gcc/tree-vectorizer.c

   1 /* Vectorizer
   2    Copyright (C) 2003-2016 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
  80
  81 /* Loop or bb location.  */
  82 source_location vect_location;
  83
  84 /* Vector mapping GIMPLE stmt to stmt_vec_info. */
  85 vec<stmt_vec_info> stmt_vec_info_vec;
  86 \f
  87 /* For mapping simduid to vectorization factor.  */
  88
  89 struct simduid_to_vf : free_ptr_hash<simduid_to_vf>
  90 {
  91   unsigned int simduid;
  92   int vf;
  93
  94   /* hash_table support.  */
  95   static inline hashval_t hash (const simduid_to_vf *);
  96   static inline int equal (const simduid_to_vf *, const simduid_to_vf *);
  97 };
  98
  99 inline hashval_t
 100 simduid_to_vf::hash (const simduid_to_vf *p)
 101 {
 102   return p->simduid;
 103 }
 104
 105 inline int
 106 simduid_to_vf::equal (const simduid_to_vf *p1, const simduid_to_vf *p2)
 107 {
 108   return p1->simduid == p2->simduid;
 109 }
 110
 111 /* This hash maps the OMP simd array to the corresponding simduid used
 112    to index into it.  Like thus,
 113
 114         _7 = GOMP_SIMD_LANE (simduid.0)
 115         ...
 116         ...
 117         D.1737[_7] = stuff;
 118
 119
 120    This hash maps from the OMP simd array (D.1737[]) to DECL_UID of
 121    simduid.0.  */
 122
 123 struct simd_array_to_simduid : free_ptr_hash<simd_array_to_simduid>
 124 {
 125   tree decl;
 126   unsigned int simduid;
 127
 128   /* hash_table support.  */
 129   static inline hashval_t hash (const simd_array_to_simduid *);
 130   static inline int equal (const simd_array_to_simduid *,
 131                            const simd_array_to_simduid *);
 132 };
 133
 134 inline hashval_t
 135 simd_array_to_simduid::hash (const simd_array_to_simduid *p)
 136 {
 137   return DECL_UID (p->decl);
 138 }
 139
 140 inline int
 141 simd_array_to_simduid::equal (const simd_array_to_simduid *p1,
 142                               const simd_array_to_simduid *p2)
 143 {
 144   return p1->decl == p2->decl;
 145 }
 146
 147 /* Fold IFN_GOMP_SIMD_LANE, IFN_GOMP_SIMD_VF, IFN_GOMP_SIMD_LAST_LANE,
 148    into their corresponding constants and remove
 149    IFN_GOMP_SIMD_ORDERED_{START,END}.  */
 150
 151 static void
 152 adjust_simduid_builtins (hash_table<simduid_to_vf> *htab)
 153 {
 154   basic_block bb;
 155
 156   FOR_EACH_BB_FN (bb, cfun)
 157     {
 158       gimple_stmt_iterator i;
 159
 160       for (i = gsi_start_bb (bb); !gsi_end_p (i); )
 161         {
 162           unsigned int vf = 1;
 163           enum internal_fn ifn;
 164           gimple *stmt = gsi_stmt (i);
 165           tree t;
 166           if (!is_gimple_call (stmt)
 167               || !gimple_call_internal_p (stmt))
 168             {
 169               gsi_next (&i);
 170               continue;
 171             }
 172           ifn = gimple_call_internal_fn (stmt);
 173           switch (ifn)
 174             {
 175             case IFN_GOMP_SIMD_LANE:
 176             case IFN_GOMP_SIMD_VF:
 177             case IFN_GOMP_SIMD_LAST_LANE:
 178               break;
 179             case IFN_GOMP_SIMD_ORDERED_START:
 180             case IFN_GOMP_SIMD_ORDERED_END:
 181               if (integer_onep (gimple_call_arg (stmt, 0)))
 182                 {
 183                   enum built_in_function bcode
 184                     = (ifn == IFN_GOMP_SIMD_ORDERED_START
 185                        ? BUILT_IN_GOMP_ORDERED_START
 186                        : BUILT_IN_GOMP_ORDERED_END);
 187                   gimple *g
 188                     = gimple_build_call (builtin_decl_explicit (bcode), 0);
 189                   tree vdef = gimple_vdef (stmt);
 190                   gimple_set_vdef (g, vdef);
 191                   SSA_NAME_DEF_STMT (vdef) = g;
 192                   gimple_set_vuse (g, gimple_vuse (stmt));
 193                   gsi_replace (&i, g, true);
 194                   continue;
 195                 }
 196               gsi_remove (&i, true);
 197               unlink_stmt_vdef (stmt);
 198               continue;
 199             default:
 200               gsi_next (&i);
