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"
46 #include "cfglayout.h"
52 #include "tree-pass.h"
53 #include "tree-flow.h"
54 #include "tree-inline.h"
57 static int count_insns (basic_block);
58 static bool ignore_bb_p (const_basic_block);
59 static bool better_p (const_edge, const_edge);
60 static edge find_best_successor (basic_block);
61 static edge find_best_predecessor (basic_block);
62 static int find_trace (basic_block, basic_block *);
64 /* Minimal outgoing edge probability considered for superblock formation. */
65 static int probability_cutoff;
66 static int branch_ratio_cutoff;
68 /* A bit BB->index is set if BB has already been seen, i.e. it is
69 connected to some trace already. */
73 mark_bb_seen (basic_block bb)
75 unsigned int size = SBITMAP_SIZE_BYTES (bb_seen) * 8;
77 if ((unsigned int)bb->index >= size)
78 bb_seen = sbitmap_resize (bb_seen, size * 2, 0);
80 SET_BIT (bb_seen, bb->index);
84 bb_seen_p (basic_block bb)
86 return TEST_BIT (bb_seen, bb->index);
89 /* Return true if we should ignore the basic block for purposes of tracing. */
91 ignore_bb_p (const_basic_block bb)
95 if (bb->index < NUM_FIXED_BLOCKS)
97 if (optimize_bb_for_size_p (bb))
100 /* A transaction is a single entry multiple exit region. It must be
101 duplicated in its entirety or not at all. */
102 g = last_stmt (CONST_CAST_BB (bb));
103 if (g && gimple_code (g) == GIMPLE_TRANSACTION)
109 /* Return number of instructions in the block. */
112 count_insns (basic_block bb)
114 gimple_stmt_iterator gsi;
118 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
120 stmt = gsi_stmt (gsi);
121 n += estimate_num_insns (stmt, &eni_size_weights);
126 /* Return true if E1 is more frequent than E2. */
128 better_p (const_edge e1, const_edge e2)
130 if (e1->count != e2->count)
131 return e1->count > e2->count;
132 if (e1->src->frequency * e1->probability !=
133 e2->src->frequency * e2->probability)
134 return (e1->src->frequency * e1->probability
135 > e2->src->frequency * e2->probability);
136 /* This is needed to avoid changes in the decision after
138 if (e1->src != e2->src)
139 return e1->src->index > e2->src->index;
140 return e1->dest->index > e2->dest->index;
143 /* Return most frequent successor of basic block BB. */
146 find_best_successor (basic_block bb)
152 FOR_EACH_EDGE (e, ei, bb->succs)
153 if (!best || better_p (e, best))
155 if (!best || ignore_bb_p (best->dest))
157 if (best->probability <= probability_cutoff)
162 /* Return most frequent predecessor of basic block BB. */
165 find_best_predecessor (basic_block bb)
171 FOR_EACH_EDGE (e, ei, bb->preds)
172 if (!best || better_p (e, best))
174 if (!best || ignore_bb_p (best->src))
176 if (EDGE_FREQUENCY (best) * REG_BR_PROB_BASE
177 < bb->frequency * branch_ratio_cutoff)
182 /* Find the trace using bb and record it in the TRACE array.
183 Return number of basic blocks recorded. */
186 find_trace (basic_block bb, basic_block *trace)
192 fprintf (dump_file, "Trace seed %i [%i]", bb->index, bb->frequency);
194 while ((e = find_best_predecessor (bb)) != NULL)
196 basic_block bb2 = e->src;
197 if (bb_seen_p (bb2) || (e->flags & (EDGE_DFS_BACK | EDGE_COMPLEX))
198 || find_best_successor (bb2) != e)
201 fprintf (dump_file, ",%i [%i]", bb->index, bb->frequency);
205 fprintf (dump_file, " forward %i [%i]", bb->index, bb->frequency);
208 /* Follow the trace in forward direction. */
209 while ((e = find_best_successor (bb)) != NULL)
212 if (bb_seen_p (bb) || (e->flags & (EDGE_DFS_BACK | EDGE_COMPLEX))
213 || find_best_predecessor (bb) != e)
216 fprintf (dump_file, ",%i [%i]", bb->index, bb->frequency);
220 fprintf (dump_file, "\n");
224 /* Look for basic blocks in frequency order, construct traces and tail duplicate
228 tail_duplicate (void)
230 fibnode_t *blocks = XCNEWVEC (fibnode_t, last_basic_block);
231 basic_block *trace = XNEWVEC (basic_block, n_basic_blocks);
232 int *counts = XNEWVEC (int, last_basic_block);
