gcc/tracer.c

   1 /* The tracer pass for the GNU compiler.
   2    Contributed by Jan Hubicka, SuSE Labs.
   3    Adapted to work on GIMPLE instead of RTL by Robert Kidd, UIUC.
   4    Copyright (C) 2001-2015 Free Software Foundation, Inc.
   5
   6    This file is part of GCC.
   7
   8    GCC is free software; you can redistribute it and/or modify it
   9    under the terms of the GNU General Public License as published by
  10    the Free Software Foundation; either version 3, or (at your option)
  11    any later version.
  12
  13    GCC is distributed in the hope that it will be useful, but WITHOUT
  14    ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
  15    or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public
  16    License 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 /* This pass performs the tail duplication needed for superblock formation.
  23    For more information see:
  24
  25      Design and Analysis of Profile-Based Optimization in Compaq's
  26      Compilation Tools for Alpha; Journal of Instruction-Level
  27      Parallelism 3 (2000) 1-25
  28
  29    Unlike Compaq's implementation we don't do the loop peeling as most
  30    probably a better job can be done by a special pass and we don't
  31    need to worry too much about the code size implications as the tail
  32    duplicates are crossjumped again if optimizations are not
  33    performed.  */
  34
  35
  36 #include "config.h"
  37 #include "system.h"
  38 #include "coretypes.h"
  39 #include "backend.h"
  40 #include "rtl.h"
  41 #include "tree.h"
  42 #include "gimple.h"
  43 #include "cfghooks.h"
  44 #include "tree-pass.h"
  45 #include "coverage.h"
  46 #include "alias.h"
  47 #include "fold-const.h"
  48 #include "profile.h"
  49 #include "cfganal.h"
  50 #include "flags.h"
  51 #include "params.h"
  52 #include "internal-fn.h"
  53 #include "gimple-iterator.h"
  54 #include "tree-cfg.h"
  55 #include "tree-ssa.h"
  56 #include "tree-inline.h"
  57 #include "cfgloop.h"
  58 #include "fibonacci_heap.h"
  59
  60 static int count_insns (basic_block);
  61 static bool ignore_bb_p (const_basic_block);
  62 static bool better_p (const_edge, const_edge);
  63 static edge find_best_successor (basic_block);
  64 static edge find_best_predecessor (basic_block);
  65 static int find_trace (basic_block, basic_block *);
  66
  67 /* Minimal outgoing edge probability considered for superblock formation.  */
  68 static int probability_cutoff;
  69 static int branch_ratio_cutoff;
  70
  71 /* A bit BB->index is set if BB has already been seen, i.e. it is
  72    connected to some trace already.  */
  73 sbitmap bb_seen;
  74
  75 static inline void
  76 mark_bb_seen (basic_block bb)
  77 {
  78   unsigned int size = SBITMAP_SIZE (bb_seen);
  79
  80   if ((unsigned int)bb->index >= size)
  81     bb_seen = sbitmap_resize (bb_seen, size * 2, 0);
  82
  83   bitmap_set_bit (bb_seen, bb->index);
  84 }
  85
  86 static inline bool
  87 bb_seen_p (basic_block bb)
  88 {
  89   return bitmap_bit_p (bb_seen, bb->index);
  90 }
  91
  92 /* Return true if we should ignore the basic block for purposes of tracing.  */
  93 static bool
  94 ignore_bb_p (const_basic_block bb)
  95 {
  96   if (bb->index < NUM_FIXED_BLOCKS)
  97     return true;
  98   if (optimize_bb_for_size_p (bb))
  99     return true;
 100
 101   if (gimple *g = last_stmt (CONST_CAST_BB (bb)))
 102     {
 103       /* A transaction is a single entry multiple exit region.  It
 104          must be duplicated in its entirety or not at all.  */
 105       if (gimple_code (g) == GIMPLE_TRANSACTION)
 106         return true;
 107
 108       /* An IFN_UNIQUE call must be duplicated as part of its group,
 109          or not at all.  */
 110       if (is_gimple_call (g)
