target.h (globalize_decl_name): New.
[gcc.git] / gcc / cfgloopanal.c
1 /* Natural loop analysis code for GNU compiler.
2 Copyright (C) 2002, 2003, 2004, 2005, 2006 Free Software Foundation, Inc.
3
4 This file is part of GCC.
5
6 GCC is free software; you can redistribute it and/or modify it under
7 the terms of the GNU General Public License as published by the Free
8 Software Foundation; either version 2, or (at your option) any later
9 version.
10
11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 for more details.
15
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING. If not, write to the Free
18 Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
19 02110-1301, USA. */
20
21 #include "config.h"
22 #include "system.h"
23 #include "coretypes.h"
24 #include "tm.h"
25 #include "rtl.h"
26 #include "hard-reg-set.h"
27 #include "obstack.h"
28 #include "basic-block.h"
29 #include "cfgloop.h"
30 #include "expr.h"
31 #include "output.h"
32
33 /* Checks whether BB is executed exactly once in each LOOP iteration. */
34
35 bool
36 just_once_each_iteration_p (const struct loop *loop, basic_block bb)
37 {
38 /* It must be executed at least once each iteration. */
39 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb))
40 return false;
41
42 /* And just once. */
43 if (bb->loop_father != loop)
44 return false;
45
46 /* But this was not enough. We might have some irreducible loop here. */
47 if (bb->flags & BB_IRREDUCIBLE_LOOP)
48 return false;
49
50 return true;
51 }
52
53 /* Structure representing edge of a graph. */
54
55 struct edge
56 {
57 int src, dest; /* Source and destination. */
58 struct edge *pred_next, *succ_next;
59 /* Next edge in predecessor and successor lists. */
60 void *data; /* Data attached to the edge. */
61 };
62
63 /* Structure representing vertex of a graph. */
64
65 struct vertex
66 {
67 struct edge *pred, *succ;
68 /* Lists of predecessors and successors. */
69 int component; /* Number of dfs restarts before reaching the
70 vertex. */
71 int post; /* Postorder number. */
72 };
73
74 /* Structure representing a graph. */
75
76 struct graph
77 {
78 int n_vertices; /* Number of vertices. */
79 struct vertex *vertices;
80 /* The vertices. */
81 };
82
83 /* Dumps graph G into F. */
84
85 extern void dump_graph (FILE *, struct graph *);
86
87 void
88 dump_graph (FILE *f, struct graph *g)
89 {
90 int i;
91 struct edge *e;
92
93 for (i = 0; i < g->n_vertices; i++)
94 {
95 if (!g->vertices[i].pred
96 && !g->vertices[i].succ)
97 continue;
98
99 fprintf (f, "%d (%d)\t<-", i, g->vertices[i].component);
100 for (e = g->vertices[i].pred; e; e = e->pred_next)
101 fprintf (f, " %d", e->src);
102 fprintf (f, "\n");
103
104 fprintf (f, "\t->");
105 for (e = g->vertices[i].succ; e; e = e->succ_next)
106 fprintf (f, " %d", e->dest);
107 fprintf (f, "\n");
108 }
109 }
110
111 /* Creates a new graph with N_VERTICES vertices. */
112
113 static struct graph *
114 new_graph (int n_vertices)
115 {
116 struct graph *g = XNEW (struct graph);
117
118 g->n_vertices = n_vertices;
119 g->vertices = XCNEWVEC (struct vertex, n_vertices);
120
121 return g;
122 }
123
124 /* Adds an edge from F to T to graph G, with DATA attached. */
125
126 static void
127 add_edge (struct graph *g, int f, int t, void *data)
128 {
129 struct edge *e = xmalloc (sizeof (struct edge));
130
131 e->src = f;
132 e->dest = t;
133 e->data = data;
134
135 e->pred_next = g->vertices[t].pred;
136 g->vertices[t].pred = e;
137
138 e->succ_next = g->vertices[f].succ;
139 g->vertices[f].succ = e;
140 }
141
142 /* Runs dfs search over vertices of G, from NQ vertices in queue QS.