 201               continue;
 202             }
 203           tree arg = gimple_call_arg (stmt, 0);
 204           gcc_assert (arg != NULL_TREE);
 205           gcc_assert (TREE_CODE (arg) == SSA_NAME);
 206           simduid_to_vf *p = NULL, data;
 207           data.simduid = DECL_UID (SSA_NAME_VAR (arg));
 208           /* Need to nullify loop safelen field since it's value is not
 209              valid after transformation.  */
 210           if (bb->loop_father && bb->loop_father->safelen > 0)
 211             bb->loop_father->safelen = 0;
 212           if (htab)
 213             {
 214               p = htab->find (&data);
 215               if (p)
 216                 vf = p->vf;
 217             }
 218           switch (ifn)
 219             {
 220             case IFN_GOMP_SIMD_VF:
 221               t = build_int_cst (unsigned_type_node, vf);
 222               break;
 223             case IFN_GOMP_SIMD_LANE:
 224               t = build_int_cst (unsigned_type_node, 0);
 225               break;
 226             case IFN_GOMP_SIMD_LAST_LANE:
 227               t = gimple_call_arg (stmt, 1);
 228               break;
 229             default:
 230               gcc_unreachable ();
 231             }
 232           update_call_from_tree (&i, t);
 233           gsi_next (&i);
 234         }
 235     }
 236 }
 237
 238 /* Helper structure for note_simd_array_uses.  */
 239
 240 struct note_simd_array_uses_struct
 241 {
 242   hash_table<simd_array_to_simduid> **htab;
 243   unsigned int simduid;
 244 };
 245
 246 /* Callback for note_simd_array_uses, called through walk_gimple_op.  */
 247
 248 static tree
 249 note_simd_array_uses_cb (tree *tp, int *walk_subtrees, void *data)
 250 {
 251   struct walk_stmt_info *wi = (struct walk_stmt_info *) data;
 252   struct note_simd_array_uses_struct *ns
 253     = (struct note_simd_array_uses_struct *) wi->info;
 254
 255   if (TYPE_P (*tp))
 256     *walk_subtrees = 0;
 257   else if (VAR_P (*tp)
 258            && lookup_attribute ("omp simd array", DECL_ATTRIBUTES (*tp))
 259            && DECL_CONTEXT (*tp) == current_function_decl)
 260     {
 261       simd_array_to_simduid data;
 262       if (!*ns->htab)
 263         *ns->htab = new hash_table<simd_array_to_simduid> (15);
 264       data.decl = *tp;
 265       data.simduid = ns->simduid;
 266       simd_array_to_simduid **slot = (*ns->htab)->find_slot (&data, INSERT);
 267       if (*slot == NULL)
 268         {
 269           simd_array_to_simduid *p = XNEW (simd_array_to_simduid);
 270           *p = data;
 271           *slot = p;
 272         }
 273       else if ((*slot)->simduid != ns->simduid)
 274         (*slot)->simduid = -1U;
 275       *walk_subtrees = 0;
 276     }
 277   return NULL_TREE;
 278 }
 279
 280 /* Find "omp simd array" temporaries and map them to corresponding
 281    simduid.  */
 282
 283 static void
 284 note_simd_array_uses (hash_table<simd_array_to_simduid> **htab)
 285 {
 286   basic_block bb;
 287   gimple_stmt_iterator gsi;
 288   struct walk_stmt_info wi;
 289   struct note_simd_array_uses_struct ns;
 290
 291   memset (&wi, 0, sizeof (wi));
 292   wi.info = &ns;
 293   ns.htab = htab;
 294
 295   FOR_EACH_BB_FN (bb, cfun)
 296     for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
 297       {
 298         gimple *stmt = gsi_stmt (gsi);
 299         if (!is_gimple_call (stmt) || !gimple_call_internal_p (stmt))
 300           continue;
 301         switch (gimple_call_internal_fn (stmt))
 302           {
 303           case IFN_GOMP_SIMD_LANE:
 304           case IFN_GOMP_SIMD_VF:
 305           case IFN_GOMP_SIMD_LAST_LANE:
 306             break;
 307           default:
 308             continue;
 309           }
 310         tree lhs = gimple_call_lhs (stmt);
 311         if (lhs == NULL_TREE)
 312           continue;
 313         imm_use_iterator use_iter;
 314         gimple *use_stmt;
 315         ns.simduid = DECL_UID (SSA_NAME_VAR (gimple_call_arg (stmt, 0)));
 316         FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, lhs)
 317           if (!is_gimple_debug (use_stmt))
