Add option for whether ceil etc. can raise "inexact", adjust x86 conditions.
[official-gcc.git] / gcc / cfgloopanal.c
blob938ac43879a530a61b7b077f774ec2fb0755958b
1 /* Natural loop analysis code for GNU compiler.
2 Copyright (C) 2002-2016 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 "rtl.h"
25 #include "tree.h"
26 #include "predict.h"
27 #include "emit-rtl.h"
28 #include "cfgloop.h"
29 #include "explow.h"
30 #include "expr.h"
31 #include "graphds.h"
32 #include "params.h"
34 struct target_cfgloop default_target_cfgloop;
35 #if SWITCHABLE_TARGET
36 struct target_cfgloop *this_target_cfgloop = &default_target_cfgloop;
37 #endif
39 /* Checks whether BB is executed exactly once in each LOOP iteration. */
41 bool
42 just_once_each_iteration_p (const struct loop *loop, const_basic_block bb)
44 /* It must be executed at least once each iteration. */
45 if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb))
46 return false;
48 /* And just once. */
49 if (bb->loop_father != loop)
50 return false;
52 /* But this was not enough. We might have some irreducible loop here. */
53 if (bb->flags & BB_IRREDUCIBLE_LOOP)
54 return false;
56 return true;
59 /* Marks blocks and edges that are part of non-recognized loops; i.e. we
60 throw away all latch edges and mark blocks inside any remaining cycle.
61 Everything is a bit complicated due to fact we do not want to do this
62 for parts of cycles that only "pass" through some loop -- i.e. for
63 each cycle, we want to mark blocks that belong directly to innermost
64 loop containing the whole cycle.
66 LOOPS is the loop tree. */
68 #define LOOP_REPR(LOOP) ((LOOP)->num + last_basic_block_for_fn (cfun))
69 #define BB_REPR(BB) ((BB)->index + 1)
71 bool
72 mark_irreducible_loops (void)
74 basic_block act;
75 struct graph_edge *ge;
76 edge e;
77 edge_iterator ei;
78 int src, dest;
79 unsigned depth;
80 struct graph *g;
81 int num = number_of_loops (cfun);
82 struct loop *cloop;
83 bool irred_loop_found = false;
84 int i;
86 gcc_assert (current_loops != NULL);
88 /* Reset the flags. */
89 FOR_BB_BETWEEN (act, ENTRY_BLOCK_PTR_FOR_FN (cfun),
90 EXIT_BLOCK_PTR_FOR_FN (cfun), next_bb)
92 act->flags &= ~BB_IRREDUCIBLE_LOOP;
93 FOR_EACH_EDGE (e, ei, act->succs)
94 e->flags &= ~EDGE_IRREDUCIBLE_LOOP;
97 /* Create the edge lists. */
98 g = new_graph (last_basic_block_for_fn (cfun) + num);
100 FOR_BB_BETWEEN (act, ENTRY_BLOCK_PTR_FOR_FN (cfun),
101 EXIT_BLOCK_PTR_FOR_FN (cfun), next_bb)
102 FOR_EACH_EDGE (e, ei, act->succs)
104 /* Ignore edges to exit. */
105 if (e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
106 continue;
108 src = BB_REPR (act);
109 dest = BB_REPR (e->dest);
111 /* Ignore latch edges. */
112 if (e->dest->loop_father->header == e->dest
113 && e->dest->loop_father->latch == act)
114 continue;
116 /* Edges inside a single loop should be left where they are. Edges
117 to subloop headers should lead to representative of the subloop,
118 but from the same place.
120 Edges exiting loops should lead from representative
121 of the son of nearest common ancestor of the loops in that
122 act lays. */
124 if (e->dest->loop_father->header == e->dest)
125 dest = LOOP_REPR (e->dest->loop_father);
127 if (!flow_bb_inside_loop_p (act->loop_father, e->dest))
129 depth = 1 + loop_depth (find_common_loop (act->loop_father,
130 e->dest->loop_father));
131 if (depth == loop_depth (act->loop_father))
132 cloop = act->loop_father;
133 else
134 cloop = (*act->loop_father->superloops)[depth];
136 src = LOOP_REPR (cloop);
139 add_edge (g, src, dest)->data = e;
142 /* Find the strongly connected components. */
143 graphds_scc (g, NULL);
145 /* Mark the irreducible loops. */
146 for (i = 0; i < g->n_vertices; i++)
147 for (ge = g->vertices[i].succ; ge; ge = ge->succ_next)
149 edge real = (edge) ge->data;
150 /* edge E in graph G is irreducible if it connects two vertices in the
151 same scc. */
153 /* All edges should lead from a component with higher number to the
154 one with lower one. */
155 gcc_assert (g->vertices[ge->src].component >= g->vertices[ge->dest].component);
157 if (g->vertices[ge->src].component != g->vertices[ge->dest].component)
158 continue;
160 real->flags |= EDGE_IRREDUCIBLE_LOOP;
161 irred_loop_found = true;
162 if (flow_bb_inside_loop_p (real->src->loop_father, real->dest))
