2015-09-25 Vladimir Makarov <vmakarov@redhat.com>
[official-gcc.git] / gcc / cfgloopanal.c
blobb1786500451799e2152a9e592f8a789bfcbe0ce4
1 /* Natural loop analysis code for GNU compiler.
2 Copyright (C) 2002-2015 Free Software Foundation, Inc.
4 This file is part of GCC.
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 3, or (at your option) any later
9 version.
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.
16 You should have received a copy of the GNU General Public License
17 along with GCC; see the file COPYING3. If not see
18 <http://www.gnu.org/licenses/>. */
20 #include "config.h"
21 #include "system.h"
22 #include "coretypes.h"
23 #include "backend.h"
24 #include "predict.h"
25 #include "rtl.h"
26 #include "cfgloop.h"
27 #include "tree.h"
28 #include "flags.h"
29 #include "alias.h"
30 #include "insn-config.h"
31 #include "expmed.h"
32 #include "dojump.h"
33 #include "explow.h"
34 #include "calls.h"
35 #include "emit-rtl.h"
36 #include "varasm.h"
37 #include "stmt.h"
38 #include "expr.h"
39 #include "graphds.h"
40 #include "params.h"
42 struct target_cfgloop default_target_cfgloop;
43 #if SWITCHABLE_TARGET
44 struct target_cfgloop *this_target_cfgloop = &default_target_cfgloop;
45 #endif
47 /* Checks whether BB is executed exactly once in each LOOP iteration. */
49 bool
50 just_once_each_iteration_p (const struct loop *loop, const_basic_block bb)
52 /* It must be executed at least once each iteration. */
53 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb))
54 return false;
56 /* And just once. */
57 if (bb->loop_father != loop)
58 return false;
60 /* But this was not enough. We might have some irreducible loop here. */
61 if (bb->flags & BB_IRREDUCIBLE_LOOP)
62 return false;
64 return true;
67 /* Marks blocks and edges that are part of non-recognized loops; i.e. we
68 throw away all latch edges and mark blocks inside any remaining cycle.
69 Everything is a bit complicated due to fact we do not want to do this
70 for parts of cycles that only "pass" through some loop -- i.e. for
71 each cycle, we want to mark blocks that belong directly to innermost
72 loop containing the whole cycle.
74 LOOPS is the loop tree. */
76 #define LOOP_REPR(LOOP) ((LOOP)->num + last_basic_block_for_fn (cfun))
77 #define BB_REPR(BB) ((BB)->index + 1)
79 bool
80 mark_irreducible_loops (void)
82 basic_block act;
83 struct graph_edge *ge;
84 edge e;
85 edge_iterator ei;
86 int src, dest;
87 unsigned depth;
88 struct graph *g;
89 int num = number_of_loops (cfun);
90 struct loop *cloop;
91 bool irred_loop_found = false;
92 int i;
94 gcc_assert (current_loops != NULL);
96 /* Reset the flags. */
97 FOR_BB_BETWEEN (act, ENTRY_BLOCK_PTR_FOR_FN (cfun),
98 EXIT_BLOCK_PTR_FOR_FN (cfun), next_bb)
100 act->flags &= ~BB_IRREDUCIBLE_LOOP;
101 FOR_EACH_EDGE (e, ei, act->succs)
102 e->flags &= ~EDGE_IRREDUCIBLE_LOOP;
105 /* Create the edge lists. */
106 g = new_graph (last_basic_block_for_fn (cfun) + num);
108 FOR_BB_BETWEEN (act, ENTRY_BLOCK_PTR_FOR_FN (cfun),
109 EXIT_BLOCK_PTR_FOR_FN (cfun), next_bb)
110 FOR_EACH_EDGE (e, ei, act->succs)
112 /* Ignore edges to exit. */
113 if (e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
114 continue;
116 src = BB_REPR (act);
117 dest = BB_REPR (e->dest);
119 /* Ignore latch edges. */
120 if (e->dest->loop_father->header == e->dest
121 && e->dest->loop_father->latch == act)
122 continue;
124 /* Edges inside a single loop should be left where they are. Edges
125 to subloop headers should lead to representative of the subloop,
126 but from the same place.