 318             walk_gimple_op (use_stmt, note_simd_array_uses_cb, &wi);
 319       }
 320 }
 321
 322 /* Shrink arrays with "omp simd array" attribute to the corresponding
 323    vectorization factor.  */
 324
 325 static void
 326 shrink_simd_arrays
 327   (hash_table<simd_array_to_simduid> *simd_array_to_simduid_htab,
 328    hash_table<simduid_to_vf> *simduid_to_vf_htab)
 329 {
 330   for (hash_table<simd_array_to_simduid>::iterator iter
 331          = simd_array_to_simduid_htab->begin ();
 332        iter != simd_array_to_simduid_htab->end (); ++iter)
 333     if ((*iter)->simduid != -1U)
 334       {
 335         tree decl = (*iter)->decl;
 336         int vf = 1;
 337         if (simduid_to_vf_htab)
 338           {
 339             simduid_to_vf *p = NULL, data;
 340             data.simduid = (*iter)->simduid;
 341             p = simduid_to_vf_htab->find (&data);
 342             if (p)
 343               vf = p->vf;
 344           }
 345         tree atype
 346           = build_array_type_nelts (TREE_TYPE (TREE_TYPE (decl)), vf);
 347         TREE_TYPE (decl) = atype;
 348         relayout_decl (decl);
 349       }
 350
 351   delete simd_array_to_simduid_htab;
 352 }
 353 \f
 354 /* A helper function to free data refs.  */
 355
 356 void
 357 vect_destroy_datarefs (vec_info *vinfo)
 358 {
 359   struct data_reference *dr;
 360   unsigned int i;
 361
 362   FOR_EACH_VEC_ELT (vinfo->datarefs, i, dr)
 363     if (dr->aux)
 364       {
 365         free (dr->aux);
 366         dr->aux = NULL;
 367       }
 368
 369   free_data_refs (vinfo->datarefs);
 370 }
 371
 372 /* A helper function to free scev and LOOP niter information, as well as
 373    clear loop constraint LOOP_C_FINITE.  */
 374
 375 void
 376 vect_free_loop_info_assumptions (struct loop *loop)
 377 {
 378   scev_reset_htab ();
 379   /* We need to explicitly reset upper bound information since they are
 380      used even after free_numbers_of_iterations_estimates_loop.  */
 381   loop->any_upper_bound = false;
 382   loop->any_likely_upper_bound = false;
 383   free_numbers_of_iterations_estimates_loop (loop);
 384   loop_constraint_clear (loop, LOOP_C_FINITE);
 385 }
 386
 387 /* Return whether STMT is inside the region we try to vectorize.  */
 388
 389 bool
 390 vect_stmt_in_region_p (vec_info *vinfo, gimple *stmt)
 391 {
 392   if (!gimple_bb (stmt))
 393     return false;
 394
 395   if (loop_vec_info loop_vinfo = dyn_cast <loop_vec_info> (vinfo))
 396     {
 397       struct loop *loop = LOOP_VINFO_LOOP (loop_vinfo);
 398       if (!flow_bb_inside_loop_p (loop, gimple_bb (stmt)))
 399         return false;
 400     }
 401   else
 402     {
 403       bb_vec_info bb_vinfo = as_a <bb_vec_info> (vinfo);
 404       if (gimple_bb (stmt) != BB_VINFO_BB (bb_vinfo)
 405           || gimple_uid (stmt) == -1U
 406           || gimple_code (stmt) == GIMPLE_PHI)
 407         return false;
 408     }
 409
 410   return true;
 411 }
 412
 413
 414 /* If LOOP has been versioned during ifcvt, return the internal call
 415    guarding it.  */
 416
 417 static gimple *
 418 vect_loop_vectorized_call (struct loop *loop)
 419 {
 420   basic_block bb = loop_preheader_edge (loop)->src;
 421   gimple *g;
 422   do
 423     {
 424       g = last_stmt (bb);
 425       if (g)
 426         break;
 427       if (!single_pred_p (bb))
 428         break;
 429       bb = single_pred (bb);
 430     }
 431   while (1);
 432   if (g && gimple_code (g) == GIMPLE_COND)
 433     {
 434       gimple_stmt_iterator gsi = gsi_for_stmt (g);
 435       gsi_prev (&gsi);
 436       if (!gsi_end_p (gsi))
 437         {
 438           g = gsi_stmt (gsi);
 439           if (gimple_call_internal_p (g, IFN_LOOP_VECTORIZED)
 440               && (tree_to_shwi (gimple_call_arg (g, 0)) == loop->num
 441                   || tree_to_shwi (gimple_call_arg (g, 1)) == loop->num))
 442             return g;
 443         }
 444     }
 445   return NULL;
 446 }
 447
 448 /* Fold LOOP_VECTORIZED internal call G to VALUE and
 449    update any immediate uses of it's LHS.  */
 450
 451 static void
