]>
Commit | Line | Data |
---|---|---|
fa99ab3d | 1 | /* The tracer pass for the GNU compiler. |
2 | Contributed by Jan Hubicka, SuSE Labs. | |
3ddccd57 | 3 | Adapted to work on GIMPLE instead of RTL by Robert Kidd, UIUC. |
cfaf579d | 4 | Copyright (C) 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008 |
3072d30e | 5 | Free Software Foundation, Inc. |
fa99ab3d | 6 | |
7 | This file is part of GCC. | |
8 | ||
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 | |
8c4c00c1 | 11 | the Free Software Foundation; either version 3, or (at your option) |
fa99ab3d | 12 | any later version. |
13 | ||
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. | |
18 | ||
19 | You should have received a copy of the GNU General Public License | |
8c4c00c1 | 20 | along with GCC; see the file COPYING3. If not see |
21 | <http://www.gnu.org/licenses/>. */ | |
fa99ab3d | 22 | |
23 | /* This pass performs the tail duplication needed for superblock formation. | |
24 | For more information see: | |
25 | ||
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 | |
29 | ||
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 | |
34 | performed. */ | |
35 | ||
36 | ||
37 | #include "config.h" | |
38 | #include "system.h" | |
805e22b2 | 39 | #include "coretypes.h" |
40 | #include "tm.h" | |
fa99ab3d | 41 | #include "tree.h" |
42 | #include "rtl.h" | |
43 | #include "hard-reg-set.h" | |
44 | #include "basic-block.h" | |
fa99ab3d | 45 | #include "fibheap.h" |
46 | #include "flags.h" | |
e7f8f0eb | 47 | #include "timevar.h" |
fa99ab3d | 48 | #include "params.h" |
44359ced | 49 | #include "coverage.h" |
77fce4cd | 50 | #include "tree-pass.h" |
3ddccd57 | 51 | #include "tree-flow.h" |
52 | #include "tree-inline.h" | |
f889f544 | 53 | #include "cfgloop.h" |
fa99ab3d | 54 | |
3ddccd57 | 55 | static int count_insns (basic_block); |
7ecb5bb2 | 56 | static bool ignore_bb_p (const_basic_block); |
57 | static bool better_p (const_edge, const_edge); | |
60b8c5b3 | 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 *); | |
fa99ab3d | 61 | |
62 | /* Minimal outgoing edge probability considered for superblock formation. */ | |
63 | static int probability_cutoff; | |
64 | static int branch_ratio_cutoff; | |
65 | ||
3ddccd57 | 66 | /* A bit BB->index is set if BB has already been seen, i.e. it is |
67 | connected to some trace already. */ | |
68 | sbitmap bb_seen; | |
fa99ab3d | 69 | |
3ddccd57 | 70 | static inline void |
71 | mark_bb_seen (basic_block bb) | |
72 | { | |
73 | unsigned int size = SBITMAP_SIZE_BYTES (bb_seen) * 8; | |
74 | ||
75 | if ((unsigned int)bb->index >= size) | |
76 | bb_seen = sbitmap_resize (bb_seen, size * 2, 0); | |
77 | ||
78 | SET_BIT (bb_seen, bb->index); | |
79 | } | |
80 | ||
81 | static inline bool | |
82 | bb_seen_p (basic_block bb) | |
83 | { | |
84 | return TEST_BIT (bb_seen, bb->index); | |
85 | } | |
fa99ab3d | 86 | |
7299020b | 87 | /* Return true if we should ignore the basic block for purposes of tracing. */ |
fa99ab3d | 88 | static bool |
7ecb5bb2 | 89 | ignore_bb_p (const_basic_block bb) |
fa99ab3d | 90 | { |
09a1f31f | 91 | gimple g; |
92 | ||
4d2e5d52 | 93 | if (bb->index < NUM_FIXED_BLOCKS) |
fa99ab3d | 94 | return true; |
0bfd8d5c | 95 | if (optimize_bb_for_size_p (bb)) |
fa99ab3d | 96 | return true; |
09a1f31f | 97 | |
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) | |
102 | return true; | |
