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) 2001-2014 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 "tm.h"
  40 #include "tree.h"
  41 #include "rtl.h"
  42 #include "hard-reg-set.h"
  43 #include "profile.h"
  44 #include "predict.h"
  45 #include "vec.h"
  46 #include "hashtab.h"
  47 #include "hash-set.h"
  48 #include "machmode.h"
  49 #include "input.h"
  50 #include "function.h"
  51 #include "dominance.h"
  52 #include "cfg.h"
  53 #include "cfganal.h"
  54 #include "basic-block.h"
  55 #include "flags.h"
  56 #include "params.h"
  57 #include "coverage.h"
  58 #include "tree-pass.h"
  59 #include "tree-ssa-alias.h"
  60 #include "internal-fn.h"
  61 #include "gimple-expr.h"
  62 #include "is-a.h"
  63 #include "gimple.h"
  64 #include "gimple-iterator.h"
  65 #include "tree-cfg.h"
  66 #include "tree-ssa.h"
  67 #include "tree-inline.h"
  68 #include "cfgloop.h"
  69 #include "fibonacci_heap.h"
  70
  71 static int count_insns (basic_block);
  72 static bool ignore_bb_p (const_basic_block);
  73 static bool better_p (const_edge, const_edge);
  74 static edge find_best_successor (basic_block);
  75 static edge find_best_predecessor (basic_block);
  76 static int find_trace (basic_block, basic_block *);
  77
  78 /* Minimal outgoing edge probability considered for superblock formation.  */
  79 static int probability_cutoff;
  80 static int branch_ratio_cutoff;
  81
  82 /* A bit BB->index is set if BB has already been seen, i.e. it is
  83    connected to some trace already.  */
  84 sbitmap bb_seen;
  85
  86 static inline void
  87 mark_bb_seen (basic_block bb)
  88 {
  89   unsigned int size = SBITMAP_SIZE (bb_seen);
  90
  91   if ((unsigned int)bb->index >= size)
  92     bb_seen = sbitmap_resize (bb_seen, size * 2, 0);
  93
  94   bitmap_set_bit (bb_seen, bb->index);
  95 }
  96
  97 static inline bool
  98 bb_seen_p (basic_block bb)
  99 {
 100   return bitmap_bit_p (bb_seen, bb->index);
 101 }
 102
 103 /* Return true if we should ignore the basic block for purposes of tracing.  */
 104 static bool
 105 ignore_bb_p (const_basic_block bb)
 106 {
 107   gimple g;
 108
 109   if (bb->index < NUM_FIXED_BLOCKS)
 110     return true;
 111   if (optimize_bb_for_size_p (bb))
 112     return true;
 113
 114   /* A transaction is a single entry multiple exit region.  It must be
 115      duplicated in its entirety or not at all.  */
 116   g = last_stmt (CONST_CAST_BB (bb));
 117   if (g && gimple_code (g) == GIMPLE_TRANSACTION)
 118     return true;
 119
 120   return false;
 121 }
 122
 123 /* Return number of instructions in the block.  */
 124
 125 static int
 126 count_insns (basic_block bb)
 127 {
 128   gimple_stmt_iterator gsi;
 129   gimple stmt;
 130   int n = 0;
 131
 132   for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
 133     {
 134       stmt = gsi_stmt (gsi);
 135       n += estimate_num_insns (stmt, &eni_size_weights);
 136     }
 137   return n;
 138 }
 139
 140 /* Return true if E1 is more frequent than E2.  */
 141 static bool
 142 better_p (const_edge e1, const_edge e2)
 143 {
 144   if (e1->count != e2->count)
 145     return e1->count > e2->count;
 146   if (e1->src->frequency * e1->probability !=
 147       e2->src->frequency * e2->probability)
 148     return (e1->src->frequency * e1->probability
 149             > e2->src->frequency * e2->probability);
 150   /* This is needed to avoid changes in the decision after
 151      CFG is modified.  */
 152   if (e1->src != e2->src)
 153     return e1->src->index > e2->src->index;
 154   return e1->dest->index > e2->dest->index;
 155 }
 156
 157 /* Return most frequent successor of basic block BB.  */
 158
 159 static edge
 160 find_best_successor (basic_block bb)
 161 {
 162   edge e;
 163   edge best = NULL;
 164   edge_iterator ei;
 165
 166   FOR_EACH_EDGE (e, ei, bb->succs)
 167     if (!best || better_p (e, best))
 168       best = e;
 169   if (!best || ignore_bb_p (best->dest))
 170     return NULL;
 171   if (best->probability <= probability_cutoff)
 172     return NULL;
 173   return best;
 174 }
 175
 176 /* Return most frequent predecessor of basic block BB.  */
 177
 178 static edge
 179 find_best_predecessor (basic_block bb)
 180 {
 181   edge e;
 182   edge best = NULL;
 183   edge_iterator ei;
 184
 185   FOR_EACH_EDGE (e, ei, bb->preds)
 186     if (!best || better_p (e, best))
 187       best = e;
 188   if (!best || ignore_bb_p (best->src))
 189     return NULL;
 190   if (EDGE_FREQUENCY (best) * REG_BR_PROB_BASE
 191       < bb->frequency * branch_ratio_cutoff)
 192     return NULL;
 193   return best;
 194 }
 195
 196 /* Find the trace using bb and record it in the TRACE array.