 452 fold_loop_vectorized_call (gimple *g, tree value)
 453 {
 454   tree lhs = gimple_call_lhs (g);
 455   use_operand_p use_p;
 456   imm_use_iterator iter;
 457   gimple *use_stmt;
 458   gimple_stmt_iterator gsi = gsi_for_stmt (g);
 459
 460   update_call_from_tree (&gsi, value);
 461   FOR_EACH_IMM_USE_STMT (use_stmt, iter, lhs)
 462     {
 463       FOR_EACH_IMM_USE_ON_STMT (use_p, iter)
 464         SET_USE (use_p, value);
 465       update_stmt (use_stmt);
 466     }
 467 }
 468 /* Set the uids of all the statements in basic blocks inside loop
 469    represented by LOOP_VINFO. LOOP_VECTORIZED_CALL is the internal
 470    call guarding the loop which has been if converted.  */
 471 static void
 472 set_uid_loop_bbs (loop_vec_info loop_vinfo, gimple *loop_vectorized_call)
 473 {
 474   tree arg = gimple_call_arg (loop_vectorized_call, 1);
 475   basic_block *bbs;
 476   unsigned int i;
 477   struct loop *scalar_loop = get_loop (cfun, tree_to_shwi (arg));
 478
 479   LOOP_VINFO_SCALAR_LOOP (loop_vinfo) = scalar_loop;
 480   gcc_checking_assert (vect_loop_vectorized_call
 481                        (LOOP_VINFO_SCALAR_LOOP (loop_vinfo))
 482                        == loop_vectorized_call);
 483   bbs = get_loop_body (scalar_loop);
 484   for (i = 0; i < scalar_loop->num_nodes; i++)
 485     {
 486       basic_block bb = bbs[i];
 487       gimple_stmt_iterator gsi;
 488       for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
 489         {
 490           gimple *phi = gsi_stmt (gsi);
 491           gimple_set_uid (phi, 0);
 492         }
 493       for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
 494         {
 495           gimple *stmt = gsi_stmt (gsi);
 496           gimple_set_uid (stmt, 0);
 497         }
 498     }
 499   free (bbs);
 500 }
 501
 502 /* Function vectorize_loops.
 503
 504    Entry point to loop vectorization phase.  */
 505
 506 unsigned
 507 vectorize_loops (void)
 508 {
 509   unsigned int i;
 510   unsigned int num_vectorized_loops = 0;
 511   unsigned int vect_loops_num;
 512   struct loop *loop;
 513   hash_table<simduid_to_vf> *simduid_to_vf_htab = NULL;
 514   hash_table<simd_array_to_simduid> *simd_array_to_simduid_htab = NULL;
 515   bool any_ifcvt_loops = false;
 516   unsigned ret = 0;
 517   struct loop *new_loop;
 518
 519   vect_loops_num = number_of_loops (cfun);
 520
 521   /* Bail out if there are no loops.  */
 522   if (vect_loops_num <= 1)
 523     return 0;
 524
 525   if (cfun->has_simduid_loops)
 526     note_simd_array_uses (&simd_array_to_simduid_htab);
 527
 528   init_stmt_vec_info_vec ();
 529
 530   /*  ----------- Analyze loops. -----------  */
 531
 532   /* If some loop was duplicated, it gets bigger number
 533      than all previously defined loops.  This fact allows us to run
 534      only over initial loops skipping newly generated ones.  */
 535   FOR_EACH_LOOP (loop, 0)
 536     if (loop->dont_vectorize)
 537       any_ifcvt_loops = true;
 538     else if ((flag_tree_loop_vectorize
 539               && optimize_loop_nest_for_speed_p (loop))
 540              || loop->force_vectorize)
 541       {
 542         loop_vec_info loop_vinfo, orig_loop_vinfo = NULL;
 543         gimple *loop_vectorized_call = vect_loop_vectorized_call (loop);
 544 vectorize_epilogue:
 545         vect_location = find_loop_location (loop);
 546         if (LOCATION_LOCUS (vect_location) != UNKNOWN_LOCATION
 547             && dump_enabled_p ())
 548           dump_printf (MSG_NOTE, "\nAnalyzing loop at %s:%d\n",
 549                        LOCATION_FILE (vect_location),
 550                        LOCATION_LINE (vect_location));
 551
 552         loop_vinfo = vect_analyze_loop (loop, orig_loop_vinfo);
 553         loop->aux = loop_vinfo;
 554
 555         if (!loop_vinfo || !LOOP_VINFO_VECTORIZABLE_P (loop_vinfo))
 556           {
 557             /* Free existing information if loop is analyzed with some
 558                assumptions.  */
 559             if (loop_constraint_set_p (loop, LOOP_C_FINITE))
 560               vect_free_loop_info_assumptions (loop);