103 | ||
fa99ab3d | 104 | return false; |
105 | } | |
106 | ||
107 | /* Return number of instructions in the block. */ | |
108 | ||
109 | static int | |
3ddccd57 | 110 | count_insns (basic_block bb) |
fa99ab3d | 111 | { |
75a70cf9 | 112 | gimple_stmt_iterator gsi; |
113 | gimple stmt; | |
fa99ab3d | 114 | int n = 0; |
115 | ||
75a70cf9 | 116 | for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi)) |
3ddccd57 | 117 | { |
75a70cf9 | 118 | stmt = gsi_stmt (gsi); |
3ddccd57 | 119 | n += estimate_num_insns (stmt, &eni_size_weights); |
120 | } | |
fa99ab3d | 121 | return n; |
122 | } | |
123 | ||
124 | /* Return true if E1 is more frequent than E2. */ | |
125 | static bool | |
7ecb5bb2 | 126 | better_p (const_edge e1, const_edge e2) |
fa99ab3d | 127 | { |
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 | |
135 | CFG is modified. */ | |
136 | if (e1->src != e2->src) | |
137 | return e1->src->index > e2->src->index; | |
138 | return e1->dest->index > e2->dest->index; | |
139 | } | |
140 | ||
141 | /* Return most frequent successor of basic block BB. */ | |
142 | ||
143 | static edge | |
60b8c5b3 | 144 | find_best_successor (basic_block bb) |
fa99ab3d | 145 | { |
146 | edge e; | |
147 | edge best = NULL; | |
cd665a06 | 148 | edge_iterator ei; |
fa99ab3d | 149 | |
cd665a06 | 150 | FOR_EACH_EDGE (e, ei, bb->succs) |
fa99ab3d | 151 | if (!best || better_p (e, best)) |
152 | best = e; | |
153 | if (!best || ignore_bb_p (best->dest)) | |
154 | return NULL; | |
155 | if (best->probability <= probability_cutoff) | |
156 | return NULL; | |
157 | return best; | |
158 | } | |
159 | ||
160 | /* Return most frequent predecessor of basic block BB. */ | |
161 | ||
162 | static edge | |
60b8c5b3 | 163 | find_best_predecessor (basic_block bb) |
fa99ab3d | 164 | { |
165 | edge e; | |
166 | edge best = NULL; | |
cd665a06 | 167 | edge_iterator ei; |
fa99ab3d | 168 | |
cd665a06 | 169 | FOR_EACH_EDGE (e, ei, bb->preds) |
fa99ab3d | 170 | if (!best || better_p (e, best)) |
171 | best = e; | |
172 | if (!best || ignore_bb_p (best->src)) | |
173 | return NULL; | |
174 | if (EDGE_FREQUENCY (best) * REG_BR_PROB_BASE | |
175 | < bb->frequency * branch_ratio_cutoff) | |
176 | return NULL; | |
177 | return best; | |
178 | } | |
179 | ||
180 | /* Find the trace using bb and record it in the TRACE array. | |
181 | Return number of basic blocks recorded. */ | |
182 | ||
183 | static int | |
60b8c5b3 | 184 | find_trace (basic_block bb, basic_block *trace) |
fa99ab3d | 185 | { |
186 | int i = 0; | |
187 | edge e; | |
188 | ||
450d042a | 189 | if (dump_file) |
190 | fprintf (dump_file, "Trace seed %i [%i]", bb->index, bb->frequency); | |
fa99ab3d | 191 | |
192 | while ((e = find_best_predecessor (bb)) != NULL) | |
193 | { | |
194 | basic_block bb2 = e->src; | |
3ddccd57 | 195 | if (bb_seen_p (bb2) || (e->flags & (EDGE_DFS_BACK | EDGE_COMPLEX)) |
fa99ab3d | 196 | || find_best_successor (bb2) != e) |
197 | break; | |
450d042a | 198 | if (dump_file) |
199 | fprintf (dump_file, ",%i [%i]", bb->index, bb->frequency); | |
fa99ab3d | 200 | bb = bb2; |
201 | } | |
450d042a | 202 | if (dump_file) |
203 | fprintf (dump_file, " forward %i [%i]", bb->index, bb->frequency); | |
fa99ab3d | 204 | trace[i++] = bb; |
205 | ||
206 | /* Follow the trace in forward direction. */ | |
207 | while ((e = find_best_successor (bb)) != NULL) | |
208 | { | |
