gcc/tree-vectorizer.c

   1 /* Vectorizer
   2    Copyright (C) 2003-2018 Free Software Foundation, Inc.
   3    Contributed by Dorit Naishlos <dorit@il.ibm.com>
   4
   5 This file is part of GCC.
   6
   7 GCC is free software; you can redistribute it and/or modify it under
   8 the terms of the GNU General Public License as published by the Free
   9 Software Foundation; either version 3, or (at your option) any later
  10 version.
  11
  12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
  13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
  14 FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  15 for more details.
  16
  17 You should have received a copy of the GNU General Public License
  18 along with GCC; see the file COPYING3.  If not see
  19 <http://www.gnu.org/licenses/>.  */
  20
  21 /* Loop and basic block vectorizer.
  22
  23   This file contains drivers for the three vectorizers:
  24   (1) loop vectorizer (inter-iteration parallelism),
  25   (2) loop-aware SLP (intra-iteration parallelism) (invoked by the loop
  26       vectorizer)
  27   (3) BB vectorizer (out-of-loops), aka SLP
  28
  29   The rest of the vectorizer's code is organized as follows:
  30   - tree-vect-loop.c - loop specific parts such as reductions, etc. These are
  31     used by drivers (1) and (2).
  32   - tree-vect-loop-manip.c - vectorizer's loop control-flow utilities, used by
  33     drivers (1) and (2).
  34   - tree-vect-slp.c - BB vectorization specific analysis and transformation,
  35     used by drivers (2) and (3).
  36   - tree-vect-stmts.c - statements analysis and transformation (used by all).
  37   - tree-vect-data-refs.c - vectorizer specific data-refs analysis and
  38     manipulations (used by all).
  39   - tree-vect-patterns.c - vectorizable code patterns detector (used by all)
  40
  41   Here's a poor attempt at illustrating that:
  42
  43      tree-vectorizer.c:
  44      loop_vect()  loop_aware_slp()  slp_vect()
  45           |        /           \          /
  46           |       /             \        /
  47           tree-vect-loop.c  tree-vect-slp.c
  48                 | \      \  /      /   |
  49                 |  \      \/      /    |
  50                 |   \     /\     /     |
  51                 |    \   /  \   /      |
  52          tree-vect-stmts.c  tree-vect-data-refs.c
  53                        \      /
  54                     tree-vect-patterns.c
  55 */
  56
  57 #include "config.h"
  58 #include "system.h"
  59 #include "coretypes.h"
  60 #include "backend.h"
  61 #include "tree.h"
  62 #include "gimple.h"
  63 #include "predict.h"
  64 #include "tree-pass.h"
  65 #include "ssa.h"
  66 #include "cgraph.h"
  67 #include "fold-const.h"
  68 #include "stor-layout.h"
  69 #include "gimple-iterator.h"
  70 #include "gimple-walk.h"
  71 #include "tree-ssa-loop-manip.h"
  72 #include "tree-ssa-loop-niter.h"
  73 #include "tree-cfg.h"
  74 #include "cfgloop.h"
  75 #include "tree-vectorizer.h"
  76 #include "tree-ssa-propagate.h"
  77 #include "dbgcnt.h"
  78 #include "tree-scalar-evolution.h"
  79 #include "stringpool.h"
  80 #include "attribs.h"
  81 #include "gimple-pretty-print.h"
  82
  83
  84 /* Loop or bb location.  */
  85 source_location vect_location;
  86
  87 /* Vector mapping GIMPLE stmt to stmt_vec_info. */
  88 vec<stmt_vec_info> *stmt_vec_info_vec;
  89
  90 /* Dump a cost entry according to args to F.  */
  91
  92 void
  93 dump_stmt_cost (FILE *f, void *data, int count, enum vect_cost_for_stmt kind,
  94                 stmt_vec_info stmt_info, int misalign,
  95                 enum vect_cost_model_location where)
  96 {
  97   fprintf (f, "%p ", data);
  98   if (stmt_info)
  99     {
 100       print_gimple_expr (f, STMT_VINFO_STMT (stmt_info), 0, TDF_SLIM);
 101       fprintf (f, " ");
 102     }
 103   else
 104     fprintf (f, "<unknown> ");
 105   fprintf (f, "%d times ", count);
 106   const char *ks = "unknown";
 107   switch (kind)
 108     {
 109     case scalar_stmt:
 110       ks = "scalar_stmt";
 111       break;
 112     case scalar_load:
 113       ks = "scalar_load";
 114       break;
 115     case scalar_store:
 116       ks = "scalar_store";
 117       break;
 118     case vector_stmt:
 119       ks = "vector_stmt";
 120       break;
 121     case vector_load:
 122       ks = "vector_load";
 123       break;
 124     case vector_gather_load:
 125       ks = "vector_gather_load";
 126       break;
 127     case unaligned_load:
 128       ks = "unaligned_load";
 129       break;
 130     case unaligned_store:
 131       ks = "unaligned_store";
 132       break;
 133     case vector_store:
 134       ks = "unaligned_store";
 135       break;
 136     case vector_scatter_store:
 137       ks = "unaligned_store";
 138       break;
 139     case vec_to_scalar:
 140       ks = "unaligned_store";
 141       break;
 142     case scalar_to_vec:
 143       ks = "unaligned_store";
 144       break;
 145     case cond_branch_not_taken:
 146       ks = "unaligned_store";
 147       break;
 148     case cond_branch_taken:
 149       ks = "unaligned_store";
 150       break;
 151     case vec_perm:
 152       ks = "unaligned_store";
 153       break;
 154     case vec_promote_demote:
 155       ks = "unaligned_store";
 156       break;
 157     case vec_construct:
 158       ks = "unaligned_store";
 159       break;
 160     }
 161   fprintf (f, "%s ", ks);
 162   if (kind == unaligned_load || kind == unaligned_store)
 163     fprintf (f, "(misalign %d) ", misalign);
 164   const char *ws = "unknown";
 165   switch (where)
 166     {
 167     case vect_prologue:
 168       ws = "prologue";
 169       break;
 170     case vect_body:
 171       ws = "body";
 172       break;
 173     case vect_epilogue:
 174       ws = "epilogue";
 175       break;
 176     }
 177   fprintf (f, "in %s\n", ws);
 178 }
 179 \f
 180 /* For mapping simduid to vectorization factor.  */
 181
 182 struct simduid_to_vf : free_ptr_hash<simduid_to_vf>
 183 {
 184   unsigned int simduid;
 185   poly_uint64 vf;
 186
 187   /* hash_table support.  */
 188   static inline hashval_t hash (const simduid_to_vf *);
 189   static inline int equal (const simduid_to_vf *, const simduid_to_vf *);
 190 };
 191
 192 inline hashval_t
 193 simduid_to_vf::hash (const simduid_to_vf *p)
 194 {
 195   return p->simduid;
 196 }
 197
 198 inline int
 199 simduid_to_vf::equal (const simduid_to_vf *p1, const simduid_to_vf *p2)
 200 {
 201   return p1->simduid == p2->simduid;
 202 }
 203
 204 /* This hash maps the OMP simd array to the corresponding simduid used
 205    to index into it.  Like thus,
 206
 207         _7 = GOMP_SIMD_LANE (simduid.0)
 208         ...