163 real->src->flags |= BB_IRREDUCIBLE_LOOP;
166 free_graph (g);
168 loops_state_set (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS);
169 return irred_loop_found;
172 /* Counts number of insns inside LOOP. */
174 num_loop_insns (const struct loop *loop)
176 basic_block *bbs, bb;
177 unsigned i, ninsns = 0;
178 rtx_insn *insn;
180 bbs = get_loop_body (loop);
181 for (i = 0; i < loop->num_nodes; i++)
183 bb = bbs[i];
184 FOR_BB_INSNS (bb, insn)
185 if (NONDEBUG_INSN_P (insn))
186 ninsns++;
188 free (bbs);
190 if (!ninsns)
191 ninsns = 1; /* To avoid division by zero. */
193 return ninsns;
196 /* Counts number of insns executed on average per iteration LOOP. */
198 average_num_loop_insns (const struct loop *loop)
200 basic_block *bbs, bb;
201 unsigned i, binsns, ninsns, ratio;
202 rtx_insn *insn;
204 ninsns = 0;
205 bbs = get_loop_body (loop);
206 for (i = 0; i < loop->num_nodes; i++)
208 bb = bbs[i];
210 binsns = 0;
211 FOR_BB_INSNS (bb, insn)
212 if (NONDEBUG_INSN_P (insn))
213 binsns++;
215 ratio = loop->header->frequency == 0
216 ? BB_FREQ_MAX
217 : (bb->frequency * BB_FREQ_MAX) / loop->header->frequency;
218 ninsns += binsns * ratio;
220 free (bbs);
222 ninsns /= BB_FREQ_MAX;
223 if (!ninsns)
224 ninsns = 1; /* To avoid division by zero. */
226 return ninsns;
229 /* Returns expected number of iterations of LOOP, according to
230 measured or guessed profile. No bounding is done on the
231 value. */
233 gcov_type
234 expected_loop_iterations_unbounded (struct loop *loop)
236 edge e;
237 edge_iterator ei;
238 gcov_type expected;
241 /* Average loop rolls about 3 times. If we have no profile at all, it is
242 best we can do. */
243 if (profile_status_for_fn (cfun) == PROFILE_ABSENT)
244 expected = 3;
245 else if (loop->latch->count || loop->header->count)
247 gcov_type count_in, count_latch;
249 count_in = 0;
250 count_latch = 0;
252 FOR_EACH_EDGE (e, ei, loop->header->preds)
253 if (e->src == loop->latch)
254 count_latch = e->count;
255 else
256 count_in += e->count;
258 if (count_in == 0)
259 expected = count_latch * 2;
260 else
261 expected = (count_latch + count_in - 1) / count_in;
263 else
265 int freq_in, freq_latch;
267 freq_in = 0;
268 freq_latch = 0;
270 FOR_EACH_EDGE (e, ei, loop->header->preds)
271 if (e->src == loop->latch)
272 freq_latch = EDGE_FREQUENCY (e);
273 else
274 freq_in += EDGE_FREQUENCY (e);
276 if (freq_in == 0)
278 /* If we have no profile at all, expect 3 iterations. */
279 if (!freq_latch)
280 expected = 3;
281 else
282 expected = freq_latch * 2;
284 else
285 expected = (freq_latch + freq_in - 1) / freq_in;
288 HOST_WIDE_INT max = get_max_loop_iterations_int (loop);
289 if (max != -1 && max < expected)
290 return max;
291 return expected;
294 /* Returns expected number of LOOP iterations. The returned value is bounded
295 by REG_BR_PROB_BASE. */
297 unsigned
298 expected_loop_iterations (struct loop *loop)
300 gcov_type expected = expected_loop_iterations_unbounded (loop);
301 return (expected > REG_BR_PROB_BASE ? REG_BR_PROB_BASE : expected);
304 /* Returns the maximum level of nesting of subloops of LOOP. */
306 unsigned
307 get_loop_level (const struct loop *loop)
309 const struct loop *ploop;
310 unsigned mx = 0, l;
312 for (ploop = loop->inner; ploop; ploop = ploop->next)
314 l = get_loop_level (ploop);
315 if (l >= mx)
316 mx = l + 1;
318 return mx;
321 /* Initialize the constants for computing set costs. */
323 void
324 init_set_costs (void)
326 int speed;
327 rtx_insn *seq;
328 rtx reg1 = gen_raw_REG (SImode, LAST_VIRTUAL_REGISTER + 1);
329 rtx reg2 = gen_raw_REG (SImode, LAST_VIRTUAL_REGISTER + 2);
330 rtx addr = gen_raw_REG (Pmode, LAST_VIRTUAL_REGISTER + 3);
331 rtx mem = validize_mem (gen_rtx_MEM (SImode, addr));
332 unsigned i;
334 target_avail_regs = 0;
335 target_clobbered_regs = 0;
336 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
337 if (TEST_HARD_REG_BIT (reg_class_contents[GENERAL_REGS], i)
338 && !fixed_regs[i])
340 target_avail_regs++;
341 if (call_used_regs[i])
342 target_clobbered_regs++;
345 target_res_regs = 3;
347 for (speed = 0; speed < 2; speed++)
349 crtl->maybe_hot_insn_p = speed;
350 /* Set up the costs for using extra registers:
352 1) If not many free registers remain, we should prefer having an
353 additional move to decreasing the number of available registers.
