1 /* Instruction scheduling pass.
2 Copyright (C) 1992-2021 Free Software Foundation, Inc.
3 Contributed by Michael Tiemann (tiemann@cygnus.com) Enhanced by,
4 and currently maintained by, Jim Wilson (wilson@cygnus.com)
5
6 This file is part of GCC.
7
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
11 version.
12
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 for more details.
17
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
21
22 /* This pass implements list scheduling within basic blocks. It is
23 run twice: (1) after flow analysis, but before register allocation,
24 and (2) after register allocation.
25
26 The first run performs interblock scheduling, moving insns between
27 different blocks in the same "region", and the second runs only
28 basic block scheduling.
29
30 Interblock motions performed are useful motions and speculative
31 motions, including speculative loads. Motions requiring code
32 duplication are not supported. The identification of motion type
33 and the check for validity of speculative motions requires
34 construction and analysis of the function's control flow graph.
35
36 The main entry point for this pass is schedule_insns(), called for
37 each function. The work of the scheduler is organized in three
38 levels: (1) function level: insns are subject to splitting,
39 control-flow-graph is constructed, regions are computed (after
40 reload, each region is of one block), (2) region level: control
41 flow graph attributes required for interblock scheduling are
42 computed (dominators, reachability, etc.), data dependences and
43 priorities are computed, and (3) block level: insns in the block
44 are actually scheduled. */
45
46 #include "config.h"
47 #include "system.h"
48 #include "coretypes.h"
49 #include "backend.h"
50 #include "target.h"
51 #include "rtl.h"
52 #include "df.h"
53 #include "memmodel.h"
54 #include "tm_p.h"
55 #include "insn-config.h"
56 #include "emit-rtl.h"
57 #include "recog.h"
58 #include "profile.h"
59 #include "insn-attr.h"
60 #include "except.h"
61 #include "cfganal.h"
62 #include "sched-int.h"
63 #include "sel-sched.h"
64 #include "tree-pass.h"
65 #include "dbgcnt.h"
66 #include "pretty-print.h"
67 #include "print-rtl.h"
68
69 /* Disable warnings about quoting issues in the pp_xxx calls below
70 that (intentionally) don't follow GCC diagnostic conventions. */
71 #if __GNUC__ >= 10
72 # pragma GCC diagnostic push
73 # pragma GCC diagnostic ignored "-Wformat-diag"
74 #endif
75
76 #ifdef INSN_SCHEDULING
77
78 /* Some accessor macros for h_i_d members only used within this file. */
79 #define FED_BY_SPEC_LOAD(INSN) (HID (INSN)->fed_by_spec_load)
80 #define IS_LOAD_INSN(INSN) (HID (insn)->is_load_insn)
81
82 /* nr_inter/spec counts interblock/speculative motion for the function. */
83 static int nr_inter, nr_spec;
84
85 static int is_cfg_nonregular (void);
86
87 /* Number of regions in the procedure. */
88 int nr_regions = 0;
89
90 /* Same as above before adding any new regions. */
91 static int nr_regions_initial = 0;
92
93 /* Table of region descriptions. */
94 region *rgn_table = NULL;
95
96 /* Array of lists of regions' blocks. */
97 int *rgn_bb_table = NULL;
98
99 /* Topological order of blocks in the region (if b2 is reachable from
100 b1, block_to_bb[b2] > block_to_bb[b1]). Note: A basic block is
101 always referred to by either block or b, while its topological
102 order name (in the region) is referred to by bb. */
103 int *block_to_bb = NULL;
104
105 /* The number of the region containing a block. */
106 int *containing_rgn = NULL;
107
108 /* ebb_head [i] - is index in rgn_bb_table of the head basic block of i'th ebb.
109 Currently we can get a ebb only through splitting of currently
110 scheduling block, therefore, we don't need ebb_head array for every region,
111 hence, its sufficient to hold it for current one only. */
112 int *ebb_head = NULL;
113
114 /* The minimum probability of reaching a source block so that it will be
115 considered for speculative scheduling. */
116 static int min_spec_prob;
117
118 static void find_single_block_region (bool);
119 static void find_rgns (void);
120 static bool too_large (int, int *, int *);
121
122 /* Blocks of the current region being scheduled. */
123 int current_nr_blocks;
124 int current_blocks;
125
126 /* A speculative motion requires checking live information on the path
127 from 'source' to 'target'. The split blocks are those to be checked.
128 After a speculative motion, live information should be modified in
129 the 'update' blocks.
130
131 Lists of split and update blocks for each candidate of the current
132 target are in array bblst_table. */
133 static basic_block *bblst_table;
134 static int bblst_size, bblst_last;
135
136 /* Arrays that hold the DFA state at the end of a basic block, to re-use
137 as the initial state at the start of successor blocks. The BB_STATE
138 array holds the actual DFA state, and BB_STATE_ARRAY[I] is a pointer
139 into BB_STATE for basic block I. FIXME: This should be a vec. */
140 static char *bb_state_array = NULL;
141 static state_t *bb_state = NULL;
142
143 /* Target info declarations.
144
145 The block currently being scheduled is referred to as the "target" block,
146 while other blocks in the region from which insns can be moved to the
147 target are called "source" blocks. The candidate structure holds info
148 about such sources: are they valid? Speculative? Etc. */
149 struct bblst
150 {
151 basic_block *first_member;
152 int nr_members;
153 };
154
155 struct candidate
156 {
157 char is_valid;
158 char is_speculative;
159 int src_prob;
160 bblst split_bbs;
161 bblst update_bbs;
162 };
163
164 static candidate *candidate_table;
165 #define IS_VALID(src) (candidate_table[src].is_valid)
166 #define IS_SPECULATIVE(src) (candidate_table[src].is_speculative)
167 #define IS_SPECULATIVE_INSN(INSN) \
168 (IS_SPECULATIVE (BLOCK_TO_BB (BLOCK_NUM (INSN))))
169 #define SRC_PROB(src) ( candidate_table[src].src_prob )
170
171 /* The bb being currently scheduled. */
172 int target_bb;
173
174 /* List of edges. */
175 struct edgelst
176 {
177 edge *first_member;
178 int nr_members;
179 };
180
181 static edge *edgelst_table;
182 static int edgelst_last;
183
184 static void extract_edgelst (sbitmap, edgelst *);
185
186 /* Target info functions. */
187 static void split_edges (int, int, edgelst *);
188 static void compute_trg_info (int);
189 void debug_candidate (int);
190 void debug_candidates (int);
191
192 /* Dominators array: dom[i] contains the sbitmap of dominators of
193 bb i in the region. */
194 static sbitmap *dom;
195
196 /* bb 0 is the only region entry. */
197 #define IS_RGN_ENTRY(bb) (!bb)
198
199 /* Is bb_src dominated by bb_trg. */
200 #define IS_DOMINATED(bb_src, bb_trg) \
201 ( bitmap_bit_p (dom[bb_src], bb_trg) )
202
203 /* Probability: Prob[i] is an int in [0, REG_BR_PROB_BASE] which is
204 the probability of bb i relative to the region entry. */
205 static int *prob;
206
207 /* Bit-set of edges, where bit i stands for edge i. */
208 typedef sbitmap edgeset;
209
210 /* Number of edges in the region. */
211 static int rgn_nr_edges;
212
213 /* Array of size rgn_nr_edges. */
214 static edge *rgn_edges;
215
216 /* Mapping from each edge in the graph to its number in the rgn. */
217 #define EDGE_TO_BIT(edge) ((int)(size_t)(edge)->aux)
218 #define SET_EDGE_TO_BIT(edge,nr) ((edge)->aux = (void *)(size_t)(nr))
219
220 /* The split edges of a source bb is different for each target
221 bb. In order to compute this efficiently, the 'potential-split edges'
222 are computed for each bb prior to scheduling a region. This is actually
223 the split edges of each bb relative to the region entry.
224
225 pot_split[bb] is the set of potential split edges of bb. */
226 static edgeset *pot_split;
227
228 /* For every bb, a set of its ancestor edges. */
229 static edgeset *ancestor_edges;
230
231 #define INSN_PROBABILITY(INSN) (SRC_PROB (BLOCK_TO_BB (BLOCK_NUM (INSN))))
232
233 /* Speculative scheduling functions. */
234 static int check_live_1 (int, rtx);
235 static void update_live_1 (int, rtx);
236 static int is_pfree (rtx, int, int);
237 static int find_conditional_protection (rtx_insn *, int);
238 static int is_conditionally_protected (rtx, int, int);
239 static int is_prisky (rtx, int, int);
240 static int is_exception_free (rtx_insn *, int, int);
241
242 static bool sets_likely_spilled (rtx);
243 static void sets_likely_spilled_1 (rtx, const_rtx, void *);
244 static void add_branch_dependences (rtx_insn *, rtx_insn *);
245 static void compute_block_dependences (int);
246
247 static void schedule_region (int);
248 static void concat_insn_mem_list (rtx_insn_list *, rtx_expr_list *,
249 rtx_insn_list **, rtx_expr_list **);
250 static void propagate_deps (int, class deps_desc *);
251 static void free_pending_lists (void);
252
253 /* Functions for construction of the control flow graph. */
254
255 /* Return 1 if control flow graph should not be constructed, 0 otherwise.
256
257 We decide not to build the control flow graph if there is possibly more
258 than one entry to the function, if computed branches exist, if we
259 have nonlocal gotos, or if we have an unreachable loop. */
260
261 static int
is_cfg_nonregular(void)262 is_cfg_nonregular (void)
263 {
264 basic_block b;
265 rtx_insn *insn;
266
267 /* If we have a label that could be the target of a nonlocal goto, then
268 the cfg is not well structured. */
269 if (nonlocal_goto_handler_labels)
270 return 1;
271
272 /* If we have any forced labels, then the cfg is not well structured. */
273 if (forced_labels)
274 return 1;
275
276 /* If we have exception handlers, then we consider the cfg not well
277 structured. ?!? We should be able to handle this now that we
278 compute an accurate cfg for EH. */
279 if (current_function_has_exception_handlers ())
280 return 1;
281
282 /* If we have insns which refer to labels as non-jumped-to operands,
283 then we consider the cfg not well structured. */
284 FOR_EACH_BB_FN (b, cfun)
285 FOR_BB_INSNS (b, insn)
286 {
287 rtx note, set, dest;
288 rtx_insn *next;
289
290 /* If this function has a computed jump, then we consider the cfg
291 not well structured. */
292 if (JUMP_P (insn) && computed_jump_p (insn))
293 return 1;
294
295 if (!INSN_P (insn))
296 continue;
297
298 note = find_reg_note (insn, REG_LABEL_OPERAND, NULL_RTX);
299 if (note == NULL_RTX)
300 continue;
301
302 /* For that label not to be seen as a referred-to label, this
303 must be a single-set which is feeding a jump *only*. This
304 could be a conditional jump with the label split off for
305 machine-specific reasons or a casesi/tablejump. */
306 next = next_nonnote_insn (insn);
307 if (next == NULL_RTX
308 || !JUMP_P (next)
309 || (JUMP_LABEL (next) != XEXP (note, 0)
310 && find_reg_note (next, REG_LABEL_TARGET,
311 XEXP (note, 0)) == NULL_RTX)
312 || BLOCK_FOR_INSN (insn) != BLOCK_FOR_INSN (next))
313 return 1;
314
315 set = single_set (insn);
316 if (set == NULL_RTX)
317 return 1;
318
319 dest = SET_DEST (set);
320 if (!REG_P (dest) || !dead_or_set_p (next, dest))
321 return 1;
322 }
323
324 /* Unreachable loops with more than one basic block are detected
325 during the DFS traversal in find_rgns.
326
327 Unreachable loops with a single block are detected here. This
328 test is redundant with the one in find_rgns, but it's much
329 cheaper to go ahead and catch the trivial case here. */
330 FOR_EACH_BB_FN (b, cfun)
331 {
332 if (EDGE_COUNT (b->preds) == 0
333 || (single_pred_p (b)
334 && single_pred (b) == b))
335 return 1;
336 }
337
338 /* All the tests passed. Consider the cfg well structured. */
339 return 0;
340 }
341
342 /* Extract list of edges from a bitmap containing EDGE_TO_BIT bits. */
343
344 static void
extract_edgelst(sbitmap set,edgelst * el)345 extract_edgelst (sbitmap set, edgelst *el)
346 {
347 unsigned int i = 0;
348 sbitmap_iterator sbi;
349
350 /* edgelst table space is reused in each call to extract_edgelst. */
351 edgelst_last = 0;
352
353 el->first_member = &edgelst_table[edgelst_last];
354 el->nr_members = 0;
355
356 /* Iterate over each word in the bitset. */
357 EXECUTE_IF_SET_IN_BITMAP (set, 0, i, sbi)
358 {
359 edgelst_table[edgelst_last++] = rgn_edges[i];
360 el->nr_members++;
361 }
362 }
363
364 /* Functions for the construction of regions. */
365
366 /* Print the regions, for debugging purposes. Callable from debugger. */
367
368 DEBUG_FUNCTION void
debug_regions(void)369 debug_regions (void)
370 {
371 int rgn, bb;
372
373 fprintf (sched_dump, "\n;; ------------ REGIONS ----------\n\n");
374 for (rgn = 0; rgn < nr_regions; rgn++)
375 {
376 fprintf (sched_dump, ";;\trgn %d nr_blocks %d:\n", rgn,
377 rgn_table[rgn].rgn_nr_blocks);
378 fprintf (sched_dump, ";;\tbb/block: ");
379
380 /* We don't have ebb_head initialized yet, so we can't use
381 BB_TO_BLOCK (). */
382 current_blocks = RGN_BLOCKS (rgn);
383
384 for (bb = 0; bb < rgn_table[rgn].rgn_nr_blocks; bb++)
385 fprintf (sched_dump, " %d/%d ", bb, rgn_bb_table[current_blocks + bb]);
386
387 fprintf (sched_dump, "\n\n");
388 }
389 }
390
391 /* Print the region's basic blocks. */
392
393 DEBUG_FUNCTION void
debug_region(int rgn)394 debug_region (int rgn)
395 {
396 int bb;
397
398 fprintf (stderr, "\n;; ------------ REGION %d ----------\n\n", rgn);
399 fprintf (stderr, ";;\trgn %d nr_blocks %d:\n", rgn,
400 rgn_table[rgn].rgn_nr_blocks);
401 fprintf (stderr, ";;\tbb/block: ");
402
403 /* We don't have ebb_head initialized yet, so we can't use
404 BB_TO_BLOCK (). */
405 current_blocks = RGN_BLOCKS (rgn);
406
407 for (bb = 0; bb < rgn_table[rgn].rgn_nr_blocks; bb++)
408 fprintf (stderr, " %d/%d ", bb, rgn_bb_table[current_blocks + bb]);
409
410 fprintf (stderr, "\n\n");
411
412 for (bb = 0; bb < rgn_table[rgn].rgn_nr_blocks; bb++)
413 {
414 dump_bb (stderr,
415 BASIC_BLOCK_FOR_FN (cfun, rgn_bb_table[current_blocks + bb]),
416 0, TDF_SLIM | TDF_BLOCKS);
417 fprintf (stderr, "\n");
418 }
419
420 fprintf (stderr, "\n");
421
422 }
423
424 /* True when a bb with index BB_INDEX contained in region RGN. */
425 static bool
bb_in_region_p(int bb_index,int rgn)426 bb_in_region_p (int bb_index, int rgn)
427 {
428 int i;
429
430 for (i = 0; i < rgn_table[rgn].rgn_nr_blocks; i++)
431 if (rgn_bb_table[current_blocks + i] == bb_index)
432 return true;
433
434 return false;
435 }
436
437 /* Dump region RGN to file F using dot syntax. */
438 void
dump_region_dot(FILE * f,int rgn)439 dump_region_dot (FILE *f, int rgn)
440 {
441 int i;
442
443 fprintf (f, "digraph Region_%d {\n", rgn);
444
445 /* We don't have ebb_head initialized yet, so we can't use
446 BB_TO_BLOCK (). */
447 current_blocks = RGN_BLOCKS (rgn);
448
449 for (i = 0; i < rgn_table[rgn].rgn_nr_blocks; i++)
450 {
451 edge e;
452 edge_iterator ei;
453 int src_bb_num = rgn_bb_table[current_blocks + i];
454 basic_block bb = BASIC_BLOCK_FOR_FN (cfun, src_bb_num);
455
456 FOR_EACH_EDGE (e, ei, bb->succs)
457 if (bb_in_region_p (e->dest->index, rgn))
458 fprintf (f, "\t%d -> %d\n", src_bb_num, e->dest->index);
459 }
460 fprintf (f, "}\n");
461 }
462
463 /* The same, but first open a file specified by FNAME. */
464 void
dump_region_dot_file(const char * fname,int rgn)465 dump_region_dot_file (const char *fname, int rgn)
466 {
467 FILE *f = fopen (fname, "wt");
468 dump_region_dot (f, rgn);
469 fclose (f);
470 }
471
472 /* Build a single block region for each basic block in the function.