128 Edges exiting loops should lead from representative
129 of the son of nearest common ancestor of the loops in that
130 act lays. */
132 if (e->dest->loop_father->header == e->dest)
133 dest = LOOP_REPR (e->dest->loop_father);
135 if (!flow_bb_inside_loop_p (act->loop_father, e->dest))
137 depth = 1 + loop_depth (find_common_loop (act->loop_father,
138 e->dest->loop_father));
139 if (depth == loop_depth (act->loop_father))
140 cloop = act->loop_father;
141 else
142 cloop = (*act->loop_father->superloops)[depth];
144 src = LOOP_REPR (cloop);
147 add_edge (g, src, dest)->data = e;
150 /* Find the strongly connected components. */
151 graphds_scc (g, NULL);
153 /* Mark the irreducible loops. */
154 for (i = 0; i < g->n_vertices; i++)
155 for (ge = g->vertices[i].succ; ge; ge = ge->succ_next)
157 edge real = (edge) ge->data;
158 /* edge E in graph G is irreducible if it connects two vertices in the
159 same scc. */
161 /* All edges should lead from a component with higher number to the
162 one with lower one. */
163 gcc_assert (g->vertices[ge->src].component >= g->vertices[ge->dest].component);
165 if (g->vertices[ge->src].component != g->vertices[ge->dest].component)
166 continue;
168 real->flags |= EDGE_IRREDUCIBLE_LOOP;
169 irred_loop_found = true;
170 if (flow_bb_inside_loop_p (real->src->loop_father, real->dest))
171 real->src->flags |= BB_IRREDUCIBLE_LOOP;
174 free_graph (g);
176 loops_state_set (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS);
177 return irred_loop_found;
180 /* Counts number of insns inside LOOP. */
182 num_loop_insns (const struct loop *loop)
184 basic_block *bbs, bb;
185 unsigned i, ninsns = 0;
186 rtx_insn *insn;
188 bbs = get_loop_body (loop);
189 for (i = 0; i < loop->num_nodes; i++)
191 bb = bbs[i];
192 FOR_BB_INSNS (bb, insn)
193 if (NONDEBUG_INSN_P (insn))
194 ninsns++;
196 free (bbs);
198 if (!ninsns)
199 ninsns = 1; /* To avoid division by zero. */
201 return ninsns;
204 /* Counts number of insns executed on average per iteration LOOP. */
206 average_num_loop_insns (const struct loop *loop)
208 basic_block *bbs, bb;
209 unsigned i, binsns, ninsns, ratio;
210 rtx_insn *insn;
212 ninsns = 0;
213 bbs = get_loop_body (loop);
214 for (i = 0; i < loop->num_nodes; i++)
216 bb = bbs[i];
218 binsns = 0;
219 FOR_BB_INSNS (bb, insn)
220 if (NONDEBUG_INSN_P (insn))
221 binsns++;
223 ratio = loop->header->frequency == 0
224 ? BB_FREQ_MAX
225 : (bb->frequency * BB_FREQ_MAX) / loop->header->frequency;
226 ninsns += binsns * ratio;
228 free (bbs);
230 ninsns /= BB_FREQ_MAX;
231 if (!ninsns)
232 ninsns = 1; /* To avoid division by zero. */
234 return ninsns;