 209         ...
 210         D.1737[_7] = stuff;
 211
 212
 213    This hash maps from the OMP simd array (D.1737[]) to DECL_UID of
 214    simduid.0.  */
 215
 216 struct simd_array_to_simduid : free_ptr_hash<simd_array_to_simduid>
 217 {
 218   tree decl;
 219   unsigned int simduid;
 220
 221   /* hash_table support.  */
 222   static inline hashval_t hash (const simd_array_to_simduid *);
 223   static inline int equal (const simd_array_to_simduid *,
 224                            const simd_array_to_simduid *);
 225 };
 226
 227 inline hashval_t
 228 simd_array_to_simduid::hash (const simd_array_to_simduid *p)
 229 {
 230   return DECL_UID (p->decl);
 231 }
 232
 233 inline int
 234 simd_array_to_simduid::equal (const simd_array_to_simduid *p1,
 235                               const simd_array_to_simduid *p2)
 236 {
 237   return p1->decl == p2->decl;
 238 }
 239
 240 /* Fold IFN_GOMP_SIMD_LANE, IFN_GOMP_SIMD_VF, IFN_GOMP_SIMD_LAST_LANE,
 241    into their corresponding constants and remove
 242    IFN_GOMP_SIMD_ORDERED_{START,END}.  */
 243
 244 static void
 245 adjust_simduid_builtins (hash_table<simduid_to_vf> *htab)
 246 {
 247   basic_block bb;
 248
 249   FOR_EACH_BB_FN (bb, cfun)
 250     {
 251       gimple_stmt_iterator i;
 252
 253       for (i = gsi_start_bb (bb); !gsi_end_p (i); )
 254         {
 255           poly_uint64 vf = 1;
 256           enum internal_fn ifn;
 257           gimple *stmt = gsi_stmt (i);
 258           tree t;
 259           if (!is_gimple_call (stmt)
 260               || !gimple_call_internal_p (stmt))
 261             {
 262               gsi_next (&i);
 263               continue;
 264             }
 265           ifn = gimple_call_internal_fn (stmt);
 266           switch (ifn)
 267             {
 268             case IFN_GOMP_SIMD_LANE:
 269             case IFN_GOMP_SIMD_VF:
 270             case IFN_GOMP_SIMD_LAST_LANE:
 271               break;
 272             case IFN_GOMP_SIMD_ORDERED_START:
 273             case IFN_GOMP_SIMD_ORDERED_END:
 274               if (integer_onep (gimple_call_arg (stmt, 0)))
 275                 {
 276                   enum built_in_function bcode
 277                     = (ifn == IFN_GOMP_SIMD_ORDERED_START
 278                        ? BUILT_IN_GOMP_ORDERED_START
 279                        : BUILT_IN_GOMP_ORDERED_END);
 280                   gimple *g
 281                     = gimple_build_call (builtin_decl_explicit (bcode), 0);
 282                   tree vdef = gimple_vdef (stmt);
 283                   gimple_set_vdef (g, vdef);
 284                   SSA_NAME_DEF_STMT (vdef) = g;
 285                   gimple_set_vuse (g, gimple_vuse (stmt));
 286                   gsi_replace (&i, g, true);
 287                   continue;
 288                 }
 289               gsi_remove (&i, true);
 290               unlink_stmt_vdef (stmt);
 291               continue;
 292             default:
 293               gsi_next (&i);
 294               continue;
 295             }
 296           tree arg = gimple_call_arg (stmt, 0);
 297           gcc_assert (arg != NULL_TREE);
 298           gcc_assert (TREE_CODE (arg) == SSA_NAME);
 299           simduid_to_vf *p = NULL, data;
 300           data.simduid = DECL_UID (SSA_NAME_VAR (arg));
 301           /* Need to nullify loop safelen field since it's value is not
 302              valid after transformation.  */
 303           if (bb->loop_father && bb->loop_father->safelen > 0)
 304             bb->loop_father->safelen = 0;
 305           if (htab)
 306             {
 307               p = htab->find (&data);
 308               if (p)
 309                 vf = p->vf;
 310             }
 311           switch (ifn)
 312             {
 313             case IFN_GOMP_SIMD_VF:
 314               t = build_int_cst (unsigned_type_node, vf);
 315               break;
 316             case IFN_GOMP_SIMD_LANE:
 317               t = build_int_cst (unsigned_type_node, 0);
 318               break;
 319             case IFN_GOMP_SIMD_LAST_LANE:
 320               t = gimple_call_arg (stmt, 1);
 321               break;
 322             default:
 323               gcc_unreachable ();
 324             }
 325           tree lhs = gimple_call_lhs (stmt);
 326           if (lhs)
 327             replace_uses_by (lhs, t);
 328           release_defs (stmt);
 329           gsi_remove (&i, true);
 330         }
 331     }
 332 }
 333
 334 /* Helper structure for note_simd_array_uses.  */
 335
 336 struct note_simd_array_uses_struct
 337 {
 338   hash_table<simd_array_to_simduid> **htab;
 339   unsigned int simduid;
 340 };
 341
 342 /* Callback for note_simd_array_uses, called through walk_gimple_op.  */
 343
 344 static tree
 345 note_simd_array_uses_cb (tree *tp, int *walk_subtrees, void *data)
 346 {
 347   struct walk_stmt_info *wi = (struct walk_stmt_info *) data;
 348   struct note_simd_array_uses_struct *ns
 349     = (struct note_simd_array_uses_struct *) wi->info;
 350
 351   if (TYPE_P (*tp))
 352     *walk_subtrees = 0;
 353   else if (VAR_P (*tp)
 354            && lookup_attribute ("omp simd array", DECL_ATTRIBUTES (*tp))
