 5225 5226 FOR_EACH_PHI_OR_STMT_DEF (def_p, stmt, op_iter, SSA_OP_DEF) 5227 { 5228 FOR_EACH_IMM_USE_STMT (ustmt, imm_iter, DEF_FROM_PTR (def_p)) 5229 { 5230 basic_block bb; 5231 5232 if (!is_gimple_debug (ustmt)) 5233 continue; 5234 5235 bb = gimple_bb (ustmt); 5236 5237 if (!flow_bb_inside_loop_p (loop, bb)) 5238 { 5239 if (gimple_debug_bind_p (ustmt)) 5240 { 5241 if (vect_print_dump_info (REPORT_DETAILS)) 5242 fprintf (vect_dump, "killing debug use"); 5243 5244 gimple_debug_bind_reset_value (ustmt); 5245 update_stmt (ustmt); 5246 } 5247 else 5248 gcc_unreachable (); 5249 } 5250 } 5251 } 5252 } 5253 5254 /* Function vect_transform_loop. 5255 5256 The analysis phase has determined that the loop is vectorizable. 5257 Vectorize the loop - created vectorized stmts to replace the scalar 5258 stmts in the loop, and update the loop exit condition. */ 5259 5260 void 5261 vect_transform_loop (loop_vec_info loop_vinfo) 5262 { 5263 struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo); 5264 basic_block *bbs = LOOP_VINFO_BBS (loop_vinfo); 5265 int nbbs = loop->num_nodes; 5266 gimple_stmt_iterator si; 5267 int i; 5268 tree ratio = NULL; 5269 int vectorization_factor = LOOP_VINFO_VECT_FACTOR (loop_vinfo); 5270 bool strided_store; 5271 bool slp_scheduled = false; 5272 unsigned int nunits; 5273 tree cond_expr = NULL_TREE; 5274 gimple_seq cond_expr_stmt_list = NULL; 5275 bool do_peeling_for_loop_bound; 5276 gimple stmt, pattern_stmt; 5277 gimple_seq pattern_def_seq = NULL; 5278 gimple_stmt_iterator pattern_def_si = gsi_start (NULL); 5279 bool transform_pattern_stmt = false; 5280 5281 if (vect_print_dump_info (REPORT_DETAILS)) 5282 fprintf (vect_dump, "=== vec_transform_loop ==="); 5283 5284 /* Peel the loop if there are data refs with unknown alignment. 5285 Only one data ref with unknown store is allowed. */ 5286 5287 if (LOOP_PEELING_FOR_ALIGNMENT (loop_vinfo)) 5288 vect_do_peeling_for_alignment (loop_vinfo); 5289 5290 do_peeling_for_loop_bound 5291 = (!LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) 5292 || (LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) 5293 && LOOP_VINFO_INT_NITERS (loop_vinfo) % vectorization_factor != 0) 5294 || LOOP_VINFO_PEELING_FOR_GAPS (loop_vinfo)); 5295 5296 if (LOOP_REQUIRES_VERSIONING_FOR_ALIGNMENT (loop_vinfo) 5297 || LOOP_REQUIRES_VERSIONING_FOR_ALIAS (loop_vinfo)) 5298 vect_loop_versioning (loop_vinfo, 5299 !do_peeling_for_loop_bound, 5300 &cond_expr, &cond_expr_stmt_list); 5301 5302 /* If the loop has a symbolic number of iterations 'n' (i.e. it's not a 5303 compile time constant), or it is a constant that doesn't divide by the 5304 vectorization factor, then an epilog loop needs to be created. 5305 We therefore duplicate the loop: the original loop will be vectorized, 5306 and will compute the first (n/VF) iterations. The second copy of the loop 5307 will remain scalar and will compute the remaining (n%VF) iterations. 5308 (VF is the vectorization factor). */ 5309 5310 if (do_peeling_for_loop_bound) 5311 vect_do_peeling_for_loop_bound (loop_vinfo, &ratio, 5312 cond_expr, cond_expr_stmt_list); 5313 else 5314 ratio = build_int_cst (TREE_TYPE (LOOP_VINFO_NITERS (loop_vinfo)), 5315 LOOP_VINFO_INT_NITERS (loop_vinfo) / vectorization_factor); 5316 5317 /* 1) Make sure the loop header has exactly two entries 5318 2) Make sure we have a preheader basic block. */ 5319 5320 gcc_assert (EDGE_COUNT (loop->header->preds) == 2); 5321 5322 split_edge (loop_preheader_edge (loop)); 5323 5324 /* FORNOW: the vectorizer supports only loops which body consist 5325 of one basic block (header + empty latch). When the vectorizer will 5326 support more involved loop forms, the order by which the BBs are 5327 traversed need to be reconsidered. */ 5328 5329 for (i = 0; i < nbbs; i++) 5330 { 5331 basic_block bb = bbs[i]; 5332 stmt_vec_info stmt_info; 5333 gimple phi; 5334 5335 for (si = gsi_start_phis (bb); !gsi_end_p (si); gsi_next (&si)) 5336 { 5337 phi = gsi_stmt (si); 5338 if (vect_print_dump_info (REPORT_DETAILS)) 5339 { 5340 fprintf (vect_dump, "------>vectorizing phi: "); 5341 print_gimple_stmt (vect_dump, phi, 0, TDF_SLIM); 5342 } 5343 stmt_info = vinfo_for_stmt (phi); 5344 if (!stmt_info) 5345 continue; 5346 5347 if (MAY_HAVE_DEBUG_STMTS && !STMT_VINFO_LIVE_P (stmt_info)) 5348 vect_loop_kill_debug_uses (loop, phi); 5349 5350 if (!STMT_VINFO_RELEVANT_P (stmt_info) 5351 && !STMT_VINFO_LIVE_P (stmt_info)) 5352 continue; 5353 5354 if ((TYPE_VECTOR_SUBPARTS (STMT_VINFO_VECTYPE (stmt_info)) 5355 != (unsigned HOST_WIDE_INT) vectorization_factor) 5356 && vect_print_dump_info (REPORT_DETAILS)) 5357 fprintf (vect_dump, "multiple-types."); 5358 5359 if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_induction_def) 5360 { 5361 if (vect_print_dump_info (REPORT_DETAILS)) 5362 fprintf (vect_dump, "transform phi."); 5363 vect_transform_stmt (phi, NULL, NULL, NULL, NULL); 5364 } 5365 } 5366 5367 pattern_stmt = NULL; 5368 for (si = gsi_start_bb (bb); !gsi_end_p (si) || transform_pattern_stmt;) 5369 { 5370 bool is_store; 5371 5372 if (transform_pattern_stmt) 5373 stmt = pattern_stmt; 5374 else 5375 stmt = gsi_stmt (si); 5376 5377 if (vect_print_dump_info (REPORT_DETAILS)) 5378 { 5379 fprintf (vect_dump, "------>vectorizing statement: "); 5380 print_gimple_stmt (vect_dump, stmt, 0, TDF_SLIM); 5381 } 5382 5383 stmt_info = vinfo_for_stmt (stmt); 5384 5385 /* vector stmts created in the outer-loop during vectorization of 5386 stmts in an inner-loop may not have a stmt_info, and do not 5387 need to be vectorized. */ 5388 if (!stmt_info) 5389 { 5390 gsi_next (&si); 5391 continue; 5392 } 5393 5394 if (MAY_HAVE_DEBUG_STMTS && !STMT_VINFO_LIVE_P (stmt_info)) 5395 vect_loop_kill_debug_uses (loop, stmt); 5396 5397 if (!STMT_VINFO_RELEVANT_P (stmt_info) 5398 && !STMT_VINFO_LIVE_P (stmt_info)) 5399 { 5400 if (STMT_VINFO_IN_PATTERN_P (stmt_info) 5401 && (pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info)) 5402 && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) 5403 || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) 5404 { 5405 stmt = pattern_stmt; 5406 stmt_info = vinfo_for_stmt (stmt); 5407 } 5408 else 5409 { 5410 gsi_next (&si); 5411 continue; 5412 } 5413 } 5414 else if (STMT_VINFO_IN_PATTERN_P (stmt_info) 5415 && (pattern_stmt = STMT_VINFO_RELATED_STMT (stmt_info)) 5416 && (STMT_VINFO_RELEVANT_P (vinfo_for_stmt (pattern_stmt)) 5417 || STMT_VINFO_LIVE_P (vinfo_for_stmt (pattern_stmt)))) 5418 transform_pattern_stmt = true; 5419 5420 /* If pattern statement has def stmts, vectorize them too. */ 5421 if (is_pattern_stmt_p (stmt_info)) 5422 { 5423 if (pattern_def_seq == NULL) 5424 { 5425 pattern_def_seq = STMT_VINFO_PATTERN_DEF_SEQ (stmt_info); 5426 pattern_def_si = gsi_start (pattern_def_seq); 5427 } 5428 else if (!gsi_end_p (pattern_def_si)) 5429 gsi_next (&pattern_def_si); 5430 if (pattern_def_seq != NULL) 5431 { 5432 gimple pattern_def_stmt = NULL; 5433 stmt_vec_info pattern_def_stmt_info = NULL; 5434 5435 while (!gsi_end_p (pattern_def_si)) 5436 { 5437 pattern_def_stmt = gsi_stmt (pattern_def_si); 5438 pattern_def_stmt_info 5439 = vinfo_for_stmt (pattern_def_stmt); 5440 if (STMT_VINFO_RELEVANT_P (pattern_def_stmt_info) 5441 || STMT_VINFO_LIVE_P (pattern_def_stmt_info)) 5442 break; 5443 gsi_next (&pattern_def_si); 5444 } 5445 5446 if (!gsi_end_p (pattern_def_si)) 5447 { 5448 if (vect_print_dump_info (REPORT_DETAILS)) 5449 { 5450 fprintf (vect_dump, "==> vectorizing pattern def" 5451 " stmt: "); 5452 print_gimple_stmt (vect_dump, pattern_def_stmt, 0, 5453 TDF_SLIM); 5454 } 5455 5456 stmt = pattern_def_stmt; 5457 stmt_info = pattern_def_stmt_info; 5458 } 5459 else 5460 { 5461 pattern_def_si = gsi_start (NULL); 5462 transform_pattern_stmt = false; 5463 } 5464 } 5465 else 5466 transform_pattern_stmt = false; 5467 } 5468 5469 gcc_assert (STMT_VINFO_VECTYPE (stmt_info)); 5470 nunits = (unsigned int) TYPE_VECTOR_SUBPARTS ( 5471 STMT_VINFO_VECTYPE (stmt_info)); 5472 if (!STMT_SLP_TYPE (stmt_info) 5473 && nunits != (unsigned int) vectorization_factor 5474 && vect_print_dump_info (REPORT_DETAILS)) 5475 /* For SLP VF is set according to unrolling factor, and not to 5476 vector size, hence for SLP this print is not valid. */ 5477 fprintf (vect_dump, "multiple-types."); 5478 5479 /* SLP. Schedule all the SLP instances when the first SLP stmt is 5480 reached. */ 5481 if (STMT_SLP_TYPE (stmt_info)) 5482 { 5483 if (!slp_scheduled) 5484 { 5485 slp_scheduled = true; 5486 5487 if (vect_print_dump_info (REPORT_DETAILS)) 5488 fprintf (vect_dump, "=== scheduling SLP instances ==="); 5489 5490 vect_schedule_slp (loop_vinfo, NULL); 5491 } 5492 5493 /* Hybrid SLP stmts must be vectorized in addition to SLP. */ 5494 if (!vinfo_for_stmt (stmt) || PURE_SLP_STMT (stmt_info)) 5495 { 5496 if (!transform_pattern_stmt && gsi_end_p (pattern_def_si)) 5497 { 5498 pattern_def_seq = NULL; 5499 gsi_next (&si); 5500 } 5501 continue; 5502 } 5503 } 5504 5505 /* -------- vectorize statement ------------ */ 5506 if (vect_print_dump_info (REPORT_DETAILS)) 5507 fprintf (vect_dump, "transform statement."); 5508 5509 strided_store = false; 5510 is_store = vect_transform_stmt (stmt, &si, &strided_store, NULL, NULL); 5511 if (is_store) 5512 { 5513 if (STMT_VINFO_STRIDED_ACCESS (stmt_info)) 5514 { 5515 /* Interleaving. If IS_STORE is TRUE, the vectorization of the 5516 interleaving chain was completed - free all the stores in 5517 the chain. */ 5518 gsi_next (&si); 5519 vect_remove_stores (GROUP_FIRST_ELEMENT (stmt_info)); 5520 continue; 5521 } 5522 else 5523 { 5524 /* Free the attached stmt_vec_info and remove the stmt. */ 5525 free_stmt_vec_info (gsi_stmt (si)); 5526 gsi_remove (&si, true); 5527 continue; 5528 } 5529 } 5530 5531 if (!transform_pattern_stmt && gsi_end_p (pattern_def_si)) 5532 { 5533 pattern_def_seq = NULL; 5534 gsi_next (&si); 5535 } 5536 } /* stmts in BB */ 5537 } /* BBs in loop */ 5538 5539 slpeel_make_loop_iterate_ntimes (loop, ratio); 5540 5541 /* The memory tags and pointers in vectorized statements need to 5542 have their SSA forms updated. FIXME, why can't this be delayed 5543 until all the loops have been transformed? */ 5544 update_ssa (TODO_update_ssa); 5545 5546 if (vect_print_dump_info (REPORT_VECTORIZED_LOCATIONS)) 5547 fprintf (vect_dump, "LOOP VECTORIZED."); 5548 if (loop->inner && vect_print_dump_info (REPORT_VECTORIZED_LOCATIONS)) 5549 fprintf (vect_dump, "OUTER LOOP VECTORIZED."); 5550 } 5551