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, 2002, 2003, 2004, 2005, 2006, 2007, 2008
5 Free Software Foundation, Inc.
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify it
10 under the terms of the GNU General Public License as published by
11 the Free Software Foundation; either version 3, or (at your option)
14 GCC is distributed in the hope that it will be useful, but WITHOUT
15 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY
16 or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public
17 License for more details.
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
23 /* This pass performs the tail duplication needed for superblock formation.
24 For more information see:
26 Design and Analysis of Profile-Based Optimization in Compaq's
27 Compilation Tools for Alpha; Journal of Instruction-Level
28 Parallelism 3 (2000) 1-25
30 Unlike Compaq's implementation we don't do the loop peeling as most
31 probably a better job can be done by a special pass and we don't
32 need to worry too much about the code size implications as the tail
33 duplicates are crossjumped again if optimizations are not
39 #include "coretypes.h"
43 #include "hard-reg-set.h"
44 #include "basic-block.h"
50 #include "tree-pass.h"
51 #include "tree-flow.h"
52 #include "tree-inline.h"
55 static int count_insns (basic_block);
56 static bool ignore_bb_p (const_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_BYTES (bb_seen) * 8;
75 if ((unsigned int)bb->index >= size)
76 bb_seen = sbitmap_resize (bb_seen, size * 2, 0);
78 SET_BIT (bb_seen, bb->index);
82 bb_seen_p (basic_block bb)
84 return TEST_BIT (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)
93 if (bb->index < NUM_FIXED_BLOCKS)
95 if (optimize_bb_for_size_p (bb))
98 /* A transaction is a single entry multiple exit region. It must be
99 duplicated in its entirety or not at all. */
100 g = last_stmt (CONST_CAST_BB (bb));
101 if (g && gimple_code (g) == GIMPLE_TRANSACTION)
107 /* Return number of instructions in the block. */
110 count_insns (basic_block bb)
112 gimple_stmt_iterator gsi;
116 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
118 stmt = gsi_stmt (gsi);
119 n += estimate_num_insns (stmt, &eni_size_weights);
124 /* Return true if E1 is more frequent than E2. */
126 better_p (const_edge e1, const_edge e2)
128 if (e1->count != e2->count)
129 return e1->count > e2->count;
130 if (e1->src->frequency * e1->probability !=
131 e2->src->frequency * e2->probability)
132 return (e1->src->frequency * e1->probability
133 > e2->src->frequency * e2->probability);
134 /* This is needed to avoid changes in the decision after
136 if (e1->src != e2->src)
137 return e1->src->index > e2->src->index;
138 return e1->dest->index > e2->dest->index;
141 /* Return most frequent successor of basic block BB. */
144 find_best_successor (basic_block bb)
150 FOR_EACH_EDGE (e, ei, bb->succs)
151 if (!best || better_p (e, best))
153 if (!best || ignore_bb_p (best->dest))
155 if (best->probability <= probability_cutoff)
160 /* Return most frequent predecessor of basic block BB. */
163 find_best_predecessor (basic_block bb)
169 FOR_EACH_EDGE (e, ei, bb->preds)
170 if (!best || better_p (e, best))
172 if (!best || ignore_bb_p (best->src))
174 if (EDGE_FREQUENCY (best) * REG_BR_PROB_BASE
175 < bb->frequency * branch_ratio_cutoff)
180 /* Find the trace using bb and record it in the TRACE array.
181 Return number of basic blocks recorded. */
184 find_trace (basic_block bb, basic_block *trace)
190 fprintf (dump_file, "Trace seed %i [%i]", bb->index, bb->frequency);
192 while ((e = find_best_predecessor (bb)) != NULL)
194 basic_block bb2 = e->src;
195 if (bb_seen_p (bb2) || (e->flags & (EDGE_DFS_BACK | EDGE_COMPLEX))
196 || find_best_successor (bb2) != e)
199 fprintf (dump_file, ",%i [%i]", bb->index, bb->frequency);
203 fprintf (dump_file, " forward %i [%i]", bb->index, bb->frequency);
206 /* Follow the trace in forward direction. */
207 while ((e = find_best_successor (bb)) != NULL)
210 if (bb_seen_p (bb) || (e->flags & (EDGE_DFS_BACK | EDGE_COMPLEX))
211 || find_best_predecessor (bb) != e)
214 fprintf (dump_file, ",%i [%i]", bb->index, bb->frequency);