 111           && gimple_call_internal_p (g)
 112           && gimple_call_internal_unique_p (g))
 113         return true;
 114     }
 115
 116   return false;
 117 }
 118
 119 /* Return number of instructions in the block.  */
 120
 121 static int
 122 count_insns (basic_block bb)
 123 {
 124   gimple_stmt_iterator gsi;
 125   gimple *stmt;
 126   int n = 0;
 127
 128   for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
 129     {
 130       stmt = gsi_stmt (gsi);
 131       n += estimate_num_insns (stmt, &eni_size_weights);
 132     }
 133   return n;
 134 }
 135
 136 /* Return true if E1 is more frequent than E2.  */
 137 static bool
 138 better_p (const_edge e1, const_edge e2)
 139 {
 140   if (e1->count != e2->count)
 141     return e1->count > e2->count;
 142   if (e1->src->frequency * e1->probability !=
 143       e2->src->frequency * e2->probability)
 144     return (e1->src->frequency * e1->probability
 145             > e2->src->frequency * e2->probability);
 146   /* This is needed to avoid changes in the decision after
 147      CFG is modified.  */
 148   if (e1->src != e2->src)
 149     return e1->src->index > e2->src->index;
 150   return e1->dest->index > e2->dest->index;
 151 }
 152
 153 /* Return most frequent successor of basic block BB.  */
 154
 155 static edge
 156 find_best_successor (basic_block bb)
 157 {
 158   edge e;
 159   edge best = NULL;
 160   edge_iterator ei;
 161
 162   FOR_EACH_EDGE (e, ei, bb->succs)
 163     if (!best || better_p (e, best))
 164       best = e;
 165   if (!best || ignore_bb_p (best->dest))
 166     return NULL;
 167   if (best->probability <= probability_cutoff)
 168     return NULL;
 169   return best;
 170 }
 171
 172 /* Return most frequent predecessor of basic block BB.  */
 173
 174 static edge
 175 find_best_predecessor (basic_block bb)
 176 {
 177   edge e;
 178   edge best = NULL;
 179   edge_iterator ei;
 180
 181   FOR_EACH_EDGE (e, ei, bb->preds)
 182     if (!best || better_p (e, best))
 183       best = e;
 184   if (!best || ignore_bb_p (best->src))
 185     return NULL;
 186   if (EDGE_FREQUENCY (best) * REG_BR_PROB_BASE
 187       < bb->frequency * branch_ratio_cutoff)
 188     return NULL;
 189   return best;
 190 }
 191
 192 /* Find the trace using bb and record it in the TRACE array.
 193    Return number of basic blocks recorded.  */
 194
 195 static int
 196 find_trace (basic_block bb, basic_block *trace)
 197 {
 198   int i = 0;
 199   edge e;
 200
 201   if (dump_file)
 202     fprintf (dump_file, "Trace seed %i [%i]", bb->index, bb->frequency);
 203
 204   while ((e = find_best_predecessor (bb)) != NULL)
 205     {
 206       basic_block bb2 = e->src;
 207       if (bb_seen_p (bb2) || (e->flags & (EDGE_DFS_BACK | EDGE_COMPLEX))
 208           || find_best_successor (bb2) != e)
 209         break;
 210       if (dump_file)
 211         fprintf (dump_file, ",%i [%i]", bb->index, bb->frequency);
 212       bb = bb2;
 213     }
 214   if (dump_file)
 215     fprintf (dump_file, " forward %i [%i]", bb->index, bb->frequency);
 216   trace[i++] = bb;
 217
 218   /* Follow the trace in forward direction.  */
 219   while ((e = find_best_successor (bb)) != NULL)
 220     {
 221       bb = e->dest;
 222       if (bb_seen_p (bb) || (e->flags & (EDGE_DFS_BACK | EDGE_COMPLEX))
 223           || find_best_predecessor (bb) != e)
 224         break;
 225       if (dump_file)
 226         fprintf (dump_file, ",%i [%i]", bb->index, bb->frequency);
 227       trace[i++] = bb;
 228     }
 229   if (dump_file)
 230     fprintf (dump_file, "\n");
 231   return i;
 232 }
 233
 234 /* Look for basic blocks in frequency order, construct traces and tail duplicate
 235    if profitable.  */
 236
 237 static bool
 238 tail_duplicate (void)
 239 {