473 This allows for using the same code for interblock and basic block
474 scheduling. */
475
476 static void
find_single_block_region(bool ebbs_p)477 find_single_block_region (bool ebbs_p)
478 {
479 basic_block bb, ebb_start;
480 int i = 0;
481
482 nr_regions = 0;
483
484 if (ebbs_p) {
485 int probability_cutoff;
486 if (profile_info && profile_status_for_fn (cfun) == PROFILE_READ)
487 probability_cutoff = param_tracer_min_branch_probability_feedback;
488 else
489 probability_cutoff = param_tracer_min_branch_probability;
490 probability_cutoff = REG_BR_PROB_BASE / 100 * probability_cutoff;
491
492 FOR_EACH_BB_FN (ebb_start, cfun)
493 {
494 RGN_NR_BLOCKS (nr_regions) = 0;
495 RGN_BLOCKS (nr_regions) = i;
496 RGN_DONT_CALC_DEPS (nr_regions) = 0;
497 RGN_HAS_REAL_EBB (nr_regions) = 0;
498
499 for (bb = ebb_start; ; bb = bb->next_bb)
500 {
501 edge e;
502
503 rgn_bb_table[i] = bb->index;
504 RGN_NR_BLOCKS (nr_regions)++;
505 CONTAINING_RGN (bb->index) = nr_regions;
506 BLOCK_TO_BB (bb->index) = i - RGN_BLOCKS (nr_regions);
507 i++;
508
509 if (bb->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
510 || LABEL_P (BB_HEAD (bb->next_bb)))
511 break;
512
513 e = find_fallthru_edge (bb->succs);
514 if (! e)
515 break;
516 if (e->probability.initialized_p ()
517 && e->probability.to_reg_br_prob_base () <= probability_cutoff)
518 break;
519 }
520
521 ebb_start = bb;
522 nr_regions++;
523 }
524 }
525 else
526 FOR_EACH_BB_FN (bb, cfun)
527 {
528 rgn_bb_table[nr_regions] = bb->index;
529 RGN_NR_BLOCKS (nr_regions) = 1;
530 RGN_BLOCKS (nr_regions) = nr_regions;
531 RGN_DONT_CALC_DEPS (nr_regions) = 0;
532 RGN_HAS_REAL_EBB (nr_regions) = 0;
533
534 CONTAINING_RGN (bb->index) = nr_regions;
535 BLOCK_TO_BB (bb->index) = 0;
536 nr_regions++;
537 }
538 }
539
540 /* Estimate number of the insns in the BB. */
541 static int
rgn_estimate_number_of_insns(basic_block bb)542 rgn_estimate_number_of_insns (basic_block bb)
543 {
544 int count;
545
546 count = INSN_LUID (BB_END (bb)) - INSN_LUID (BB_HEAD (bb));
547
548 if (MAY_HAVE_DEBUG_INSNS)
549 {
550 rtx_insn *insn;
551
552 FOR_BB_INSNS (bb, insn)
553 if (DEBUG_INSN_P (insn))
554 count--;
555 }
556
557 return count;
558 }
559
560 /* Update number of blocks and the estimate for number of insns
561 in the region. Return true if the region is "too large" for interblock
562 scheduling (compile time considerations). */
563
564 static bool
too_large(int block,int * num_bbs,int * num_insns)565 too_large (int block, int *num_bbs, int *num_insns)
566 {
567 (*num_bbs)++;
568 (*num_insns) += (common_sched_info->estimate_number_of_insns
569 (BASIC_BLOCK_FOR_FN (cfun, block)));
570
571 return ((*num_bbs > param_max_sched_region_blocks)
572 || (*num_insns > param_max_sched_region_insns));
573 }
574
575 /* Update_loop_relations(blk, hdr): Check if the loop headed by max_hdr[blk]
576 is still an inner loop. Put in max_hdr[blk] the header of the most inner
577 loop containing blk. */
578 #define UPDATE_LOOP_RELATIONS(blk, hdr) \
579 { \
580 if (max_hdr[blk] == -1) \
581 max_hdr[blk] = hdr; \
582 else if (dfs_nr[max_hdr[blk]] > dfs_nr[hdr]) \
583 bitmap_clear_bit (inner, hdr); \
584 else if (dfs_nr[max_hdr[blk]] < dfs_nr[hdr]) \
585 { \
586 bitmap_clear_bit (inner,max_hdr[blk]); \
587 max_hdr[blk] = hdr; \
588 } \
589 }
590
591 /* Find regions for interblock scheduling.
592
593 A region for scheduling can be:
594
595 * A loop-free procedure, or
596
597 * A reducible inner loop, or
598
599 * A basic block not contained in any other region.
600
601 ?!? In theory we could build other regions based on extended basic
602 blocks or reverse extended basic blocks. Is it worth the trouble?
603
604 Loop blocks that form a region are put into the region's block list
605 in topological order.
606
607 This procedure stores its results into the following global (ick) variables
608
609 * rgn_nr
610 * rgn_table
611 * rgn_bb_table
612 * block_to_bb
613 * containing region
614
615 We use dominator relationships to avoid making regions out of non-reducible
616 loops.
617
618 This procedure needs to be converted to work on pred/succ lists instead
619 of edge tables. That would simplify it somewhat. */
620
621 static void
haifa_find_rgns(void)622 haifa_find_rgns (void)
623 {
624 int *max_hdr, *dfs_nr, *degree;
625 char no_loops = 1;
626 int node, child, loop_head, i, head, tail;
627 int count = 0, sp, idx = 0;
628 edge_iterator current_edge;
629 edge_iterator *stack;
630 int num_bbs, num_insns, unreachable;
631 int too_large_failure;
632 basic_block bb;
633
634 /* Perform a DFS traversal of the cfg. Identify loop headers, inner loops
635 and a mapping from block to its loop header (if the block is contained
636 in a loop, else -1).
637
638 Store results in HEADER, INNER, and MAX_HDR respectively, these will
639 be used as inputs to the second traversal.
640
641 STACK, SP and DFS_NR are only used during the first traversal. */
642
643 /* Allocate and initialize variables for the first traversal. */
644 max_hdr = XNEWVEC (int, last_basic_block_for_fn (cfun));
645 dfs_nr = XCNEWVEC (int, last_basic_block_for_fn (cfun));
646 stack = XNEWVEC (edge_iterator, n_edges_for_fn (cfun));
647
648 /* Note if a block is a natural inner loop header. */
649 auto_sbitmap inner (last_basic_block_for_fn (cfun));
650 bitmap_ones (inner);
651
652 /* Note if a block is a natural loop header. */
653 auto_sbitmap header (last_basic_block_for_fn (cfun));
654 bitmap_clear (header);
655
656 /* Note if a block is in the block queue. */
657 auto_sbitmap in_queue (last_basic_block_for_fn (cfun));
658 bitmap_clear (in_queue);
659
660 /* Note if a block is in the block queue. */
661 auto_sbitmap in_stack (last_basic_block_for_fn (cfun));
662 bitmap_clear (in_stack);
663
664 for (i = 0; i < last_basic_block_for_fn (cfun); i++)
665 max_hdr[i] = -1;
666
667 #define EDGE_PASSED(E) (ei_end_p ((E)) || ei_edge ((E))->aux)
668 #define SET_EDGE_PASSED(E) (ei_edge ((E))->aux = ei_edge ((E)))
669
670 /* DFS traversal to find inner loops in the cfg. */
671
672 current_edge = ei_start (single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun))->succs);
673 sp = -1;
674
675 while (1)
676 {
677 if (EDGE_PASSED (current_edge))
678 {
679 /* We have reached a leaf node or a node that was already
680 processed. Pop edges off the stack until we find
681 an edge that has not yet been processed. */
682 while (sp >= 0 && EDGE_PASSED (current_edge))
683 {
684 /* Pop entry off the stack. */
685 current_edge = stack[sp--];
686 node = ei_edge (current_edge)->src->index;
687 gcc_assert (node != ENTRY_BLOCK);
688 child = ei_edge (current_edge)->dest->index;
689 gcc_assert (child != EXIT_BLOCK);
690 bitmap_clear_bit (in_stack, child);
691 if (max_hdr[child] >= 0 && bitmap_bit_p (in_stack, max_hdr[child]))
692 UPDATE_LOOP_RELATIONS (node, max_hdr[child]);
693 ei_next (¤t_edge);
694 }
695
696 /* See if have finished the DFS tree traversal. */
697 if (sp < 0 && EDGE_PASSED (current_edge))
698 break;
699
700 /* Nope, continue the traversal with the popped node. */
701 continue;
702 }
703
704 /* Process a node. */
705 node = ei_edge (current_edge)->src->index;
706 gcc_assert (node != ENTRY_BLOCK);
707 bitmap_set_bit (in_stack, node);
708 dfs_nr[node] = ++count;
709
710 /* We don't traverse to the exit block. */
711 child = ei_edge (current_edge)->dest->index;
712 if (child == EXIT_BLOCK)
713 {
714 SET_EDGE_PASSED (current_edge);
715 ei_next (¤t_edge);
716 continue;
717 }
718
719 /* If the successor is in the stack, then we've found a loop.
720 Mark the loop, if it is not a natural loop, then it will
721 be rejected during the second traversal. */
722 if (bitmap_bit_p (in_stack, child))
723 {
724 no_loops = 0;
725 bitmap_set_bit (header, child);
726 UPDATE_LOOP_RELATIONS (node, child);
727 SET_EDGE_PASSED (current_edge);
728 ei_next (¤t_edge);
729 continue;
730 }
731
732 /* If the child was already visited, then there is no need to visit
733 it again. Just update the loop relationships and restart
734 with a new edge. */
735 if (dfs_nr[child])
736 {
737 if (max_hdr[child] >= 0 && bitmap_bit_p (in_stack, max_hdr[child]))
738 UPDATE_LOOP_RELATIONS (node, max_hdr[child]);
739 SET_EDGE_PASSED (current_edge);
740 ei_next (¤t_edge);
741 continue;
742 }
743
744 /* Push an entry on the stack and continue DFS traversal. */
745 stack[++sp] = current_edge;
746 SET_EDGE_PASSED (current_edge);
747 current_edge = ei_start (ei_edge (current_edge)->dest->succs);
748 }
749
750 /* Reset ->aux field used by EDGE_PASSED. */
751 FOR_ALL_BB_FN (bb, cfun)
752 {
753 edge_iterator ei;
754 edge e;
755 FOR_EACH_EDGE (e, ei, bb->succs)
756 e->aux = NULL;
757 }
758
759
760 /* Another check for unreachable blocks. The earlier test in
761 is_cfg_nonregular only finds unreachable blocks that do not
762 form a loop.
763
764 The DFS traversal will mark every block that is reachable from
765 the entry node by placing a nonzero value in dfs_nr. Thus if
766 dfs_nr is zero for any block, then it must be unreachable. */
767 unreachable = 0;
768 FOR_EACH_BB_FN (bb, cfun)
769 if (dfs_nr[bb->index] == 0)
770 {
771 unreachable = 1;
772 break;
773 }
774
775 /* Gross. To avoid wasting memory, the second pass uses the dfs_nr array
776 to hold degree counts. */
777 degree = dfs_nr;
778
779 FOR_EACH_BB_FN (bb, cfun)
780 degree[bb->index] = EDGE_COUNT (bb->preds);
781
782 /* Do not perform region scheduling if there are any unreachable
783 blocks. */
784 if (!unreachable)
785 {
786 int *queue, *degree1 = NULL;
787 /* We use EXTENDED_RGN_HEADER as an addition to HEADER and put
788 there basic blocks, which are forced to be region heads.
789 This is done to try to assemble few smaller regions
790 from a too_large region. */
791 sbitmap extended_rgn_header = NULL;
792 bool extend_regions_p;
793
794 if (no_loops)
795 bitmap_set_bit (header, 0);
796
797 /* Second traversal:find reducible inner loops and topologically sort
798 block of each region. */
799
800 queue = XNEWVEC (int, n_basic_blocks_for_fn (cfun));
801
802 extend_regions_p = param_max_sched_extend_regions_iters > 0;
803 if (extend_regions_p)
804 {
805 degree1 = XNEWVEC (int, last_basic_block_for_fn (cfun));
806 extended_rgn_header =
807 sbitmap_alloc (last_basic_block_for_fn (cfun));
808 bitmap_clear (extended_rgn_header);
809 }
810
811 /* Find blocks which are inner loop headers. We still have non-reducible
812 loops to consider at this point. */
813 FOR_EACH_BB_FN (bb, cfun)
814 {
815 if (bitmap_bit_p (header, bb->index) && bitmap_bit_p (inner, bb->index))
816 {
817 edge e;
818 edge_iterator ei;
819 basic_block jbb;
820
821 /* Now check that the loop is reducible. We do this separate
822 from finding inner loops so that we do not find a reducible
823 loop which contains an inner non-reducible loop.
824
825 A simple way to find reducible/natural loops is to verify
826 that each block in the loop is dominated by the loop
827 header.
828
829 If there exists a block that is not dominated by the loop
830 header, then the block is reachable from outside the loop
831 and thus the loop is not a natural loop. */
832 FOR_EACH_BB_FN (jbb, cfun)
833 {
834 /* First identify blocks in the loop, except for the loop
835 entry block. */
836 if (bb->index == max_hdr[jbb->index] && bb != jbb)
837 {
838 /* Now verify that the block is dominated by the loop
839 header. */
840 if (!dominated_by_p (CDI_DOMINATORS, jbb, bb))
841 break;
842 }
843 }
844
845 /* If we exited the loop early, then I is the header of
846 a non-reducible loop and we should quit processing it
847 now. */
848 if (jbb != EXIT_BLOCK_PTR_FOR_FN (cfun))
849 continue;
850
851 /* I is a header of an inner loop, or block 0 in a subroutine
852 with no loops at all. */
853 head = tail = -1;
854 too_large_failure = 0;
855 loop_head = max_hdr[bb->index];
856
857 if (extend_regions_p)
858 /* We save degree in case when we meet a too_large region
859 and cancel it. We need a correct degree later when
860 calling extend_rgns. */
861 memcpy (degree1, degree,
862 last_basic_block_for_fn (cfun) * sizeof (int));
863
864 /* Decrease degree of all I's successors for topological
865 ordering. */
866 FOR_EACH_EDGE (e, ei, bb->succs)
867 if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun))
868 --degree[e->dest->index];
869
870 /* Estimate # insns, and count # blocks in the region. */
871 num_bbs = 1;
872 num_insns = common_sched_info->estimate_number_of_insns (bb);
873
874 /* Find all loop latches (blocks with back edges to the loop
875 header) or all the leaf blocks in the cfg has no loops.
876
877 Place those blocks into the queue. */
878 if (no_loops)
879 {
880 FOR_EACH_BB_FN (jbb, cfun)
881 /* Leaf nodes have only a single successor which must
882 be EXIT_BLOCK. */
883 if (single_succ_p (jbb)
884 && single_succ (jbb) == EXIT_BLOCK_PTR_FOR_FN (cfun))
885 {
886 queue[++tail] = jbb->index;
887 bitmap_set_bit (in_queue, jbb->index);
888
889 if (too_large (jbb->index, &num_bbs, &num_insns))
890 {
891 too_large_failure = 1;
892 break;
893 }
894 }
895 }
896 else
897 {
898 edge e;
899
900 FOR_EACH_EDGE (e, ei, bb->preds)
901 {
902 if (e->src == ENTRY_BLOCK_PTR_FOR_FN (cfun))
903 continue;
904
905 node = e->src->index;
906
907 if (max_hdr[node] == loop_head && node != bb->index)
908 {
909 /* This is a loop latch. */
910 queue[++tail] = node;
911 bitmap_set_bit (in_queue, node);
912
913 if (too_large (node, &num_bbs, &num_insns))
914 {
915 too_large_failure = 1;
916 break;
917 }
918 }
919 }
920 }
921
922 /* Now add all the blocks in the loop to the queue.