 355            && DECL_CONTEXT (*tp) == current_function_decl)
 356     {
 357       simd_array_to_simduid data;
 358       if (!*ns->htab)
 359         *ns->htab = new hash_table<simd_array_to_simduid> (15);
 360       data.decl = *tp;
 361       data.simduid = ns->simduid;
 362       simd_array_to_simduid **slot = (*ns->htab)->find_slot (&data, INSERT);
 363       if (*slot == NULL)
 364         {
 365           simd_array_to_simduid *p = XNEW (simd_array_to_simduid);
 366           *p = data;
 367           *slot = p;
 368         }
 369       else if ((*slot)->simduid != ns->simduid)
 370         (*slot)->simduid = -1U;
 371       *walk_subtrees = 0;
 372     }
 373   return NULL_TREE;
 374 }
 375
 376 /* Find "omp simd array" temporaries and map them to corresponding
 377    simduid.  */
 378
 379 static void
 380 note_simd_array_uses (hash_table<simd_array_to_simduid> **htab)
 381 {
 382   basic_block bb;
 383   gimple_stmt_iterator gsi;
 384   struct walk_stmt_info wi;
 385   struct note_simd_array_uses_struct ns;
 386
 387   memset (&wi, 0, sizeof (wi));
 388   wi.info = &ns;
 389   ns.htab = htab;
 390
 391   FOR_EACH_BB_FN (bb, cfun)
 392     for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
 393       {
 394         gimple *stmt = gsi_stmt (gsi);
 395         if (!is_gimple_call (stmt) || !gimple_call_internal_p (stmt))
 396           continue;
 397         switch (gimple_call_internal_fn (stmt))
 398           {
 399           case IFN_GOMP_SIMD_LANE:
 400           case IFN_GOMP_SIMD_VF:
 401           case IFN_GOMP_SIMD_LAST_LANE:
 402             break;
 403           default:
 404             continue;
 405           }
 406         tree lhs = gimple_call_lhs (stmt);
 407         if (lhs == NULL_TREE)
 408           continue;
 409         imm_use_iterator use_iter;
 410         gimple *use_stmt;
 411         ns.simduid = DECL_UID (SSA_NAME_VAR (gimple_call_arg (stmt, 0)));
 412         FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, lhs)
 413           if (!is_gimple_debug (use_stmt))
 414             walk_gimple_op (use_stmt, note_simd_array_uses_cb, &wi);
 415       }
 416 }
 417
 418 /* Shrink arrays with "omp simd array" attribute to the corresponding
 419    vectorization factor.  */
 420
 421 static void
 422 shrink_simd_arrays
 423   (hash_table<simd_array_to_simduid> *simd_array_to_simduid_htab,
 424    hash_table<simduid_to_vf> *simduid_to_vf_htab)
 425 {
 426   for (hash_table<simd_array_to_simduid>::iterator iter
 427          = simd_array_to_simduid_htab->begin ();
 428        iter != simd_array_to_simduid_htab->end (); ++iter)
 429     if ((*iter)->simduid != -1U)
 430       {
 431         tree decl = (*iter)->decl;
 432         poly_uint64 vf = 1;
 433         if (simduid_to_vf_htab)
 434           {
 435             simduid_to_vf *p = NULL, data;
 436             data.simduid = (*iter)->simduid;
 437             p = simduid_to_vf_htab->find (&data);
 438             if (p)
 439               vf = p->vf;
 440           }
 441         tree atype
 442           = build_array_type_nelts (TREE_TYPE (TREE_TYPE (decl)), vf);
 443         TREE_TYPE (decl) = atype;
 444         relayout_decl (decl);
 445       }
 446
 447   delete simd_array_to_simduid_htab;
 448 }
 449 \f
 450 /* Initialize the vec_info with kind KIND_IN and target cost data
 451    TARGET_COST_DATA_IN.  */
 452
 453 vec_info::vec_info (vec_info::vec_kind kind_in, void *target_cost_data_in)
 454   : kind (kind_in),
 455     datarefs (vNULL),
 456     ddrs (vNULL),
 457     target_cost_data (target_cost_data_in)
 458 {
 459   stmt_vec_infos.create (50);
 460   set_stmt_vec_info_vec (&stmt_vec_infos);
 461 }
 462
 463 vec_info::~vec_info ()
 464 {
 465   slp_instance instance;
 466   struct data_reference *dr;
 467   unsigned int i;
 468
 469   FOR_EACH_VEC_ELT (datarefs, i, dr)
 470     if (dr->aux)
 471       {
 472         free (dr->aux);
 473         dr->aux = NULL;
 474       }
 475
 476   FOR_EACH_VEC_ELT (slp_instances, i, instance)
 477     vect_free_slp_instance (instance);
 478
 479   free_data_refs (datarefs);
 480   free_dependence_relations (ddrs);
 481   destroy_cost_data (target_cost_data);
 482   free_stmt_vec_infos (&stmt_vec_infos);
 483 }
 484
 485 /* A helper function to free scev and LOOP niter information, as well as
 486    clear loop constraint LOOP_C_FINITE.  */
 487
 488 void
 489 vect_free_loop_info_assumptions (struct loop *loop)
 490 {
 491   scev_reset_htab ();
 492   /* We need to explicitly reset upper bound information since they are
 493      used even after free_numbers_of_iterations_estimates.  */
 494   loop->any_upper_bound = false;
 495   loop->any_likely_upper_bound = false;
 496   free_numbers_of_iterations_estimates (loop);
 497   loop_constraint_clear (loop, LOOP_C_FINITE);
 498 }
 499
 500 /* Return whether STMT is inside the region we try to vectorize.  */
 501
 502 bool
 503 vect_stmt_in_region_p (vec_info *vinfo, gimple *stmt)
 504 {
 505   if (!gimple_bb (stmt))
 506     return false;
 507
 508   if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