 561
 562             /* If we applied if-conversion then try to vectorize the
 563                BB of innermost loops.
 564                ???  Ideally BB vectorization would learn to vectorize
 565                control flow by applying if-conversion on-the-fly, the
 566                following retains the if-converted loop body even when
 567                only non-if-converted parts took part in BB vectorization.  */
 568             if (flag_tree_slp_vectorize != 0
 569                 && loop_vectorized_call
 570                 && ! loop->inner)
 571               {
 572                 basic_block bb = loop->header;
 573                 for (gimple_stmt_iterator gsi = gsi_start_bb (bb);
 574                      !gsi_end_p (gsi); gsi_next (&gsi))
 575                   {
 576                     gimple *stmt = gsi_stmt (gsi);
 577                     gimple_set_uid (stmt, -1);
 578                     gimple_set_visited (stmt, false);
 579                   }
 580                 if (vect_slp_bb (bb))
 581                   {
 582                     dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location,
 583                                      "basic block vectorized\n");
 584                     fold_loop_vectorized_call (loop_vectorized_call,
 585                                                boolean_true_node);
 586                     ret |= TODO_cleanup_cfg;
 587                   }
 588               }
 589             continue;
 590           }
 591
 592         if (!dbg_cnt (vect_loop))
 593           {
 594             /* We may miss some if-converted loops due to
 595                debug counter.  Set any_ifcvt_loops to visit
 596                them at finalization.  */
 597             any_ifcvt_loops = true;
 598             /* Free existing information if loop is analyzed with some
 599                assumptions.  */
 600             if (loop_constraint_set_p (loop, LOOP_C_FINITE))
 601               vect_free_loop_info_assumptions (loop);
 602
 603             break;
 604           }
 605
 606         if (loop_vectorized_call)
 607           set_uid_loop_bbs (loop_vinfo, loop_vectorized_call);
 608         if (LOCATION_LOCUS (vect_location) != UNKNOWN_LOCATION
 609             && dump_enabled_p ())
 610           dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location,
 611                            "loop vectorized\n");
 612         new_loop = vect_transform_loop (loop_vinfo);
 613         num_vectorized_loops++;
 614         /* Now that the loop has been vectorized, allow it to be unrolled
 615            etc.  */
 616         loop->force_vectorize = false;
 617
 618         if (loop->simduid)
 619           {
 620             simduid_to_vf *simduid_to_vf_data = XNEW (simduid_to_vf);
 621             if (!simduid_to_vf_htab)
 622               simduid_to_vf_htab = new hash_table<simduid_to_vf> (15);
 623             simduid_to_vf_data->simduid = DECL_UID (loop->simduid);
 624             simduid_to_vf_data->vf = loop_vinfo->vectorization_factor;
 625             *simduid_to_vf_htab->find_slot (simduid_to_vf_data, INSERT)
 626               = simduid_to_vf_data;
 627           }
 628
 629         if (loop_vectorized_call)
 630           {
 631             fold_loop_vectorized_call (loop_vectorized_call, boolean_true_node);
 632             ret |= TODO_cleanup_cfg;
 633           }
 634
 635         if (new_loop)
 636           {
 637             /* Epilogue of vectorized loop must be vectorized too.  */
 638             vect_loops_num = number_of_loops (cfun);
 639             loop = new_loop;
 640             orig_loop_vinfo = loop_vinfo;  /* To pass vect_analyze_loop.  */
 641             goto vectorize_epilogue;
 642           }
 643       }
 644
 645   vect_location = UNKNOWN_LOCATION;
 646
 647   statistics_counter_event (cfun, "Vectorized loops", num_vectorized_loops);