143 The vertices in postorder are stored into QT. If FORWARD is false,
144 backward dfs is run. */
145
146 static void
147 dfs (struct graph *g, int *qs, int nq, int *qt, bool forward)
148 {
149 int i, tick = 0, v, comp = 0, top;
150 struct edge *e;
151 struct edge **stack = xmalloc (sizeof (struct edge *) * g->n_vertices);
152
153 for (i = 0; i < g->n_vertices; i++)
154 {
155 g->vertices[i].component = -1;
156 g->vertices[i].post = -1;
157 }
158
159 #define FST_EDGE(V) (forward ? g->vertices[(V)].succ : g->vertices[(V)].pred)
160 #define NEXT_EDGE(E) (forward ? (E)->succ_next : (E)->pred_next)
161 #define EDGE_SRC(E) (forward ? (E)->src : (E)->dest)
162 #define EDGE_DEST(E) (forward ? (E)->dest : (E)->src)
163
164 for (i = 0; i < nq; i++)
165 {
166 v = qs[i];
167 if (g->vertices[v].post != -1)
168 continue;
169
170 g->vertices[v].component = comp++;
171 e = FST_EDGE (v);
172 top = 0;
173
174 while (1)
175 {
176 while (e && g->vertices[EDGE_DEST (e)].component != -1)
177 e = NEXT_EDGE (e);
178
179 if (!e)
180 {
181 if (qt)
182 qt[tick] = v;
183 g->vertices[v].post = tick++;
184
185 if (!top)
186 break;
187
188 e = stack[--top];
189 v = EDGE_SRC (e);
190 e = NEXT_EDGE (e);
191 continue;
192 }
193
194 stack[top++] = e;
195 v = EDGE_DEST (e);
196 e = FST_EDGE (v);
197 g->vertices[v].component = comp - 1;
198 }
199 }
200
201 free (stack);
202 }
203
204 /* Marks the edge E in graph G irreducible if it connects two vertices in the
205 same scc. */
206
207 static void
208 check_irred (struct graph *g, struct edge *e)
209 {
210 edge real = e->data;
211
212 /* All edges should lead from a component with higher number to the
213 one with lower one. */
214 gcc_assert (g->vertices[e->src].component >= g->vertices[e->dest].component);
215
216 if (g->vertices[e->src].component != g->vertices[e->dest].component)
217 return;
218
219 real->flags |= EDGE_IRREDUCIBLE_LOOP;
220 if (flow_bb_inside_loop_p (real->src->loop_father, real->dest))
221 real->src->flags |= BB_IRREDUCIBLE_LOOP;
222 }
223
224 /* Runs CALLBACK for all edges in G. */
225
226 static void
227 for_each_edge (struct graph *g,
228 void (callback) (struct graph *, struct edge *))
229 {
230 struct edge *e;
231 int i;
232
233 for (i = 0; i < g->n_vertices; i++)
234 for (e = g->vertices[i].succ; e; e = e->succ_next)
235 callback (g, e);
236 }
237
238 /* Releases the memory occupied by G. */
239
240 static void
241 free_graph (struct graph *g)
242 {
243 struct edge *e, *n;
244 int i;
245
246 for (i = 0; i < g->n_vertices; i++)
247 for (e = g->vertices[i].succ; e; e = n)
248 {
249 n = e->succ_next;
250 free (e);
251 }
252 free (g->vertices);
253 free (g);
254 }
255
256 /* Marks blocks and edges that are part of non-recognized loops; i.e. we
257 throw away all latch edges and mark blocks inside any remaining cycle.
258 Everything is a bit complicated due to fact we do not want to do this
259 for parts of cycles that only "pass" through some loop -- i.e. for
260 each cycle, we want to mark blocks that belong directly to innermost
261 loop containing the whole cycle.
262
263 LOOPS is the loop tree. */
264
265 #define LOOP_REPR(LOOP) ((LOOP)->num + last_basic_block)
266 #define BB_REPR(BB) ((BB)->index + 1)
267
268 void
269 mark_irreducible_loops (void)
270 {
271 basic_block act;
272 edge e;
273 edge_iterator ei;
274 int i, src, dest;
275 struct graph *g;
276 int num = current_loops ? number_of_loops () : 1;
277 int *queue1 = XNEWVEC (int, last_basic_block + num);
278 int *queue2 = XNEWVEC (int, last_basic_block + num);
279 int nq, depth;
280 struct loop *cloop, *loop;
281 loop_iterator li;
282
283 /* Reset the flags. */
284 FOR_BB_BETWEEN (act, ENTRY_BLOCK_PTR, EXIT_BLOCK_PTR, next_bb)
285 {
286 act->flags &= ~BB_IRREDUCIBLE_LOOP;
287 FOR_EACH_EDGE (e, ei, act->succs)
288 e->flags &= ~EDGE_IRREDUCIBLE_LOOP;
289 }
290
291 /* Create the edge lists. */
292 g = new_graph (last_basic_block + num);
293
294 FOR_BB_BETWEEN (act, ENTRY_BLOCK_PTR, EXIT_BLOCK_PTR, next_bb)
295 FOR_EACH_EDGE (e, ei, act->succs)
296 {
297 /* Ignore edges to exit. */
298 if (e->dest == EXIT_BLOCK_PTR)
299 continue;
300
301 src = BB_REPR (act);
302 dest = BB_REPR (e->dest);
303
304 if (current_loops)
305 {
306 /* Ignore latch edges. */
307 if (e->dest->loop_father->header == e->dest
308 && e->dest->loop_father->latch == act)
309 continue;
310
311 /* Edges inside a single loop should be left where they are. Edges
312 to subloop headers should lead to representative of the subloop,
313 but from the same place.
