1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
|
/* Natural loop analysis code for GNU compiler.
Copyright (C) 2002-2019 Free Software Foundation, Inc.
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify it under
the terms of the GNU General Public License as published by the Free
Software Foundation; either version 3, or (at your option) any later
version.
GCC is distributed in the hope that it will be useful, but WITHOUT ANY
WARRANTY; without even the implied warranty of MERCHANTABILITY or
FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
for more details.
You should have received a copy of the GNU General Public License
along with GCC; see the file COPYING3. If not see
<http://www.gnu.org/licenses/>. */
#include "config.h"
#include "system.h"
#include "coretypes.h"
#include "backend.h"
#include "rtl.h"
#include "tree.h"
#include "predict.h"
#include "memmodel.h"
#include "emit-rtl.h"
#include "cfgloop.h"
#include "explow.h"
#include "expr.h"
#include "graphds.h"
#include "params.h"
#include "sreal.h"
#include "regs.h"
#include "function-abi.h"
struct target_cfgloop default_target_cfgloop;
#if SWITCHABLE_TARGET
struct target_cfgloop *this_target_cfgloop = &default_target_cfgloop;
#endif
/* Checks whether BB is executed exactly once in each LOOP iteration. */
bool
just_once_each_iteration_p (const class loop *loop, const_basic_block bb)
{
/* It must be executed at least once each iteration. */
if (!dominated_by_p (CDI_DOMINATORS, loop->latch, bb))
return false;
/* And just once. */
if (bb->loop_father != loop)
return false;
/* But this was not enough. We might have some irreducible loop here. */
if (bb->flags & BB_IRREDUCIBLE_LOOP)
return false;
return true;
}
/* Marks blocks and edges that are part of non-recognized loops; i.e. we
throw away all latch edges and mark blocks inside any remaining cycle.
Everything is a bit complicated due to fact we do not want to do this
for parts of cycles that only "pass" through some loop -- i.e. for
each cycle, we want to mark blocks that belong directly to innermost
loop containing the whole cycle.
LOOPS is the loop tree. */
#define LOOP_REPR(LOOP) ((LOOP)->num + last_basic_block_for_fn (cfun))
#define BB_REPR(BB) ((BB)->index + 1)
bool
mark_irreducible_loops (void)
{
basic_block act;
struct graph_edge *ge;
edge e;
edge_iterator ei;
int src, dest;
unsigned depth;
struct graph *g;
int num = number_of_loops (cfun);
class loop *cloop;
bool irred_loop_found = false;
int i;
gcc_assert (current_loops != NULL);
/* Reset the flags. */
FOR_BB_BETWEEN (act, ENTRY_BLOCK_PTR_FOR_FN (cfun),
EXIT_BLOCK_PTR_FOR_FN (cfun), next_bb)
{
act->flags &= ~BB_IRREDUCIBLE_LOOP;
FOR_EACH_EDGE (e, ei, act->succs)
e->flags &= ~EDGE_IRREDUCIBLE_LOOP;
}
/* Create the edge lists. */
g = new_graph (last_basic_block_for_fn (cfun) + num);
FOR_BB_BETWEEN (act, ENTRY_BLOCK_PTR_FOR_FN (cfun),
EXIT_BLOCK_PTR_FOR_FN (cfun), next_bb)
FOR_EACH_EDGE (e, ei, act->succs)
{
/* Ignore edges to exit. */
if (e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun))
continue;
src = BB_REPR (act);
dest = BB_REPR (e->dest);
/* Ignore latch edges. */
if (e->dest->loop_father->header == e->dest
&& e->dest->loop_father->latch == act)
continue;
/* Edges inside a single loop should be left where they are. Edges
to subloop headers should lead to representative of the subloop,
but from the same place.