 509     {
 510       struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
 511       if (!flow_bb_inside_loop_p (loop, gimple_bb (stmt)))
 512         return false;
 513     }
 514   else
 515     {
 516       bb_vec_info bb_vinfo = as_a <bb_vec_info> (vinfo);
 517       if (gimple_bb (stmt) != BB_VINFO_BB (bb_vinfo)
 518           || gimple_uid (stmt) == -1U
 519           || gimple_code (stmt) == GIMPLE_PHI)
 520         return false;
 521     }
 522
 523   return true;
 524 }
 525
 526
 527 /* If LOOP has been versioned during ifcvt, return the internal call
 528    guarding it.  */
 529
 530 static gimple *
 531 vect_loop_vectorized_call (struct loop *loop)
 532 {
 533   basic_block bb = loop_preheader_edge (loop)->src;
 534   gimple *g;
 535   do
 536     {
 537       g = last_stmt (bb);
 538       if (g)
 539         break;
 540       if (!single_pred_p (bb))
 541         break;
 542       bb = single_pred (bb);
 543     }
 544   while (1);
 545   if (g && gimple_code (g) == GIMPLE_COND)
 546     {
 547       gimple_stmt_iterator gsi = gsi_for_stmt (g);
 548       gsi_prev (&gsi);
 549       if (!gsi_end_p (gsi))
 550         {
 551           g = gsi_stmt (gsi);
 552           if (gimple_call_internal_p (g, IFN_LOOP_VECTORIZED)
 553               && (tree_to_shwi (gimple_call_arg (g, 0)) == loop->num
 554                   || tree_to_shwi (gimple_call_arg (g, 1)) == loop->num))
 555             return g;
 556         }
 557     }
 558   return NULL;
 559 }
 560
 561 /* If LOOP has been versioned during loop distribution, return the gurading
 562    internal call.  */
 563
 564 static gimple *
 565 vect_loop_dist_alias_call (struct loop *loop)
 566 {
 567   basic_block bb;
 568   basic_block entry;
 569   struct loop *outer, *orig;
 570   gimple_stmt_iterator gsi;
 571   gimple *g;
 572
 573   if (loop->orig_loop_num == 0)
 574     return NULL;
 575
 576   orig = get_loop (cfun, loop->orig_loop_num);
 577   if (orig == NULL)
 578     {
 579       /* The original loop is somehow destroyed.  Clear the information.  */
 580       loop->orig_loop_num = 0;
 581       return NULL;
 582     }
 583
 584   if (loop != orig)
 585     bb = nearest_common_dominator (CDI_DOMINATORS, loop->header, orig->header);
 586   else
 587     bb = loop_preheader_edge (loop)->src;
 588
 589   outer = bb->loop_father;
 590   entry = ENTRY_BLOCK_PTR_FOR_FN (cfun);
 591
 592   /* Look upward in dominance tree.  */
 593   for (; bb != entry && flow_bb_inside_loop_p (outer, bb);
 594        bb = get_immediate_dominator (CDI_DOMINATORS, bb))
 595     {
 596       g = last_stmt (bb);
 597       if (g == NULL || gimple_code (g) != GIMPLE_COND)
 598         continue;
 599
 600       gsi = gsi_for_stmt (g);
 601       gsi_prev (&gsi);
 602       if (gsi_end_p (gsi))
 603         continue;
 604
 605       g = gsi_stmt (gsi);
 606       /* The guarding internal function call must have the same distribution
 607          alias id.  */
 608       if (gimple_call_internal_p (g, IFN_LOOP_DIST_ALIAS)
 609           && (tree_to_shwi (gimple_call_arg (g, 0)) == loop->orig_loop_num))
 610         return g;
 611     }
 612   return NULL;
 613 }
 614
 615 /* Set the uids of all the statements in basic blocks inside loop
 616    represented by LOOP_VINFO. LOOP_VECTORIZED_CALL is the internal
 617    call guarding the loop which has been if converted.  */
 618 static void
 619 set_uid_loop_bbs (loop_vec_info loop_vinfo, gimple *loop_vectorized_call)
 620 {
 621   tree arg = gimple_call_arg (loop_vectorized_call, 1);
 622   basic_block *bbs;
 623   unsigned int i;
 624   struct loop *scalar_loop = get_loop (cfun, tree_to_shwi (arg));
 625
 626   LOOP_VINFO_SCALAR_LOOP (loop_vinfo) = scalar_loop;
 627   gcc_checking_assert (vect_loop_vectorized_call (scalar_loop)
 628                        == loop_vectorized_call);
 629   /* If we are going to vectorize outer loop, prevent vectorization
 630      of the inner loop in the scalar loop - either the scalar loop is
 631      thrown away, so it is a wasted work, or is used only for
 632      a few iterations.  */
 633   if (scalar_loop->inner)
 634     {
 635       gimple *g = vect_loop_vectorized_call (scalar_loop->inner);
 636       if (g)
 637         {
 638           arg = gimple_call_arg (g, 0);
 639           get_loop (cfun, tree_to_shwi (arg))->dont_vectorize = true;
 640           fold_loop_internal_call (g, boolean_false_node);
 641         }
 642     }
 643   bbs = get_loop_body (scalar_loop);
 644   for (i = 0; i < scalar_loop->num_nodes; i++)
 645     {
 646       basic_block bb = bbs[i];
 647       gimple_stmt_iterator gsi;
 648       for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
 649         {
 650           gimple *phi = gsi_stmt (gsi);
 651           gimple_set_uid (phi, 0);
 652         }
 653       for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
 654         {
 655           gimple *stmt = gsi_stmt (gsi);
 656           gimple_set_uid (stmt, 0);
 657         }
 658     }
 659   free (bbs);
 660 }
 661
 662 /* Function vectorize_loops.
