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) 2001-2021 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"
47 #include "gimple-iterator.h"
50 #include "tree-inline.h"
52 #include "alloc-pool.h"
53 #include "fibonacci_heap.h"
56 static void analyze_bb (basic_block, int *);
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;
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 static sbitmap can_duplicate_bb;
89 /* Cache VAL as value of can_duplicate_bb_p for BB. */
91 cache_can_duplicate_bb_p (const_basic_block bb, bool val)
94 bitmap_set_bit (can_duplicate_bb, bb->index);
97 /* Return cached value of can_duplicate_bb_p for BB. */
99 cached_can_duplicate_bb_p (const_basic_block bb)
101 if (can_duplicate_bb)
103 unsigned int size = SBITMAP_SIZE (can_duplicate_bb);
104 if ((unsigned int)bb->index < size)
105 return bitmap_bit_p (can_duplicate_bb, bb->index);
107 /* Assume added bb's should not be duplicated. */
111 return can_duplicate_block_p (bb);
114 /* Return true if we should ignore the basic block for purposes of tracing. */
116 ignore_bb_p (const_basic_block bb)
118 if (bb->index < NUM_FIXED_BLOCKS)
120 if (optimize_bb_for_size_p (bb))
123 return !cached_can_duplicate_bb_p (bb);
126 /* Return number of instructions in the block. */
129 analyze_bb (basic_block bb, int *count)
131 gimple_stmt_iterator gsi;
135 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
137 stmt = gsi_stmt (gsi);
138 n += estimate_num_insns (stmt, &eni_size_weights);
142 cache_can_duplicate_bb_p (bb, can_duplicate_block_p (CONST_CAST_BB (bb)));
145 /* Return true if E1 is more frequent than E2. */
147 better_p (const_edge e1, const_edge e2)
149 if ((e1->count () > e2->count ()) || (e1->count () < e2->count ()))
150 return e1->count () > e2->count ();
151 /* This is needed to avoid changes in the decision after
153 if (e1->src != e2->src)
154 return e1->src->index > e2->src->index;
155 return e1->dest->index > e2->dest->index;
158 /* Return most frequent successor of basic block BB. */
161 find_best_successor (basic_block bb)
167 FOR_EACH_EDGE (e, ei, bb->succs)
169 if (!e->count ().initialized_p ())
171 if (!best || better_p (e, best))
174 if (!best || ignore_bb_p (best->dest))
176 if (!best->probability.initialized_p ()
177 || best->probability.to_reg_br_prob_base () <= 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)
193 if (!e->count ().initialized_p ())
195 if (!best || better_p (e, best))
198 if (!best || ignore_bb_p (best->src))
200 if (bb->count.initialized_p ()
201 && (best->count ().to_frequency (cfun) * REG_BR_PROB_BASE
202 < bb->count.to_frequency (cfun) * branch_ratio_cutoff))
207 /* Find the trace using bb and record it in the TRACE array.
208 Return number of basic blocks recorded. */
211 find_trace (basic_block bb, basic_block *trace)
217 fprintf (dump_file, "Trace seed %i [%i]", bb->index, bb->count.to_frequency (cfun));
219 while ((e = find_best_predecessor (bb)) != NULL)
221 basic_block bb2 = e->src;
222 if (bb_seen_p (bb2) || (e->flags & (EDGE_DFS_BACK | EDGE_COMPLEX))
223 || find_best_successor (bb2) != e)
226 fprintf (dump_file, ",%i [%i]", bb->index, bb->count.to_frequency (cfun));
230 fprintf (dump_file, " forward %i [%i]", bb->index, bb->count.to_frequency (cfun));
233 /* Follow the trace in forward direction. */
234 while ((e = find_best_successor (bb)) != NULL)
237 if (bb_seen_p (bb) || (e->flags & (EDGE_DFS_BACK | EDGE_COMPLEX))
238 || find_best_predecessor (bb) != e)
241 fprintf (dump_file, ",%i [%i]", bb->index, bb->count.to_frequency (cfun));
245 fprintf (dump_file, "\n");
249 /* Duplicate block BB2, placing it after BB in the CFG. Return the
250 newly created block. */
252 transform_duplicate (basic_block bb, basic_block bb2)
257 e = find_edge (bb, bb2);
259 copy = duplicate_block (bb2, e, bb);
260 flush_pending_stmts (e);
262 add_phi_args_after_copy (©, 1, NULL);
267 /* Look for basic blocks in frequency order, construct traces and tail duplicate
271 tail_duplicate (void)
273 auto_vec<fibonacci_node<long, basic_block_def>*> blocks;
274 blocks.safe_grow_cleared (last_basic_block_for_fn (cfun), true);
276 basic_block *trace = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun));
277 int *counts = XNEWVEC (int, last_basic_block_for_fn (cfun));
278 int ninsns = 0, nduplicated = 0;
279 gcov_type weighted_insns = 0, traced_insns = 0;