Edges exiting loops should lead from representative
of the son of nearest common ancestor of the loops in that
act lays. */
if (e->dest->loop_father->header == e->dest)
dest = LOOP_REPR (e->dest->loop_father);
if (!flow_bb_inside_loop_p (act->loop_father, e->dest))
{
depth = 1 + loop_depth (find_common_loop (act->loop_father,
e->dest->loop_father));
if (depth == loop_depth (act->loop_father))
cloop = act->loop_father;
else
cloop = (*act->loop_father->superloops)[depth];
src = LOOP_REPR (cloop);
}
add_edge (g, src, dest)->data = e;
}
/* Find the strongly connected components. */
graphds_scc (g, NULL);
/* Mark the irreducible loops. */
for (i = 0; i < g->n_vertices; i++)
for (ge = g->vertices[i].succ; ge; ge = ge->succ_next)
{
edge real = (edge) ge->data;
/* edge E in graph G is irreducible if it connects two vertices in the
same scc. */
/* All edges should lead from a component with higher number to the
one with lower one. */
gcc_assert (g->vertices[ge->src].component >= g->vertices[ge->dest].component);
if (g->vertices[ge->src].component != g->vertices[ge->dest].component)
continue;
real->flags |= EDGE_IRREDUCIBLE_LOOP;
irred_loop_found = true;
if (flow_bb_inside_loop_p (real->src->loop_father, real->dest))
real->src->flags |= BB_IRREDUCIBLE_LOOP;
}
free_graph (g);
loops_state_set (LOOPS_HAVE_MARKED_IRREDUCIBLE_REGIONS);
return irred_loop_found;
}
/* Counts number of insns inside LOOP. */
int
num_loop_insns (const class loop *loop)
{
basic_block *bbs, bb;
unsigned i, ninsns = 0;
rtx_insn *insn;
bbs = get_loop_body (loop);
for (i = 0; i < loop->num_nodes; i++)
{
bb = bbs[i];
FOR_BB_INSNS (bb, insn)
if (NONDEBUG_INSN_P (insn))
ninsns++;
}
free (bbs);
if (!ninsns)
ninsns = 1; /* To avoid division by zero. */
return ninsns;
}
/* Counts number of insns executed on average per iteration LOOP. */
int
average_num_loop_insns (const class loop *loop)
{
basic_block *bbs, bb;
unsigned i, binsns;
sreal ninsns;
rtx_insn *insn;
ninsns = 0;
bbs = get_loop_body (loop);
for (i = 0; i < loop->num_nodes; i++)
{
bb = bbs[i];
binsns = 0;
FOR_BB_INSNS (bb, insn)
if (NONDEBUG_INSN_P (insn))
binsns++;
ninsns += (sreal)binsns * bb->count.to_sreal_scale (loop->header->count);
/* Avoid overflows. */
if (ninsns > 1000000)
return 100000;
}
free (bbs);
int64_t ret = ninsns.to_int ();
if (!ret)
ret = 1; /* To avoid division by zero. */
return ret;
}
/* Returns expected number of iterations of LOOP, according to
measured or guessed profile.
This functions attempts to return "sane" value even if profile
information is not good enough to derive osmething.
If BY_PROFILE_ONLY is set, this logic is bypassed and function
return -1 in those scenarios. */
gcov_type
expected_loop_iterations_unbounded (const class loop *loop,
bool *read_profile_p,
bool by_profile_only)
{
edge e;
edge_iterator ei;
gcov_type expected = -1;
if (read_profile_p)
*read_profile_p = false;
/* If we have no profile at all, use AVG_LOOP_NITER. */
if (profile_status_for_fn (cfun) == PROFILE_ABSENT)
{
if (by_profile_only)
return -1;
expected = param_avg_loop_niter;
}
else if (loop->latch && (loop->latch->count.initialized_p ()
|| loop->header->count.initialized_p ()))
{
profile_count count_in = profile_count::zero (),
count_latch = profile_count::zero ();
FOR_EACH_EDGE (e, ei, loop->header->preds)
if (e->src == loop->latch)
count_latch = e->count ();
else
count_in += e->count ();
if (!count_latch.initialized_p ())
{
if (by_profile_only)
return -1;
expected = param_avg_loop_niter;
}
else if (!count_in.nonzero_p ())
{
if (by_profile_only)
return -1;
expected = count_latch.to_gcov_type () * 2;
}
else
{
expected = (count_latch.to_gcov_type () + count_in.to_gcov_type ()
- 1) / count_in.to_gcov_type ();
if (read_profile_p
&& count_latch.reliable_p () && count_in.reliable_p ())
*read_profile_p = true;
}
}
else
{
if (by_profile_only)
return -1;
expected = param_avg_loop_niter;
}
if (!by_profile_only)
{
HOST_WIDE_INT max = get_max_loop_iterations_int (loop);
if (max != -1 && max < expected)
return max;
}
return expected;
}
/* Returns expected number of LOOP iterations. The returned value is bounded
by REG_BR_PROB_BASE. */
unsigned
expected_loop_iterations (class loop *loop)
{
gcov_type expected = expected_loop_iterations_unbounded (loop);
return (expected > REG_BR_PROB_BASE ? REG_BR_PROB_BASE : expected);
}
/* Returns the maximum level of nesting of subloops of LOOP. */
unsigned
get_loop_level (const class loop *loop)
{
const class loop *ploop;
unsigned mx = 0, l;
for (ploop = loop->inner; ploop; ploop = ploop->next)
{
l = get_loop_level (ploop);
if (l >= mx)
mx = l + 1;
}
return mx;
}
/* Initialize the constants for computing set costs. */
void
init_set_costs (void)
{
int speed;
rtx_insn *seq;
rtx reg1 = gen_raw_REG (SImode, LAST_VIRTUAL_REGISTER + 1);
rtx reg2 = gen_raw_REG (SImode, LAST_VIRTUAL_REGISTER + 2);
rtx addr = gen_raw_REG (Pmode, LAST_VIRTUAL_REGISTER + 3);
rtx mem = validize_mem (gen_rtx_MEM (SImode, addr));
unsigned i;
target_avail_regs = 0;
target_clobbered_regs = 0;
for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
if (TEST_HARD_REG_BIT (reg_class_contents[GENERAL_REGS], i)
&& !fixed_regs[i])
{
target_avail_regs++;
/* ??? This is only a rough heuristic. It doesn't cope well
with alternative ABIs, but that's an optimization rather than
correctness issue. */
if (default_function_abi.clobbers_full_reg_p (i))
target_clobbered_regs++;
}
target_res_regs = 3;
for (speed = 0; speed < 2; speed++)
{
crtl->maybe_hot_insn_p = speed;
/* Set up the costs for using extra registers:
1) If not many free registers remain, we should prefer having an
additional move to decreasing the number of available registers.