 663
 664    Entry point to loop vectorization phase.  */
 665
 666 unsigned
 667 vectorize_loops (void)
 668 {
 669   unsigned int i;
 670   unsigned int num_vectorized_loops = 0;
 671   unsigned int vect_loops_num;
 672   struct loop *loop;
 673   hash_table<simduid_to_vf> *simduid_to_vf_htab = NULL;
 674   hash_table<simd_array_to_simduid> *simd_array_to_simduid_htab = NULL;
 675   bool any_ifcvt_loops = false;
 676   unsigned ret = 0;
 677   struct loop *new_loop;
 678
 679   vect_loops_num = number_of_loops (cfun);
 680
 681   /* Bail out if there are no loops.  */
 682   if (vect_loops_num <= 1)
 683     return 0;
 684
 685   if (cfun->has_simduid_loops)
 686     note_simd_array_uses (&simd_array_to_simduid_htab);
 687
 688   set_stmt_vec_info_vec (NULL);
 689
 690   /*  ----------- Analyze loops. -----------  */
 691
 692   /* If some loop was duplicated, it gets bigger number
 693      than all previously defined loops.  This fact allows us to run
 694      only over initial loops skipping newly generated ones.  */
 695   FOR_EACH_LOOP (loop, 0)
 696     if (loop->dont_vectorize)
 697       {
 698         any_ifcvt_loops = true;
 699         /* If-conversion sometimes versions both the outer loop
 700            (for the case when outer loop vectorization might be
 701            desirable) as well as the inner loop in the scalar version
 702            of the loop.  So we have:
 703             if (LOOP_VECTORIZED (1, 3))
 704               {
 705                 loop1
 706                   loop2
 707               }
 708             else
 709               loop3 (copy of loop1)
 710                 if (LOOP_VECTORIZED (4, 5))
 711                   loop4 (copy of loop2)
 712                 else
 713                   loop5 (copy of loop4)
 714            If FOR_EACH_LOOP gives us loop3 first (which has
 715            dont_vectorize set), make sure to process loop1 before loop4;
 716            so that we can prevent vectorization of loop4 if loop1
 717            is successfully vectorized.  */
 718         if (loop->inner)
 719           {
 720             gimple *loop_vectorized_call
 721               = vect_loop_vectorized_call (loop);
 722             if (loop_vectorized_call
 723                 && vect_loop_vectorized_call (loop->inner))
 724               {
 725                 tree arg = gimple_call_arg (loop_vectorized_call, 0);
 726                 struct loop *vector_loop
 727                   = get_loop (cfun, tree_to_shwi (arg));
 728                 if (vector_loop && vector_loop != loop)
 729                   {
 730                     loop = vector_loop;
 731                     /* Make sure we don't vectorize it twice.  */
 732                     loop->dont_vectorize = true;
 733                     goto try_vectorize;
 734                   }
 735               }
 736           }
 737       }
 738     else
 739       {
 740         loop_vec_info loop_vinfo, orig_loop_vinfo;
 741         gimple *loop_vectorized_call, *loop_dist_alias_call;
 742        try_vectorize:
 743         if (!((flag_tree_loop_vectorize
 744                && optimize_loop_nest_for_speed_p (loop))
 745               || loop->force_vectorize))
 746           continue;
 747         orig_loop_vinfo = NULL;
 748         loop_vectorized_call = vect_loop_vectorized_call (loop);
 749         loop_dist_alias_call = vect_loop_dist_alias_call (loop);
 750        vectorize_epilogue:
 751         vect_location = find_loop_location (loop);
 752         if (LOCATION_LOCUS (vect_location) != UNKNOWN_LOCATION
 753             && dump_enabled_p ())
 754           dump_printf (MSG_NOTE, "\nAnalyzing loop at %s:%d\n",
 755                        LOCATION_FILE (vect_location),
 756                        LOCATION_LINE (vect_location));
 757
 758         loop_vinfo = vect_analyze_loop (loop, orig_loop_vinfo);
 759         loop->aux = loop_vinfo;
 760
 761         if (!loop_vinfo || !LOOP_VINFO_VECTORIZABLE_P (loop_vinfo))
 762           {
 763             /* Free existing information if loop is analyzed with some
 764                assumptions.  */
 765             if (loop_constraint_set_p (loop, LOOP_C_FINITE))
 766               vect_free_loop_info_assumptions (loop);