923
924 We know the loop is a natural loop; however the algorithm
925 above will not always mark certain blocks as being in the
926 loop. Consider:
927 node children
928 a b,c
929 b c
930 c a,d
931 d b
932
933 The algorithm in the DFS traversal may not mark B & D as part
934 of the loop (i.e. they will not have max_hdr set to A).
935
936 We know they cannot be loop latches (else they would have
937 had max_hdr set since they'd have a backedge to a dominator
938 block). So we don't need them on the initial queue.
939
940 We know they are part of the loop because they are dominated
941 by the loop header and can be reached by a backwards walk of
942 the edges starting with nodes on the initial queue.
943
944 It is safe and desirable to include those nodes in the
945 loop/scheduling region. To do so we would need to decrease
946 the degree of a node if it is the target of a backedge
947 within the loop itself as the node is placed in the queue.
948
949 We do not do this because I'm not sure that the actual
950 scheduling code will properly handle this case. ?!? */
951
952 while (head < tail && !too_large_failure)
953 {
954 edge e;
955 child = queue[++head];
956
957 FOR_EACH_EDGE (e, ei,
958 BASIC_BLOCK_FOR_FN (cfun, child)->preds)
959 {
960 node = e->src->index;
961
962 /* See discussion above about nodes not marked as in
963 this loop during the initial DFS traversal. */
964 if (e->src == ENTRY_BLOCK_PTR_FOR_FN (cfun)
965 || max_hdr[node] != loop_head)
966 {
967 tail = -1;
968 break;
969 }
970 else if (!bitmap_bit_p (in_queue, node) && node != bb->index)
971 {
972 queue[++tail] = node;
973 bitmap_set_bit (in_queue, node);
974
975 if (too_large (node, &num_bbs, &num_insns))
976 {
977 too_large_failure = 1;
978 break;
979 }
980 }
981 }
982 }
983
984 if (tail >= 0 && !too_large_failure)
985 {
986 /* Place the loop header into list of region blocks. */
987 degree[bb->index] = -1;
988 rgn_bb_table[idx] = bb->index;
989 RGN_NR_BLOCKS (nr_regions) = num_bbs;
990 RGN_BLOCKS (nr_regions) = idx++;
991 RGN_DONT_CALC_DEPS (nr_regions) = 0;
992 RGN_HAS_REAL_EBB (nr_regions) = 0;
993 CONTAINING_RGN (bb->index) = nr_regions;
994 BLOCK_TO_BB (bb->index) = count = 0;
995
996 /* Remove blocks from queue[] when their in degree
997 becomes zero. Repeat until no blocks are left on the
998 list. This produces a topological list of blocks in
999 the region. */
1000 while (tail >= 0)
1001 {
1002 if (head < 0)
1003 head = tail;
1004 child = queue[head];
1005 if (degree[child] == 0)
1006 {
1007 edge e;
1008
1009 degree[child] = -1;
1010 rgn_bb_table[idx++] = child;
1011 BLOCK_TO_BB (child) = ++count;
1012 CONTAINING_RGN (child) = nr_regions;
1013 queue[head] = queue[tail--];
1014
1015 FOR_EACH_EDGE (e, ei,
1016 BASIC_BLOCK_FOR_FN (cfun,
1017 child)->succs)
1018 if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun))
1019 --degree[e->dest->index];
1020 }
1021 else
1022 --head;
1023 }
1024 ++nr_regions;
1025 }
1026 else if (extend_regions_p)
1027 {
1028 /* Restore DEGREE. */
1029 int *t = degree;
1030
1031 degree = degree1;
1032 degree1 = t;
1033
1034 /* And force successors of BB to be region heads.
1035 This may provide several smaller regions instead
1036 of one too_large region. */
1037 FOR_EACH_EDGE (e, ei, bb->succs)
1038 if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun))
1039 bitmap_set_bit (extended_rgn_header, e->dest->index);
1040 }
1041 }
1042 }
1043 free (queue);
1044
1045 if (extend_regions_p)
1046 {
1047 free (degree1);
1048
1049 bitmap_ior (header, header, extended_rgn_header);
1050 sbitmap_free (extended_rgn_header);
1051
1052 extend_rgns (degree, &idx, header, max_hdr);
1053 }
1054 }
1055
1056 /* Any block that did not end up in a region is placed into a region
1057 by itself. */
1058 FOR_EACH_BB_FN (bb, cfun)
1059 if (degree[bb->index] >= 0)
1060 {
1061 rgn_bb_table[idx] = bb->index;
1062 RGN_NR_BLOCKS (nr_regions) = 1;
1063 RGN_BLOCKS (nr_regions) = idx++;
1064 RGN_DONT_CALC_DEPS (nr_regions) = 0;
1065 RGN_HAS_REAL_EBB (nr_regions) = 0;
1066 CONTAINING_RGN (bb->index) = nr_regions++;
1067 BLOCK_TO_BB (bb->index) = 0;
1068 }
1069
1070 free (max_hdr);
1071 free (degree);
1072 free (stack);
1073 }
1074
1075
1076 /* Wrapper function.
1077 If FLAG_SEL_SCHED_PIPELINING is set, then use custom function to form
1078 regions. Otherwise just call find_rgns_haifa. */
1079 static void
find_rgns(void)1080 find_rgns (void)
1081 {
1082 if (sel_sched_p () && flag_sel_sched_pipelining)
1083 sel_find_rgns ();
1084 else
1085 haifa_find_rgns ();
1086 }
1087
1088 static int gather_region_statistics (int **);
1089 static void print_region_statistics (int *, int, int *, int);
1090
1091 /* Calculate the histogram that shows the number of regions having the
1092 given number of basic blocks, and store it in the RSP array. Return
1093 the size of this array. */
1094 static int
gather_region_statistics(int ** rsp)1095 gather_region_statistics (int **rsp)
1096 {
1097 int i, *a = 0, a_sz = 0;
1098
1099 /* a[i] is the number of regions that have (i + 1) basic blocks. */
1100 for (i = 0; i < nr_regions; i++)
1101 {
1102 int nr_blocks = RGN_NR_BLOCKS (i);
1103
1104 gcc_assert (nr_blocks >= 1);
1105
1106 if (nr_blocks > a_sz)
1107 {
1108 a = XRESIZEVEC (int, a, nr_blocks);
1109 do
1110 a[a_sz++] = 0;
1111 while (a_sz != nr_blocks);
1112 }
1113
1114 a[nr_blocks - 1]++;
1115 }
1116
1117 *rsp = a;
1118 return a_sz;
1119 }
1120
1121 /* Print regions statistics. S1 and S2 denote the data before and after
1122 calling extend_rgns, respectively. */
1123 static void
print_region_statistics(int * s1,int s1_sz,int * s2,int s2_sz)1124 print_region_statistics (int *s1, int s1_sz, int *s2, int s2_sz)
1125 {
1126 int i;
1127
1128 /* We iterate until s2_sz because extend_rgns does not decrease
1129 the maximal region size. */
1130 for (i = 1; i < s2_sz; i++)
1131 {
1132 int n1, n2;
1133
1134 n2 = s2[i];
1135
1136 if (n2 == 0)
1137 continue;
1138
1139 if (i >= s1_sz)
1140 n1 = 0;
1141 else
1142 n1 = s1[i];
1143
1144 fprintf (sched_dump, ";; Region extension statistics: size %d: " \
1145 "was %d + %d more\n", i + 1, n1, n2 - n1);
1146 }
1147 }
1148
1149 /* Extend regions.
1150 DEGREE - Array of incoming edge count, considering only
1151 the edges, that don't have their sources in formed regions yet.
1152 IDXP - pointer to the next available index in rgn_bb_table.
1153 HEADER - set of all region heads.
1154 LOOP_HDR - mapping from block to the containing loop
1155 (two blocks can reside within one region if they have
1156 the same loop header). */
1157 void
extend_rgns(int * degree,int * idxp,sbitmap header,int * loop_hdr)1158 extend_rgns (int *degree, int *idxp, sbitmap header, int *loop_hdr)
1159 {
1160 int *order, i, rescan = 0, idx = *idxp, iter = 0, max_iter, *max_hdr;
1161 int nblocks = n_basic_blocks_for_fn (cfun) - NUM_FIXED_BLOCKS;
1162
1163 max_iter = param_max_sched_extend_regions_iters;
1164
1165 max_hdr = XNEWVEC (int, last_basic_block_for_fn (cfun));
1166
1167 order = XNEWVEC (int, last_basic_block_for_fn (cfun));
1168 post_order_compute (order, false, false);
1169
1170 for (i = nblocks - 1; i >= 0; i--)
1171 {
1172 int bbn = order[i];
1173 if (degree[bbn] >= 0)
1174 {
1175 max_hdr[bbn] = bbn;
1176 rescan = 1;
1177 }
1178 else
1179 /* This block already was processed in find_rgns. */
1180 max_hdr[bbn] = -1;
1181 }
1182
1183 /* The idea is to topologically walk through CFG in top-down order.
1184 During the traversal, if all the predecessors of a node are
1185 marked to be in the same region (they all have the same max_hdr),
1186 then current node is also marked to be a part of that region.
1187 Otherwise the node starts its own region.
1188 CFG should be traversed until no further changes are made. On each
1189 iteration the set of the region heads is extended (the set of those
1190 blocks that have max_hdr[bbi] == bbi). This set is upper bounded by the
1191 set of all basic blocks, thus the algorithm is guaranteed to
1192 terminate. */
1193
1194 while (rescan && iter < max_iter)
1195 {
1196 rescan = 0;
1197
1198 for (i = nblocks - 1; i >= 0; i--)
1199 {
1200 edge e;
1201 edge_iterator ei;
1202 int bbn = order[i];
1203
1204 if (max_hdr[bbn] != -1 && !bitmap_bit_p (header, bbn))
1205 {
1206 int hdr = -1;
1207
1208 FOR_EACH_EDGE (e, ei, BASIC_BLOCK_FOR_FN (cfun, bbn)->preds)
1209 {
1210 int predn = e->src->index;
1211
1212 if (predn != ENTRY_BLOCK
1213 /* If pred wasn't processed in find_rgns. */
1214 && max_hdr[predn] != -1
1215 /* And pred and bb reside in the same loop.
1216 (Or out of any loop). */
1217 && loop_hdr[bbn] == loop_hdr[predn])
1218 {
1219 if (hdr == -1)
1220 /* Then bb extends the containing region of pred. */
1221 hdr = max_hdr[predn];
1222 else if (hdr != max_hdr[predn])
1223 /* Too bad, there are at least two predecessors
1224 that reside in different regions. Thus, BB should
1225 begin its own region. */
1226 {
1227 hdr = bbn;
1228 break;
1229 }
1230 }
1231 else
1232 /* BB starts its own region. */
1233 {
1234 hdr = bbn;
1235 break;
1236 }
1237 }
1238
1239 if (hdr == bbn)
1240 {
1241 /* If BB start its own region,
1242 update set of headers with BB. */
1243 bitmap_set_bit (header, bbn);
1244 rescan = 1;
1245 }
1246 else
1247 gcc_assert (hdr != -1);
1248
1249 max_hdr[bbn] = hdr;
1250 }
1251 }
1252
1253 iter++;
1254 }
1255
1256 /* Statistics were gathered on the SPEC2000 package of tests with
1257 mainline weekly snapshot gcc-4.1-20051015 on ia64.
1258
1259 Statistics for SPECint:
1260 1 iteration : 1751 cases (38.7%)
1261 2 iterations: 2770 cases (61.3%)
1262 Blocks wrapped in regions by find_rgns without extension: 18295 blocks
1263 Blocks wrapped in regions by 2 iterations in extend_rgns: 23821 blocks
1264 (We don't count single block regions here).
1265
1266 Statistics for SPECfp:
1267 1 iteration : 621 cases (35.9%)
1268 2 iterations: 1110 cases (64.1%)
1269 Blocks wrapped in regions by find_rgns without extension: 6476 blocks
1270 Blocks wrapped in regions by 2 iterations in extend_rgns: 11155 blocks
1271 (We don't count single block regions here).
1272
1273 By default we do at most 2 iterations.
1274 This can be overridden with max-sched-extend-regions-iters parameter:
1275 0 - disable region extension,
1276 N > 0 - do at most N iterations. */
1277
1278 if (sched_verbose && iter != 0)
1279 fprintf (sched_dump, ";; Region extension iterations: %d%s\n", iter,
1280 rescan ? "... failed" : "");
1281
1282 if (!rescan && iter != 0)
1283 {
1284 int *s1 = NULL, s1_sz = 0;
1285
1286 /* Save the old statistics for later printout. */
1287 if (sched_verbose >= 6)
1288 s1_sz = gather_region_statistics (&s1);
1289
1290 /* We have succeeded. Now assemble the regions. */
1291 for (i = nblocks - 1; i >= 0; i--)
1292 {
1293 int bbn = order[i];
1294
1295 if (max_hdr[bbn] == bbn)
1296 /* BBN is a region head. */
1297 {
1298 edge e;
1299 edge_iterator ei;
1300 int num_bbs = 0, j, num_insns = 0, large;
1301
1302 large = too_large (bbn, &num_bbs, &num_insns);
1303
1304 degree[bbn] = -1;
1305 rgn_bb_table[idx] = bbn;
1306 RGN_BLOCKS (nr_regions) = idx++;
1307 RGN_DONT_CALC_DEPS (nr_regions) = 0;
1308 RGN_HAS_REAL_EBB (nr_regions) = 0;
1309 CONTAINING_RGN (bbn) = nr_regions;
1310 BLOCK_TO_BB (bbn) = 0;
1311
1312 FOR_EACH_EDGE (e, ei, BASIC_BLOCK_FOR_FN (cfun, bbn)->succs)
1313 if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun))
1314 degree[e->dest->index]--;
1315
1316 if (!large)
1317 /* Here we check whether the region is too_large. */
1318 for (j = i - 1; j >= 0; j--)
1319 {
1320 int succn = order[j];
1321 if (max_hdr[succn] == bbn)
1322 {
1323 if ((large = too_large (succn, &num_bbs, &num_insns)))
1324 break;
1325 }
1326 }
1327
1328 if (large)
1329 /* If the region is too_large, then wrap every block of
1330 the region into single block region.
1331 Here we wrap region head only. Other blocks are
1332 processed in the below cycle. */
1333 {
1334 RGN_NR_BLOCKS (nr_regions) = 1;
1335 nr_regions++;
1336 }
1337
1338 num_bbs = 1;
1339
1340 for (j = i - 1; j >= 0; j--)
1341 {
1342 int succn = order[j];
1343
1344 if (max_hdr[succn] == bbn)
1345 /* This cycle iterates over all basic blocks, that
1346 are supposed to be in the region with head BBN,
1347 and wraps them into that region (or in single
1348 block region). */
1349 {
1350 gcc_assert (degree[succn] == 0);
1351
1352 degree[succn] = -1;
1353 rgn_bb_table[idx] = succn;
1354 BLOCK_TO_BB (succn) = large ? 0 : num_bbs++;
1355 CONTAINING_RGN (succn) = nr_regions;
1356
1357 if (large)
1358 /* Wrap SUCCN into single block region. */
1359 {
1360 RGN_BLOCKS (nr_regions) = idx;
1361 RGN_NR_BLOCKS (nr_regions) = 1;
1362 RGN_DONT_CALC_DEPS (nr_regions) = 0;
1363 RGN_HAS_REAL_EBB (nr_regions) = 0;
1364 nr_regions++;
1365 }
1366
1367 idx++;
1368
1369 FOR_EACH_EDGE (e, ei,
1370 BASIC_BLOCK_FOR_FN (cfun, succn)->succs)
1371 if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun))
1372 degree[e->dest->index]--;
1373 }
1374 }
1375
1376 if (!large)
1377 {
1378 RGN_NR_BLOCKS (nr_regions) = num_bbs;
1379 nr_regions++;
1380 }
1381 }
1382 }
1383
1384 if (sched_verbose >= 6)
1385 {
1386 int *s2, s2_sz;
1387
1388 /* Get the new statistics and print the comparison with the
1389 one before calling this function. */
1390 s2_sz = gather_region_statistics (&s2);
1391 print_region_statistics (s1, s1_sz, s2, s2_sz);
1392 free (s1);
1393 free (s2);
1394 }
1395 }
1396
1397 free (order);
1398 free (max_hdr);
1399
1400 *idxp = idx;
1401 }
1402
1403 /* Functions for regions scheduling information. */
1404
1405 /* Compute dominators, probability, and potential-split-edges of bb.