209 | bb = e->dest; | |
3ddccd57 | 210 | if (bb_seen_p (bb) || (e->flags & (EDGE_DFS_BACK | EDGE_COMPLEX)) |
fa99ab3d | 211 | || find_best_predecessor (bb) != e) |
212 | break; | |
450d042a | 213 | if (dump_file) |
214 | fprintf (dump_file, ",%i [%i]", bb->index, bb->frequency); | |
fa99ab3d | 215 | trace[i++] = bb; |
216 | } | |
450d042a | 217 | if (dump_file) |
218 | fprintf (dump_file, "\n"); | |
fa99ab3d | 219 | return i; |
220 | } | |
221 | ||
222 | /* Look for basic blocks in frequency order, construct traces and tail duplicate | |
223 | if profitable. */ | |
224 | ||
35c67c83 | 225 | static bool |
60b8c5b3 | 226 | tail_duplicate (void) |
fa99ab3d | 227 | { |
4c36ffe6 | 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); | |
fa99ab3d | 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; | |
235 | int max_dup_insns; | |
236 | basic_block bb; | |
35c67c83 | 237 | bool changed = false; |
fa99ab3d | 238 | |
3ddccd57 | 239 | /* Create an oversized sbitmap to reduce the chance that we need to |
240 | resize it. */ | |
241 | bb_seen = sbitmap_alloc (last_basic_block * 2); | |
242 | sbitmap_zero (bb_seen); | |
243 | initialize_original_copy_tables (); | |
244 | ||
ab6a34f2 | 245 | if (profile_info && flag_branch_probabilities) |
fa99ab3d | 246 | probability_cutoff = PARAM_VALUE (TRACER_MIN_BRANCH_PROBABILITY_FEEDBACK); |
247 | else | |
248 | probability_cutoff = PARAM_VALUE (TRACER_MIN_BRANCH_PROBABILITY); | |
249 | probability_cutoff = REG_BR_PROB_BASE / 100 * probability_cutoff; | |
250 | ||
251 | branch_ratio_cutoff = | |
252 | (REG_BR_PROB_BASE / 100 * PARAM_VALUE (TRACER_MIN_BRANCH_RATIO)); | |
253 | ||
254 | FOR_EACH_BB (bb) | |
255 | { | |
256 | int n = count_insns (bb); | |
257 | if (!ignore_bb_p (bb)) | |
258 | blocks[bb->index] = fibheap_insert (heap, -bb->frequency, | |
259 | bb); | |
260 | ||
261 | counts [bb->index] = n; | |
262 | ninsns += n; | |
263 | weighted_insns += n * bb->frequency; | |
264 | } | |
265 | ||
ab6a34f2 | 266 | if (profile_info && flag_branch_probabilities) |
fa99ab3d | 267 | cover_insns = PARAM_VALUE (TRACER_DYNAMIC_COVERAGE_FEEDBACK); |
268 | else | |
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; | |
272 | ||
273 | while (traced_insns < cover_insns && nduplicated < max_dup_insns | |
274 | && !fibheap_empty (heap)) | |
275 | { | |
45ba1503 | 276 | basic_block bb = (basic_block) fibheap_extract_min (heap); |
fa99ab3d | 277 | int n, pos; |
278 | ||
279 | if (!bb) | |
280 | break; | |
281 | ||
282 | blocks[bb->index] = NULL; | |
283 | ||
284 | if (ignore_bb_p (bb)) | |
285 | continue; | |
3ddccd57 | 286 | gcc_assert (!bb_seen_p (bb)); |
fa99ab3d | 287 | |
288 | n = find_trace (bb, trace); | |
289 | ||
290 | bb = trace[0]; | |
291 | traced_insns += bb->frequency * counts [bb->index]; | |
292 | if (blocks[bb->index]) | |
293 | { | |
294 | fibheap_delete_node (heap, blocks[bb->index]); | |
295 | blocks[bb->index] = NULL; | |
296 | } | |
297 | ||
298 | for (pos = 1; pos < n; pos++) | |
299 | { | |
300 | basic_block bb2 = trace[pos]; | |
301 | ||
302 | if (blocks[bb2->index]) | |
303 | { | |
304 | fibheap_delete_node (heap, blocks[bb2->index]); | |
305 | blocks[bb2->index] = NULL; | |
306 | } | |
307 | traced_insns += bb2->frequency * counts [bb2->index]; | |
cd665a06 | 308 | if (EDGE_COUNT (bb2->preds) > 1 |
f889f544 | 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. */ | |