 767
 768             /* If we applied if-conversion then try to vectorize the
 769                BB of innermost loops.
 770                ???  Ideally BB vectorization would learn to vectorize
 771                control flow by applying if-conversion on-the-fly, the
 772                following retains the if-converted loop body even when
 773                only non-if-converted parts took part in BB vectorization.  */
 774             if (flag_tree_slp_vectorize != 0
 775                 && loop_vectorized_call
 776                 && ! loop->inner)
 777               {
 778                 basic_block bb = loop->header;
 779                 bool has_mask_load_store = false;
 780                 for (gimple_stmt_iterator gsi = gsi_start_bb (bb);
 781                      !gsi_end_p (gsi); gsi_next (&gsi))
 782                   {
 783                     gimple *stmt = gsi_stmt (gsi);
 784                     if (is_gimple_call (stmt)
 785                         && gimple_call_internal_p (stmt)
 786                         && (gimple_call_internal_fn (stmt) == IFN_MASK_LOAD
 787                             || gimple_call_internal_fn (stmt) == IFN_MASK_STORE))
 788                       {
 789                         has_mask_load_store = true;
 790                         break;
 791                       }
 792                     gimple_set_uid (stmt, -1);
 793                     gimple_set_visited (stmt, false);
 794                   }
 795                 if (! has_mask_load_store && vect_slp_bb (bb))
 796                   {
 797                     dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location,
 798                                      "basic block vectorized\n");
 799                     fold_loop_internal_call (loop_vectorized_call,
 800                                              boolean_true_node);
 801                     loop_vectorized_call = NULL;
 802                     ret |= TODO_cleanup_cfg;
 803                   }
 804               }
 805             /* If outer loop vectorization fails for LOOP_VECTORIZED guarded
 806                loop, don't vectorize its inner loop; we'll attempt to
 807                vectorize LOOP_VECTORIZED guarded inner loop of the scalar
 808                loop version.  */
 809             if (loop_vectorized_call && loop->inner)
 810               loop->inner->dont_vectorize = true;
 811             continue;
 812           }
 813
 814         if (!dbg_cnt (vect_loop))
 815           {
 816             /* We may miss some if-converted loops due to
 817                debug counter.  Set any_ifcvt_loops to visit
 818                them at finalization.  */
 819             any_ifcvt_loops = true;
 820             /* Free existing information if loop is analyzed with some
 821                assumptions.  */
 822             if (loop_constraint_set_p (loop, LOOP_C_FINITE))
 823               vect_free_loop_info_assumptions (loop);
 824
 825             break;
 826           }
 827
 828         if (loop_vectorized_call)
 829           set_uid_loop_bbs (loop_vinfo, loop_vectorized_call);
 830         if (LOCATION_LOCUS (vect_location) != UNKNOWN_LOCATION
 831             && dump_enabled_p ())
 832           dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location,
 833                            "loop vectorized\n");
 834         new_loop = vect_transform_loop (loop_vinfo);
 835         num_vectorized_loops++;
 836         /* Now that the loop has been vectorized, allow it to be unrolled
 837            etc.  */
 838         loop->force_vectorize = false;
 839
 840         if (loop->simduid)
 841           {
 842             simduid_to_vf *simduid_to_vf_data = XNEW (simduid_to_vf);
 843             if (!simduid_to_vf_htab)
 844               simduid_to_vf_htab = new hash_table<simduid_to_vf> (15);
 845             simduid_to_vf_data->simduid = DECL_UID (loop->simduid);
 846             simduid_to_vf_data->vf = loop_vinfo->vectorization_factor;
 847             *simduid_to_vf_htab->find_slot (simduid_to_vf_data, INSERT)
 848               = simduid_to_vf_data;
 849           }
 850
 851         if (loop_vectorized_call)
 852           {
 853             fold_loop_internal_call (loop_vectorized_call, boolean_true_node);
 854             loop_vectorized_call = NULL;
 855             ret |= TODO_cleanup_cfg;
 856           }
 857         if (loop_dist_alias_call)
 858           {
 859             tree value = gimple_call_arg (loop_dist_alias_call, 1);
 860             fold_loop_internal_call (loop_dist_alias_call, value);
 861             loop_dist_alias_call = NULL;
 862             ret |= TODO_cleanup_cfg;
 863           }
 864
 865         if (new_loop)
 866           {
 867             /* Epilogue of vectorized loop must be vectorized too.  */
 868             vect_loops_num = number_of_loops (cfun);
 869             loop = new_loop;
 870             orig_loop_vinfo = loop_vinfo;  /* To pass vect_analyze_loop.  */
 871             goto vectorize_epilogue;
 872           }
 873       }
 874
 875   vect_location = UNKNOWN_LOCATION;
 876
 877   statistics_counter_event (cfun, "Vectorized loops", num_vectorized_loops);
 878   if (dump_enabled_p ()
 879       || (num_vectorized_loops > 0 && dump_enabled_p ()))
 880     dump_printf_loc (MSG_NOTE, vect_location,
 881                      "vectorized %u loops in function.\n",
 882                      num_vectorized_loops);
 883
 884   /*  ----------- Finalize. -----------  */
 885
 886   if (any_ifcvt_loops)
 887     for (i = 1; i < vect_loops_num; i++)
 888       {
 889         loop = get_loop (cfun, i);
 890         if (loop && loop->dont_vectorize)
 891           {
 892             gimple *g = vect_loop_vectorized_call (loop);
 893             if (g)
 894               {
 895                 fold_loop_internal_call (g, boolean_false_node);
 896                 ret |= TODO_cleanup_cfg;
 897                 g = NULL;
 898               }
 899             else
 900               g = vect_loop_dist_alias_call (loop);
 901
 902             if (g)
 903               {
 904                 fold_loop_internal_call (g, boolean_false_node);
 905                 ret |= TODO_cleanup_cfg;
 906               }
 907           }
 908       }
 909
 910   for (i = 1; i < vect_loops_num; i++)
 911     {
 912       loop_vec_info loop_vinfo;
 913       bool has_mask_store;
 914
 915       loop = get_loop (cfun, i);
 916       if (!loop)
 917         continue;
 918       loop_vinfo = (loop_vec_info) loop->aux;
 919       has_mask_store = false;
 920       if (loop_vinfo)
 921         has_mask_store = LOOP_VINFO_HAS_MASK_STORE (loop_vinfo);
 922       delete loop_vinfo;
 923       if (has_mask_store
 924           && targetm.vectorize.empty_mask_is_expensive (IFN_MASK_STORE))
 925         optimize_mask_stores (loop);
 926       loop->aux = NULL;
 927     }
 928
 929   /* Fold IFN_GOMP_SIMD_{VF,LANE,LAST_LANE,ORDERED_{START,END}} builtins.  */
 930   if (cfun->has_simduid_loops)
 931     adjust_simduid_builtins (simduid_to_vf_htab);
 932
 933   /* Shrink any "omp array simd" temporary arrays to the
 934      actual vectorization factors.  */
 935   if (simd_array_to_simduid_htab)
 936     shrink_simd_arrays (simd_array_to_simduid_htab, simduid_to_vf_htab);
 937   delete simduid_to_vf_htab;
 938   cfun->has_simduid_loops = false;
 939
 940   if (num_vectorized_loops > 0)
 941     {
 942       /* If we vectorized any loop only virtual SSA form needs to be updated.