1406 Assume that these values were already computed for bb's predecessors. */
1407
1408 static void
compute_dom_prob_ps(int bb)1409 compute_dom_prob_ps (int bb)
1410 {
1411 edge_iterator in_ei;
1412 edge in_edge;
1413
1414 /* We shouldn't have any real ebbs yet. */
1415 gcc_assert (ebb_head [bb] == bb + current_blocks);
1416
1417 if (IS_RGN_ENTRY (bb))
1418 {
1419 bitmap_set_bit (dom[bb], 0);
1420 prob[bb] = REG_BR_PROB_BASE;
1421 return;
1422 }
1423
1424 prob[bb] = 0;
1425
1426 /* Initialize dom[bb] to '111..1'. */
1427 bitmap_ones (dom[bb]);
1428
1429 FOR_EACH_EDGE (in_edge, in_ei,
1430 BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (bb))->preds)
1431 {
1432 int pred_bb;
1433 edge out_edge;
1434 edge_iterator out_ei;
1435
1436 if (in_edge->src == ENTRY_BLOCK_PTR_FOR_FN (cfun))
1437 continue;
1438
1439 pred_bb = BLOCK_TO_BB (in_edge->src->index);
1440 bitmap_and (dom[bb], dom[bb], dom[pred_bb]);
1441 bitmap_ior (ancestor_edges[bb],
1442 ancestor_edges[bb], ancestor_edges[pred_bb]);
1443
1444 bitmap_set_bit (ancestor_edges[bb], EDGE_TO_BIT (in_edge));
1445
1446 bitmap_ior (pot_split[bb], pot_split[bb], pot_split[pred_bb]);
1447
1448 FOR_EACH_EDGE (out_edge, out_ei, in_edge->src->succs)
1449 bitmap_set_bit (pot_split[bb], EDGE_TO_BIT (out_edge));
1450
1451 prob[bb] += combine_probabilities
1452 (prob[pred_bb],
1453 in_edge->probability.initialized_p ()
1454 ? in_edge->probability.to_reg_br_prob_base ()
1455 : 0);
1456 // The rounding divide in combine_probabilities can result in an extra
1457 // probability increment propagating along 50-50 edges. Eventually when
1458 // the edges re-merge, the accumulated probability can go slightly above
1459 // REG_BR_PROB_BASE.
1460 if (prob[bb] > REG_BR_PROB_BASE)
1461 prob[bb] = REG_BR_PROB_BASE;
1462 }
1463
1464 bitmap_set_bit (dom[bb], bb);
1465 bitmap_and_compl (pot_split[bb], pot_split[bb], ancestor_edges[bb]);
1466
1467 if (sched_verbose >= 2)
1468 fprintf (sched_dump, ";; bb_prob(%d, %d) = %3d\n", bb, BB_TO_BLOCK (bb),
1469 (100 * prob[bb]) / REG_BR_PROB_BASE);
1470 }
1471
1472 /* Functions for target info. */
1473
1474 /* Compute in BL the list of split-edges of bb_src relatively to bb_trg.
1475 Note that bb_trg dominates bb_src. */
1476
1477 static void
split_edges(int bb_src,int bb_trg,edgelst * bl)1478 split_edges (int bb_src, int bb_trg, edgelst *bl)
1479 {
1480 auto_sbitmap src (SBITMAP_SIZE (pot_split[bb_src]));
1481 bitmap_copy (src, pot_split[bb_src]);
1482
1483 bitmap_and_compl (src, src, pot_split[bb_trg]);
1484 extract_edgelst (src, bl);
1485 }
1486
1487 /* Find the valid candidate-source-blocks for the target block TRG, compute
1488 their probability, and check if they are speculative or not.
1489 For speculative sources, compute their update-blocks and split-blocks. */
1490
1491 static void
compute_trg_info(int trg)1492 compute_trg_info (int trg)
1493 {
1494 candidate *sp;
1495 edgelst el = { NULL, 0 };
1496 int i, j, k, update_idx;
1497 basic_block block;
1498 edge_iterator ei;
1499 edge e;
1500
1501 candidate_table = XNEWVEC (candidate, current_nr_blocks);
1502
1503 bblst_last = 0;
1504 /* bblst_table holds split blocks and update blocks for each block after
1505 the current one in the region. split blocks and update blocks are
1506 the TO blocks of region edges, so there can be at most rgn_nr_edges
1507 of them. */
1508 bblst_size = (current_nr_blocks - target_bb) * rgn_nr_edges;
1509 bblst_table = XNEWVEC (basic_block, bblst_size);
1510
1511 edgelst_last = 0;
1512 edgelst_table = XNEWVEC (edge, rgn_nr_edges);
1513
1514 /* Define some of the fields for the target bb as well. */
1515 sp = candidate_table + trg;
1516 sp->is_valid = 1;
1517 sp->is_speculative = 0;
1518 sp->src_prob = REG_BR_PROB_BASE;
1519
1520 auto_sbitmap visited (last_basic_block_for_fn (cfun));
1521
1522 for (i = trg + 1; i < current_nr_blocks; i++)
1523 {
1524 sp = candidate_table + i;
1525
1526 sp->is_valid = IS_DOMINATED (i, trg);
1527 if (sp->is_valid)
1528 {
1529 int tf = prob[trg], cf = prob[i];
1530
1531 /* In CFGs with low probability edges TF can possibly be zero. */
1532 sp->src_prob = (tf ? GCOV_COMPUTE_SCALE (cf, tf) : 0);
1533 sp->is_valid = (sp->src_prob >= min_spec_prob);
1534 }
1535
1536 if (sp->is_valid)
1537 {
1538 split_edges (i, trg, &el);
1539 sp->is_speculative = (el.nr_members) ? 1 : 0;
1540 if (sp->is_speculative && !flag_schedule_speculative)
1541 sp->is_valid = 0;
1542 }
1543
1544 if (sp->is_valid)
1545 {
1546 /* Compute split blocks and store them in bblst_table.
1547 The TO block of every split edge is a split block. */
1548 sp->split_bbs.first_member = &bblst_table[bblst_last];
1549 sp->split_bbs.nr_members = el.nr_members;
1550 for (j = 0; j < el.nr_members; bblst_last++, j++)
1551 bblst_table[bblst_last] = el.first_member[j]->dest;
1552 sp->update_bbs.first_member = &bblst_table[bblst_last];
1553
1554 /* Compute update blocks and store them in bblst_table.
1555 For every split edge, look at the FROM block, and check
1556 all out edges. For each out edge that is not a split edge,
1557 add the TO block to the update block list. This list can end
1558 up with a lot of duplicates. We need to weed them out to avoid
1559 overrunning the end of the bblst_table. */
1560
1561 update_idx = 0;
1562 bitmap_clear (visited);
1563 for (j = 0; j < el.nr_members; j++)
1564 {
1565 block = el.first_member[j]->src;
1566 FOR_EACH_EDGE (e, ei, block->succs)
1567 {
1568 if (!bitmap_bit_p (visited, e->dest->index))
1569 {
1570 for (k = 0; k < el.nr_members; k++)
1571 if (e == el.first_member[k])
1572 break;
1573
1574 if (k >= el.nr_members)
1575 {
1576 bblst_table[bblst_last++] = e->dest;
1577 bitmap_set_bit (visited, e->dest->index);
1578 update_idx++;
1579 }
1580 }
1581 }
1582 }
1583 sp->update_bbs.nr_members = update_idx;
1584
1585 /* Make sure we didn't overrun the end of bblst_table. */
1586 gcc_assert (bblst_last <= bblst_size);
1587 }
1588 else
1589 {
1590 sp->split_bbs.nr_members = sp->update_bbs.nr_members = 0;
1591
1592 sp->is_speculative = 0;
1593 sp->src_prob = 0;
1594 }
1595 }
1596 }
1597
1598 /* Free the computed target info. */
1599 static void
free_trg_info(void)1600 free_trg_info (void)
1601 {
1602 free (candidate_table);
1603 free (bblst_table);
1604 free (edgelst_table);
1605 }
1606
1607 /* Print candidates info, for debugging purposes. Callable from debugger. */
1608
1609 DEBUG_FUNCTION void
debug_candidate(int i)1610 debug_candidate (int i)
1611 {
1612 if (!candidate_table[i].is_valid)
1613 return;
1614
1615 if (candidate_table[i].is_speculative)
1616 {
1617 int j;
1618 fprintf (sched_dump, "src b %d bb %d speculative \n", BB_TO_BLOCK (i), i);
1619
1620 fprintf (sched_dump, "split path: ");
1621 for (j = 0; j < candidate_table[i].split_bbs.nr_members; j++)
1622 {
1623 int b = candidate_table[i].split_bbs.first_member[j]->index;
1624
1625 fprintf (sched_dump, " %d ", b);
1626 }
1627 fprintf (sched_dump, "\n");
1628
1629 fprintf (sched_dump, "update path: ");
1630 for (j = 0; j < candidate_table[i].update_bbs.nr_members; j++)
1631 {
1632 int b = candidate_table[i].update_bbs.first_member[j]->index;
1633
1634 fprintf (sched_dump, " %d ", b);
1635 }
1636 fprintf (sched_dump, "\n");
1637 }
1638 else
1639 {
1640 fprintf (sched_dump, " src %d equivalent\n", BB_TO_BLOCK (i));
1641 }
1642 }
1643
1644 /* Print candidates info, for debugging purposes. Callable from debugger. */
1645
1646 DEBUG_FUNCTION void
debug_candidates(int trg)1647 debug_candidates (int trg)
1648 {
1649 int i;
1650
1651 fprintf (sched_dump, "----------- candidate table: target: b=%d bb=%d ---\n",
1652 BB_TO_BLOCK (trg), trg);
1653 for (i = trg + 1; i < current_nr_blocks; i++)
1654 debug_candidate (i);
1655 }
1656
1657 /* Functions for speculative scheduling. */
1658
1659 static bitmap_head not_in_df;
1660
1661 /* Return 0 if x is a set of a register alive in the beginning of one
1662 of the split-blocks of src, otherwise return 1. */
1663
1664 static int
check_live_1(int src,rtx x)1665 check_live_1 (int src, rtx x)
1666 {
1667 int i;
1668 int regno;
1669 rtx reg = SET_DEST (x);
1670
1671 if (reg == 0)
1672 return 1;
1673
1674 while (GET_CODE (reg) == SUBREG
1675 || GET_CODE (reg) == ZERO_EXTRACT
1676 || GET_CODE (reg) == STRICT_LOW_PART)
1677 reg = XEXP (reg, 0);
1678
1679 if (GET_CODE (reg) == PARALLEL)
1680 {
1681 int i;
1682
1683 for (i = XVECLEN (reg, 0) - 1; i >= 0; i--)
1684 if (XEXP (XVECEXP (reg, 0, i), 0) != 0)
1685 if (check_live_1 (src, XEXP (XVECEXP (reg, 0, i), 0)))
1686 return 1;
1687
1688 return 0;
1689 }
1690
1691 if (!REG_P (reg))
1692 return 1;
1693
1694 regno = REGNO (reg);
1695
1696 if (regno < FIRST_PSEUDO_REGISTER && global_regs[regno])
1697 {
1698 /* Global registers are assumed live. */
1699 return 0;
1700 }
1701 else
1702 {
1703 if (regno < FIRST_PSEUDO_REGISTER)
1704 {
1705 /* Check for hard registers. */
1706 int j = REG_NREGS (reg);
1707 while (--j >= 0)
1708 {
1709 for (i = 0; i < candidate_table[src].split_bbs.nr_members; i++)
1710 {
1711 basic_block b = candidate_table[src].split_bbs.first_member[i];
1712 int t = bitmap_bit_p (¬_in_df, b->index);
1713
1714 /* We can have split blocks, that were recently generated.