280 fibonacci_heap<long, basic_block_def> heap (LONG_MIN);
281 gcov_type cover_insns;
284 bool changed = false;
286 /* Create an oversized sbitmap to reduce the chance that we need to
288 bb_seen = sbitmap_alloc (last_basic_block_for_fn (cfun) * 2);
289 bitmap_clear (bb_seen);
290 can_duplicate_bb = sbitmap_alloc (last_basic_block_for_fn (cfun));
291 bitmap_clear (can_duplicate_bb);
292 initialize_original_copy_tables ();
294 if (profile_info && profile_status_for_fn (cfun) == PROFILE_READ)
295 probability_cutoff = param_tracer_min_branch_probability_feedback;
297 probability_cutoff = param_tracer_min_branch_probability;
298 probability_cutoff = REG_BR_PROB_BASE / 100 * probability_cutoff;
300 branch_ratio_cutoff =
301 (REG_BR_PROB_BASE / 100 * param_tracer_min_branch_ratio);
303 FOR_EACH_BB_FN (bb, cfun)
307 if (!ignore_bb_p (bb))
308 blocks[bb->index] = heap.insert (-bb->count.to_frequency (cfun), bb);
310 counts [bb->index] = n;
312 weighted_insns += n * bb->count.to_frequency (cfun);
315 if (profile_info && profile_status_for_fn (cfun) == PROFILE_READ)
316 cover_insns = param_tracer_dynamic_coverage_feedback;
318 cover_insns = param_tracer_dynamic_coverage;
319 cover_insns = (weighted_insns * cover_insns + 50) / 100;
320 max_dup_insns = (ninsns * param_tracer_max_code_growth + 50) / 100;
322 while (traced_insns < cover_insns && nduplicated < max_dup_insns
325 basic_block bb = heap.extract_min ();
331 blocks[bb->index] = NULL;
333 if (ignore_bb_p (bb))
335 gcc_assert (!bb_seen_p (bb));
337 n = find_trace (bb, trace);
340 traced_insns += bb->count.to_frequency (cfun) * counts [bb->index];
341 if (blocks[bb->index])
343 heap.delete_node (blocks[bb->index]);
344 blocks[bb->index] = NULL;
347 for (pos = 1; pos < n; pos++)
349 basic_block bb2 = trace[pos];
351 if (blocks[bb2->index])
353 heap.delete_node (blocks[bb2->index]);
354 blocks[bb2->index] = NULL;
356 traced_insns += bb2->count.to_frequency (cfun) * counts [bb2->index];
357 if (EDGE_COUNT (bb2->preds) > 1
358 && can_duplicate_block_p (bb2)
359 /* We have the tendency to duplicate the loop header
360 of all do { } while loops. Do not do that - it is
361 not profitable and it might create a loop with multiple
362 entries or at least rotate the loop. */
363 && bb2->loop_father->header != bb2)
365 nduplicated += counts [bb2->index];
366 basic_block copy = transform_duplicate (bb, bb2);
368 /* Reconsider the original copy of block we've duplicated.
369 Removing the most common predecessor may make it to be
371 blocks[bb2->index] = heap.insert (-bb2->count.to_frequency (cfun), bb2);
374 fprintf (dump_file, "Duplicated %i as %i [%i]\n",
375 bb2->index, copy->index, copy->count.to_frequency (cfun));
382 /* In case the trace became infrequent, stop duplicating. */
383 if (ignore_bb_p (bb))
387 fprintf (dump_file, " covered now %.1f\n\n",
388 traced_insns * 100.0 / weighted_insns);
391 fprintf (dump_file, "Duplicated %i insns (%i%%)\n", nduplicated,
392 nduplicated * 100 / ninsns);
394 free_original_copy_tables ();
395 sbitmap_free (bb_seen);
396 sbitmap_free (can_duplicate_bb);
397 can_duplicate_bb = NULL;
406 const pass_data pass_data_tracer =
408 GIMPLE_PASS, /* type */
410 OPTGROUP_NONE, /* optinfo_flags */
411 TV_TRACER, /* tv_id */
412 0, /* properties_required */
413 0, /* properties_provided */
414 0, /* properties_destroyed */
415 0, /* todo_flags_start */
416 TODO_update_ssa, /* todo_flags_finish */
419 class pass_tracer : public gimple_opt_pass
422 pass_tracer (gcc::context *ctxt)
423 : gimple_opt_pass (pass_data_tracer, ctxt)
426 /* opt_pass methods: */
427 virtual bool gate (function *)
429 return (optimize > 0 && flag_tracer && flag_reorder_blocks);
432 virtual unsigned int execute (function *);
434 }; // class pass_tracer
437 pass_tracer::execute (function *fun)
441 if (n_basic_blocks_for_fn (fun) <= NUM_FIXED_BLOCKS + 1)
444 mark_dfs_back_edges ();
446 brief_dump_cfg (dump_file, dump_flags);
448 /* Trace formation is done on the fly inside tail_duplicate */
449 changed = tail_duplicate ();
452 free_dominance_info (CDI_DOMINATORS);
453 /* If we changed the CFG schedule loops for fixup by cleanup_cfg. */
454 loops_state_set (LOOPS_NEED_FIXUP);
458 brief_dump_cfg (dump_file, dump_flags);
460 return changed ? TODO_cleanup_cfg : 0;
465 make_pass_tracer (gcc::context *ctxt)
467 return new pass_tracer (ctxt);