 197    Return number of basic blocks recorded.  */
 198
 199 static int
 200 find_trace (basic_block bb, basic_block *trace)
 201 {
 202   int i = 0;
 203   edge e;
 204
 205   if (dump_file)
 206     fprintf (dump_file, "Trace seed %i [%i]", bb->index, bb->frequency);
 207
 208   while ((e = find_best_predecessor (bb)) != NULL)
 209     {
 210       basic_block bb2 = e->src;
 211       if (bb_seen_p (bb2) || (e->flags & (EDGE_DFS_BACK | EDGE_COMPLEX))
 212           || find_best_successor (bb2) != e)
 213         break;
 214       if (dump_file)
 215         fprintf (dump_file, ",%i [%i]", bb->index, bb->frequency);
 216       bb = bb2;
 217     }
 218   if (dump_file)
 219     fprintf (dump_file, " forward %i [%i]", bb->index, bb->frequency);
 220   trace[i++] = bb;
 221
 222   /* Follow the trace in forward direction.  */
 223   while ((e = find_best_successor (bb)) != NULL)
 224     {
 225       bb = e->dest;
 226       if (bb_seen_p (bb) || (e->flags & (EDGE_DFS_BACK | EDGE_COMPLEX))
 227           || find_best_predecessor (bb) != e)
 228         break;
 229       if (dump_file)
 230         fprintf (dump_file, ",%i [%i]", bb->index, bb->frequency);
 231       trace[i++] = bb;
 232     }
 233   if (dump_file)
 234     fprintf (dump_file, "\n");
 235   return i;
 236 }
 237
 238 /* Look for basic blocks in frequency order, construct traces and tail duplicate
 239    if profitable.  */
 240
 241 static bool
 242 tail_duplicate (void)
 243 {
 244   auto_vec<fibonacci_node<long, basic_block_def>*> blocks;
 245   blocks.safe_grow_cleared (last_basic_block_for_fn (cfun));
 246
 247   basic_block *trace = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun));
 248   int *counts = XNEWVEC (int, last_basic_block_for_fn (cfun));
 249   int ninsns = 0, nduplicated = 0;
 250   gcov_type weighted_insns = 0, traced_insns = 0;
 251   fibonacci_heap<long, basic_block_def> heap (LONG_MIN);
 252   gcov_type cover_insns;
 253   int max_dup_insns;
 254   basic_block bb;
 255   bool changed = false;
 256
 257   /* Create an oversized sbitmap to reduce the chance that we need to
 258      resize it.  */
 259   bb_seen = sbitmap_alloc (last_basic_block_for_fn (cfun) * 2);
 260   bitmap_clear (bb_seen);
 261   initialize_original_copy_tables ();
 262
 263   if (profile_info && flag_branch_probabilities)
 264     probability_cutoff = PARAM_VALUE (TRACER_MIN_BRANCH_PROBABILITY_FEEDBACK);
 265   else
 266     probability_cutoff = PARAM_VALUE (TRACER_MIN_BRANCH_PROBABILITY);
 267   probability_cutoff = REG_BR_PROB_BASE / 100 * probability_cutoff;
 268
 269   branch_ratio_cutoff =
 270     (REG_BR_PROB_BASE / 100 * PARAM_VALUE (TRACER_MIN_BRANCH_RATIO));
 271
 272   FOR_EACH_BB_FN (bb, cfun)
 273     {
 274       int n = count_insns (bb);
 275       if (!ignore_bb_p (bb))
 276         blocks[bb->index] = heap.insert (-bb->frequency, bb);
 277
 278       counts [bb->index] = n;
 279       ninsns += n;
 280       weighted_insns += n * bb->frequency;
 281     }
 282
 283   if (profile_info && flag_branch_probabilities)
 284     cover_insns = PARAM_VALUE (TRACER_DYNAMIC_COVERAGE_FEEDBACK);
 285   else
 286     cover_insns = PARAM_VALUE (TRACER_DYNAMIC_COVERAGE);
 287   cover_insns = (weighted_insns * cover_insns + 50) / 100;
 288   max_dup_insns = (ninsns * PARAM_VALUE (TRACER_MAX_CODE_GROWTH) + 50) / 100;
 289
 290   while (traced_insns < cover_insns && nduplicated < max_dup_insns
 291          && !heap.empty ())
 292     {
 293       basic_block bb = heap.extract_min ();
 294       int n, pos;
 295
 296       if (!bb)
 297         break;
 298
 299       blocks[bb->index] = NULL;
 300
 301       if (ignore_bb_p (bb))
 302         continue;
 303       gcc_assert (!bb_seen_p (bb));
 304
 305       n = find_trace (bb, trace);
 306
 307       bb = trace[0];
 308       traced_insns += bb->frequency * counts [bb->index];
 309       if (blocks[bb->index])
 310         {
 311           heap.delete_node (blocks[bb->index]);
 312           blocks[bb->index] = NULL;
 313         }
 314
 315       for (pos = 1; pos < n; pos++)