 648   if (dump_enabled_p ()
 649       || (num_vectorized_loops > 0 && dump_enabled_p ()))
 650     dump_printf_loc (MSG_NOTE, vect_location,
 651                      "vectorized %u loops in function.\n",
 652                      num_vectorized_loops);
 653
 654   /*  ----------- Finalize. -----------  */
 655
 656   if (any_ifcvt_loops)
 657     for (i = 1; i < vect_loops_num; i++)
 658       {
 659         loop = get_loop (cfun, i);
 660         if (loop && loop->dont_vectorize)
 661           {
 662             gimple *g = vect_loop_vectorized_call (loop);
 663             if (g)
 664               {
 665                 fold_loop_vectorized_call (g, boolean_false_node);
 666                 ret |= TODO_cleanup_cfg;
 667               }
 668           }
 669       }
 670
 671   for (i = 1; i < vect_loops_num; i++)
 672     {
 673       loop_vec_info loop_vinfo;
 674       bool has_mask_store;
 675
 676       loop = get_loop (cfun, i);
 677       if (!loop)
 678         continue;
 679       loop_vinfo = (loop_vec_info) loop->aux;
 680       has_mask_store = false;
 681       if (loop_vinfo)
 682         has_mask_store = LOOP_VINFO_HAS_MASK_STORE (loop_vinfo);
 683       destroy_loop_vec_info (loop_vinfo, true);
 684       if (has_mask_store)
 685         optimize_mask_stores (loop);
 686       loop->aux = NULL;
 687     }
 688
 689   free_stmt_vec_info_vec ();
 690
 691   /* Fold IFN_GOMP_SIMD_{VF,LANE,LAST_LANE,ORDERED_{START,END}} builtins.  */
 692   if (cfun->has_simduid_loops)
 693     adjust_simduid_builtins (simduid_to_vf_htab);
 694
 695   /* Shrink any "omp array simd" temporary arrays to the
 696      actual vectorization factors.  */
 697   if (simd_array_to_simduid_htab)
 698     shrink_simd_arrays (simd_array_to_simduid_htab, simduid_to_vf_htab);
 699   delete simduid_to_vf_htab;
 700   cfun->has_simduid_loops = false;
 701
 702   if (num_vectorized_loops > 0)
 703     {
 704       /* If we vectorized any loop only virtual SSA form needs to be updated.
 705          ???  Also while we try hard to update loop-closed SSA form we fail
 706          to properly do this in some corner-cases (see PR56286).  */
 707       rewrite_into_loop_closed_ssa (NULL, TODO_update_ssa_only_virtuals);
 708       return TODO_cleanup_cfg;
 709     }
 710
 711   return ret;
 712 }
 713
 714
 715 /* Entry point to the simduid cleanup pass.  */
 716
 717 namespace {
 718
 719 const pass_data pass_data_simduid_cleanup =
 720 {
 721   GIMPLE_PASS, /* type */
 722   "simduid", /* name */
 723   OPTGROUP_NONE, /* optinfo_flags */
 724   TV_NONE, /* tv_id */
 725   ( PROP_ssa | PROP_cfg ), /* properties_required */
 726   0, /* properties_provided */
 727   0, /* properties_destroyed */
 728   0, /* todo_flags_start */
 729   0, /* todo_flags_finish */
 730 };
 731
 732 class pass_simduid_cleanup : public gimple_opt_pass
 733 {
 734 public:
 735   pass_simduid_cleanup (gcc::context *ctxt)
 736     : gimple_opt_pass (pass_data_simduid_cleanup, ctxt)
 737   {}
 738
 739   /* opt_pass methods: */
 740   opt_pass * clone () { return new pass_simduid_cleanup (m_ctxt); }
 741   virtual bool gate (function *fun) { return fun->has_simduid_loops; }
 742   virtual unsigned int execute (function *);
 743
 744 }; // class pass_simduid_cleanup
 745
 746 unsigned int
 747 pass_simduid_cleanup::execute (function *fun)
 748 {
 749   hash_table<simd_array_to_simduid> *simd_array_to_simduid_htab = NULL;
 750
 751   note_simd_array_uses (&simd_array_to_simduid_htab);
 752
 753   /* Fold IFN_GOMP_SIMD_{VF,LANE,LAST_LANE,ORDERED_{START,END}} builtins.  */
 754   adjust_simduid_builtins (NULL);
 755
 756   /* Shrink any "omp array simd" temporary arrays to the
 757      actual vectorization factors.  */
 758   if (simd_array_to_simduid_htab)
 759     shrink_simd_arrays (simd_array_to_simduid_htab, NULL);