1715 Such blocks are always outside current region. */
1716 gcc_assert (!t || (CONTAINING_RGN (b->index)
1717 != CONTAINING_RGN (BB_TO_BLOCK (src))));
1718
1719 if (t || REGNO_REG_SET_P (df_get_live_in (b), regno + j))
1720 return 0;
1721 }
1722 }
1723 }
1724 else
1725 {
1726 /* Check for pseudo registers. */
1727 for (i = 0; i < candidate_table[src].split_bbs.nr_members; i++)
1728 {
1729 basic_block b = candidate_table[src].split_bbs.first_member[i];
1730 int t = bitmap_bit_p (¬_in_df, b->index);
1731
1732 gcc_assert (!t || (CONTAINING_RGN (b->index)
1733 != CONTAINING_RGN (BB_TO_BLOCK (src))));
1734
1735 if (t || REGNO_REG_SET_P (df_get_live_in (b), regno))
1736 return 0;
1737 }
1738 }
1739 }
1740
1741 return 1;
1742 }
1743
1744 /* If x is a set of a register R, mark that R is alive in the beginning
1745 of every update-block of src. */
1746
1747 static void
update_live_1(int src,rtx x)1748 update_live_1 (int src, rtx x)
1749 {
1750 int i;
1751 int regno;
1752 rtx reg = SET_DEST (x);
1753
1754 if (reg == 0)
1755 return;
1756
1757 while (GET_CODE (reg) == SUBREG
1758 || GET_CODE (reg) == ZERO_EXTRACT
1759 || GET_CODE (reg) == STRICT_LOW_PART)
1760 reg = XEXP (reg, 0);
1761
1762 if (GET_CODE (reg) == PARALLEL)
1763 {
1764 int i;
1765
1766 for (i = XVECLEN (reg, 0) - 1; i >= 0; i--)
1767 if (XEXP (XVECEXP (reg, 0, i), 0) != 0)
1768 update_live_1 (src, XEXP (XVECEXP (reg, 0, i), 0));
1769
1770 return;
1771 }
1772
1773 if (!REG_P (reg))
1774 return;
1775
1776 /* Global registers are always live, so the code below does not apply
1777 to them. */
1778
1779 regno = REGNO (reg);
1780
1781 if (! HARD_REGISTER_NUM_P (regno)
1782 || !global_regs[regno])
1783 {
1784 for (i = 0; i < candidate_table[src].update_bbs.nr_members; i++)
1785 {
1786 basic_block b = candidate_table[src].update_bbs.first_member[i];
1787 bitmap_set_range (df_get_live_in (b), regno, REG_NREGS (reg));
1788 }
1789 }
1790 }
1791
1792 /* Return 1 if insn can be speculatively moved from block src to trg,
1793 otherwise return 0. Called before first insertion of insn to
1794 ready-list or before the scheduling. */
1795
1796 static int
check_live(rtx_insn * insn,int src)1797 check_live (rtx_insn *insn, int src)
1798 {
1799 /* Find the registers set by instruction. */
1800 if (GET_CODE (PATTERN (insn)) == SET
1801 || GET_CODE (PATTERN (insn)) == CLOBBER)
1802 return check_live_1 (src, PATTERN (insn));
1803 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
1804 {
1805 int j;
1806 for (j = XVECLEN (PATTERN (insn), 0) - 1; j >= 0; j--)
1807 if ((GET_CODE (XVECEXP (PATTERN (insn), 0, j)) == SET
1808 || GET_CODE (XVECEXP (PATTERN (insn), 0, j)) == CLOBBER)
1809 && !check_live_1 (src, XVECEXP (PATTERN (insn), 0, j)))
1810 return 0;
1811
1812 return 1;
1813 }
1814
1815 return 1;
1816 }
1817
1818 /* Update the live registers info after insn was moved speculatively from
1819 block src to trg. */
1820
1821 static void
update_live(rtx_insn * insn,int src)1822 update_live (rtx_insn *insn, int src)
1823 {
1824 /* Find the registers set by instruction. */
1825 if (GET_CODE (PATTERN (insn)) == SET
1826 || GET_CODE (PATTERN (insn)) == CLOBBER)
1827 update_live_1 (src, PATTERN (insn));
1828 else if (GET_CODE (PATTERN (insn)) == PARALLEL)
1829 {
1830 int j;
1831 for (j = XVECLEN (PATTERN (insn), 0) - 1; j >= 0; j--)
1832 if (GET_CODE (XVECEXP (PATTERN (insn), 0, j)) == SET
1833 || GET_CODE (XVECEXP (PATTERN (insn), 0, j)) == CLOBBER)
1834 update_live_1 (src, XVECEXP (PATTERN (insn), 0, j));
1835 }
1836 }
1837
1838 /* Nonzero if block bb_to is equal to, or reachable from block bb_from. */
1839 #define IS_REACHABLE(bb_from, bb_to) \
1840 (bb_from == bb_to \
1841 || IS_RGN_ENTRY (bb_from) \
1842 || (bitmap_bit_p (ancestor_edges[bb_to], \
1843 EDGE_TO_BIT (single_pred_edge (BASIC_BLOCK_FOR_FN (cfun, \
1844 BB_TO_BLOCK (bb_from)))))))
1845
1846 /* Turns on the fed_by_spec_load flag for insns fed by load_insn. */
1847
1848 static void
set_spec_fed(rtx load_insn)1849 set_spec_fed (rtx load_insn)
1850 {
1851 sd_iterator_def sd_it;
1852 dep_t dep;
1853
1854 FOR_EACH_DEP (load_insn, SD_LIST_FORW, sd_it, dep)
1855 if (DEP_TYPE (dep) == REG_DEP_TRUE)
1856 FED_BY_SPEC_LOAD (DEP_CON (dep)) = 1;
1857 }
1858
1859 /* On the path from the insn to load_insn_bb, find a conditional
1860 branch depending on insn, that guards the speculative load. */
1861
1862 static int
find_conditional_protection(rtx_insn * insn,int load_insn_bb)1863 find_conditional_protection (rtx_insn *insn, int load_insn_bb)
1864 {
1865 sd_iterator_def sd_it;
1866 dep_t dep;
1867
1868 /* Iterate through DEF-USE forward dependences. */
1869 FOR_EACH_DEP (insn, SD_LIST_FORW, sd_it, dep)
1870 {
1871 rtx_insn *next = DEP_CON (dep);
1872
1873 if ((CONTAINING_RGN (BLOCK_NUM (next)) ==
1874 CONTAINING_RGN (BB_TO_BLOCK (load_insn_bb)))
1875 && IS_REACHABLE (INSN_BB (next), load_insn_bb)
1876 && load_insn_bb != INSN_BB (next)
1877 && DEP_TYPE (dep) == REG_DEP_TRUE
1878 && (JUMP_P (next)
1879 || find_conditional_protection (next, load_insn_bb)))
1880 return 1;
1881 }
1882 return 0;
1883 } /* find_conditional_protection */
1884
1885 /* Returns 1 if the same insn1 that participates in the computation
1886 of load_insn's address is feeding a conditional branch that is
1887 guarding on load_insn. This is true if we find two DEF-USE
1888 chains:
1889 insn1 -> ... -> conditional-branch
1890 insn1 -> ... -> load_insn,
1891 and if a flow path exists:
1892 insn1 -> ... -> conditional-branch -> ... -> load_insn,
1893 and if insn1 is on the path
1894 region-entry -> ... -> bb_trg -> ... load_insn.
1895
1896 Locate insn1 by climbing on INSN_BACK_DEPS from load_insn.
1897 Locate the branch by following INSN_FORW_DEPS from insn1. */
1898
1899 static int
is_conditionally_protected(rtx load_insn,int bb_src,int bb_trg)1900 is_conditionally_protected (rtx load_insn, int bb_src, int bb_trg)
1901 {
1902 sd_iterator_def sd_it;
1903 dep_t dep;
1904
1905 FOR_EACH_DEP (load_insn, SD_LIST_BACK, sd_it, dep)
1906 {
1907 rtx_insn *insn1 = DEP_PRO (dep);
1908
1909 /* Must be a DEF-USE dependence upon non-branch. */
1910 if (DEP_TYPE (dep) != REG_DEP_TRUE
1911 || JUMP_P (insn1))
1912 continue;
1913
1914 /* Must exist a path: region-entry -> ... -> bb_trg -> ... load_insn. */
1915 if (INSN_BB (insn1) == bb_src
1916 || (CONTAINING_RGN (BLOCK_NUM (insn1))
1917 != CONTAINING_RGN (BB_TO_BLOCK (bb_src)))
1918 || (!IS_REACHABLE (bb_trg, INSN_BB (insn1))
1919 && !IS_REACHABLE (INSN_BB (insn1), bb_trg)))
1920 continue;
1921
1922 /* Now search for the conditional-branch. */
1923 if (find_conditional_protection (insn1, bb_src))
1924 return 1;
1925
1926 /* Recursive step: search another insn1, "above" current insn1. */
1927 return is_conditionally_protected (insn1, bb_src, bb_trg);
1928 }
1929
1930 /* The chain does not exist. */
1931 return 0;
1932 } /* is_conditionally_protected */
1933
1934 /* Returns 1 if a clue for "similar load" 'insn2' is found, and hence
1935 load_insn can move speculatively from bb_src to bb_trg. All the
1936 following must hold:
1937
1938 (1) both loads have 1 base register (PFREE_CANDIDATEs).
1939 (2) load_insn and load1 have a def-use dependence upon
1940 the same insn 'insn1'.
1941 (3) either load2 is in bb_trg, or:
1942 - there's only one split-block, and
1943 - load1 is on the escape path, and
1944
1945 From all these we can conclude that the two loads access memory
1946 addresses that differ at most by a constant, and hence if moving
1947 load_insn would cause an exception, it would have been caused by
1948 load2 anyhow. */
1949
1950 static int
is_pfree(rtx load_insn,int bb_src,int bb_trg)1951 is_pfree (rtx load_insn, int bb_src, int bb_trg)
1952 {
1953 sd_iterator_def back_sd_it;
1954 dep_t back_dep;
1955 candidate *candp = candidate_table + bb_src;
1956
1957 if (candp->split_bbs.nr_members != 1)
1958 /* Must have exactly one escape block. */
1959 return 0;
1960
1961 FOR_EACH_DEP (load_insn, SD_LIST_BACK, back_sd_it, back_dep)
1962 {
1963 rtx_insn *insn1 = DEP_PRO (back_dep);
1964
1965 if (DEP_TYPE (back_dep) == REG_DEP_TRUE)
1966 /* Found a DEF-USE dependence (insn1, load_insn). */
1967 {
1968 sd_iterator_def fore_sd_it;
1969 dep_t fore_dep;
1970
1971 FOR_EACH_DEP (insn1, SD_LIST_FORW, fore_sd_it, fore_dep)
1972 {
1973 rtx_insn *insn2 = DEP_CON (fore_dep);
1974
1975 if (DEP_TYPE (fore_dep) == REG_DEP_TRUE)
1976 {
1977 /* Found a DEF-USE dependence (insn1, insn2). */
1978 if (haifa_classify_insn (insn2) != PFREE_CANDIDATE)
1979 /* insn2 not guaranteed to be a 1 base reg load. */
1980 continue;
1981
1982 if (INSN_BB (insn2) == bb_trg)
1983 /* insn2 is the similar load, in the target block. */
1984 return 1;
1985
1986 if (*(candp->split_bbs.first_member) == BLOCK_FOR_INSN (insn2))
1987 /* insn2 is a similar load, in a split-block. */
1988 return 1;
1989 }
1990 }
1991 }
1992 }
1993
1994 /* Couldn't find a similar load. */
1995 return 0;
1996 } /* is_pfree */
1997
1998 /* Return 1 if load_insn is prisky (i.e. if load_insn is fed by
1999 a load moved speculatively, or if load_insn is protected by
2000 a compare on load_insn's address). */
2001
2002 static int
is_prisky(rtx load_insn,int bb_src,int bb_trg)2003 is_prisky (rtx load_insn, int bb_src, int bb_trg)
2004 {
2005 if (FED_BY_SPEC_LOAD (load_insn))
2006 return 1;
2007
2008 if (sd_lists_empty_p (load_insn, SD_LIST_BACK))
2009 /* Dependence may 'hide' out of the region. */
2010 return 1;
2011
2012 if (is_conditionally_protected (load_insn, bb_src, bb_trg))
2013 return 1;
2014
2015 return 0;
2016 }
2017
2018 /* Insn is a candidate to be moved speculatively from bb_src to bb_trg.
2019 Return 1 if insn is exception-free (and the motion is valid)
2020 and 0 otherwise. */
2021
2022 static int
is_exception_free(rtx_insn * insn,int bb_src,int bb_trg)2023 is_exception_free (rtx_insn *insn, int bb_src, int bb_trg)
2024 {
2025 int insn_class = haifa_classify_insn (insn);
2026
2027 /* Handle non-load insns. */
2028 switch (insn_class)
2029 {
2030 case TRAP_FREE:
2031 return 1;
2032 case TRAP_RISKY:
2033 return 0;
2034 default:;
2035 }
2036
2037 /* Handle loads. */
2038 if (!flag_schedule_speculative_load)
2039 return 0;
2040 IS_LOAD_INSN (insn) = 1;
2041 switch (insn_class)
2042 {
2043 case IFREE:
2044 return (1);
2045 case IRISKY:
2046 return 0;
2047 case PFREE_CANDIDATE:
2048 if (is_pfree (insn, bb_src, bb_trg))
2049 return 1;
2050 /* Don't 'break' here: PFREE-candidate is also PRISKY-candidate. */
2051 /* FALLTHRU */
2052 case PRISKY_CANDIDATE:
2053 if (!flag_schedule_speculative_load_dangerous
2054 || is_prisky (insn, bb_src, bb_trg))
2055 return 0;
2056 break;
2057 default:;
2058 }
2059
2060 return flag_schedule_speculative_load_dangerous;
2061 }
2062
2063 /* The number of insns from the current block scheduled so far. */
2064 static int sched_target_n_insns;
2065 /* The number of insns from the current block to be scheduled in total. */
2066 static int target_n_insns;
2067 /* The number of insns from the entire region scheduled so far. */
2068 static int sched_n_insns;
2069
2070 /* Implementations of the sched_info functions for region scheduling. */
2071 static void init_ready_list (void);
2072 static int can_schedule_ready_p (rtx_insn *);
2073 static void begin_schedule_ready (rtx_insn *);
2074 static ds_t new_ready (rtx_insn *, ds_t);
2075 static int schedule_more_p (void);
2076 static const char *rgn_print_insn (const rtx_insn *, int);
2077 static int rgn_rank (rtx_insn *, rtx_insn *);
2078 static void compute_jump_reg_dependencies (rtx, regset);
2079
2080 /* Functions for speculative scheduling. */
2081 static void rgn_add_remove_insn (rtx_insn *, int);
2082 static void rgn_add_block (basic_block, basic_block);
2083 static void rgn_fix_recovery_cfg (int, int, int);
2084 static basic_block advance_target_bb (basic_block, rtx_insn *);
2085
2086 /* Return nonzero if there are more insns that should be scheduled. */
2087
2088 static int
schedule_more_p(void)2089 schedule_more_p (void)
2090 {
2091 return sched_target_n_insns < target_n_insns;
2092 }
2093
2094 /* Add all insns that are initially ready to the ready list READY. Called
2095 once before scheduling a set of insns. */
2096
2097 static void
init_ready_list(void)2098 init_ready_list (void)
2099 {
2100 rtx_insn *prev_head = current_sched_info->prev_head;
2101 rtx_insn *next_tail = current_sched_info->next_tail;
2102 int bb_src;
2103 rtx_insn *insn;
2104
2105 target_n_insns = 0;
2106 sched_target_n_insns = 0;
2107 sched_n_insns = 0;
2108
2109 /* Print debugging information. */
2110 if (sched_verbose >= 5)
2111 debug_rgn_dependencies (target_bb);
2112
2113 /* Prepare current target block info. */
2114 if (current_nr_blocks > 1)
2115 compute_trg_info (target_bb);
2116
2117 /* Initialize ready list with all 'ready' insns in target block.
2118 Count number of insns in the target block being scheduled. */
2119 for (insn = NEXT_INSN (prev_head); insn != next_tail; insn = NEXT_INSN (insn))
2120 {
2121 gcc_assert (TODO_SPEC (insn) == HARD_DEP || TODO_SPEC (insn) == DEP_POSTPONED);
2122 TODO_SPEC (insn) = HARD_DEP;
2123 try_ready (insn);
2124 target_n_insns++;
2125
2126 gcc_assert (!(TODO_SPEC (insn) & BEGIN_CONTROL));
2127 }
2128
2129 /* Add to ready list all 'ready' insns in valid source blocks.
2130 For speculative insns, check-live, exception-free, and
2131 issue-delay. */
2132 for (bb_src = target_bb + 1; bb_src < current_nr_blocks; bb_src++)
2133 if (IS_VALID (bb_src))
2134 {
2135 rtx_insn *src_head;
2136 rtx_insn *src_next_tail;
2137 rtx_insn *tail, *head;
2138
2139 get_ebb_head_tail (EBB_FIRST_BB (bb_src), EBB_LAST_BB (bb_src),
2140 &head, &tail);
2141 src_next_tail = NEXT_INSN (tail);
2142 src_head = head;
2143
2144 for (insn = src_head; insn != src_next_tail; insn = NEXT_INSN (insn))
2145 if (INSN_P (insn))
2146 {
2147 gcc_assert (TODO_SPEC (insn) == HARD_DEP || TODO_SPEC (insn) == DEP_POSTPONED);
2148 TODO_SPEC (insn) = HARD_DEP;
2149 try_ready (insn);
2150 }
2151 }
2152 }
2153
2154 /* Called after taking INSN from the ready list. Returns nonzero if this
2155 insn can be scheduled, nonzero if we should silently discard it. */
2156
2157 static int
can_schedule_ready_p(rtx_insn * insn)2158 can_schedule_ready_p (rtx_insn *insn)
2159 {
2160 /* An interblock motion? */
2161 if (INSN_BB (insn) != target_bb && IS_SPECULATIVE_INSN (insn))
2162 {
2163 /* Cannot schedule this insn unless all operands are live. */
2164 if (!check_live (insn, INSN_BB (insn)))
2165 return 0;
2166
2167 /* Should not move expensive instructions speculatively. */
2168 if (GET_CODE (PATTERN (insn)) != CLOBBER
2169 && !targetm.sched.can_speculate_insn (insn))
2170 return 0;
2171 }
2172
2173 return 1;
2174 }
2175
2176 /* Updates counter and other information. Split from can_schedule_ready_p ()
2177 because when we schedule insn speculatively then insn passed to
2178 can_schedule_ready_p () differs from the one passed to
2179 begin_schedule_ready (). */
2180 static void
begin_schedule_ready(rtx_insn * insn)2181 begin_schedule_ready (rtx_insn *insn)
2182 {
2183 /* An interblock motion? */
2184 if (INSN_BB (insn) != target_bb)
2185 {
2186 if (IS_SPECULATIVE_INSN (insn))
2187 {
2188 gcc_assert (check_live (insn, INSN_BB (insn)));
2189
2190 update_live (insn, INSN_BB (insn));
2191
2192 /* For speculative load, mark insns fed by it. */
2193 if (IS_LOAD_INSN (insn) || FED_BY_SPEC_LOAD (insn))
2194 set_spec_fed (insn);
2195
2196 nr_spec++;
2197 }
2198 nr_inter++;
2199 }
2200 else
2201 {
2202 /* In block motion. */
2203 sched_target_n_insns++;
2204 }
2205 sched_n_insns++;
2206 }
2207
2208 /* Called after INSN has all its hard dependencies resolved and the speculation
2209 of type TS is enough to overcome them all.