314
315 Edges exiting loops should lead from representative
316 of the son of nearest common ancestor of the loops in that
317 act lays. */
318
319 if (e->dest->loop_father->header == e->dest)
320 dest = LOOP_REPR (e->dest->loop_father);
321
322 if (!flow_bb_inside_loop_p (act->loop_father, e->dest))
323 {
324 depth = find_common_loop (act->loop_father,
325 e->dest->loop_father)->depth + 1;
326 if (depth == act->loop_father->depth)
327 cloop = act->loop_father;
328 else
329 cloop = act->loop_father->pred[depth];
330
331 src = LOOP_REPR (cloop);
332 }
333 }
334
335 add_edge (g, src, dest, e);
336 }
337
338 /* Find the strongly connected components. Use the algorithm of Tarjan --
339 first determine the postorder dfs numbering in reversed graph, then
340 run the dfs on the original graph in the order given by decreasing
341 numbers assigned by the previous pass. */
342 nq = 0;
343 FOR_BB_BETWEEN (act, ENTRY_BLOCK_PTR, EXIT_BLOCK_PTR, next_bb)
344 {
345 queue1[nq++] = BB_REPR (act);
346 }
347
348 if (current_loops)
349 {
350 FOR_EACH_LOOP (li, loop, 0)
351 {
352 queue1[nq++] = LOOP_REPR (loop);
353 }
354 }
355 dfs (g, queue1, nq, queue2, false);
356 for (i = 0; i < nq; i++)
357 queue1[i] = queue2[nq - i - 1];
358 dfs (g, queue1, nq, NULL, true);
359
360 /* Mark the irreducible loops. */
361 for_each_edge (g, check_irred);
362
363 free_graph (g);
364 free (queue1);
365 free (queue2);
366
367 if (current_loops)
368 current_loops->state |= LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS;
369 }
370
371 /* Counts number of insns inside LOOP. */
372 int
373 num_loop_insns (struct loop *loop)
374 {
375 basic_block *bbs, bb;
376 unsigned i, ninsns = 0;
377 rtx insn;
378
379 bbs = get_loop_body (loop);
380 for (i = 0; i < loop->num_nodes; i++)
381 {
382 bb = bbs[i];
383 ninsns++;
384 for (insn = BB_HEAD (bb); insn != BB_END (bb); insn = NEXT_INSN (insn))
385 if (INSN_P (insn))
386 ninsns++;
387 }
388 free(bbs);
389
390 return ninsns;
391 }
392
393 /* Counts number of insns executed on average per iteration LOOP. */
394 int
395 average_num_loop_insns (struct loop *loop)
396 {
397 basic_block *bbs, bb;
398 unsigned i, binsns, ninsns, ratio;
399 rtx insn;
400
401 ninsns = 0;
402 bbs = get_loop_body (loop);
403 for (i = 0; i < loop->num_nodes; i++)
404 {
405 bb = bbs[i];
406
407 binsns = 1;
408 for (insn = BB_HEAD (bb); insn != BB_END (bb); insn = NEXT_INSN (insn))
409 if (INSN_P (insn))
410 binsns++;
411
412 ratio = loop->header->frequency == 0
413 ? BB_FREQ_MAX
414 : (bb->frequency * BB_FREQ_MAX) / loop->header->frequency;
415 ninsns += binsns * ratio;
416 }
417 free(bbs);
418
419 ninsns /= BB_FREQ_MAX;
420 if (!ninsns)
421 ninsns = 1; /* To avoid division by zero. */
422
423 return ninsns;
424 }
425
426 /* Returns expected number of LOOP iterations.
