Daily bump.
[official-gcc.git] / gcc / tracer.c
blob05300a289eacfbe3da8d61dce67c1bff58e5b695
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) 2001-2018 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)
11 any later version.
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
33 performed. */
36 #include "config.h"
37 #include "system.h"
38 #include "coretypes.h"
39 #include "backend.h"
40 #include "rtl.h"
41 #include "tree.h"
42 #include "gimple.h"
43 #include "cfghooks.h"
44 #include "tree-pass.h"
45 #include "profile.h"
46 #include "cfganal.h"
47 #include "params.h"
48 #include "gimple-iterator.h"
49 #include "tree-cfg.h"
50 #include "tree-ssa.h"
51 #include "tree-inline.h"
52 #include "cfgloop.h"
53 #include "fibonacci_heap.h"
54 #include "tracer.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. */
68 static sbitmap bb_seen;
70 static inline void
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);
81 static inline bool
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. */
88 bool
89 ignore_bb_p (const_basic_block bb)
91 if (bb->index < NUM_FIXED_BLOCKS)
92 return true;
93 if (optimize_bb_for_size_p (bb))
94 return true;
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)
101 return true;
103 /* An IFN_UNIQUE call must be duplicated as part of its group,
104 or not at all. */
105 if (is_gimple_call (g)
106 && gimple_call_internal_p (g)
107 && gimple_call_internal_unique_p (g))
108 return true;
111 return false;
114 /* Return number of instructions in the block. */
116 static int
117 count_insns (basic_block bb)
119 gimple_stmt_iterator gsi;
120 gimple *stmt;
121 int n = 0;
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);
128 return n;
131 /* Return true if E1 is more frequent than E2. */
132 static bool
133 better_p (const_edge e1, const_edge e2)
135 if ((e1->count () > e2->count ()) || (e1->count () < e2->count ()))
136 return e1->count () > e2->count ();
137 /* This is needed to avoid changes in the decision after
138 CFG is modified. */
139 if (e1->src != e2->src)
140 return e1->src->index > e2->src->index;
141 return e1->dest->index > e2->dest->index;
144 /* Return most frequent successor of basic block BB. */
146 static edge
147 find_best_successor (basic_block bb)
149 edge e;
150 edge best = NULL;
151 edge_iterator ei;
153 FOR_EACH_EDGE (e, ei, bb->succs)
155 if (!e->count ().initialized_p ())
156 return NULL;
157 if (!best || better_p (e, best))
158 best = e;
160 if (!best || ignore_bb_p (best->dest))
161 return NULL;
162 if (!best->probability.initialized_p ()
163 || best->probability.to_reg_br_prob_base () <= probability_cutoff)
164 return NULL;
165 return best;
168 /* Return most frequent predecessor of basic block BB. */
170 static edge
171 find_best_predecessor (basic_block bb)
173 edge e;
174 edge best = NULL;
175 edge_iterator ei;
177 FOR_EACH_EDGE (e, ei, bb->preds)
179 if (!e->count ().initialized_p ())
180 return NULL;
181 if (!best || better_p (e, best))
182 best = e;
184 if (!best || ignore_bb_p (best->src))
185 return NULL;
186 if (bb->count.initialized_p ()
187 && (best->count ().to_frequency (cfun) * REG_BR_PROB_BASE
188 < bb->count.to_frequency (cfun) * branch_ratio_cutoff))
189 return NULL;
190 return best;
193 /* Find the trace using bb and record it in the TRACE array.
194 Return number of basic blocks recorded. */
196 static int
197 find_trace (basic_block bb, basic_block *trace)
199 int i = 0;
200 edge e;
202 if (dump_file)
203 fprintf (dump_file, "Trace seed %i [%i]", bb->index, bb->count.to_frequency (cfun));
205 while ((e = find_best_predecessor (bb)) != NULL)
207 basic_block bb2 = e->src;
208 if (bb_seen_p (bb2) || (e->flags & (EDGE_DFS_BACK | EDGE_COMPLEX))
209 || find_best_successor (bb2) != e)
210 break;
211 if (dump_file)
212 fprintf (dump_file, ",%i [%i]", bb->index, bb->count.to_frequency (cfun));
213 bb = bb2;
215 if (dump_file)
216 fprintf (dump_file, " forward %i [%i]", bb->index, bb->count.to_frequency (cfun));
217 trace[i++] = bb;
219 /* Follow the trace in forward direction. */
220 while ((e = find_best_successor (bb)) != NULL)
222 bb = e->dest;
223 if (bb_seen_p (bb) || (e->flags & (EDGE_DFS_BACK | EDGE_COMPLEX))
224 || find_best_predecessor (bb) != e)
225 break;
226 if (dump_file)
227 fprintf (dump_file, ",%i [%i]", bb->index, bb->count.to_frequency (cfun));
228 trace[i++] = bb;
230 if (dump_file)
231 fprintf (dump_file, "\n");
232 return i;
235 /* Duplicate block BB2, placing it after BB in the CFG. Return the
236 newly created block. */
237 basic_block
238 transform_duplicate (basic_block bb, basic_block bb2)
240 edge e;
241 basic_block copy;
243 e = find_edge (bb, bb2);
245 copy = duplicate_block (bb2, e, bb);
246 flush_pending_stmts (e);
248 add_phi_args_after_copy (&copy, 1, NULL);
250 return (copy);