 943          ???  Also while we try hard to update loop-closed SSA form we fail
 944          to properly do this in some corner-cases (see PR56286).  */
 945       rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa_only_virtuals);
 946       return TODO_cleanup_cfg;
 947     }
 948
 949   return ret;
 950 }
 951
 952
 953 /* Entry point to the simduid cleanup pass.  */
 954
 955 namespace {
 956
 957 const pass_data pass_data_simduid_cleanup =
 958 {
 959   GIMPLE_PASS, /* type */
 960   "simduid", /* name */
 961   OPTGROUP_NONE, /* optinfo_flags */
 962   TV_NONE, /* tv_id */
 963   ( PROP_ssa | PROP_cfg ), /* properties_required */
 964   0, /* properties_provided */
 965   0, /* properties_destroyed */
 966   0, /* todo_flags_start */
 967   0, /* todo_flags_finish */
 968 };
 969
 970 class pass_simduid_cleanup : public gimple_opt_pass
 971 {
 972 public:
 973   pass_simduid_cleanup (gcc::context *ctxt)
 974     : gimple_opt_pass (pass_data_simduid_cleanup, ctxt)
 975   {}
 976
 977   /* opt_pass methods: */
 978   opt_pass * clone () { return new pass_simduid_cleanup (m_ctxt); }
 979   virtual bool gate (function *fun) { return fun->has_simduid_loops; }
 980   virtual unsigned int execute (function *);
 981
 982 }; // class pass_simduid_cleanup
 983
 984 unsigned int
 985 pass_simduid_cleanup::execute (function *fun)
 986 {
 987   hash_table<simd_array_to_simduid> *simd_array_to_simduid_htab = NULL;
 988
 989   note_simd_array_uses (&simd_array_to_simduid_htab);
 990
 991   /* Fold IFN_GOMP_SIMD_{VF,LANE,LAST_LANE,ORDERED_{START,END}} builtins.  */
 992   adjust_simduid_builtins (NULL);
 993
 994   /* Shrink any "omp array simd" temporary arrays to the
 995      actual vectorization factors.  */
 996   if (simd_array_to_simduid_htab)
 997     shrink_simd_arrays (simd_array_to_simduid_htab, NULL);
 998   fun->has_simduid_loops = false;
 999   return 0;
1000 }
1001
1002 }  // anon namespace
1003
1004 gimple_opt_pass *
1005 make_pass_simduid_cleanup (gcc::context *ctxt)
1006 {
1007   return new pass_simduid_cleanup (ctxt);
1008 }
1009
1010
1011 /*  Entry point to basic block SLP phase.  */
1012
1013 namespace {
1014
1015 const pass_data pass_data_slp_vectorize =
1016 {
1017   GIMPLE_PASS, /* type */
1018   "slp", /* name */
1019   OPTGROUP_LOOP | OPTGROUP_VEC, /* optinfo_flags */
1020   TV_TREE_SLP_VECTORIZATION, /* tv_id */
1021   ( PROP_ssa | PROP_cfg ), /* properties_required */
1022   0, /* properties_provided */
1023   0, /* properties_destroyed */
1024   0, /* todo_flags_start */
1025   TODO_update_ssa, /* todo_flags_finish */
1026 };
1027
1028 class pass_slp_vectorize : public gimple_opt_pass
1029 {
1030 public:
1031   pass_slp_vectorize (gcc::context *ctxt)
1032     : gimple_opt_pass (pass_data_slp_vectorize, ctxt)
1033   {}
1034
1035   /* opt_pass methods: */
1036   opt_pass * clone () { return new pass_slp_vectorize (m_ctxt); }
1037   virtual bool gate (function *) { return flag_tree_slp_vectorize != 0; }
1038   virtual unsigned int execute (function *);
1039
1040 }; // class pass_slp_vectorize
1041
1042 unsigned int
1043 pass_slp_vectorize::execute (function *fun)
1044 {
1045   basic_block bb;
1046
1047   bool in_loop_pipeline = scev_initialized_p ();
1048   if (!in_loop_pipeline)
1049     {
1050       loop_optimizer_init (LOOPS_NORMAL);
1051       scev_initialize ();
1052     }
1053
1054   /* Mark all stmts as not belonging to the current region and unvisited.  */
1055   FOR_EACH_BB_FN (bb, fun)
1056     {
1057       for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi);
1058            gsi_next (&gsi))
1059         {
1060           gimple *stmt = gsi_stmt (gsi);
1061           gimple_set_uid (stmt, -1);
1062           gimple_set_visited (stmt, false);
1063         }
1064     }
1065
1066   FOR_EACH_BB_FN (bb, fun)
1067     {
1068       if (vect_slp_bb (bb))
1069         dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location,
1070                          "basic block vectorized\n");
1071     }
1072
1073   if (!in_loop_pipeline)
1074     {
1075       scev_finalize ();
1076       loop_optimizer_finalize ();
1077     }
1078
1079   return 0;
1080 }
1081
1082 } // anon namespace
1083
1084 gimple_opt_pass *
1085 make_pass_slp_vectorize (gcc::context *ctxt)
1086 {
1087   return new pass_slp_vectorize (ctxt);
1088 }
1089
1090
1091 /* Increase alignment of global arrays to improve vectorization potential.
1092    TODO:
1093    - Consider also structs that have an array field.
1094    - Use ipa analysis to prune arrays that can't be vectorized?
1095      This should involve global alignment analysis and in the future also
1096      array padding.  */
1097
1098 static unsigned get_vec_alignment_for_type (tree);
1099 static hash_map<tree, unsigned> *type_align_map;
1100
1101 /* Return alignment of array's vector type corresponding to scalar type.
1102    0 if no vector type exists.  */
1103 static unsigned
1104 get_vec_alignment_for_array_type (tree type)
1105 {
1106   gcc_assert (TREE_CODE (type) == ARRAY_TYPE);
1107   poly_uint64 array_size, vector_size;
1108
1109   tree vectype = get_vectype_for_scalar_type (strip_array_types (type));
1110   if (!vectype
1111       || !poly_int_tree_p (TYPE_SIZE (type), &array_size)
1112       || !poly_int_tree_p (TYPE_SIZE (vectype), &vector_size)
1113       || maybe_lt (array_size, vector_size))
1114     return 0;
1115
1116   return TYPE_ALIGN (vectype);
1117 }
1118
1119 /* Return alignment of field having maximum alignment of vector type
1120    corresponding to it's scalar type. For now, we only consider fields whose
1121    offset is a multiple of it's vector alignment.