237 /* Returns expected number of iterations of LOOP, according to
238 measured or guessed profile. No bounding is done on the
239 value. */
241 gcov_type
242 expected_loop_iterations_unbounded (const struct loop *loop)
244 edge e;
245 edge_iterator ei;
247 if (loop->latch->count || loop->header->count)
249 gcov_type count_in, count_latch, expected;
251 count_in = 0;
252 count_latch = 0;
254 FOR_EACH_EDGE (e, ei, loop->header->preds)
255 if (e->src == loop->latch)
256 count_latch = e->count;
257 else
258 count_in += e->count;
260 if (count_in == 0)
261 expected = count_latch * 2;
262 else
263 expected = (count_latch + count_in - 1) / count_in;
265 return expected;
267 else
269 int freq_in, freq_latch;
271 freq_in = 0;
272 freq_latch = 0;
274 FOR_EACH_EDGE (e, ei, loop->header->preds)
275 if (e->src == loop->latch)
276 freq_latch = EDGE_FREQUENCY (e);
277 else
278 freq_in += EDGE_FREQUENCY (e);
280 if (freq_in == 0)
281 return freq_latch * 2;
283 return (freq_latch + freq_in - 1) / freq_in;
287 /* Returns expected number of LOOP iterations. The returned value is bounded
288 by REG_BR_PROB_BASE. */
290 unsigned
291 expected_loop_iterations (const struct loop *loop)
293 gcov_type expected = expected_loop_iterations_unbounded (loop);
294 return (expected > REG_BR_PROB_BASE ? REG_BR_PROB_BASE : expected);
297 /* Returns the maximum level of nesting of subloops of LOOP. */
299 unsigned
300 get_loop_level (const struct loop *loop)
302 const struct loop *ploop;
303 unsigned mx = 0, l;
305 for (ploop = loop->inner; ploop; ploop = ploop->next)
307 l = get_loop_level (ploop);
308 if (l >= mx)
309 mx = l + 1;
311 return mx;
314 /* Initialize the constants for computing set costs. */
316 void
317 init_set_costs (void)
319 int speed;
320 rtx_insn *seq;
321 rtx reg1 = gen_raw_REG (SImode, LAST_VIRTUAL_REGISTER + 1);
322 rtx reg2 = gen_raw_REG (SImode, LAST_VIRTUAL_REGISTER + 2);
323 rtx addr = gen_raw_REG (Pmode, LAST_VIRTUAL_REGISTER + 3);
324 rtx mem = validize_mem (gen_rtx_MEM (SImode, addr));
325 unsigned i;
327 target_avail_regs = 0;
328 target_clobbered_regs = 0;
329 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
330 if (TEST_HARD_REG_BIT (reg_class_contents[GENERAL_REGS], i)
331 && !fixed_regs[i])
333 target_avail_regs++;
334 if (call_used_regs[i])
335 target_clobbered_regs++;
338 target_res_regs = 3;
340 for (speed = 0; speed < 2; speed++)
342 crtl->maybe_hot_insn_p = speed;
343 /* Set up the costs for using extra registers:
345 1) If not many free registers remain, we should prefer having an
346 additional move to decreasing the number of available registers.
347 (TARGET_REG_COST).