 760   fun->has_simduid_loops = false;
 761   return 0;
 762 }
 763
 764 }  // anon namespace
 765
 766 gimple_opt_pass *
 767 make_pass_simduid_cleanup (gcc::context *ctxt)
 768 {
 769   return new pass_simduid_cleanup (ctxt);
 770 }
 771
 772
 773 /*  Entry point to basic block SLP phase.  */
 774
 775 namespace {
 776
 777 const pass_data pass_data_slp_vectorize =
 778 {
 779   GIMPLE_PASS, /* type */
 780   "slp", /* name */
 781   OPTGROUP_LOOP | OPTGROUP_VEC, /* optinfo_flags */
 782   TV_TREE_SLP_VECTORIZATION, /* tv_id */
 783   ( PROP_ssa | PROP_cfg ), /* properties_required */
 784   0, /* properties_provided */
 785   0, /* properties_destroyed */
 786   0, /* todo_flags_start */
 787   TODO_update_ssa, /* todo_flags_finish */
 788 };
 789
 790 class pass_slp_vectorize : public gimple_opt_pass
 791 {
 792 public:
 793   pass_slp_vectorize (gcc::context *ctxt)
 794     : gimple_opt_pass (pass_data_slp_vectorize, ctxt)
 795   {}
 796
 797   /* opt_pass methods: */
 798   opt_pass * clone () { return new pass_slp_vectorize (m_ctxt); }
 799   virtual bool gate (function *) { return flag_tree_slp_vectorize != 0; }
 800   virtual unsigned int execute (function *);
 801
 802 }; // class pass_slp_vectorize
 803
 804 unsigned int
 805 pass_slp_vectorize::execute (function *fun)
 806 {
 807   basic_block bb;
 808
 809   bool in_loop_pipeline = scev_initialized_p ();
 810   if (!in_loop_pipeline)
 811     {
 812       loop_optimizer_init (LOOPS_NORMAL);
 813       scev_initialize ();
 814     }
 815
 816   /* Mark all stmts as not belonging to the current region and unvisited.  */
 817   FOR_EACH_BB_FN (bb, fun)
 818     {
 819       for (gimple_stmt_iterator gsi = gsi_start_bb (bb); !gsi_end_p (gsi);
 820            gsi_next (&gsi))
 821         {
 822           gimple *stmt = gsi_stmt (gsi);
 823           gimple_set_uid (stmt, -1);
 824           gimple_set_visited (stmt, false);
 825         }
 826     }
 827
 828   init_stmt_vec_info_vec ();
 829
 830   FOR_EACH_BB_FN (bb, fun)
 831     {
 832       if (vect_slp_bb (bb))
 833         dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location,
 834                          "basic block vectorized\n");
 835     }
 836
 837   free_stmt_vec_info_vec ();
 838
 839   if (!in_loop_pipeline)
 840     {
 841       scev_finalize ();
 842       loop_optimizer_finalize ();
 843     }
 844
 845   return 0;
 846 }
 847
 848 } // anon namespace
 849
 850 gimple_opt_pass *
 851 make_pass_slp_vectorize (gcc::context *ctxt)
 852 {
 853   return new pass_slp_vectorize (ctxt);
 854 }
 855
 856
 857 /* Increase alignment of global arrays to improve vectorization potential.
 858    TODO:
 859    - Consider also structs that have an array field.
 860    - Use ipa analysis to prune arrays that can't be vectorized?
 861      This should involve global alignment analysis and in the future also
 862      array padding.  */
 863
 864 static unsigned get_vec_alignment_for_type (tree);
 865 static hash_map<tree, unsigned> *type_align_map;
 866
 867 /* Return alignment of array's vector type corresponding to scalar type.
 868    0 if no vector type exists.  */
 869 static unsigned
 870 get_vec_alignment_for_array_type (tree type)
 871 {
 872   gcc_assert (TREE_CODE (type) == ARRAY_TYPE);
 873
 874   tree vectype = get_vectype_for_scalar_type (strip_array_types (type));
 875   if (!vectype
 876       || !TYPE_SIZE (type)
 877       || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST
 878       || tree_int_cst_lt (TYPE_SIZE (type), TYPE_SIZE (vectype)))
 879     return 0;
 880
 881   return TYPE_ALIGN (vectype);
 882 }
 883
 884 /* Return alignment of field having maximum alignment of vector type
 885    corresponding to it's scalar type. For now, we only consider fields whose
 886    offset is a multiple of it's vector alignment.