2210 Return nonzero if it should be moved to the ready list or the queue, or zero
2211 if we should silently discard it. */
2212 static ds_t
new_ready(rtx_insn * next,ds_t ts)2213 new_ready (rtx_insn *next, ds_t ts)
2214 {
2215 if (INSN_BB (next) != target_bb)
2216 {
2217 int not_ex_free = 0;
2218
2219 /* For speculative insns, before inserting to ready/queue,
2220 check live, exception-free, and issue-delay. */
2221 if (!IS_VALID (INSN_BB (next))
2222 || CANT_MOVE (next)
2223 || (IS_SPECULATIVE_INSN (next)
2224 && ((recog_memoized (next) >= 0
2225 && min_insn_conflict_delay (curr_state, next, next)
2226 > param_max_sched_insn_conflict_delay)
2227 || IS_SPECULATION_CHECK_P (next)
2228 || !check_live (next, INSN_BB (next))
2229 || (not_ex_free = !is_exception_free (next, INSN_BB (next),
2230 target_bb)))))
2231 {
2232 if (not_ex_free
2233 /* We are here because is_exception_free () == false.
2234 But we possibly can handle that with control speculation. */
2235 && sched_deps_info->generate_spec_deps
2236 && spec_info->mask & BEGIN_CONTROL)
2237 {
2238 ds_t new_ds;
2239
2240 /* Add control speculation to NEXT's dependency type. */
2241 new_ds = set_dep_weak (ts, BEGIN_CONTROL, MAX_DEP_WEAK);
2242
2243 /* Check if NEXT can be speculated with new dependency type. */
2244 if (sched_insn_is_legitimate_for_speculation_p (next, new_ds))
2245 /* Here we got new control-speculative instruction. */
2246 ts = new_ds;
2247 else
2248 /* NEXT isn't ready yet. */
2249 ts = DEP_POSTPONED;
2250 }
2251 else
2252 /* NEXT isn't ready yet. */
2253 ts = DEP_POSTPONED;
2254 }
2255 }
2256
2257 return ts;
2258 }
2259
2260 /* Return a string that contains the insn uid and optionally anything else
2261 necessary to identify this insn in an output. It's valid to use a
2262 static buffer for this. The ALIGNED parameter should cause the string
2263 to be formatted so that multiple output lines will line up nicely. */
2264
2265 static const char *
rgn_print_insn(const rtx_insn * insn,int aligned)2266 rgn_print_insn (const rtx_insn *insn, int aligned)
2267 {
2268 static char tmp[80];
2269
2270 if (aligned)
2271 sprintf (tmp, "b%3d: i%4d", INSN_BB (insn), INSN_UID (insn));
2272 else
2273 {
2274 if (current_nr_blocks > 1 && INSN_BB (insn) != target_bb)
2275 sprintf (tmp, "%d/b%d", INSN_UID (insn), INSN_BB (insn));
2276 else
2277 sprintf (tmp, "%d", INSN_UID (insn));
2278 }
2279 return tmp;
2280 }
2281
2282 /* Compare priority of two insns. Return a positive number if the second
2283 insn is to be preferred for scheduling, and a negative one if the first
2284 is to be preferred. Zero if they are equally good. */
2285
2286 static int
rgn_rank(rtx_insn * insn1,rtx_insn * insn2)2287 rgn_rank (rtx_insn *insn1, rtx_insn *insn2)
2288 {
2289 /* Some comparison make sense in interblock scheduling only. */
2290 if (INSN_BB (insn1) != INSN_BB (insn2))
2291 {
2292 int spec_val, prob_val;
2293
2294 /* Prefer an inblock motion on an interblock motion. */
2295 if ((INSN_BB (insn2) == target_bb) && (INSN_BB (insn1) != target_bb))
2296 return 1;
2297 if ((INSN_BB (insn1) == target_bb) && (INSN_BB (insn2) != target_bb))
2298 return -1;
2299
2300 /* Prefer a useful motion on a speculative one. */
2301 spec_val = IS_SPECULATIVE_INSN (insn1) - IS_SPECULATIVE_INSN (insn2);
2302 if (spec_val)
2303 return spec_val;
2304
2305 /* Prefer a more probable (speculative) insn. */
2306 prob_val = INSN_PROBABILITY (insn2) - INSN_PROBABILITY (insn1);
2307 if (prob_val)
2308 return prob_val;
2309 }
2310 return 0;
2311 }
2312
2313 /* NEXT is an instruction that depends on INSN (a backward dependence);
2314 return nonzero if we should include this dependence in priority
2315 calculations. */
2316
2317 int
contributes_to_priority(rtx_insn * next,rtx_insn * insn)2318 contributes_to_priority (rtx_insn *next, rtx_insn *insn)
2319 {
2320 /* NEXT and INSN reside in one ebb. */
2321 return BLOCK_TO_BB (BLOCK_NUM (next)) == BLOCK_TO_BB (BLOCK_NUM (insn));
2322 }
2323
2324 /* INSN is a JUMP_INSN. Store the set of registers that must be
2325 considered as used by this jump in USED. */
2326
2327 static void
compute_jump_reg_dependencies(rtx insn ATTRIBUTE_UNUSED,regset used ATTRIBUTE_UNUSED)2328 compute_jump_reg_dependencies (rtx insn ATTRIBUTE_UNUSED,
2329 regset used ATTRIBUTE_UNUSED)
2330 {
2331 /* Nothing to do here, since we postprocess jumps in
2332 add_branch_dependences. */
2333 }
2334
2335 /* This variable holds common_sched_info hooks and data relevant to
2336 the interblock scheduler. */
2337 static struct common_sched_info_def rgn_common_sched_info;
2338
2339
2340 /* This holds data for the dependence analysis relevant to
2341 the interblock scheduler. */
2342 static struct sched_deps_info_def rgn_sched_deps_info;
2343
2344 /* This holds constant data used for initializing the above structure
2345 for the Haifa scheduler. */
2346 static const struct sched_deps_info_def rgn_const_sched_deps_info =
2347 {
2348 compute_jump_reg_dependencies,
2349 NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL,
2350 0, 0, 0
2351 };
2352
2353 /* Same as above, but for the selective scheduler. */
2354 static const struct sched_deps_info_def rgn_const_sel_sched_deps_info =
2355 {
2356 compute_jump_reg_dependencies,
2357 NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL,
2358 0, 0, 0
2359 };
2360
2361 /* Return true if scheduling INSN will trigger finish of scheduling
2362 current block. */
2363 static bool
rgn_insn_finishes_block_p(rtx_insn * insn)2364 rgn_insn_finishes_block_p (rtx_insn *insn)
2365 {
2366 if (INSN_BB (insn) == target_bb
2367 && sched_target_n_insns + 1 == target_n_insns)
2368 /* INSN is the last not-scheduled instruction in the current block. */
2369 return true;
2370
2371 return false;
2372 }
2373
2374 /* Used in schedule_insns to initialize current_sched_info for scheduling
2375 regions (or single basic blocks). */
2376
2377 static const struct haifa_sched_info rgn_const_sched_info =
2378 {
2379 init_ready_list,
2380 can_schedule_ready_p,
2381 schedule_more_p,
2382 new_ready,
2383 rgn_rank,
2384 rgn_print_insn,
2385 contributes_to_priority,
2386 rgn_insn_finishes_block_p,
2387
2388 NULL, NULL,
2389 NULL, NULL,
2390 0, 0,
2391
2392 rgn_add_remove_insn,
2393 begin_schedule_ready,
2394 NULL,
2395 advance_target_bb,
2396 NULL, NULL,
2397 SCHED_RGN
2398 };
2399
2400 /* This variable holds the data and hooks needed to the Haifa scheduler backend
2401 for the interblock scheduler frontend. */
2402 static struct haifa_sched_info rgn_sched_info;
2403
2404 /* Returns maximum priority that an insn was assigned to. */
2405
2406 int
get_rgn_sched_max_insns_priority(void)2407 get_rgn_sched_max_insns_priority (void)
2408 {
2409 return rgn_sched_info.sched_max_insns_priority;
2410 }
2411
2412 /* Determine if PAT sets a TARGET_CLASS_LIKELY_SPILLED_P register. */
2413
2414 static bool
sets_likely_spilled(rtx pat)2415 sets_likely_spilled (rtx pat)
2416 {
2417 bool ret = false;
2418 note_pattern_stores (pat, sets_likely_spilled_1, &ret);
2419 return ret;
2420 }
2421
2422 static void
sets_likely_spilled_1(rtx x,const_rtx pat,void * data)2423 sets_likely_spilled_1 (rtx x, const_rtx pat, void *data)
2424 {
2425 bool *ret = (bool *) data;
2426
2427 if (GET_CODE (pat) == SET
2428 && REG_P (x)
2429 && HARD_REGISTER_P (x)
2430 && targetm.class_likely_spilled_p (REGNO_REG_CLASS (REGNO (x))))
2431 *ret = true;
2432 }
2433
2434 /* A bitmap to note insns that participate in any dependency. Used in
2435 add_branch_dependences. */
2436 static sbitmap insn_referenced;
2437
2438 /* Add dependences so that branches are scheduled to run last in their
2439 block. */
2440 static void
add_branch_dependences(rtx_insn * head,rtx_insn * tail)2441 add_branch_dependences (rtx_insn *head, rtx_insn *tail)
2442 {
2443 rtx_insn *insn, *last;
2444
2445 /* For all branches, calls, uses, clobbers, cc0 setters, and instructions
2446 that can throw exceptions, force them to remain in order at the end of
2447 the block by adding dependencies and giving the last a high priority.
2448 There may be notes present, and prev_head may also be a note.
2449
2450 Branches must obviously remain at the end. Calls should remain at the
2451 end since moving them results in worse register allocation. Uses remain
2452 at the end to ensure proper register allocation.
2453
2454 cc0 setters remain at the end because they can't be moved away from
2455 their cc0 user.
2456
2457 Predecessors of SCHED_GROUP_P instructions at the end remain at the end.
2458
2459 COND_EXEC insns cannot be moved past a branch (see e.g. PR17808).
2460
2461 Insns setting TARGET_CLASS_LIKELY_SPILLED_P registers (usually return
2462 values) are not moved before reload because we can wind up with register
2463 allocation failures. */
2464
2465 while (tail != head && DEBUG_INSN_P (tail))
2466 tail = PREV_INSN (tail);
2467
2468 insn = tail;
2469 last = 0;
2470 while (CALL_P (insn)
2471 || JUMP_P (insn) || JUMP_TABLE_DATA_P (insn)
2472 || (NONJUMP_INSN_P (insn)
2473 && (GET_CODE (PATTERN (insn)) == USE
2474 || GET_CODE (PATTERN (insn)) == CLOBBER
2475 || can_throw_internal (insn)
2476 || (HAVE_cc0 && sets_cc0_p (PATTERN (insn)))
2477 || (!reload_completed
2478 && sets_likely_spilled (PATTERN (insn)))))
2479 || NOTE_P (insn)
2480 || (last != 0 && SCHED_GROUP_P (last)))
2481 {
2482 if (!NOTE_P (insn))
2483 {
2484 if (last != 0
2485 && sd_find_dep_between (insn, last, false) == NULL)
2486 {
2487 if (! sched_insns_conditions_mutex_p (last, insn))
2488 add_dependence (last, insn, REG_DEP_ANTI);
2489 bitmap_set_bit (insn_referenced, INSN_LUID (insn));
2490 }
2491
2492 CANT_MOVE (insn) = 1;
2493
2494 last = insn;
2495 }
2496
2497 /* Don't overrun the bounds of the basic block. */
2498 if (insn == head)
2499 break;
2500
2501 do
2502 insn = PREV_INSN (insn);
2503 while (insn != head && DEBUG_INSN_P (insn));
2504 }
2505
2506 /* Selective scheduling handles control dependencies by itself, and
2507 CANT_MOVE flags ensure that other insns will be kept in place. */
2508 if (sel_sched_p ())
2509 return;
2510
2511 /* Make sure these insns are scheduled last in their block. */
2512 insn = last;
2513 if (insn != 0)
2514 while (insn != head)
2515 {
2516 insn = prev_nonnote_insn (insn);
2517
2518 if (bitmap_bit_p (insn_referenced, INSN_LUID (insn))
2519 || DEBUG_INSN_P (insn))
2520 continue;
2521
2522 if (! sched_insns_conditions_mutex_p (last, insn))
2523 add_dependence (last, insn, REG_DEP_ANTI);
2524 }
2525
2526 if (!targetm.have_conditional_execution ())
2527 return;
2528
2529 /* Finally, if the block ends in a jump, and we are doing intra-block
2530 scheduling, make sure that the branch depends on any COND_EXEC insns
2531 inside the block to avoid moving the COND_EXECs past the branch insn.
2532
2533 We only have to do this after reload, because (1) before reload there
2534 are no COND_EXEC insns, and (2) the region scheduler is an intra-block
2535 scheduler after reload.
2536
2537 FIXME: We could in some cases move COND_EXEC insns past the branch if
2538 this scheduler would be a little smarter. Consider this code:
2539
2540 T = [addr]
2541 C ? addr += 4
2542 !C ? X += 12
2543 C ? T += 1
2544 C ? jump foo
2545
2546 On a target with a one cycle stall on a memory access the optimal
2547 sequence would be:
2548
2549 T = [addr]
2550 C ? addr += 4
2551 C ? T += 1
2552 C ? jump foo
2553 !C ? X += 12
2554
2555 We don't want to put the 'X += 12' before the branch because it just
2556 wastes a cycle of execution time when the branch is taken.