314 | && (!current_loops | |
315 | || bb2->loop_father->header != bb2)) | |
fa99ab3d | 316 | { |
cd665a06 | 317 | edge e; |
3ddccd57 | 318 | basic_block copy; |
319 | ||
320 | nduplicated += counts [bb2->index]; | |
fa99ab3d | 321 | |
c6356c17 | 322 | e = find_edge (bb, bb2); |
48e1416a | 323 | |
3ddccd57 | 324 | copy = duplicate_block (bb2, e, bb); |
325 | flush_pending_stmts (e); | |
cd665a06 | 326 | |
28c92cbb | 327 | add_phi_args_after_copy (©, 1, NULL); |
fa99ab3d | 328 | |
329 | /* Reconsider the original copy of block we've duplicated. | |
de132707 | 330 | Removing the most common predecessor may make it to be |
fa99ab3d | 331 | head. */ |
3ddccd57 | 332 | blocks[bb2->index] = |
333 | fibheap_insert (heap, -bb2->frequency, bb2); | |
fa99ab3d | 334 | |
450d042a | 335 | if (dump_file) |
336 | fprintf (dump_file, "Duplicated %i as %i [%i]\n", | |
3ddccd57 | 337 | bb2->index, copy->index, copy->frequency); |
338 | ||
339 | bb2 = copy; | |
35c67c83 | 340 | changed = true; |
fa99ab3d | 341 | } |
3ddccd57 | 342 | mark_bb_seen (bb2); |
fa99ab3d | 343 | bb = bb2; |
344 | /* In case the trace became infrequent, stop duplicating. */ | |
345 | if (ignore_bb_p (bb)) | |
346 | break; | |
347 | } | |
450d042a | 348 | if (dump_file) |
349 | fprintf (dump_file, " covered now %.1f\n\n", | |
fa99ab3d | 350 | traced_insns * 100.0 / weighted_insns); |
351 | } | |
450d042a | 352 | if (dump_file) |
353 | fprintf (dump_file, "Duplicated %i insns (%i%%)\n", nduplicated, | |
fa99ab3d | 354 | nduplicated * 100 / ninsns); |
355 | ||
3ddccd57 | 356 | free_original_copy_tables (); |
357 | sbitmap_free (bb_seen); | |
fa99ab3d | 358 | free (blocks); |
359 | free (trace); | |
360 | free (counts); | |
361 | fibheap_delete (heap); | |
35c67c83 | 362 | |
363 | return changed; | |
fa99ab3d | 364 | } |
365 | ||
207c7ab2 | 366 | /* Main entry point to this file. */ |
fa99ab3d | 367 | |
3ddccd57 | 368 | static unsigned int |
207c7ab2 | 369 | tracer (void) |
fa99ab3d | 370 | { |
35c67c83 | 371 | bool changed; |
372 | ||
4d2e5d52 | 373 | if (n_basic_blocks <= NUM_FIXED_BLOCKS + 1) |
3ddccd57 | 374 | return 0; |
e7f8f0eb | 375 | |
fa99ab3d | 376 | mark_dfs_back_edges (); |
450d042a | 377 | if (dump_file) |
4a020a8c | 378 | brief_dump_cfg (dump_file); |
3ddccd57 | 379 | |
380 | /* Trace formation is done on the fly inside tail_duplicate */ | |
35c67c83 | 381 | changed = tail_duplicate (); |
382 | if (changed) | |
383 | free_dominance_info (CDI_DOMINATORS); | |
3ddccd57 | 384 | |
450d042a | 385 | if (dump_file) |
4a020a8c | 386 | brief_dump_cfg (dump_file); |
e7f8f0eb | 387 | |
35c67c83 | 388 | return changed ? TODO_cleanup_cfg : 0; |
77fce4cd | 389 | } |
390 | \f | |
391 | static bool | |
3ddccd57 | 392 | gate_tracer (void) |
77fce4cd | 393 | { |
3ddccd57 | 394 | return (optimize > 0 && flag_tracer && flag_reorder_blocks); |
fa99ab3d | 395 | } |
77fce4cd | 396 | |
20099e35 | 397 | struct gimple_opt_pass pass_tracer = |
77fce4cd | 398 | { |
20099e35 | 399 | { |
400 | GIMPLE_PASS, | |
77fce4cd | 401 | "tracer", /* name */ |
3ddccd57 | 402 | gate_tracer, /* gate */ |
403 | tracer, /* execute */ | |
77fce4cd | 404 | NULL, /* sub */ |
405 | NULL, /* next */ | |
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 */ | |
771e2890 | 412 | TODO_update_ssa |
20099e35 | 413 | | TODO_verify_ssa /* todo_flags_finish */ |
414 | } | |
77fce4cd | 415 | }; |