348 2) If no registers are available, we need to spill, which may require
349 storing the old value to memory and loading it back
350 (TARGET_SPILL_COST). */
352 start_sequence ();
353 emit_move_insn (reg1, reg2);
354 seq = get_insns ();
355 end_sequence ();
356 target_reg_cost [speed] = seq_cost (seq, speed);
358 start_sequence ();
359 emit_move_insn (mem, reg1);
360 emit_move_insn (reg2, mem);
361 seq = get_insns ();
362 end_sequence ();
363 target_spill_cost [speed] = seq_cost (seq, speed);
365 default_rtl_profile ();
368 /* Estimates cost of increased register pressure caused by making N_NEW new
369 registers live around the loop. N_OLD is the number of registers live
370 around the loop. If CALL_P is true, also take into account that
371 call-used registers may be clobbered in the loop body, reducing the
372 number of available registers before we spill. */
374 unsigned
375 estimate_reg_pressure_cost (unsigned n_new, unsigned n_old, bool speed,
376 bool call_p)
378 unsigned cost;
379 unsigned regs_needed = n_new + n_old;
380 unsigned available_regs = target_avail_regs;
382 /* If there is a call in the loop body, the call-clobbered registers
383 are not available for loop invariants. */
384 if (call_p)
385 available_regs = available_regs - target_clobbered_regs;
387 /* If we have enough registers, we should use them and not restrict
388 the transformations unnecessarily. */
389 if (regs_needed + target_res_regs <= available_regs)
390 return 0;
392 if (regs_needed <= available_regs)
393 /* If we are close to running out of registers, try to preserve
394 them. */
395 cost = target_reg_cost [speed] * n_new;
396 else
397 /* If we run out of registers, it is very expensive to add another
398 one. */
399 cost = target_spill_cost [speed] * n_new;
401 if (optimize && (flag_ira_region == IRA_REGION_ALL
402 || flag_ira_region == IRA_REGION_MIXED)
403 && number_of_loops (cfun) <= (unsigned) IRA_MAX_LOOPS_NUM)
404 /* IRA regional allocation deals with high register pressure
405 better. So decrease the cost (to do more accurate the cost
406 calculation for IRA, we need to know how many registers lives
407 through the loop transparently). */
408 cost /= 2;
410 return cost;
413 /* Sets EDGE_LOOP_EXIT flag for all loop exits. */
415 void
416 mark_loop_exit_edges (void)
418 basic_block bb;
419 edge e;
421 if (number_of_loops (cfun) <= 1)
422 return;
424 FOR_EACH_BB_FN (bb, cfun)
426 edge_iterator ei;
428 FOR_EACH_EDGE (e, ei, bb->succs)
430 if (loop_outer (bb->loop_father)
431 && loop_exit_edge_p (bb->loop_father, e))
432 e->flags |= EDGE_LOOP_EXIT;
433 else
434 e->flags &= ~EDGE_LOOP_EXIT;
439 /* Return exit edge if loop has only one exit that is likely
440 to be executed on runtime (i.e. it is not EH or leading
441 to noreturn call. */
443 edge
444 single_likely_exit (struct loop *loop)
446 edge found = single_exit (loop);
447 vec<edge> exits;
448 unsigned i;
449 edge ex;
451 if (found)
452 return found;
453 exits = get_loop_exit_edges (loop);
454 FOR_EACH_VEC_ELT (exits, i, ex)
456 if (ex->flags & (EDGE_EH | EDGE_ABNORMAL_CALL))
457 continue;
458 /* The constant of 5 is set in a way so noreturn calls are
459 ruled out by this test. The static branch prediction algorithm
460 will not assign such a low probability to conditionals for usual
461 reasons. */
462 if (profile_status_for_fn (cfun) != PROFILE_ABSENT
463 && ex->probability < 5 && !ex->count)
464 continue;
465 if (!found)
466 found = ex;
467 else
469 exits.release ();
470 return NULL;
473 exits.release ();
474 return found;
478 /* Gets basic blocks of a LOOP. Header is the 0-th block, rest is in dfs
479 order against direction of edges from latch. Specially, if
480 header != latch, latch is the 1-st block. */
482 vec<basic_block>
483 get_loop_hot_path (const struct loop *loop)
485 basic_block bb = loop->header;
486 vec<basic_block> path = vNULL;
487 bitmap visited = BITMAP_ALLOC (NULL);
489 while (true)
491 edge_iterator ei;
492 edge e;
493 edge best = NULL;
495 path.safe_push (bb);
496 bitmap_set_bit (visited, bb->index);
497 FOR_EACH_EDGE (e, ei, bb->succs)
498 if ((!best || e->probability > best->probability)
499 && !loop_exit_edge_p (loop, e)
500 && !bitmap_bit_p (visited, e->dest->index))
501 best = e;
502 if (!best || best->dest == loop->header)
503 break;
504 bb = best->dest;
506 BITMAP_FREE (visited);
507 return path;