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) 2001-2016 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"
44 #include "tree-pass.h"
48 #include "gimple-iterator.h"
51 #include "tree-inline.h"
53 #include "fibonacci_heap.h"
56 static int count_insns (basic_block);
57 static bool better_p (const_edge, const_edge);
58 static edge find_best_successor (basic_block);
59 static edge find_best_predecessor (basic_block);
60 static int find_trace (basic_block, basic_block *);
62 /* Minimal outgoing edge probability considered for superblock formation. */
63 static int probability_cutoff;
64 static int branch_ratio_cutoff;
66 /* A bit BB->index is set if BB has already been seen, i.e. it is
67 connected to some trace already. */
71 mark_bb_seen (basic_block bb)
73 unsigned int size = SBITMAP_SIZE (bb_seen);
75 if ((unsigned int)bb->index >= size)
76 bb_seen = sbitmap_resize (bb_seen, size * 2, 0);
78 bitmap_set_bit (bb_seen, bb->index);
82 bb_seen_p (basic_block bb)
84 return bitmap_bit_p (bb_seen, bb->index);
87 /* Return true if we should ignore the basic block for purposes of tracing. */
89 ignore_bb_p (const_basic_block bb)
91 if (bb->index < NUM_FIXED_BLOCKS)
93 if (optimize_bb_for_size_p (bb))
96 if (gimple *g = last_stmt (CONST_CAST_BB (bb)))
98 /* A transaction is a single entry multiple exit region. It
99 must be duplicated in its entirety or not at all. */
100 if (gimple_code (g) == GIMPLE_TRANSACTION)
103 /* An IFN_UNIQUE call must be duplicated as part of its group,
105 if (is_gimple_call (g)
106 && gimple_call_internal_p (g)
107 && gimple_call_internal_unique_p (g))
114 /* Return number of instructions in the block. */
117 count_insns (basic_block bb)
119 gimple_stmt_iterator gsi;
123 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
125 stmt = gsi_stmt (gsi);
126 n += estimate_num_insns (stmt, &eni_size_weights);
131 /* Return true if E1 is more frequent than E2. */
133 better_p (const_edge e1, const_edge e2)
135 if (e1->count != e2->count)
136 return e1->count > e2->count;
137 if (e1->src->frequency * e1->probability !=
138 e2->src->frequency * e2->probability)
139 return (e1->src->frequency * e1->probability
140 > e2->src->frequency * e2->probability);
141 /* This is needed to avoid changes in the decision after
143 if (e1->src != e2->src)
144 return e1->src->index > e2->src->index;
145 return e1->dest->index > e2->dest->index;
148 /* Return most frequent successor of basic block BB. */
151 find_best_successor (basic_block bb)
157 FOR_EACH_EDGE (e, ei, bb->succs)
158 if (!best || better_p (e, best))
160 if (!best || ignore_bb_p (best->dest))
162 if (best->probability <= probability_cutoff)
167 /* Return most frequent predecessor of basic block BB. */
170 find_best_predecessor (basic_block bb)
176 FOR_EACH_EDGE (e, ei, bb->preds)
177 if (!best || better_p (e, best))
179 if (!best || ignore_bb_p (best->src))
181 if (EDGE_FREQUENCY (best) * REG_BR_PROB_BASE
182 < bb->frequency * branch_ratio_cutoff)
187 /* Find the trace using bb and record it in the TRACE array.
188 Return number of basic blocks recorded. */
191 find_trace (basic_block bb, basic_block *trace)
197 fprintf (dump_file, "Trace seed %i [%i]", bb->index, bb->frequency);
199 while ((e = find_best_predecessor (bb)) != NULL)
201 basic_block bb2 = e->src;
202 if (bb_seen_p (bb2) || (e->flags & (EDGE_DFS_BACK | EDGE_COMPLEX))
203 || find_best_successor (bb2) != e)
206 fprintf (dump_file, ",%i [%i]", bb->index, bb->frequency);
210 fprintf (dump_file, " forward %i [%i]", bb->index, bb->frequency);
213 /* Follow the trace in forward direction. */
214 while ((e = find_best_successor (bb)) != NULL)
217 if (bb_seen_p (bb) || (e->flags & (EDGE_DFS_BACK | EDGE_COMPLEX))
218 || find_best_predecessor (bb) != e)
221 fprintf (dump_file, ",%i [%i]", bb->index, bb->frequency);
225 fprintf (dump_file, "\n");
229 /* Duplicate block BB2, placing it after BB in the CFG. Return the
230 newly created block. */
232 transform_duplicate (basic_block bb, basic_block bb2)
237 e = find_edge (bb, bb2);
239 copy = duplicate_block (bb2, e, bb);
240 flush_pending_stmts (e);
242 add_phi_args_after_copy (©, 1, NULL);
247 /* Look for basic blocks in frequency order, construct traces and tail duplicate