354 (TARGET_REG_COST).
355 2) If no registers are available, we need to spill, which may require
356 storing the old value to memory and loading it back
357 (TARGET_SPILL_COST). */
359 start_sequence ();
360 emit_move_insn (reg1, reg2);
361 seq = get_insns ();
362 end_sequence ();
363 target_reg_cost [speed] = seq_cost (seq, speed);
365 start_sequence ();
366 emit_move_insn (mem, reg1);
367 emit_move_insn (reg2, mem);
368 seq = get_insns ();
369 end_sequence ();
370 target_spill_cost [speed] = seq_cost (seq, speed);
372 default_rtl_profile ();
375 /* Estimates cost of increased register pressure caused by making N_NEW new
376 registers live around the loop. N_OLD is the number of registers live
377 around the loop. If CALL_P is true, also take into account that
378 call-used registers may be clobbered in the loop body, reducing the
379 number of available registers before we spill. */
381 unsigned
382 estimate_reg_pressure_cost (unsigned n_new, unsigned n_old, bool speed,
383 bool call_p)
385 unsigned cost;
386 unsigned regs_needed = n_new + n_old;
387 unsigned available_regs = target_avail_regs;
389 /* If there is a call in the loop body, the call-clobbered registers
390 are not available for loop invariants. */
391 if (call_p)
392 available_regs = available_regs - target_clobbered_regs;
394 /* If we have enough registers, we should use them and not restrict
395 the transformations unnecessarily. */
396 if (regs_needed + target_res_regs <= available_regs)
397 return 0;
399 if (regs_needed <= available_regs)
400 /* If we are close to running out of registers, try to preserve
401 them. */
402 cost = target_reg_cost [speed] * n_new;
403 else
404 /* If we run out of registers, it is very expensive to add another
405 one. */
406 cost = target_spill_cost [speed] * n_new;
408 if (optimize && (flag_ira_region == IRA_REGION_ALL
409 || flag_ira_region == IRA_REGION_MIXED)
410 && number_of_loops (cfun) <= (unsigned) IRA_MAX_LOOPS_NUM)
411 /* IRA regional allocation deals with high register pressure
412 better. So decrease the cost (to do more accurate the cost
413 calculation for IRA, we need to know how many registers lives
414 through the loop transparently). */
415 cost /= 2;
417 return cost;
420 /* Sets EDGE_LOOP_EXIT flag for all loop exits. */
422 void
423 mark_loop_exit_edges (void)
425 basic_block bb;
426 edge e;
428 if (number_of_loops (cfun) <= 1)
429 return;
431 FOR_EACH_BB_FN (bb, cfun)
433 edge_iterator ei;
435 FOR_EACH_EDGE (e, ei, bb->succs)
437 if (loop_outer (bb->loop_father)
438 && loop_exit_edge_p (bb->loop_father, e))
439 e->flags |= EDGE_LOOP_EXIT;
440 else
441 e->flags &= ~EDGE_LOOP_EXIT;
446 /* Return exit edge if loop has only one exit that is likely
447 to be executed on runtime (i.e. it is not EH or leading
448 to noreturn call. */
450 edge
451 single_likely_exit (struct loop *loop)
453 edge found = single_exit (loop);
454 vec<edge> exits;
455 unsigned i;
456 edge ex;
458 if (found)
459 return found;
460 exits = get_loop_exit_edges (loop);
461 FOR_EACH_VEC_ELT (exits, i, ex)
463 if (ex->flags & (EDGE_EH | EDGE_ABNORMAL_CALL))
464 continue;
465 /* The constant of 5 is set in a way so noreturn calls are
466 ruled out by this test. The static branch prediction algorithm
467 will not assign such a low probability to conditionals for usual
468 reasons. */
469 if (profile_status_for_fn (cfun) != PROFILE_ABSENT
470 && ex->probability < 5 && !ex->count)
471 continue;
472 if (!found)
473 found = ex;
474 else
476 exits.release ();
477 return NULL;
480 exits.release ();
481 return found;
485 /* Gets basic blocks of a LOOP. Header is the 0-th block, rest is in dfs
486 order against direction of edges from latch. Specially, if
487 header != latch, latch is the 1-st block. */
489 vec<basic_block>
490 get_loop_hot_path (const struct loop *loop)
492 basic_block bb = loop->header;
493 vec<basic_block> path = vNULL;
494 bitmap visited = BITMAP_ALLOC (NULL);
496 while (true)
498 edge_iterator ei;
499 edge e;
500 edge best = NULL;
502 path.safe_push (bb);
503 bitmap_set_bit (visited, bb->index);
504 FOR_EACH_EDGE (e, ei, bb->succs)
505 if ((!best || e->probability > best->probability)
506 && !loop_exit_edge_p (loop, e)
507 && !bitmap_bit_p (visited, e->dest->index))
508 best = e;
509 if (!best || best->dest == loop->header)
510 break;
511 bb = best->dest;
513 BITMAP_FREE (visited);
514 return path;