 240   auto_vec<fibonacci_node<long, basic_block_def>*> blocks;
 241   blocks.safe_grow_cleared (last_basic_block_for_fn (cfun));
 242
 243   basic_block *trace = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun));
 244   int *counts = XNEWVEC (int, last_basic_block_for_fn (cfun));
 245   int ninsns = 0, nduplicated = 0;
 246   gcov_type weighted_insns = 0, traced_insns = 0;
 247   fibonacci_heap<long, basic_block_def> heap (LONG_MIN);
 248   gcov_type cover_insns;
 249   int max_dup_insns;
 250   basic_block bb;
 251   bool changed = false;
 252
 253   /* Create an oversized sbitmap to reduce the chance that we need to
 254      resize it.  */
 255   bb_seen = sbitmap_alloc (last_basic_block_for_fn (cfun) * 2);
 256   bitmap_clear (bb_seen);
 257   initialize_original_copy_tables ();
 258
 259   if (profile_info && flag_branch_probabilities)
 260     probability_cutoff = PARAM_VALUE (TRACER_MIN_BRANCH_PROBABILITY_FEEDBACK);
 261   else
 262     probability_cutoff = PARAM_VALUE (TRACER_MIN_BRANCH_PROBABILITY);
 263   probability_cutoff = REG_BR_PROB_BASE / 100 * probability_cutoff;
 264
 265   branch_ratio_cutoff =
 266     (REG_BR_PROB_BASE / 100 * PARAM_VALUE (TRACER_MIN_BRANCH_RATIO));
 267
 268   FOR_EACH_BB_FN (bb, cfun)
 269     {
 270       int n = count_insns (bb);
 271       if (!ignore_bb_p (bb))
 272         blocks[bb->index] = heap.insert (-bb->frequency, bb);
 273
 274       counts [bb->index] = n;
 275       ninsns += n;
 276       weighted_insns += n * bb->frequency;
 277     }
 278
 279   if (profile_info && flag_branch_probabilities)
 280     cover_insns = PARAM_VALUE (TRACER_DYNAMIC_COVERAGE_FEEDBACK);
 281   else
 282     cover_insns = PARAM_VALUE (TRACER_DYNAMIC_COVERAGE);
 283   cover_insns = (weighted_insns * cover_insns + 50) / 100;
 284   max_dup_insns = (ninsns * PARAM_VALUE (TRACER_MAX_CODE_GROWTH) + 50) / 100;
 285
 286   while (traced_insns < cover_insns && nduplicated < max_dup_insns
 287          && !heap.empty ())
 288     {
 289       basic_block bb = heap.extract_min ();
 290       int n, pos;
 291
 292       if (!bb)
 293         break;
 294
 295       blocks[bb->index] = NULL;
 296
 297       if (ignore_bb_p (bb))
 298         continue;
 299       gcc_assert (!bb_seen_p (bb));
 300
 301       n = find_trace (bb, trace);
 302
 303       bb = trace[0];
 304       traced_insns += bb->frequency * counts [bb->index];
 305       if (blocks[bb->index])
 306         {
 307           heap.delete_node (blocks[bb->index]);
 308           blocks[bb->index] = NULL;
 309         }
 310
 311       for (pos = 1; pos < n; pos++)
 312         {
 313           basic_block bb2 = trace[pos];
 314
 315           if (blocks[bb2->index])
 316             {
 317               heap.delete_node (blocks[bb2->index]);
 318               blocks[bb2->index] = NULL;
 319             }
 320           traced_insns += bb2->frequency * counts [bb2->index];
 321           if (EDGE_COUNT (bb2->preds) > 1
 322               && can_duplicate_block_p (bb2)
 323               /* We have the tendency to duplicate the loop header
 324                  of all do { } while loops.  Do not do that - it is
 325                  not profitable and it might create a loop with multiple
 326                  entries or at least rotate the loop.  */
 327               && bb2->loop_father->header != bb2)
 328             {
 329               edge e;
 330               basic_block copy;
 331
 332               nduplicated += counts [bb2->index];
 333
 334               e = find_edge (bb, bb2);
 335
 336               copy = duplicate_block (bb2, e, bb);
 337               flush_pending_stmts (e);
 338
 339               add_phi_args_after_copy (&copy, 1, NULL);
 340
 341               /* Reconsider the original copy of block we've duplicated.