1122    0 if no suitable field is found.  */
1123 static unsigned
1124 get_vec_alignment_for_record_type (tree type)
1125 {
1126   gcc_assert (TREE_CODE (type) == RECORD_TYPE);
1127
1128   unsigned max_align = 0, alignment;
1129   HOST_WIDE_INT offset;
1130   tree offset_tree;
1131
1132   if (TYPE_PACKED (type))
1133     return 0;
1134
1135   unsigned *slot = type_align_map->get (type);
1136   if (slot)
1137     return *slot;
1138
1139   for (tree field = first_field (type);
1140        field != NULL_TREE;
1141        field = DECL_CHAIN (field))
1142     {
1143       /* Skip if not FIELD_DECL or if alignment is set by user.  */
1144       if (TREE_CODE (field) != FIELD_DECL
1145           || DECL_USER_ALIGN (field)
1146           || DECL_ARTIFICIAL (field))
1147         continue;
1148
1149       /* We don't need to process the type further if offset is variable,
1150          since the offsets of remaining members will also be variable.  */
1151       if (TREE_CODE (DECL_FIELD_OFFSET (field)) != INTEGER_CST
1152           || TREE_CODE (DECL_FIELD_BIT_OFFSET (field)) != INTEGER_CST)
1153         break;
1154
1155       /* Similarly stop processing the type if offset_tree
1156          does not fit in unsigned HOST_WIDE_INT.  */
1157       offset_tree = bit_position (field);
1158       if (!tree_fits_uhwi_p (offset_tree))
1159         break;
1160
1161       offset = tree_to_uhwi (offset_tree);
1162       alignment = get_vec_alignment_for_type (TREE_TYPE (field));
1163
1164       /* Get maximum alignment of vectorized field/array among those members
1165          whose offset is multiple of the vector alignment.  */
1166       if (alignment
1167           && (offset % alignment == 0)
1168           && (alignment > max_align))
1169         max_align = alignment;
1170     }
1171
1172   type_align_map->put (type, max_align);
1173   return max_align;
1174 }
1175
1176 /* Return alignment of vector type corresponding to decl's scalar type
1177    or 0 if it doesn't exist or the vector alignment is lesser than
1178    decl's alignment.  */
1179 static unsigned
1180 get_vec_alignment_for_type (tree type)
1181 {
1182   if (type == NULL_TREE)
1183     return 0;
1184
1185   gcc_assert (TYPE_P (type));
1186
1187   static unsigned alignment = 0;
1188   switch (TREE_CODE (type))
1189     {
1190       case ARRAY_TYPE:
1191         alignment = get_vec_alignment_for_array_type (type);
1192         break;
1193       case RECORD_TYPE:
1194         alignment = get_vec_alignment_for_record_type (type);
1195         break;
1196       default:
1197         alignment = 0;
1198         break;
1199     }
1200
1201   return (alignment > TYPE_ALIGN (type)) ? alignment : 0;
1202 }
1203
1204 /* Entry point to increase_alignment pass.  */
1205 static unsigned int
1206 increase_alignment (void)
1207 {
1208   varpool_node *vnode;
1209
1210   vect_location = UNKNOWN_LOCATION;
1211   type_align_map = new hash_map<tree, unsigned>;
1212
1213   /* Increase the alignment of all global arrays for vectorization.  */
1214   FOR_EACH_DEFINED_VARIABLE (vnode)
1215     {
1216       tree decl = vnode->decl;
1217       unsigned int alignment;
1218
1219       if ((decl_in_symtab_p (decl)
1220           && !symtab_node::get (decl)->can_increase_alignment_p ())
1221           || DECL_USER_ALIGN (decl) || DECL_ARTIFICIAL (decl))
1222         continue;
1223
1224       alignment = get_vec_alignment_for_type (TREE_TYPE (decl));
1225       if (alignment && vect_can_force_dr_alignment_p (decl, alignment))
1226         {
1227           vnode->increase_alignment (alignment);
1228           dump_printf (MSG_NOTE, "Increasing alignment of decl: ");
1229           dump_generic_expr (MSG_NOTE, TDF_SLIM, decl);
1230           dump_printf (MSG_NOTE, "\n");
1231         }
1232     }
1233
1234   delete type_align_map;
1235   return 0;
1236 }
1237
1238
1239 namespace {
1240
1241 const pass_data pass_data_ipa_increase_alignment =
1242 {
1243   SIMPLE_IPA_PASS, /* type */
1244   "increase_alignment", /* name */
1245   OPTGROUP_LOOP | OPTGROUP_VEC, /* optinfo_flags */
1246   TV_IPA_OPT, /* tv_id */
1247   0, /* properties_required */
1248   0, /* properties_provided */
1249   0, /* properties_destroyed */
1250   0, /* todo_flags_start */
1251   0, /* todo_flags_finish */
1252 };
1253
1254 class pass_ipa_increase_alignment : public simple_ipa_opt_pass
1255 {
1256 public:
1257   pass_ipa_increase_alignment (gcc::context *ctxt)
1258     : simple_ipa_opt_pass (pass_data_ipa_increase_alignment, ctxt)
1259   {}
1260
1261   /* opt_pass methods: */
1262   virtual bool gate (function *)
1263     {
1264       return flag_section_anchors && flag_tree_loop_vectorize;
1265     }
1266
1267   virtual unsigned int execute (function *) { return increase_alignment (); }
1268
1269 }; // class pass_ipa_increase_alignment
1270
1271 } // anon namespace
1272
1273 simple_ipa_opt_pass *
1274 make_pass_ipa_increase_alignment (gcc::context *ctxt)
1275 {
1276   return new pass_ipa_increase_alignment (ctxt);
1277 }