2557
2558 Note that in the example "!C" will always be true. That is another
2559 possible improvement for handling COND_EXECs in this scheduler: it
2560 could remove always-true predicates. */
2561
2562 if (!reload_completed || ! (JUMP_P (tail) || JUMP_TABLE_DATA_P (tail)))
2563 return;
2564
2565 insn = tail;
2566 while (insn != head)
2567 {
2568 insn = PREV_INSN (insn);
2569
2570 /* Note that we want to add this dependency even when
2571 sched_insns_conditions_mutex_p returns true. The whole point
2572 is that we _want_ this dependency, even if these insns really
2573 are independent. */
2574 if (INSN_P (insn) && GET_CODE (PATTERN (insn)) == COND_EXEC)
2575 add_dependence (tail, insn, REG_DEP_ANTI);
2576 }
2577 }
2578
2579 /* Data structures for the computation of data dependences in a regions. We
2580 keep one `deps' structure for every basic block. Before analyzing the
2581 data dependences for a bb, its variables are initialized as a function of
2582 the variables of its predecessors. When the analysis for a bb completes,
2583 we save the contents to the corresponding bb_deps[bb] variable. */
2584
2585 static class deps_desc *bb_deps;
2586
2587 static void
concat_insn_mem_list(rtx_insn_list * copy_insns,rtx_expr_list * copy_mems,rtx_insn_list ** old_insns_p,rtx_expr_list ** old_mems_p)2588 concat_insn_mem_list (rtx_insn_list *copy_insns,
2589 rtx_expr_list *copy_mems,
2590 rtx_insn_list **old_insns_p,
2591 rtx_expr_list **old_mems_p)
2592 {
2593 rtx_insn_list *new_insns = *old_insns_p;
2594 rtx_expr_list *new_mems = *old_mems_p;
2595
2596 while (copy_insns)
2597 {
2598 new_insns = alloc_INSN_LIST (copy_insns->insn (), new_insns);
2599 new_mems = alloc_EXPR_LIST (VOIDmode, copy_mems->element (), new_mems);
2600 copy_insns = copy_insns->next ();
2601 copy_mems = copy_mems->next ();
2602 }
2603
2604 *old_insns_p = new_insns;
2605 *old_mems_p = new_mems;
2606 }
2607
2608 /* Join PRED_DEPS to the SUCC_DEPS. */
2609 void
deps_join(class deps_desc * succ_deps,class deps_desc * pred_deps)2610 deps_join (class deps_desc *succ_deps, class deps_desc *pred_deps)
2611 {
2612 unsigned reg;
2613 reg_set_iterator rsi;
2614
2615 /* The reg_last lists are inherited by successor. */
2616 EXECUTE_IF_SET_IN_REG_SET (&pred_deps->reg_last_in_use, 0, reg, rsi)
2617 {
2618 struct deps_reg *pred_rl = &pred_deps->reg_last[reg];
2619 struct deps_reg *succ_rl = &succ_deps->reg_last[reg];
2620
2621 succ_rl->uses = concat_INSN_LIST (pred_rl->uses, succ_rl->uses);
2622 succ_rl->sets = concat_INSN_LIST (pred_rl->sets, succ_rl->sets);
2623 succ_rl->implicit_sets
2624 = concat_INSN_LIST (pred_rl->implicit_sets, succ_rl->implicit_sets);
2625 succ_rl->clobbers = concat_INSN_LIST (pred_rl->clobbers,
2626 succ_rl->clobbers);
2627 succ_rl->uses_length += pred_rl->uses_length;
2628 succ_rl->clobbers_length += pred_rl->clobbers_length;
2629 }
2630 IOR_REG_SET (&succ_deps->reg_last_in_use, &pred_deps->reg_last_in_use);
2631
2632 /* Mem read/write lists are inherited by successor. */
2633 concat_insn_mem_list (pred_deps->pending_read_insns,
2634 pred_deps->pending_read_mems,
2635 &succ_deps->pending_read_insns,
2636 &succ_deps->pending_read_mems);
2637 concat_insn_mem_list (pred_deps->pending_write_insns,
2638 pred_deps->pending_write_mems,
2639 &succ_deps->pending_write_insns,
2640 &succ_deps->pending_write_mems);
2641
2642 succ_deps->pending_jump_insns
2643 = concat_INSN_LIST (pred_deps->pending_jump_insns,
2644 succ_deps->pending_jump_insns);
2645 succ_deps->last_pending_memory_flush
2646 = concat_INSN_LIST (pred_deps->last_pending_memory_flush,
2647 succ_deps->last_pending_memory_flush);
2648
2649 succ_deps->pending_read_list_length += pred_deps->pending_read_list_length;
2650 succ_deps->pending_write_list_length += pred_deps->pending_write_list_length;
2651 succ_deps->pending_flush_length += pred_deps->pending_flush_length;
2652
2653 /* last_function_call is inherited by successor. */
2654 succ_deps->last_function_call
2655 = concat_INSN_LIST (pred_deps->last_function_call,
2656 succ_deps->last_function_call);
2657
2658 /* last_function_call_may_noreturn is inherited by successor. */
2659 succ_deps->last_function_call_may_noreturn
2660 = concat_INSN_LIST (pred_deps->last_function_call_may_noreturn,
2661 succ_deps->last_function_call_may_noreturn);
2662
2663 /* sched_before_next_call is inherited by successor. */
2664 succ_deps->sched_before_next_call
2665 = concat_INSN_LIST (pred_deps->sched_before_next_call,
2666 succ_deps->sched_before_next_call);
2667 }
2668
2669 /* After computing the dependencies for block BB, propagate the dependencies
2670 found in TMP_DEPS to the successors of the block. */
2671 static void
propagate_deps(int bb,class deps_desc * pred_deps)2672 propagate_deps (int bb, class deps_desc *pred_deps)
2673 {
2674 basic_block block = BASIC_BLOCK_FOR_FN (cfun, BB_TO_BLOCK (bb));
2675 edge_iterator ei;
2676 edge e;
2677
2678 /* bb's structures are inherited by its successors. */
2679 FOR_EACH_EDGE (e, ei, block->succs)
2680 {
2681 /* Only bbs "below" bb, in the same region, are interesting. */
2682 if (e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun)
2683 || CONTAINING_RGN (block->index) != CONTAINING_RGN (e->dest->index)
2684 || BLOCK_TO_BB (e->dest->index) <= bb)
2685 continue;
2686
2687 deps_join (bb_deps + BLOCK_TO_BB (e->dest->index), pred_deps);
2688 }
2689
2690 /* These lists should point to the right place, for correct
2691 freeing later. */
2692 bb_deps[bb].pending_read_insns = pred_deps->pending_read_insns;
2693 bb_deps[bb].pending_read_mems = pred_deps->pending_read_mems;
2694 bb_deps[bb].pending_write_insns = pred_deps->pending_write_insns;
2695 bb_deps[bb].pending_write_mems = pred_deps->pending_write_mems;
2696 bb_deps[bb].pending_jump_insns = pred_deps->pending_jump_insns;
2697
2698 /* Can't allow these to be freed twice. */
2699 pred_deps->pending_read_insns = 0;
2700 pred_deps->pending_read_mems = 0;
2701 pred_deps->pending_write_insns = 0;
2702 pred_deps->pending_write_mems = 0;
2703 pred_deps->pending_jump_insns = 0;
2704 }
2705
2706 /* Compute dependences inside bb. In a multiple blocks region:
2707 (1) a bb is analyzed after its predecessors, and (2) the lists in
2708 effect at the end of bb (after analyzing for bb) are inherited by
2709 bb's successors.
2710
2711 Specifically for reg-reg data dependences, the block insns are
2712 scanned by sched_analyze () top-to-bottom. Three lists are
2713 maintained by sched_analyze (): reg_last[].sets for register DEFs,
2714 reg_last[].implicit_sets for implicit hard register DEFs, and
2715 reg_last[].uses for register USEs.
2716
2717 When analysis is completed for bb, we update for its successors:
2718 ; - DEFS[succ] = Union (DEFS [succ], DEFS [bb])
2719 ; - IMPLICIT_DEFS[succ] = Union (IMPLICIT_DEFS [succ], IMPLICIT_DEFS [bb])
2720 ; - USES[succ] = Union (USES [succ], DEFS [bb])
2721
2722 The mechanism for computing mem-mem data dependence is very
2723 similar, and the result is interblock dependences in the region. */
2724
2725 static void
compute_block_dependences(int bb)2726 compute_block_dependences (int bb)
2727 {
2728 rtx_insn *head, *tail;
2729 class deps_desc tmp_deps;
2730
2731 tmp_deps = bb_deps[bb];
2732
2733 /* Do the analysis for this block. */
2734 gcc_assert (EBB_FIRST_BB (bb) == EBB_LAST_BB (bb));
2735 get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail);
2736
2737 sched_analyze (&tmp_deps, head, tail);
2738
2739 add_branch_dependences (head, tail);
2740
2741 if (current_nr_blocks > 1)
2742 propagate_deps (bb, &tmp_deps);
2743
2744 /* Free up the INSN_LISTs. */
2745 free_deps (&tmp_deps);
2746
2747 if (targetm.sched.dependencies_evaluation_hook)
2748 targetm.sched.dependencies_evaluation_hook (head, tail);
2749 }
2750
2751 /* Free dependencies of instructions inside BB. */
2752 static void
free_block_dependencies(int bb)2753 free_block_dependencies (int bb)
2754 {
2755 rtx_insn *head;
2756 rtx_insn *tail;
2757
2758 get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail);
2759
2760 if (no_real_insns_p (head, tail))
2761 return;
2762
2763 sched_free_deps (head, tail, true);
2764 }
2765
2766 /* Remove all INSN_LISTs and EXPR_LISTs from the pending lists and add
2767 them to the unused_*_list variables, so that they can be reused. */
2768
2769 static void
free_pending_lists(void)2770 free_pending_lists (void)
2771 {
2772 int bb;
2773
2774 for (bb = 0; bb < current_nr_blocks; bb++)
2775 {
2776 free_INSN_LIST_list (&bb_deps[bb].pending_read_insns);
2777 free_INSN_LIST_list (&bb_deps[bb].pending_write_insns);
2778 free_EXPR_LIST_list (&bb_deps[bb].pending_read_mems);
2779 free_EXPR_LIST_list (&bb_deps[bb].pending_write_mems);
2780 free_INSN_LIST_list (&bb_deps[bb].pending_jump_insns);
2781 }
2782 }
2783
2784 /* Print dependences for debugging starting from FROM_BB.
2785 Callable from debugger. */
2786 /* Print dependences for debugging starting from FROM_BB.
2787 Callable from debugger. */
2788 DEBUG_FUNCTION void
debug_rgn_dependencies(int from_bb)2789 debug_rgn_dependencies (int from_bb)
2790 {
2791 int bb;
2792
2793 fprintf (sched_dump,
2794 ";; --------------- forward dependences: ------------ \n");
2795
2796 for (bb = from_bb; bb < current_nr_blocks; bb++)
2797 {
2798 rtx_insn *head, *tail;
2799
2800 get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail);
2801 fprintf (sched_dump, "\n;; --- Region Dependences --- b %d bb %d \n",
2802 BB_TO_BLOCK (bb), bb);
2803
2804 debug_dependencies (head, tail);
2805 }
2806 }
2807
2808 /* Print dependencies information for instructions between HEAD and TAIL.
2809 ??? This function would probably fit best in haifa-sched.c. */
debug_dependencies(rtx_insn * head,rtx_insn * tail)2810 void debug_dependencies (rtx_insn *head, rtx_insn *tail)
2811 {
2812 rtx_insn *insn;
2813 rtx_insn *next_tail = NEXT_INSN (tail);
2814
2815 fprintf (sched_dump, ";; %7s%6s%6s%6s%6s%6s%14s\n",
2816 "insn", "code", "bb", "dep", "prio", "cost",
2817 "reservation");
2818 fprintf (sched_dump, ";; %7s%6s%6s%6s%6s%6s%14s\n",
2819 "----", "----", "--", "---", "----", "----",
2820 "-----------");
2821
2822 for (insn = head; insn != next_tail; insn = NEXT_INSN (insn))
2823 {
2824 if (! INSN_P (insn))
2825 {
2826 int n;
2827 fprintf (sched_dump, ";; %6d ", INSN_UID (insn));
2828 if (NOTE_P (insn))
2829 {
2830 n = NOTE_KIND (insn);
2831 fprintf (sched_dump, "%s\n", GET_NOTE_INSN_NAME (n));
2832 }
2833 else
2834 fprintf (sched_dump, " {%s}\n", GET_RTX_NAME (GET_CODE (insn)));
2835 continue;
2836 }
2837
2838 fprintf (sched_dump,
2839 ";; %s%5d%6d%6d%6d%6d%6d ",
2840 (SCHED_GROUP_P (insn) ? "+" : " "),
2841 INSN_UID (insn),
2842 INSN_CODE (insn),
2843 BLOCK_NUM (insn),
2844 sched_emulate_haifa_p ? -1 : sd_lists_size (insn, SD_LIST_BACK),
2845 (sel_sched_p () ? (sched_emulate_haifa_p ? -1
2846 : INSN_PRIORITY (insn))
2847 : INSN_PRIORITY (insn)),
2848 (sel_sched_p () ? (sched_emulate_haifa_p ? -1
2849 : insn_sched_cost (insn))
2850 : insn_sched_cost (insn)));
2851
2852 if (recog_memoized (insn) < 0)
2853 fprintf (sched_dump, "nothing");
2854 else
2855 print_reservation (sched_dump, insn);
2856
2857 fprintf (sched_dump, "\t: ");
2858 {
2859 sd_iterator_def sd_it;
2860 dep_t dep;
2861
2862 FOR_EACH_DEP (insn, SD_LIST_FORW, sd_it, dep)
2863 fprintf (sched_dump, "%d%s%s ", INSN_UID (DEP_CON (dep)),
2864 DEP_NONREG (dep) ? "n" : "",
2865 DEP_MULTIPLE (dep) ? "m" : "");
2866 }
2867 fprintf (sched_dump, "\n");
2868 }
2869
2870 fprintf (sched_dump, "\n");
2871 }
2872
2873 /* Dump dependency graph for the current region to a file using dot syntax. */
2874
2875 void
dump_rgn_dependencies_dot(FILE * file)2876 dump_rgn_dependencies_dot (FILE *file)
2877 {
2878 rtx_insn *head, *tail, *con, *pro;
2879 sd_iterator_def sd_it;
2880 dep_t dep;
2881 int bb;
2882 pretty_printer pp;
2883
2884 pp.buffer->stream = file;
2885 pp_printf (&pp, "digraph SchedDG {\n");
2886
2887 for (bb = 0; bb < current_nr_blocks; ++bb)
2888 {
2889 /* Begin subgraph (basic block). */
2890 pp_printf (&pp, "subgraph cluster_block_%d {\n", bb);
2891 pp_printf (&pp, "\t" "color=blue;" "\n");
2892 pp_printf (&pp, "\t" "style=bold;" "\n");
2893 pp_printf (&pp, "\t" "label=\"BB #%d\";\n", BB_TO_BLOCK (bb));
2894
2895 /* Setup head and tail (no support for EBBs). */
2896 gcc_assert (EBB_FIRST_BB (bb) == EBB_LAST_BB (bb));
2897 get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail);
2898 tail = NEXT_INSN (tail);
2899
2900 /* Dump all insns. */
2901 for (con = head; con != tail; con = NEXT_INSN (con))
2902 {
2903 if (!INSN_P (con))
2904 continue;
2905
2906 /* Pretty print the insn. */
2907 pp_printf (&pp, "\t%d [label=\"{", INSN_UID (con));
2908 pp_write_text_to_stream (&pp);
2909 print_insn (&pp, con, /*verbose=*/false);
2910 pp_write_text_as_dot_label_to_stream (&pp, /*for_record=*/true);
2911 pp_write_text_to_stream (&pp);
2912
2913 /* Dump instruction attributes. */
2914 pp_printf (&pp, "|{ uid:%d | luid:%d | prio:%d }}\",shape=record]\n",
2915 INSN_UID (con), INSN_LUID (con), INSN_PRIORITY (con));
2916
2917 /* Dump all deps. */
2918 FOR_EACH_DEP (con, SD_LIST_BACK, sd_it, dep)
2919 {
2920 int weight = 0;
2921 const char *color;
2922 pro = DEP_PRO (dep);
2923
2924 switch (DEP_TYPE (dep))
2925 {
2926 case REG_DEP_TRUE:
2927 color = "black";
2928 weight = 1;
2929 break;
2930 case REG_DEP_OUTPUT:
2931 case REG_DEP_ANTI:
2932 color = "orange";
2933 break;
2934 case REG_DEP_CONTROL:
2935 color = "blue";
2936 break;
2937 default:
2938 gcc_unreachable ();
2939 }
2940
2941 pp_printf (&pp, "\t%d -> %d [color=%s",
2942 INSN_UID (pro), INSN_UID (con), color);
2943 if (int cost = dep_cost (dep))
2944 pp_printf (&pp, ",label=%d", cost);
2945 pp_printf (&pp, ",weight=%d", weight);
2946 pp_printf (&pp, "];\n");
2947 }
2948 }
2949 pp_printf (&pp, "}\n");
2950 }
2951
2952 pp_printf (&pp, "}\n");
2953 pp_flush (&pp);
2954 }
2955
2956 /* Dump dependency graph for the current region to a file using dot syntax. */
2957
2958 DEBUG_FUNCTION void
dump_rgn_dependencies_dot(const char * fname)2959 dump_rgn_dependencies_dot (const char *fname)
2960 {
2961 FILE *fp;
2962
2963 fp = fopen (fname, "w");
2964 if (!fp)
2965 {
2966 perror ("fopen");
2967 return;
2968 }
2969
2970 dump_rgn_dependencies_dot (fp);
2971 fclose (fp);
2972 }
2973
2974
2975 /* Returns true if all the basic blocks of the current region have
2976 NOTE_DISABLE_SCHED_OF_BLOCK which means not to schedule that region. */
2977 bool
sched_is_disabled_for_current_region_p(void)2978 sched_is_disabled_for_current_region_p (void)
2979 {
2980 int bb;
2981
2982 for (bb = 0; bb < current_nr_blocks; bb++)
2983 if (!(BASIC_BLOCK_FOR_FN (cfun,
2984 BB_TO_BLOCK (bb))->flags & BB_DISABLE_SCHEDULE))
2985 return false;
2986
2987 return true;
2988 }
2989
2990 /* Free all region dependencies saved in INSN_BACK_DEPS and
2991 INSN_RESOLVED_BACK_DEPS. The Haifa scheduler does this on the fly
2992 when scheduling, so this function is supposed to be called from
2993 the selective scheduling only. */
2994 void
free_rgn_deps(void)2995 free_rgn_deps (void)
2996 {
2997 int bb;
2998
2999 for (bb = 0; bb < current_nr_blocks; bb++)
3000 {
3001 rtx_insn *head, *tail;
3002
3003 gcc_assert (EBB_FIRST_BB (bb) == EBB_LAST_BB (bb));
3004 get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail);
3005
3006 sched_free_deps (head, tail, false);
3007 }
3008 }
3009
3010 static int rgn_n_insns;
3011
3012 /* Compute insn priority for a current region. */
3013 void
compute_priorities(void)3014 compute_priorities (void)
3015 {
3016 int bb;
3017
3018 current_sched_info->sched_max_insns_priority = 0;
3019 for (bb = 0; bb < current_nr_blocks; bb++)
3020 {
3021 rtx_insn *head, *tail;
3022
3023 gcc_assert (EBB_FIRST_BB (bb) == EBB_LAST_BB (bb));
3024 get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail);
3025
3026 if (no_real_insns_p (head, tail))
3027 continue;
3028
3029 rgn_n_insns += set_priorities (head, tail);
3030 }
3031 current_sched_info->sched_max_insns_priority++;
3032 }
3033
3034 /* (Re-)initialize the arrays of DFA states at the end of each basic block.