 316         {
 317           basic_block bb2 = trace[pos];
 318
 319           if (blocks[bb2->index])
 320             {
 321               heap.delete_node (blocks[bb2->index]);
 322               blocks[bb2->index] = NULL;
 323             }
 324           traced_insns += bb2->frequency * counts [bb2->index];
 325           if (EDGE_COUNT (bb2->preds) > 1
 326               && can_duplicate_block_p (bb2)
 327               /* We have the tendency to duplicate the loop header
 328                  of all do { } while loops.  Do not do that - it is
 329                  not profitable and it might create a loop with multiple
 330                  entries or at least rotate the loop.  */
 331               && bb2->loop_father->header != bb2)
 332             {
 333               edge e;
 334               basic_block copy;
 335
 336               nduplicated += counts [bb2->index];
 337
 338               e = find_edge (bb, bb2);
 339
 340               copy = duplicate_block (bb2, e, bb);
 341               flush_pending_stmts (e);
 342
 343               add_phi_args_after_copy (&copy, 1, NULL);
 344
 345               /* Reconsider the original copy of block we've duplicated.
 346                  Removing the most common predecessor may make it to be
 347                  head.  */
 348               blocks[bb2->index] = heap.insert (-bb2->frequency, bb2);
 349
 350               if (dump_file)
 351                 fprintf (dump_file, "Duplicated %i as %i [%i]\n",
 352                          bb2->index, copy->index, copy->frequency);
 353
 354               bb2 = copy;
 355               changed = true;
 356             }
 357           mark_bb_seen (bb2);
 358           bb = bb2;
 359           /* In case the trace became infrequent, stop duplicating.  */
 360           if (ignore_bb_p (bb))
 361             break;
 362         }
 363       if (dump_file)
 364         fprintf (dump_file, " covered now %.1f\n\n",
 365                  traced_insns * 100.0 / weighted_insns);
 366     }
 367   if (dump_file)
 368     fprintf (dump_file, "Duplicated %i insns (%i%%)\n", nduplicated,
 369              nduplicated * 100 / ninsns);
 370
 371   free_original_copy_tables ();
 372   sbitmap_free (bb_seen);
 373   free (trace);
 374   free (counts);
 375
 376   return changed;
 377 }
 378 \f
 379 namespace {
 380
 381 const pass_data pass_data_tracer =
 382 {
 383   GIMPLE_PASS, /* type */
 384   "tracer", /* name */
 385   OPTGROUP_NONE, /* optinfo_flags */
 386   TV_TRACER, /* tv_id */
 387   0, /* properties_required */
 388   0, /* properties_provided */
 389   0, /* properties_destroyed */
 390   0, /* todo_flags_start */
 391   TODO_update_ssa, /* todo_flags_finish */
 392 };
 393
 394 class pass_tracer : public gimple_opt_pass
 395 {
 396 public:
 397   pass_tracer (gcc::context *ctxt)
 398     : gimple_opt_pass (pass_data_tracer, ctxt)
 399   {}
 400
 401   /* opt_pass methods: */
 402   virtual bool gate (function *)
 403     {
 404       return (optimize > 0 && flag_tracer && flag_reorder_blocks);
 405     }
 406
 407   virtual unsigned int execute (function *);
 408
 409 }; // class pass_tracer
 410
 411 unsigned int
 412 pass_tracer::execute (function *fun)
 413 {
 414   bool changed;
 415
 416   if (n_basic_blocks_for_fn (fun) <= NUM_FIXED_BLOCKS + 1)
 417     return 0;
 418
 419   mark_dfs_back_edges ();
 420   if (dump_file)
 421     brief_dump_cfg (dump_file, dump_flags);
 422
 423   /* Trace formation is done on the fly inside tail_duplicate */
 424   changed = tail_duplicate ();
 425   if (changed)
 426     {
 427       free_dominance_info (CDI_DOMINATORS);
 428       /* If we changed the CFG schedule loops for fixup by cleanup_cfg.  */
 429       loops_state_set (LOOPS_NEED_FIXUP);
 430     }
 431
 432   if (dump_file)
 433     brief_dump_cfg (dump_file, dump_flags);
 434
 435   return changed ? TODO_cleanup_cfg : 0;
 436 }
 437 } // anon namespace
 438
 439 gimple_opt_pass *
 440 make_pass_tracer (gcc::context *ctxt)
 441 {
 442   return new pass_tracer (ctxt);
 443 }