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.
6 This file is part of GCC.
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)
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.
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/>. */
22 /* This pass performs the tail duplication needed for superblock formation.
23 For more information see:
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
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
38 #include "coretypes.h"
43 #include "double-int.h"
50 #include "fold-const.h"
52 #include "hard-reg-set.h"
57 #include "dominance.h"
60 #include "basic-block.h"
64 #include "tree-pass.h"
65 #include "tree-ssa-alias.h"
66 #include "internal-fn.h"
67 #include "gimple-expr.h"
70 #include "gimple-iterator.h"
73 #include "tree-inline.h"
75 #include "fibonacci_heap.h"
77 static int count_insns (basic_block);
78 static bool ignore_bb_p (const_basic_block);
79 static bool better_p (const_edge, const_edge);
80 static edge find_best_successor (basic_block);
81 static edge find_best_predecessor (basic_block);
82 static int find_trace (basic_block, basic_block *);
84 /* Minimal outgoing edge probability considered for superblock formation. */
85 static int probability_cutoff;
86 static int branch_ratio_cutoff;
88 /* A bit BB->index is set if BB has already been seen, i.e. it is
89 connected to some trace already. */
93 mark_bb_seen (basic_block bb)
95 unsigned int size = SBITMAP_SIZE (bb_seen);
97 if ((unsigned int)bb->index >= size)
98 bb_seen = sbitmap_resize (bb_seen, size * 2, 0);
100 bitmap_set_bit (bb_seen, bb->index);
104 bb_seen_p (basic_block bb)
106 return bitmap_bit_p (bb_seen, bb->index);
109 /* Return true if we should ignore the basic block for purposes of tracing. */
111 ignore_bb_p (const_basic_block bb)
115 if (bb->index < NUM_FIXED_BLOCKS)
117 if (optimize_bb_for_size_p (bb))
120 /* A transaction is a single entry multiple exit region. It must be
121 duplicated in its entirety or not at all. */
122 g = last_stmt (CONST_CAST_BB (bb));
123 if (g && gimple_code (g) == GIMPLE_TRANSACTION)
129 /* Return number of instructions in the block. */
132 count_insns (basic_block bb)
134 gimple_stmt_iterator gsi;
138 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
140 stmt = gsi_stmt (gsi);
141 n += estimate_num_insns (stmt, &eni_size_weights);
146 /* Return true if E1 is more frequent than E2. */
148 better_p (const_edge e1, const_edge e2)
150 if (e1->count != e2->count)
151 return e1->count > e2->count;
152 if (e1->src->frequency * e1->probability !=
153 e2->src->frequency * e2->probability)
154 return (e1->src->frequency * e1->probability
155 > e2->src->frequency * e2->probability);
156 /* This is needed to avoid changes in the decision after
158 if (e1->src != e2->src)
159 return e1->src->index > e2->src->index;
160 return e1->dest->index > e2->dest->index;
163 /* Return most frequent successor of basic block BB. */
166 find_best_successor (basic_block bb)
172 FOR_EACH_EDGE (e, ei, bb->succs)
173 if (!best || better_p (e, best))
175 if (!best || ignore_bb_p (best->dest))
177 if (best->probability <= probability_cutoff)
182 /* Return most frequent predecessor of basic block BB. */
185 find_best_predecessor (basic_block bb)
191 FOR_EACH_EDGE (e, ei, bb->preds)
192 if (!best || better_p (e, best))
194 if (!best || ignore_bb_p (best->src))
196 if (EDGE_FREQUENCY (best) * REG_BR_PROB_BASE
197 < bb->frequency * branch_ratio_cutoff)
202 /* Find the trace using bb and record it in the TRACE array.
203 Return number of basic blocks recorded. */
206 find_trace (basic_block bb, basic_block *trace)
212 fprintf (dump_file, "Trace seed %i [%i]", bb->index, bb->frequency);
214 while ((e = find_best_predecessor (bb)) != NULL)
216 basic_block bb2 = e->src;
217 if (bb_seen_p (bb2) || (e->flags & (EDGE_DFS_BACK | EDGE_COMPLEX))
218 || find_best_successor (bb2) != e)
221 fprintf (dump_file, ",%i [%i]", bb->index, bb->frequency);
225 fprintf (dump_file, " forward %i [%i]", bb->index, bb->frequency);
228 /* Follow the trace in forward direction. */
229 while ((e = find_best_successor (bb)) != NULL)
232 if (bb_seen_p (bb) || (e->flags & (EDGE_DFS_BACK | EDGE_COMPLEX))
233 || find_best_predecessor (bb) != e)
236 fprintf (dump_file, ",%i [%i]", bb->index, bb->frequency);
240 fprintf (dump_file, "\n");
244 /* Look for basic blocks in frequency order, construct traces and tail duplicate
248 tail_duplicate (void)
250 auto_vec<fibonacci_node<long, basic_block_def>*> blocks;