 887    0 if no suitable field is found.  */
 888 static unsigned
 889 get_vec_alignment_for_record_type (tree type)
 890 {
 891   gcc_assert (TREE_CODE (type) == RECORD_TYPE);
 892
 893   unsigned max_align = 0, alignment;
 894   HOST_WIDE_INT offset;
 895   tree offset_tree;
 896
 897   if (TYPE_PACKED (type))
 898     return 0;
 899
 900   unsigned *slot = type_align_map->get (type);
 901   if (slot)
 902     return *slot;
 903
 904   for (tree field = first_field (type);
 905        field != NULL_TREE;
 906        field = DECL_CHAIN (field))
 907     {
 908       /* Skip if not FIELD_DECL or if alignment is set by user.  */
 909       if (TREE_CODE (field) != FIELD_DECL
 910           || DECL_USER_ALIGN (field)
 911           || DECL_ARTIFICIAL (field))
 912         continue;
 913
 914       /* We don't need to process the type further if offset is variable,
 915          since the offsets of remaining members will also be variable.  */
 916       if (TREE_CODE (DECL_FIELD_OFFSET (field)) != INTEGER_CST
 917           || TREE_CODE (DECL_FIELD_BIT_OFFSET (field)) != INTEGER_CST)
 918         break;
 919
 920       /* Similarly stop processing the type if offset_tree
 921          does not fit in unsigned HOST_WIDE_INT.  */
 922       offset_tree = bit_position (field);
 923       if (!tree_fits_uhwi_p (offset_tree))
 924         break;
 925
 926       offset = tree_to_uhwi (offset_tree);
 927       alignment = get_vec_alignment_for_type (TREE_TYPE (field));
 928
 929       /* Get maximum alignment of vectorized field/array among those members
 930          whose offset is multiple of the vector alignment.  */
 931       if (alignment
 932           && (offset % alignment == 0)
 933           && (alignment > max_align))
 934         max_align = alignment;
 935     }
 936
 937   type_align_map->put (type, max_align);
 938   return max_align;
 939 }
 940
 941 /* Return alignment of vector type corresponding to decl's scalar type
 942    or 0 if it doesn't exist or the vector alignment is lesser than
 943    decl's alignment.  */
 944 static unsigned
 945 get_vec_alignment_for_type (tree type)
 946 {
 947   if (type == NULL_TREE)
 948     return 0;
 949
 950   gcc_assert (TYPE_P (type));
 951
 952   static unsigned alignment = 0;
 953   switch (TREE_CODE (type))
 954     {
 955       case ARRAY_TYPE:
 956         alignment = get_vec_alignment_for_array_type (type);
 957         break;
 958       case RECORD_TYPE:
 959         alignment = get_vec_alignment_for_record_type (type);
 960         break;
 961       default:
 962         alignment = 0;
 963         break;
 964     }
 965
 966   return (alignment > TYPE_ALIGN (type)) ? alignment : 0;
 967 }
 968
 969 /* Entry point to increase_alignment pass.  */
 970 static unsigned int
 971 increase_alignment (void)
 972 {
 973   varpool_node *vnode;
 974
 975   vect_location = UNKNOWN_LOCATION;
 976   type_align_map = new hash_map<tree, unsigned>;
 977
 978   /* Increase the alignment of all global arrays for vectorization.  */
 979   FOR_EACH_DEFINED_VARIABLE (vnode)
 980     {
 981       tree decl = vnode->decl;
 982       unsigned int alignment;
 983
 984       if ((decl_in_symtab_p (decl)
 985           && !symtab_node::get (decl)->can_increase_alignment_p ())
 986           || DECL_USER_ALIGN (decl) || DECL_ARTIFICIAL (decl))
 987         continue;
 988
 989       alignment = get_vec_alignment_for_type (TREE_TYPE (decl));
 990       if (alignment && vect_can_force_dr_alignment_p (decl, alignment))
 991         {
 992           vnode->increase_alignment (alignment);
 993           dump_printf (MSG_NOTE, "Increasing alignment of decl: ");
 994           dump_generic_expr (MSG_NOTE, TDF_SLIM, decl);
 995           dump_printf (MSG_NOTE, "\n");
 996         }
 997     }
 998
 999   delete type_align_map;
1000   return 0;
1001 }
1002
1003
1004 namespace {
1005
1006 const pass_data pass_data_ipa_increase_alignment =
1007 {
1008   SIMPLE_IPA_PASS, /* type */
1009   "increase_alignment", /* name */
1010   OPTGROUP_LOOP | OPTGROUP_VEC, /* optinfo_flags */
1011   TV_IPA_OPT, /* tv_id */
1012   0, /* properties_required */
1013   0, /* properties_provided */
1014   0, /* properties_destroyed */
1015   0, /* todo_flags_start */
1016   0, /* todo_flags_finish */
1017 };
1018
1019 class pass_ipa_increase_alignment : public simple_ipa_opt_pass
1020 {
1021 public:
1022   pass_ipa_increase_alignment (gcc::context *ctxt)
1023     : simple_ipa_opt_pass (pass_data_ipa_increase_alignment, ctxt)
1024   {}
1025
1026   /* opt_pass methods: */
1027   virtual bool gate (function *)
1028     {
1029       return flag_section_anchors && flag_tree_loop_vectorize;
1030     }
1031
1032   virtual unsigned int execute (function *) { return increase_alignment (); }
1033
1034 }; // class pass_ipa_increase_alignment
1035
1036 } // anon namespace
1037
1038 simple_ipa_opt_pass *
1039 make_pass_ipa_increase_alignment (gcc::context *ctxt)
1040 {
1041   return new pass_ipa_increase_alignment (ctxt);
1042 }