gcc/tree-vectorizer.c

   1 /* Vectorizer
   2    Copyright (C) 2003, 2004, 2005, 2006, 2007, 2008, 2009, 2010
   3    Free Software Foundation, Inc.
   4    Contributed by Dorit Naishlos <dorit@il.ibm.com>
   5
   6 This file is part of GCC.
   7
   8 GCC is free software; you can redistribute it and/or modify it under
   9 the terms of the GNU General Public License as published by the Free
  10 Software Foundation; either version 3, or (at your option) any later
  11 version.
  12
  13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
  14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
  15 FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  16 for more details.
  17
  18 You should have received a copy of the GNU General Public License
  19 along with GCC; see the file COPYING3.  If not see
  20 <http://www.gnu.org/licenses/>.  */
  21
  22 /* Loop and basic block vectorizer.
  23
  24   This file contains drivers for the three vectorizers:
  25   (1) loop vectorizer (inter-iteration parallelism),
  26   (2) loop-aware SLP (intra-iteration parallelism) (invoked by the loop
  27       vectorizer)
  28   (3) BB vectorizer (out-of-loops), aka SLP
  29
  30   The rest of the vectorizer's code is organized as follows:
  31   - tree-vect-loop.c - loop specific parts such as reductions, etc. These are
  32     used by drivers (1) and (2).
  33   - tree-vect-loop-manip.c - vectorizer's loop control-flow utilities, used by
  34     drivers (1) and (2).
  35   - tree-vect-slp.c - BB vectorization specific analysis and transformation,
  36     used by drivers (2) and (3).
  37   - tree-vect-stmts.c - statements analysis and transformation (used by all).
  38   - tree-vect-data-refs.c - vectorizer specific data-refs analysis and
  39     manipulations (used by all).
  40   - tree-vect-patterns.c - vectorizable code patterns detector (used by all)
  41
  42   Here's a poor attempt at illustrating that:
  43
  44      tree-vectorizer.c:
  45      loop_vect()  loop_aware_slp()  slp_vect()
  46           |        /           \          /
  47           |       /             \        /
  48           tree-vect-loop.c  tree-vect-slp.c
  49                 | \      \  /      /   |
  50                 |  \      \/      /    |
  51                 |   \     /\     /     |
  52                 |    \   /  \   /      |
  53          tree-vect-stmts.c  tree-vect-data-refs.c
  54                        \      /
  55                     tree-vect-patterns.c
  56 */
  57
  58 #include "config.h"
  59 #include "system.h"
  60 #include "coretypes.h"
  61 #include "dumpfile.h"
  62 #include "tm.h"
  63 #include "ggc.h"
  64 #include "tree.h"
  65 #include "tree-pretty-print.h"
  66 #include "tree-flow.h"
  67 #include "cfgloop.h"
  68 #include "tree-vectorizer.h"
  69 #include "tree-pass.h"
  70
  71 /* Loop or bb location.  */
  72 LOC vect_location;
  73
  74 /* Vector mapping GIMPLE stmt to stmt_vec_info. */
  75 vec<vec_void_p> stmt_vec_info_vec;
  76
  77 \f
  78 /* Function vectorize_loops.
  79
  80    Entry point to loop vectorization phase.  */
  81
  82 unsigned
  83 vectorize_loops (void)
  84 {
  85   unsigned int i;
  86   unsigned int num_vectorized_loops = 0;
  87   unsigned int vect_loops_num;
  88   loop_iterator li;
  89   struct loop *loop;
  90
  91   vect_loops_num = number_of_loops ();
  92
  93   /* Bail out if there are no loops.  */
  94   if (vect_loops_num <= 1)
  95     return 0;
  96
  97   init_stmt_vec_info_vec ();
  98
  99   /*  ----------- Analyze loops. -----------  */
 100
 101   /* If some loop was duplicated, it gets bigger number
 102      than all previously defined loops.  This fact allows us to run
 103      only over initial loops skipping newly generated ones.  */
 104   FOR_EACH_LOOP (li, loop, 0)
 105     if (optimize_loop_nest_for_speed_p (loop))
 106       {
 107         loop_vec_info loop_vinfo;
 108         vect_location = find_loop_location (loop);
 109         if (LOCATION_LOCUS (vect_location) != UNKNOWN_LOC
 110             && dump_enabled_p ())
 111           dump_printf (MSG_ALL, "\nAnalyzing loop at %s:%d\n",
 112                        LOC_FILE (vect_location), LOC_LINE (vect_location));
 113
 114         loop_vinfo = vect_analyze_loop (loop);
 115         loop->aux = loop_vinfo;
 116
 117         if (!loop_vinfo || !LOOP_VINFO_VECTORIZABLE_P (loop_vinfo))
 118           continue;
 119
 120         if (LOCATION_LOCUS (vect_location) != UNKNOWN_LOC
 121             && dump_enabled_p ())
 122           dump_printf (MSG_ALL, "\n\nVectorizing loop at %s:%d\n",
 123                        LOC_FILE (vect_location), LOC_LINE (vect_location));
 124         vect_transform_loop (loop_vinfo);
 125         num_vectorized_loops++;
 126       }
 127
 128   vect_location = UNKNOWN_LOC;
 129
 130   statistics_counter_event (cfun, "Vectorized loops", num_vectorized_loops);
 131   if (dump_enabled_p ()
 132       || (num_vectorized_loops > 0 && dump_enabled_p ()))
 133     dump_printf_loc (MSG_ALL, vect_location,
 134                      "vectorized %u loops in function.\n",