251 tail_duplicate (void)
253 auto_vec<fibonacci_node<long, basic_block_def>*> blocks;
254 blocks.safe_grow_cleared (last_basic_block_for_fn (cfun));
256 basic_block *trace = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun));
257 int *counts = XNEWVEC (int, last_basic_block_for_fn (cfun));
258 int ninsns = 0, nduplicated = 0;
259 gcov_type weighted_insns = 0, traced_insns = 0;
260 fibonacci_heap<long, basic_block_def> heap (LONG_MIN);
261 gcov_type cover_insns;
264 bool changed = false;
266 /* Create an oversized sbitmap to reduce the chance that we need to
268 bb_seen = sbitmap_alloc (last_basic_block_for_fn (cfun) * 2);
269 bitmap_clear (bb_seen);
270 initialize_original_copy_tables ();
272 if (profile_info && flag_branch_probabilities)
273 probability_cutoff = PARAM_VALUE (TRACER_MIN_BRANCH_PROBABILITY_FEEDBACK);
275 probability_cutoff = PARAM_VALUE (TRACER_MIN_BRANCH_PROBABILITY);
276 probability_cutoff = REG_BR_PROB_BASE / 100 * probability_cutoff;
278 branch_ratio_cutoff =
279 (REG_BR_PROB_BASE / 100 * PARAM_VALUE (TRACER_MIN_BRANCH_RATIO));
281 FOR_EACH_BB_FN (bb, cfun)
283 int n = count_insns (bb);
284 if (!ignore_bb_p (bb))
285 blocks[bb->index] = heap.insert (-bb->frequency, bb);
287 counts [bb->index] = n;
289 weighted_insns += n * bb->frequency;
292 if (profile_info && flag_branch_probabilities)
293 cover_insns = PARAM_VALUE (TRACER_DYNAMIC_COVERAGE_FEEDBACK);
295 cover_insns = PARAM_VALUE (TRACER_DYNAMIC_COVERAGE);
296 cover_insns = (weighted_insns * cover_insns + 50) / 100;
297 max_dup_insns = (ninsns * PARAM_VALUE (TRACER_MAX_CODE_GROWTH) + 50) / 100;
299 while (traced_insns < cover_insns && nduplicated < max_dup_insns
302 basic_block bb = heap.extract_min ();
308 blocks[bb->index] = NULL;
310 if (ignore_bb_p (bb))
312 gcc_assert (!bb_seen_p (bb));
314 n = find_trace (bb, trace);
317 traced_insns += bb->frequency * counts [bb->index];
318 if (blocks[bb->index])
320 heap.delete_node (blocks[bb->index]);
321 blocks[bb->index] = NULL;
324 for (pos = 1; pos < n; pos++)
326 basic_block bb2 = trace[pos];
328 if (blocks[bb2->index])
330 heap.delete_node (blocks[bb2->index]);
331 blocks[bb2->index] = NULL;
333 traced_insns += bb2->frequency * counts [bb2->index];
334 if (EDGE_COUNT (bb2->preds) > 1
335 && can_duplicate_block_p (bb2)
336 /* We have the tendency to duplicate the loop header
337 of all do { } while loops. Do not do that - it is
338 not profitable and it might create a loop with multiple
339 entries or at least rotate the loop. */
340 && bb2->loop_father->header != bb2)
342 nduplicated += counts [bb2->index];
343 basic_block copy = transform_duplicate (bb, bb2);
345 /* Reconsider the original copy of block we've duplicated.
346 Removing the most common predecessor may make it to be
348 blocks[bb2->index] = heap.insert (-bb2->frequency, bb2);
351 fprintf (dump_file, "Duplicated %i as %i [%i]\n",
352 bb2->index, copy->index, copy->frequency);
359 /* In case the trace became infrequent, stop duplicating. */
360 if (ignore_bb_p (bb))
364 fprintf (dump_file, " covered now %.1f\n\n",
365 traced_insns * 100.0 / weighted_insns);
368 fprintf (dump_file, "Duplicated %i insns (%i%%)\n", nduplicated,
369 nduplicated * 100 / ninsns);
371 free_original_copy_tables ();
372 sbitmap_free (bb_seen);
381 const pass_data pass_data_tracer =
383 GIMPLE_PASS, /* type */
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 */
394 class pass_tracer : public gimple_opt_pass
397 pass_tracer (gcc::context *ctxt)
398 : gimple_opt_pass (pass_data_tracer, ctxt)
401 /* opt_pass methods: */
402 virtual bool gate (function *)
404 return (optimize > 0 && flag_tracer && flag_reorder_blocks);
407 virtual unsigned int execute (function *);
409 }; // class pass_tracer
412 pass_tracer::execute (function *fun)
416 if (n_basic_blocks_for_fn (fun) <= NUM_FIXED_BLOCKS + 1)
419 mark_dfs_back_edges ();
421 brief_dump_cfg (dump_file, dump_flags);
423 /* Trace formation is done on the fly inside tail_duplicate */
424 changed = tail_duplicate ();
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);
433 brief_dump_cfg (dump_file, dump_flags);
435 return changed ? TODO_cleanup_cfg : 0;
440 make_pass_tracer (gcc::context *ctxt)
442 return new pass_tracer (ctxt);