(TARGET_REG_COST).
2) If no registers are available, we need to spill, which may require
storing the old value to memory and loading it back
(TARGET_SPILL_COST). */
start_sequence ();
emit_move_insn (reg1, reg2);
seq = get_insns ();
end_sequence ();
target_reg_cost [speed] = seq_cost (seq, speed);
start_sequence ();
emit_move_insn (mem, reg1);
emit_move_insn (reg2, mem);
seq = get_insns ();
end_sequence ();
target_spill_cost [speed] = seq_cost (seq, speed);
}
default_rtl_profile ();
}
/* Estimates cost of increased register pressure caused by making N_NEW new
registers live around the loop. N_OLD is the number of registers live
around the loop. If CALL_P is true, also take into account that
call-used registers may be clobbered in the loop body, reducing the
number of available registers before we spill. */
unsigned
estimate_reg_pressure_cost (unsigned n_new, unsigned n_old, bool speed,
bool call_p)
{
unsigned cost;
unsigned regs_needed = n_new + n_old;
unsigned available_regs = target_avail_regs;
/* If there is a call in the loop body, the call-clobbered registers
are not available for loop invariants. */
if (call_p)
available_regs = available_regs - target_clobbered_regs;
/* If we have enough registers, we should use them and not restrict
the transformations unnecessarily. */
if (regs_needed + target_res_regs <= available_regs)
return 0;
if (regs_needed <= available_regs)
/* If we are close to running out of registers, try to preserve
them. */
cost = target_reg_cost [speed] * n_new;
else
/* If we run out of registers, it is very expensive to add another
one. */
cost = target_spill_cost [speed] * n_new;
if (optimize && (flag_ira_region == IRA_REGION_ALL
|| flag_ira_region == IRA_REGION_MIXED)
&& number_of_loops (cfun) <= (unsigned) param_ira_max_loops_num)
/* IRA regional allocation deals with high register pressure
better. So decrease the cost (to do more accurate the cost
calculation for IRA, we need to know how many registers lives
through the loop transparently). */
cost /= 2;
return cost;
}
/* Sets EDGE_LOOP_EXIT flag for all loop exits. */
void
mark_loop_exit_edges (void)
{
basic_block bb;
edge e;
if (number_of_loops (cfun) <= 1)
return;
FOR_EACH_BB_FN (bb, cfun)
{
edge_iterator ei;
FOR_EACH_EDGE (e, ei, bb->succs)
{
if (loop_outer (bb->loop_father)
&& loop_exit_edge_p (bb->loop_father, e))
e->flags |= EDGE_LOOP_EXIT;
else
e->flags &= ~EDGE_LOOP_EXIT;
}
}
}
/* Return exit edge if loop has only one exit that is likely
to be executed on runtime (i.e. it is not EH or leading
to noreturn call. */
edge
single_likely_exit (class loop *loop)
{
edge found = single_exit (loop);
vec<edge> exits;
unsigned i;
edge ex;
if (found)
return found;
exits = get_loop_exit_edges (loop);
FOR_EACH_VEC_ELT (exits, i, ex)
{
if (probably_never_executed_edge_p (cfun, ex)
/* We want to rule out paths to noreturns but not low probabilities
resulting from adjustments or combining.
FIXME: once we have better quality tracking, make this more
robust. */
|| ex->probability <= profile_probability::very_unlikely ())
continue;
if (!found)
found = ex;
else
{
exits.release ();
return NULL;
}
}
exits.release ();
return found;
}
/* Gets basic blocks of a LOOP. Header is the 0-th block, rest is in dfs
order against direction of edges from latch. Specially, if
header != latch, latch is the 1-st block. */
vec<basic_block>
get_loop_hot_path (const class loop *loop)
{
basic_block bb = loop->header;
vec<basic_block> path = vNULL;
bitmap visited = BITMAP_ALLOC (NULL);
while (true)
{
edge_iterator ei;
edge e;
edge best = NULL;
path.safe_push (bb);
bitmap_set_bit (visited, bb->index);
FOR_EACH_EDGE (e, ei, bb->succs)
if ((!best || e->probability > best->probability)
&& !loop_exit_edge_p (loop, e)
&& !bitmap_bit_p (visited, e->dest->index))
best = e;
if (!best || best->dest == loop->header)
break;
bb = best->dest;
}
BITMAP_FREE (visited);
return path;
}
|