 135                      num_vectorized_loops);
 136
 137   /*  ----------- Finalize. -----------  */
 138
 139   for (i = 1; i < vect_loops_num; i++)
 140     {
 141       loop_vec_info loop_vinfo;
 142
 143       loop = get_loop (i);
 144       if (!loop)
 145         continue;
 146       loop_vinfo = (loop_vec_info) loop->aux;
 147       destroy_loop_vec_info (loop_vinfo, true);
 148       loop->aux = NULL;
 149     }
 150
 151   free_stmt_vec_info_vec ();
 152
 153   return num_vectorized_loops > 0 ? TODO_cleanup_cfg : 0;
 154 }
 155
 156
 157 /*  Entry point to basic block SLP phase.  */
 158
 159 static unsigned int
 160 execute_vect_slp (void)
 161 {
 162   basic_block bb;
 163
 164   init_stmt_vec_info_vec ();
 165
 166   FOR_EACH_BB (bb)
 167     {
 168       vect_location = find_bb_location (bb);
 169
 170       if (vect_slp_analyze_bb (bb))
 171         {
 172           vect_slp_transform_bb (bb);
 173           if (dump_enabled_p ())
 174             dump_printf_loc (MSG_OPTIMIZED_LOCATIONS, vect_location,
 175                              "basic block vectorized using SLP\n");
 176         }
 177     }
 178
 179   free_stmt_vec_info_vec ();
 180   return 0;
 181 }
 182
 183 static bool
 184 gate_vect_slp (void)
 185 {
 186   /* Apply SLP either if the vectorizer is on and the user didn't specify
 187      whether to run SLP or not, or if the SLP flag was set by the user.  */
 188   return ((flag_tree_vectorize != 0 && flag_tree_slp_vectorize != 0)
 189           || flag_tree_slp_vectorize == 1);
 190 }
 191
 192 struct gimple_opt_pass pass_slp_vectorize =
 193 {
 194  {
 195   GIMPLE_PASS,
 196   "slp",                                /* name */
 197   OPTGROUP_LOOP
 198   | OPTGROUP_VEC,                       /* optinfo_flags */
 199   gate_vect_slp,                        /* gate */
 200   execute_vect_slp,                     /* execute */
 201   NULL,                                 /* sub */
 202   NULL,                                 /* next */
 203   0,                                    /* static_pass_number */
 204   TV_TREE_SLP_VECTORIZATION,            /* tv_id */
 205   PROP_ssa | PROP_cfg,                  /* properties_required */
 206   0,                                    /* properties_provided */
 207   0,                                    /* properties_destroyed */
 208   0,                                    /* todo_flags_start */
 209   TODO_ggc_collect
 210     | TODO_verify_ssa
 211     | TODO_update_ssa
 212     | TODO_verify_stmts                 /* todo_flags_finish */
 213  }
 214 };
 215
 216
 217 /* Increase alignment of global arrays to improve vectorization potential.
 218    TODO:
 219    - Consider also structs that have an array field.
 220    - Use ipa analysis to prune arrays that can't be vectorized?
 221      This should involve global alignment analysis and in the future also
 222      array padding.  */
 223
 224 static unsigned int
 225 increase_alignment (void)
 226 {
 227   struct varpool_node *vnode;
 228
 229   /* Increase the alignment of all global arrays for vectorization.  */
 230   FOR_EACH_DEFINED_VARIABLE (vnode)
 231     {
 232       tree vectype, decl = vnode->symbol.decl;
 233       tree t;
 234       unsigned int alignment;
 235
 236       t = TREE_TYPE(decl);
 237       if (TREE_CODE (t) != ARRAY_TYPE)
 238         continue;
 239       vectype = get_vectype_for_scalar_type (strip_array_types (t));
 240       if (!vectype)
 241         continue;
 242       alignment = TYPE_ALIGN (vectype);
 243       if (DECL_ALIGN (decl) >= alignment)
 244         continue;
 245
 246       if (vect_can_force_dr_alignment_p (decl, alignment))
 247         {
 248           DECL_ALIGN (decl) = TYPE_ALIGN (vectype);
 249           DECL_USER_ALIGN (decl) = 1;
 250           dump_printf (MSG_NOTE, "Increasing alignment of decl: ");
 251           dump_generic_expr (MSG_NOTE, TDF_SLIM, decl);
 252           dump_printf (MSG_NOTE, "\n");
 253         }
 254     }
 255   return 0;
 256 }
 257
 258
 259 static bool
 260 gate_increase_alignment (void)
 261 {
 262   return flag_section_anchors && flag_tree_vectorize;
 263 }
 264
 265
 266 struct simple_ipa_opt_pass pass_ipa_increase_alignment =
 267 {
 268  {
 269   SIMPLE_IPA_PASS,
 270   "increase_alignment",                 /* name */
 271   OPTGROUP_LOOP
 272   | OPTGROUP_VEC,                       /* optinfo_flags */
 273   gate_increase_alignment,              /* gate */
 274   increase_alignment,                   /* execute */
 275   NULL,                                 /* sub */
 276   NULL,                                 /* next */
 277   0,                                    /* static_pass_number */
 278   TV_IPA_OPT,                           /* tv_id */
 279   0,                                    /* properties_required */
 280   0,                                    /* properties_provided */
 281   0,                                    /* properties_destroyed */
 282   0,                                    /* todo_flags_start */
 283   0                                     /* todo_flags_finish */
 284  }
 285 };