251 blocks.safe_grow_cleared (last_basic_block_for_fn (cfun));
253 basic_block *trace = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun));
254 int *counts = XNEWVEC (int, last_basic_block_for_fn (cfun));
255 int ninsns = 0, nduplicated = 0;
256 gcov_type weighted_insns = 0, traced_insns = 0;
257 fibonacci_heap<long, basic_block_def> heap (LONG_MIN);
258 gcov_type cover_insns;
261 bool changed = false;
263 /* Create an oversized sbitmap to reduce the chance that we need to
265 bb_seen = sbitmap_alloc (last_basic_block_for_fn (cfun) * 2);
266 bitmap_clear (bb_seen);
267 initialize_original_copy_tables ();
269 if (profile_info && flag_branch_probabilities)
270 probability_cutoff = PARAM_VALUE (TRACER_MIN_BRANCH_PROBABILITY_FEEDBACK);
272 probability_cutoff = PARAM_VALUE (TRACER_MIN_BRANCH_PROBABILITY);
273 probability_cutoff = REG_BR_PROB_BASE / 100 * probability_cutoff;
275 branch_ratio_cutoff =
276 (REG_BR_PROB_BASE / 100 * PARAM_VALUE (TRACER_MIN_BRANCH_RATIO));
278 FOR_EACH_BB_FN (bb, cfun)
280 int n = count_insns (bb);
281 if (!ignore_bb_p (bb))
282 blocks[bb->index] = heap.insert (-bb->frequency, bb);
284 counts [bb->index] = n;
286 weighted_insns += n * bb->frequency;
289 if (profile_info && flag_branch_probabilities)
290 cover_insns = PARAM_VALUE (TRACER_DYNAMIC_COVERAGE_FEEDBACK);
292 cover_insns = PARAM_VALUE (TRACER_DYNAMIC_COVERAGE);
293 cover_insns = (weighted_insns * cover_insns + 50) / 100;
294 max_dup_insns = (ninsns * PARAM_VALUE (TRACER_MAX_CODE_GROWTH) + 50) / 100;
296 while (traced_insns < cover_insns && nduplicated < max_dup_insns
299 basic_block bb = heap.extract_min ();
305 blocks[bb->index] = NULL;
307 if (ignore_bb_p (bb))
309 gcc_assert (!bb_seen_p (bb));
311 n = find_trace (bb, trace);
314 traced_insns += bb->frequency * counts [bb->index];
315 if (blocks[bb->index])
317 heap.delete_node (blocks[bb->index]);
318 blocks[bb->index] = NULL;
321 for (pos = 1; pos < n; pos++)
323 basic_block bb2 = trace[pos];
325 if (blocks[bb2->index])
327 heap.delete_node (blocks[bb2->index]);
328 blocks[bb2->index] = NULL;
330 traced_insns += bb2->frequency * counts [bb2->index];
331 if (EDGE_COUNT (bb2->preds) > 1
332 && can_duplicate_block_p (bb2)
333 /* We have the tendency to duplicate the loop header
334 of all do { } while loops. Do not do that - it is
335 not profitable and it might create a loop with multiple
336 entries or at least rotate the loop. */
337 && bb2->loop_father->header != bb2)
342 nduplicated += counts [bb2->index];
344 e = find_edge (bb, bb2);
346 copy = duplicate_block (bb2, e, bb);
347 flush_pending_stmts (e);
349 add_phi_args_after_copy (©, 1, NULL);
351 /* Reconsider the original copy of block we've duplicated.
352 Removing the most common predecessor may make it to be
354 blocks[bb2->index] = heap.insert (-bb2->frequency, bb2);
357 fprintf (dump_file, "Duplicated %i as %i [%i]\n",
358 bb2->index, copy->index, copy->frequency);
365 /* In case the trace became infrequent, stop duplicating. */
366 if (ignore_bb_p (bb))
370 fprintf (dump_file, " covered now %.1f\n\n",
371 traced_insns * 100.0 / weighted_insns);
374 fprintf (dump_file, "Duplicated %i insns (%i%%)\n", nduplicated,
375 nduplicated * 100 / ninsns);
377 free_original_copy_tables ();
378 sbitmap_free (bb_seen);
387 const pass_data pass_data_tracer =
389 GIMPLE_PASS, /* type */
391 OPTGROUP_NONE, /* optinfo_flags */
392 TV_TRACER, /* tv_id */
393 0, /* properties_required */
394 0, /* properties_provided */
395 0, /* properties_destroyed */
396 0, /* todo_flags_start */
397 TODO_update_ssa, /* todo_flags_finish */
400 class pass_tracer : public gimple_opt_pass
403 pass_tracer (gcc::context *ctxt)
404 : gimple_opt_pass (pass_data_tracer, ctxt)
407 /* opt_pass methods: */
408 virtual bool gate (function *)
410 return (optimize > 0 && flag_tracer && flag_reorder_blocks);
413 virtual unsigned int execute (function *);
415 }; // class pass_tracer
418 pass_tracer::execute (function *fun)
422 if (n_basic_blocks_for_fn (fun) <= NUM_FIXED_BLOCKS + 1)
425 mark_dfs_back_edges ();
427 brief_dump_cfg (dump_file, dump_flags);
429 /* Trace formation is done on the fly inside tail_duplicate */
430 changed = tail_duplicate ();
433 free_dominance_info (CDI_DOMINATORS);
434 /* If we changed the CFG schedule loops for fixup by cleanup_cfg. */
435 loops_state_set (LOOPS_NEED_FIXUP);
439 brief_dump_cfg (dump_file, dump_flags);
441 return changed ? TODO_cleanup_cfg : 0;
446 make_pass_tracer (gcc::context *ctxt)
448 return new pass_tracer (ctxt);