218 fprintf (dump_file, "\n");
222 /* Look for basic blocks in frequency order, construct traces and tail duplicate
226 tail_duplicate (void)
228 fibnode_t *blocks = XCNEWVEC (fibnode_t, last_basic_block);
229 basic_block *trace = XNEWVEC (basic_block, n_basic_blocks);
230 int *counts = XNEWVEC (int, last_basic_block);
231 int ninsns = 0, nduplicated = 0;
232 gcov_type weighted_insns = 0, traced_insns = 0;
233 fibheap_t heap = fibheap_new ();
234 gcov_type cover_insns;
237 bool changed = false;
239 /* Create an oversized sbitmap to reduce the chance that we need to
241 bb_seen = sbitmap_alloc (last_basic_block * 2);
242 sbitmap_zero (bb_seen);
243 initialize_original_copy_tables ();
245 if (profile_info && flag_branch_probabilities)
246 probability_cutoff = PARAM_VALUE (TRACER_MIN_BRANCH_PROBABILITY_FEEDBACK);
248 probability_cutoff = PARAM_VALUE (TRACER_MIN_BRANCH_PROBABILITY);
249 probability_cutoff = REG_BR_PROB_BASE / 100 * probability_cutoff;
251 branch_ratio_cutoff =
252 (REG_BR_PROB_BASE / 100 * PARAM_VALUE (TRACER_MIN_BRANCH_RATIO));
256 int n = count_insns (bb);
257 if (!ignore_bb_p (bb))
258 blocks[bb->index] = fibheap_insert (heap, -bb->frequency,
261 counts [bb->index] = n;
263 weighted_insns += n * bb->frequency;
266 if (profile_info && flag_branch_probabilities)
267 cover_insns = PARAM_VALUE (TRACER_DYNAMIC_COVERAGE_FEEDBACK);
269 cover_insns = PARAM_VALUE (TRACER_DYNAMIC_COVERAGE);
270 cover_insns = (weighted_insns * cover_insns + 50) / 100;
271 max_dup_insns = (ninsns * PARAM_VALUE (TRACER_MAX_CODE_GROWTH) + 50) / 100;
273 while (traced_insns < cover_insns && nduplicated < max_dup_insns
274 && !fibheap_empty (heap))
276 basic_block bb = (basic_block) fibheap_extract_min (heap);
282 blocks[bb->index] = NULL;
284 if (ignore_bb_p (bb))
286 gcc_assert (!bb_seen_p (bb));
288 n = find_trace (bb, trace);
291 traced_insns += bb->frequency * counts [bb->index];
292 if (blocks[bb->index])
294 fibheap_delete_node (heap, blocks[bb->index]);
295 blocks[bb->index] = NULL;
298 for (pos = 1; pos < n; pos++)
300 basic_block bb2 = trace[pos];
302 if (blocks[bb2->index])
304 fibheap_delete_node (heap, blocks[bb2->index]);
305 blocks[bb2->index] = NULL;
307 traced_insns += bb2->frequency * counts [bb2->index];
308 if (EDGE_COUNT (bb2->preds) > 1
309 && can_duplicate_block_p (bb2)
310 /* We have the tendency to duplicate the loop header
311 of all do { } while loops. Do not do that - it is
312 not profitable and it might create a loop with multiple
313 entries or at least rotate the loop. */
315 || bb2->loop_father->header != bb2))
320 nduplicated += counts [bb2->index];
322 e = find_edge (bb, bb2);
324 copy = duplicate_block (bb2, e, bb);
325 flush_pending_stmts (e);
327 add_phi_args_after_copy (©, 1, NULL);
329 /* Reconsider the original copy of block we've duplicated.
330 Removing the most common predecessor may make it to be
333 fibheap_insert (heap, -bb2->frequency, bb2);
336 fprintf (dump_file, "Duplicated %i as %i [%i]\n",
337 bb2->index, copy->index, copy->frequency);
344 /* In case the trace became infrequent, stop duplicating. */
345 if (ignore_bb_p (bb))
349 fprintf (dump_file, " covered now %.1f\n\n",
350 traced_insns * 100.0 / weighted_insns);
353 fprintf (dump_file, "Duplicated %i insns (%i%%)\n", nduplicated,
354 nduplicated * 100 / ninsns);
356 free_original_copy_tables ();
357 sbitmap_free (bb_seen);
361 fibheap_delete (heap);
366 /* Main entry point to this file. */
373 if (n_basic_blocks <= NUM_FIXED_BLOCKS + 1)
376 mark_dfs_back_edges ();
378 brief_dump_cfg (dump_file);
380 /* Trace formation is done on the fly inside tail_duplicate */
381 changed = tail_duplicate ();
383 free_dominance_info (CDI_DOMINATORS);
386 brief_dump_cfg (dump_file);
388 return changed ? TODO_cleanup_cfg : 0;
394 return (optimize > 0 && flag_tracer && flag_reorder_blocks);
397 struct gimple_opt_pass pass_tracer =
402 gate_tracer, /* gate */
403 tracer, /* execute */
406 0, /* static_pass_number */
407 TV_TRACER, /* tv_id */
408 0, /* properties_required */
409 0, /* properties_provided */
410 0, /* properties_destroyed */
411 0, /* todo_flags_start */
413 | TODO_verify_ssa /* todo_flags_finish */