253 /* Look for basic blocks in frequency order, construct traces and tail duplicate
254 if profitable. */
256 static bool
257 tail_duplicate (void)
259 auto_vec<fibonacci_node<long, basic_block_def>*> blocks;
260 blocks.safe_grow_cleared (last_basic_block_for_fn (cfun));
262 basic_block *trace = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun));
263 int *counts = XNEWVEC (int, last_basic_block_for_fn (cfun));
264 int ninsns = 0, nduplicated = 0;
265 gcov_type weighted_insns = 0, traced_insns = 0;
266 fibonacci_heap<long, basic_block_def> heap (LONG_MIN);
267 gcov_type cover_insns;
268 int max_dup_insns;
269 basic_block bb;
270 bool changed = false;
272 /* Create an oversized sbitmap to reduce the chance that we need to
273 resize it. */
274 bb_seen = sbitmap_alloc (last_basic_block_for_fn (cfun) * 2);
275 bitmap_clear (bb_seen);
276 initialize_original_copy_tables ();
278 if (profile_info && profile_status_for_fn (cfun) == PROFILE_READ)
279 probability_cutoff = PARAM_VALUE (TRACER_MIN_BRANCH_PROBABILITY_FEEDBACK);
280 else
281 probability_cutoff = PARAM_VALUE (TRACER_MIN_BRANCH_PROBABILITY);
282 probability_cutoff = REG_BR_PROB_BASE / 100 * probability_cutoff;
284 branch_ratio_cutoff =
285 (REG_BR_PROB_BASE / 100 * PARAM_VALUE (TRACER_MIN_BRANCH_RATIO));
287 FOR_EACH_BB_FN (bb, cfun)
289 int n = count_insns (bb);
290 if (!ignore_bb_p (bb))
291 blocks[bb->index] = heap.insert (-bb->count.to_frequency (cfun), bb);
293 counts [bb->index] = n;
294 ninsns += n;
295 weighted_insns += n * bb->count.to_frequency (cfun);
298 if (profile_info && profile_status_for_fn (cfun) == PROFILE_READ)
299 cover_insns = PARAM_VALUE (TRACER_DYNAMIC_COVERAGE_FEEDBACK);
300 else
301 cover_insns = PARAM_VALUE (TRACER_DYNAMIC_COVERAGE);
302 cover_insns = (weighted_insns * cover_insns + 50) / 100;
303 max_dup_insns = (ninsns * PARAM_VALUE (TRACER_MAX_CODE_GROWTH) + 50) / 100;
305 while (traced_insns < cover_insns && nduplicated < max_dup_insns
306 && !heap.empty ())
308 basic_block bb = heap.extract_min ();
309 int n, pos;
311 if (!bb)
312 break;
314 blocks[bb->index] = NULL;
316 if (ignore_bb_p (bb))
317 continue;
318 gcc_assert (!bb_seen_p (bb));
320 n = find_trace (bb, trace);
322 bb = trace[0];
323 traced_insns += bb->count.to_frequency (cfun) * counts [bb->index];
324 if (blocks[bb->index])
326 heap.delete_node (blocks[bb->index]);
327 blocks[bb->index] = NULL;
330 for (pos = 1; pos < n; pos++)
332 basic_block bb2 = trace[pos];
334 if (blocks[bb2->index])
336 heap.delete_node (blocks[bb2->index]);
337 blocks[bb2->index] = NULL;
339 traced_insns += bb2->count.to_frequency (cfun) * counts [bb2->index];
340 if (EDGE_COUNT (bb2->preds) > 1
341 && can_duplicate_block_p (bb2)
342 /* We have the tendency to duplicate the loop header
343 of all do { } while loops. Do not do that - it is
344 not profitable and it might create a loop with multiple
345 entries or at least rotate the loop. */
346 && bb2->loop_father->header != bb2)
348 nduplicated += counts [bb2->index];
349 basic_block copy = transform_duplicate (bb, bb2);
351 /* Reconsider the original copy of block we've duplicated.
352 Removing the most common predecessor may make it to be
353 head. */
354 blocks[bb2->index] = heap.insert (-bb2->count.to_frequency (cfun), bb2);
356 if (dump_file)
357 fprintf (dump_file, "Duplicated %i as %i [%i]\n",
358 bb2->index, copy->index, copy->count.to_frequency (cfun));
360 bb2 = copy;
361 changed = true;
363 mark_bb_seen (bb2);
364 bb = bb2;
365 /* In case the trace became infrequent, stop duplicating. */
366 if (ignore_bb_p (bb))
367 break;
369 if (dump_file)
370 fprintf (dump_file, " covered now %.1f\n\n",
371 traced_insns * 100.0 / weighted_insns);
373 if (dump_file)
374 fprintf (dump_file, "Duplicated %i insns (%i%%)\n", nduplicated,
375 nduplicated * 100 / ninsns);
377 free_original_copy_tables ();
378 sbitmap_free (bb_seen);
379 free (trace);
380 free (counts);
382 return changed;
385 namespace {
387 const pass_data pass_data_tracer =
389 GIMPLE_PASS, /* type */
390 "tracer", /* name */
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
402 public:
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
417 unsigned int
418 pass_tracer::execute (function *fun)
420 bool changed;
422 if (n_basic_blocks_for_fn (fun) <= NUM_FIXED_BLOCKS + 1)
423 return 0;
425 mark_dfs_back_edges ();
426 if (dump_file)
427 brief_dump_cfg (dump_file, dump_flags);
429 /* Trace formation is done on the fly inside tail_duplicate */
430 changed = tail_duplicate ();
431 if (changed)
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);
438 if (dump_file)
439 brief_dump_cfg (dump_file, dump_flags);
441 return changed ? TODO_cleanup_cfg : 0;
443 } // anon namespace
445 gimple_opt_pass *
446 make_pass_tracer (gcc::context *ctxt)
448 return new pass_tracer (ctxt);