427 Compute upper bound on number of iterations in case they do not fit integer
428 to help loop peeling heuristics. Use exact counts if at all possible. */
429 unsigned
430 expected_loop_iterations (const struct loop *loop)
431 {
432 edge e;
433 edge_iterator ei;
434
435 if (loop->latch->count || loop->header->count)
436 {
437 gcov_type count_in, count_latch, expected;
438
439 count_in = 0;
440 count_latch = 0;
441
442 FOR_EACH_EDGE (e, ei, loop->header->preds)
443 if (e->src == loop->latch)
444 count_latch = e->count;
445 else
446 count_in += e->count;
447
448 if (count_in == 0)
449 expected = count_latch * 2;
450 else
451 expected = (count_latch + count_in - 1) / count_in;
452
453 /* Avoid overflows. */
454 return (expected > REG_BR_PROB_BASE ? REG_BR_PROB_BASE : expected);
455 }
456 else
457 {
458 int freq_in, freq_latch;
459
460 freq_in = 0;
461 freq_latch = 0;
462
463 FOR_EACH_EDGE (e, ei, loop->header->preds)
464 if (e->src == loop->latch)
465 freq_latch = EDGE_FREQUENCY (e);
466 else
467 freq_in += EDGE_FREQUENCY (e);
468
469 if (freq_in == 0)
470 return freq_latch * 2;
471
472 return (freq_latch + freq_in - 1) / freq_in;
473 }
474 }
475
476 /* Returns the maximum level of nesting of subloops of LOOP. */
477
478 unsigned
479 get_loop_level (const struct loop *loop)
480 {
481 const struct loop *ploop;
482 unsigned mx = 0, l;
483
484 for (ploop = loop->inner; ploop; ploop = ploop->next)
485 {
486 l = get_loop_level (ploop);
487 if (l >= mx)
488 mx = l + 1;
489 }
490 return mx;
491 }
492
493 /* Returns estimate on cost of computing SEQ. */
494
495 static unsigned
496 seq_cost (rtx seq)
497 {
498 unsigned cost = 0;
499 rtx set;
500
501 for (; seq; seq = NEXT_INSN (seq))
502 {
503 set = single_set (seq);
504 if (set)
505 cost += rtx_cost (set, SET);
506 else
507 cost++;
508 }
509
510 return cost;
511 }
512
513 /* The properties of the target. */
514
515 unsigned target_avail_regs; /* Number of available registers. */
516 unsigned target_res_regs; /* Number of reserved registers. */
517 unsigned target_small_cost; /* The cost for register when there is a free one. */
518 unsigned target_pres_cost; /* The cost for register when there are not too many
519 free ones. */
520 unsigned target_spill_cost; /* The cost for register when we need to spill. */
521
522 /* Initialize the constants for computing set costs. */
523
524 void
525 init_set_costs (void)
526 {
527 rtx seq;
528 rtx reg1 = gen_raw_REG (SImode, FIRST_PSEUDO_REGISTER);
529 rtx reg2 = gen_raw_REG (SImode, FIRST_PSEUDO_REGISTER + 1);
530 rtx addr = gen_raw_REG (Pmode, FIRST_PSEUDO_REGISTER + 2);
531 rtx mem = validize_mem (gen_rtx_MEM (SImode, addr));
532 unsigned i;
533
534 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
535 if (TEST_HARD_REG_BIT (reg_class_contents[GENERAL_REGS], i)
536 && !fixed_regs[i])
537 target_avail_regs++;
538
539 target_res_regs = 3;
540
541 /* These are really just heuristic values. */
542
543 start_sequence ();
544 emit_move_insn (reg1, reg2);
545 seq = get_insns ();
546 end_sequence ();
547 target_small_cost = seq_cost (seq);
548 target_pres_cost = 2 * target_small_cost;
549
550 start_sequence ();
551 emit_move_insn (mem, reg1);
552 emit_move_insn (reg2, mem);
553 seq = get_insns ();
554 end_sequence ();
555 target_spill_cost = seq_cost (seq);
556 }
557
558 /* Calculates cost for having SIZE new loop global variables. REGS_USED is the
559 number of global registers used in loop. N_USES is the number of relevant
560 variable uses. */
561
562 unsigned
563 global_cost_for_size (unsigned size, unsigned regs_used, unsigned n_uses)
564 {
565 unsigned regs_needed = regs_used + size;
566 unsigned cost = 0;
567
568 if (regs_needed + target_res_regs <= target_avail_regs)
569 cost += target_small_cost * size;
570 else if (regs_needed <= target_avail_regs)
571 cost += target_pres_cost * size;
572 else
573 {
574 cost += target_pres_cost * size;
575 cost += target_spill_cost * n_uses * (regs_needed - target_avail_regs) / regs_needed;
576 }
577
578 return cost;
579 }
580
581 /* Sets EDGE_LOOP_EXIT flag for all loop exits. */
582
583 void
584 mark_loop_exit_edges (void)
585 {
586 basic_block bb;
587 edge e;
588
589 if (!current_loops)
590 return;
591
592 FOR_EACH_BB (bb)
593 {
594 edge_iterator ei;
595
596 FOR_EACH_EDGE (e, ei, bb->succs)
597 {
598 if (bb->loop_father->outer
599 && loop_exit_edge_p (bb->loop_father, e))
600 e->flags |= EDGE_LOOP_EXIT;
601 else
602 e->flags &= ~EDGE_LOOP_EXIT;
603 }
604 }
605 }
606