 342                  Removing the most common predecessor may make it to be
 343                  head.  */
 344               blocks[bb2->index] = heap.insert (-bb2->frequency, bb2);
 345
 346               if (dump_file)
 347                 fprintf (dump_file, "Duplicated %i as %i [%i]\n",
 348                          bb2->index, copy->index, copy->frequency);
 349
 350               bb2 = copy;
 351               changed = true;
 352             }
 353           mark_bb_seen (bb2);
 354           bb = bb2;
 355           /* In case the trace became infrequent, stop duplicating.  */
 356           if (ignore_bb_p (bb))
 357             break;
 358         }
 359       if (dump_file)
 360         fprintf (dump_file, " covered now %.1f\n\n",
 361                  traced_insns * 100.0 / weighted_insns);
 362     }
 363   if (dump_file)
 364     fprintf (dump_file, "Duplicated %i insns (%i%%)\n", nduplicated,
 365              nduplicated * 100 / ninsns);
 366
 367   free_original_copy_tables ();
 368   sbitmap_free (bb_seen);
 369   free (trace);
 370   free (counts);
 371
 372   return changed;
 373 }
 374 \f
 375 namespace {
 376
 377 const pass_data pass_data_tracer =
 378 {
 379   GIMPLE_PASS, /* type */
 380   "tracer", /* name */
 381   OPTGROUP_NONE, /* optinfo_flags */
 382   TV_TRACER, /* tv_id */
 383   0, /* properties_required */
 384   0, /* properties_provided */
 385   0, /* properties_destroyed */
 386   0, /* todo_flags_start */
 387   TODO_update_ssa, /* todo_flags_finish */
 388 };
 389
 390 class pass_tracer : public gimple_opt_pass
 391 {
 392 public:
 393   pass_tracer (gcc::context *ctxt)
 394     : gimple_opt_pass (pass_data_tracer, ctxt)
 395   {}
 396
 397   /* opt_pass methods: */
 398   virtual bool gate (function *)
 399     {
 400       return (optimize > 0 && flag_tracer && flag_reorder_blocks);
 401     }
 402
 403   virtual unsigned int execute (function *);
 404
 405 }; // class pass_tracer
 406
 407 unsigned int
 408 pass_tracer::execute (function *fun)
 409 {
 410   bool changed;
 411
 412   if (n_basic_blocks_for_fn (fun) <= NUM_FIXED_BLOCKS + 1)
 413     return 0;
 414
 415   mark_dfs_back_edges ();
 416   if (dump_file)
 417     brief_dump_cfg (dump_file, dump_flags);
 418
 419   /* Trace formation is done on the fly inside tail_duplicate */
 420   changed = tail_duplicate ();
 421   if (changed)
 422     {
 423       free_dominance_info (CDI_DOMINATORS);
 424       /* If we changed the CFG schedule loops for fixup by cleanup_cfg.  */
 425       loops_state_set (LOOPS_NEED_FIXUP);
 426     }
 427
 428   if (dump_file)
 429     brief_dump_cfg (dump_file, dump_flags);
 430
 431   return changed ? TODO_cleanup_cfg : 0;
 432 }
 433 } // anon namespace
 434
 435 gimple_opt_pass *
 436 make_pass_tracer (gcc::context *ctxt)
 437 {
 438   return new pass_tracer (ctxt);
 439 }