233 int ninsns = 0, nduplicated = 0;
234 gcov_type weighted_insns = 0, traced_insns = 0;
235 fibheap_t heap = fibheap_new ();
236 gcov_type cover_insns;
239 bool changed = false;
241 /* Create an oversized sbitmap to reduce the chance that we need to
243 bb_seen = sbitmap_alloc (last_basic_block * 2);
244 sbitmap_zero (bb_seen);
245 initialize_original_copy_tables ();
247 if (profile_info && flag_branch_probabilities)
248 probability_cutoff = PARAM_VALUE (TRACER_MIN_BRANCH_PROBABILITY_FEEDBACK);
250 probability_cutoff = PARAM_VALUE (TRACER_MIN_BRANCH_PROBABILITY);
251 probability_cutoff = REG_BR_PROB_BASE / 100 * probability_cutoff;
253 branch_ratio_cutoff =
254 (REG_BR_PROB_BASE / 100 * PARAM_VALUE (TRACER_MIN_BRANCH_RATIO));
258 int n = count_insns (bb);
259 if (!ignore_bb_p (bb))
260 blocks[bb->index] = fibheap_insert (heap, -bb->frequency,
263 counts [bb->index] = n;
265 weighted_insns += n * bb->frequency;
268 if (profile_info && flag_branch_probabilities)
269 cover_insns = PARAM_VALUE (TRACER_DYNAMIC_COVERAGE_FEEDBACK);
271 cover_insns = PARAM_VALUE (TRACER_DYNAMIC_COVERAGE);
272 cover_insns = (weighted_insns * cover_insns + 50) / 100;
273 max_dup_insns = (ninsns * PARAM_VALUE (TRACER_MAX_CODE_GROWTH) + 50) / 100;
275 while (traced_insns < cover_insns && nduplicated < max_dup_insns
276 && !fibheap_empty (heap))
278 basic_block bb = (basic_block) fibheap_extract_min (heap);
284 blocks[bb->index] = NULL;
286 if (ignore_bb_p (bb))
288 gcc_assert (!bb_seen_p (bb));
290 n = find_trace (bb, trace);
293 traced_insns += bb->frequency * counts [bb->index];
294 if (blocks[bb->index])
296 fibheap_delete_node (heap, blocks[bb->index]);
297 blocks[bb->index] = NULL;
300 for (pos = 1; pos < n; pos++)
302 basic_block bb2 = trace[pos];
304 if (blocks[bb2->index])
306 fibheap_delete_node (heap, blocks[bb2->index]);
307 blocks[bb2->index] = NULL;
309 traced_insns += bb2->frequency * counts [bb2->index];
310 if (EDGE_COUNT (bb2->preds) > 1
311 && can_duplicate_block_p (bb2)
312 /* We have the tendency to duplicate the loop header
313 of all do { } while loops. Do not do that - it is
314 not profitable and it might create a loop with multiple
315 entries or at least rotate the loop. */
317 || bb2->loop_father->header != bb2))
322 nduplicated += counts [bb2->index];
324 e = find_edge (bb, bb2);
326 copy = duplicate_block (bb2, e, bb);
327 flush_pending_stmts (e);
329 add_phi_args_after_copy (©, 1, NULL);
331 /* Reconsider the original copy of block we've duplicated.
332 Removing the most common predecessor may make it to be
335 fibheap_insert (heap, -bb2->frequency, bb2);
338 fprintf (dump_file, "Duplicated %i as %i [%i]\n",
339 bb2->index, copy->index, copy->frequency);
346 /* In case the trace became infrequent, stop duplicating. */
347 if (ignore_bb_p (bb))
351 fprintf (dump_file, " covered now %.1f\n\n",
352 traced_insns * 100.0 / weighted_insns);
355 fprintf (dump_file, "Duplicated %i insns (%i%%)\n", nduplicated,
356 nduplicated * 100 / ninsns);
358 free_original_copy_tables ();
359 sbitmap_free (bb_seen);
363 fibheap_delete (heap);
368 /* Main entry point to this file. */
375 gcc_assert (current_ir_type () == IR_GIMPLE);
377 if (n_basic_blocks <= NUM_FIXED_BLOCKS + 1)
380 mark_dfs_back_edges ();
382 dump_flow_info (dump_file, dump_flags);
384 /* Trace formation is done on the fly inside tail_duplicate */
385 changed = tail_duplicate ();
387 free_dominance_info (CDI_DOMINATORS);
390 dump_flow_info (dump_file, dump_flags);
392 return changed ? TODO_cleanup_cfg : 0;
398 return (optimize > 0 && flag_tracer && flag_reorder_blocks);
401 struct gimple_opt_pass pass_tracer =
406 gate_tracer, /* gate */
407 tracer, /* execute */
410 0, /* static_pass_number */
411 TV_TRACER, /* tv_id */
412 0, /* properties_required */
413 0, /* properties_provided */
414 0, /* properties_destroyed */
415 0, /* todo_flags_start */
417 | TODO_verify_ssa /* todo_flags_finish */