3035
3036 SAVED_LAST_BASIC_BLOCK is the previous length of the arrays. It must be
3037 zero for the first call to this function, to allocate the arrays for the
3038 first time.
3039
3040 This function is called once during initialization of the scheduler, and
3041 called again to resize the arrays if new basic blocks have been created,
3042 for example for speculation recovery code. */
3043
3044 static void
realloc_bb_state_array(int saved_last_basic_block)3045 realloc_bb_state_array (int saved_last_basic_block)
3046 {
3047 char *old_bb_state_array = bb_state_array;
3048 size_t lbb = (size_t) last_basic_block_for_fn (cfun);
3049 size_t slbb = (size_t) saved_last_basic_block;
3050
3051 /* Nothing to do if nothing changed since the last time this was called. */
3052 if (saved_last_basic_block == last_basic_block_for_fn (cfun))
3053 return;
3054
3055 /* The selective scheduler doesn't use the state arrays. */
3056 if (sel_sched_p ())
3057 {
3058 gcc_assert (bb_state_array == NULL && bb_state == NULL);
3059 return;
3060 }
3061
3062 gcc_checking_assert (saved_last_basic_block == 0
3063 || (bb_state_array != NULL && bb_state != NULL));
3064
3065 bb_state_array = XRESIZEVEC (char, bb_state_array, lbb * dfa_state_size);
3066 bb_state = XRESIZEVEC (state_t, bb_state, lbb);
3067
3068 /* If BB_STATE_ARRAY has moved, fixup all the state pointers array.
3069 Otherwise only fixup the newly allocated ones. For the state
3070 array itself, only initialize the new entries. */
3071 bool bb_state_array_moved = (bb_state_array != old_bb_state_array);
3072 for (size_t i = bb_state_array_moved ? 0 : slbb; i < lbb; i++)
3073 bb_state[i] = (state_t) (bb_state_array + i * dfa_state_size);
3074 for (size_t i = slbb; i < lbb; i++)
3075 state_reset (bb_state[i]);
3076 }
3077
3078 /* Free the arrays of DFA states at the end of each basic block. */
3079
3080 static void
free_bb_state_array(void)3081 free_bb_state_array (void)
3082 {
3083 free (bb_state_array);
3084 free (bb_state);
3085 bb_state_array = NULL;
3086 bb_state = NULL;
3087 }
3088
3089 /* Schedule a region. A region is either an inner loop, a loop-free
3090 subroutine, or a single basic block. Each bb in the region is
3091 scheduled after its flow predecessors. */
3092
3093 static void
schedule_region(int rgn)3094 schedule_region (int rgn)
3095 {
3096 int bb;
3097 int sched_rgn_n_insns = 0;
3098
3099 rgn_n_insns = 0;
3100
3101 /* Do not support register pressure sensitive scheduling for the new regions
3102 as we don't update the liveness info for them. */
3103 if (sched_pressure != SCHED_PRESSURE_NONE
3104 && rgn >= nr_regions_initial)
3105 {
3106 free_global_sched_pressure_data ();
3107 sched_pressure = SCHED_PRESSURE_NONE;
3108 }
3109
3110 rgn_setup_region (rgn);
3111
3112 /* Don't schedule region that is marked by
3113 NOTE_DISABLE_SCHED_OF_BLOCK. */
3114 if (sched_is_disabled_for_current_region_p ())
3115 return;
3116
3117 sched_rgn_compute_dependencies (rgn);
3118
3119 sched_rgn_local_init (rgn);
3120
3121 /* Set priorities. */
3122 compute_priorities ();
3123
3124 sched_extend_ready_list (rgn_n_insns);
3125
3126 if (sched_pressure == SCHED_PRESSURE_WEIGHTED)
3127 {
3128 sched_init_region_reg_pressure_info ();
3129 for (bb = 0; bb < current_nr_blocks; bb++)
3130 {
3131 basic_block first_bb, last_bb;
3132 rtx_insn *head, *tail;
3133
3134 first_bb = EBB_FIRST_BB (bb);
3135 last_bb = EBB_LAST_BB (bb);
3136
3137 get_ebb_head_tail (first_bb, last_bb, &head, &tail);
3138
3139 if (no_real_insns_p (head, tail))
3140 {
3141 gcc_assert (first_bb == last_bb);
3142 continue;
3143 }
3144 sched_setup_bb_reg_pressure_info (first_bb, PREV_INSN (head));
3145 }
3146 }
3147
3148 /* Now we can schedule all blocks. */
3149 for (bb = 0; bb < current_nr_blocks; bb++)
3150 {
3151 basic_block first_bb, last_bb, curr_bb;
3152 rtx_insn *head, *tail;
3153
3154 first_bb = EBB_FIRST_BB (bb);
3155 last_bb = EBB_LAST_BB (bb);
3156
3157 get_ebb_head_tail (first_bb, last_bb, &head, &tail);
3158
3159 if (no_real_insns_p (head, tail))
3160 {
3161 gcc_assert (first_bb == last_bb);
3162 continue;
3163 }
3164
3165 current_sched_info->prev_head = PREV_INSN (head);
3166 current_sched_info->next_tail = NEXT_INSN (tail);
3167
3168 remove_notes (head, tail);
3169
3170 unlink_bb_notes (first_bb, last_bb);
3171
3172 target_bb = bb;
3173
3174 gcc_assert (flag_schedule_interblock || current_nr_blocks == 1);
3175 current_sched_info->queue_must_finish_empty = current_nr_blocks == 1;
3176
3177 curr_bb = first_bb;
3178 if (dbg_cnt (sched_block))
3179 {
3180 edge f;
3181 int saved_last_basic_block = last_basic_block_for_fn (cfun);
3182
3183 schedule_block (&curr_bb, bb_state[first_bb->index]);
3184 gcc_assert (EBB_FIRST_BB (bb) == first_bb);
3185 sched_rgn_n_insns += sched_n_insns;
3186 realloc_bb_state_array (saved_last_basic_block);
3187 f = find_fallthru_edge (last_bb->succs);
3188 if (f
3189 && (!f->probability.initialized_p ()
3190 || (f->probability.to_reg_br_prob_base () * 100
3191 / REG_BR_PROB_BASE
3192 >= param_sched_state_edge_prob_cutoff)))
3193 {
3194 memcpy (bb_state[f->dest->index], curr_state,
3195 dfa_state_size);
3196 if (sched_verbose >= 5)
3197 fprintf (sched_dump, "saving state for edge %d->%d\n",
3198 f->src->index, f->dest->index);
3199 }
3200 }
3201 else
3202 {
3203 sched_rgn_n_insns += rgn_n_insns;
3204 }
3205
3206 /* Clean up. */
3207 if (current_nr_blocks > 1)
3208 free_trg_info ();
3209 }
3210
3211 /* Sanity check: verify that all region insns were scheduled. */
3212 gcc_assert (sched_rgn_n_insns == rgn_n_insns);
3213
3214 sched_finish_ready_list ();
3215
3216 /* Done with this region. */
3217 sched_rgn_local_finish ();
3218
3219 /* Free dependencies. */
3220 for (bb = 0; bb < current_nr_blocks; ++bb)
3221 free_block_dependencies (bb);
3222
3223 gcc_assert (haifa_recovery_bb_ever_added_p
3224 || deps_pools_are_empty_p ());
3225 }
3226
3227 /* Initialize data structures for region scheduling. */
3228
3229 void
sched_rgn_init(bool single_blocks_p)3230 sched_rgn_init (bool single_blocks_p)
3231 {
3232 min_spec_prob = ((param_min_spec_prob * REG_BR_PROB_BASE)
3233 / 100);
3234
3235 nr_inter = 0;
3236 nr_spec = 0;
3237
3238 extend_regions ();
3239
3240 CONTAINING_RGN (ENTRY_BLOCK) = -1;
3241 CONTAINING_RGN (EXIT_BLOCK) = -1;
3242
3243 realloc_bb_state_array (0);
3244
3245 /* Compute regions for scheduling. */
3246 if (single_blocks_p
3247 || n_basic_blocks_for_fn (cfun) == NUM_FIXED_BLOCKS + 1
3248 || !flag_schedule_interblock
3249 || is_cfg_nonregular ())
3250 {
3251 find_single_block_region (sel_sched_p ());
3252 }
3253 else
3254 {
3255 /* Compute the dominators and post dominators. */
3256 if (!sel_sched_p ())
3257 calculate_dominance_info (CDI_DOMINATORS);
3258
3259 /* Find regions. */
3260 find_rgns ();
3261
3262 if (sched_verbose >= 3)
3263 debug_regions ();
3264
3265 /* For now. This will move as more and more of haifa is converted
3266 to using the cfg code. */
3267 if (!sel_sched_p ())
3268 free_dominance_info (CDI_DOMINATORS);
3269 }
3270
3271 gcc_assert (nr_regions > 0 && nr_regions <= n_basic_blocks_for_fn (cfun));
3272
3273 RGN_BLOCKS (nr_regions) = (RGN_BLOCKS (nr_regions - 1)
3274 + RGN_NR_BLOCKS (nr_regions - 1));
3275 nr_regions_initial = nr_regions;
3276 }
3277
3278 /* Free data structures for region scheduling. */
3279 void
sched_rgn_finish(void)3280 sched_rgn_finish (void)
3281 {
3282 free_bb_state_array ();
3283
3284 /* Reposition the prologue and epilogue notes in case we moved the
3285 prologue/epilogue insns. */
3286 if (reload_completed)
3287 reposition_prologue_and_epilogue_notes ();
3288
3289 if (sched_verbose)
3290 {
3291 if (reload_completed == 0
3292 && flag_schedule_interblock)
3293 {
3294 fprintf (sched_dump,
3295 "\n;; Procedure interblock/speculative motions == %d/%d \n",
3296 nr_inter, nr_spec);
3297 }
3298 else
3299 gcc_assert (nr_inter <= 0);
3300 fprintf (sched_dump, "\n\n");
3301 }
3302
3303 nr_regions = 0;
3304
3305 free (rgn_table);
3306 rgn_table = NULL;
3307
3308 free (rgn_bb_table);
3309 rgn_bb_table = NULL;
3310
3311 free (block_to_bb);
3312 block_to_bb = NULL;
3313
3314 free (containing_rgn);
3315 containing_rgn = NULL;
3316
3317 free (ebb_head);
3318 ebb_head = NULL;
3319 }
3320
3321 /* Setup global variables like CURRENT_BLOCKS and CURRENT_NR_BLOCK to
3322 point to the region RGN. */
3323 void
rgn_setup_region(int rgn)3324 rgn_setup_region (int rgn)
3325 {
3326 int bb;
3327
3328 /* Set variables for the current region. */
3329 current_nr_blocks = RGN_NR_BLOCKS (rgn);
3330 current_blocks = RGN_BLOCKS (rgn);
3331
3332 /* EBB_HEAD is a region-scope structure. But we realloc it for
3333 each region to save time/memory/something else.
3334 See comments in add_block1, for what reasons we allocate +1 element. */
3335 ebb_head = XRESIZEVEC (int, ebb_head, current_nr_blocks + 1);
3336 for (bb = 0; bb <= current_nr_blocks; bb++)
3337 ebb_head[bb] = current_blocks + bb;
3338 }
3339
3340 /* Compute instruction dependencies in region RGN. */
3341 void
sched_rgn_compute_dependencies(int rgn)3342 sched_rgn_compute_dependencies (int rgn)
3343 {
3344 if (!RGN_DONT_CALC_DEPS (rgn))
3345 {
3346 int bb;
3347
3348 if (sel_sched_p ())
3349 sched_emulate_haifa_p = 1;
3350
3351 init_deps_global ();
3352
3353 /* Initializations for region data dependence analysis. */
3354 bb_deps = XNEWVEC (class deps_desc, current_nr_blocks);
3355 for (bb = 0; bb < current_nr_blocks; bb++)
3356 init_deps (bb_deps + bb, false);
3357
3358 /* Initialize bitmap used in add_branch_dependences. */
3359 insn_referenced = sbitmap_alloc (sched_max_luid);
3360 bitmap_clear (insn_referenced);
3361
3362 /* Compute backward dependencies. */
3363 for (bb = 0; bb < current_nr_blocks; bb++)
3364 compute_block_dependences (bb);
3365
3366 sbitmap_free (insn_referenced);
3367 free_pending_lists ();
3368 finish_deps_global ();
3369 free (bb_deps);
3370
3371 /* We don't want to recalculate this twice. */
3372 RGN_DONT_CALC_DEPS (rgn) = 1;
3373
3374 if (sel_sched_p ())
3375 sched_emulate_haifa_p = 0;
3376 }
3377 else
3378 /* (This is a recovery block. It is always a single block region.)
3379 OR (We use selective scheduling.) */
3380 gcc_assert (current_nr_blocks == 1 || sel_sched_p ());
3381 }
3382
3383 /* Init region data structures. Returns true if this region should
3384 not be scheduled. */
3385 void
sched_rgn_local_init(int rgn)3386 sched_rgn_local_init (int rgn)
3387 {
3388 int bb;
3389
3390 /* Compute interblock info: probabilities, split-edges, dominators, etc. */
3391 if (current_nr_blocks > 1)
3392 {
3393 basic_block block;
3394 edge e;
3395 edge_iterator ei;
3396
3397 prob = XNEWVEC (int, current_nr_blocks);
3398
3399 dom = sbitmap_vector_alloc (current_nr_blocks, current_nr_blocks);
3400 bitmap_vector_clear (dom, current_nr_blocks);
3401
3402 /* Use ->aux to implement EDGE_TO_BIT mapping. */
3403 rgn_nr_edges = 0;
3404 FOR_EACH_BB_FN (block, cfun)
3405 {
3406 if (CONTAINING_RGN (block->index) != rgn)
3407 continue;
3408 FOR_EACH_EDGE (e, ei, block->succs)
3409 SET_EDGE_TO_BIT (e, rgn_nr_edges++);
3410 }
3411
3412 rgn_edges = XNEWVEC (edge, rgn_nr_edges);
3413 rgn_nr_edges = 0;
3414 FOR_EACH_BB_FN (block, cfun)
3415 {
3416 if (CONTAINING_RGN (block->index) != rgn)
3417 continue;
3418 FOR_EACH_EDGE (e, ei, block->succs)
3419 rgn_edges[rgn_nr_edges++] = e;
3420 }
3421
3422 /* Split edges. */
3423 pot_split = sbitmap_vector_alloc (current_nr_blocks, rgn_nr_edges);
3424 bitmap_vector_clear (pot_split, current_nr_blocks);
3425 ancestor_edges = sbitmap_vector_alloc (current_nr_blocks, rgn_nr_edges);
3426 bitmap_vector_clear (ancestor_edges, current_nr_blocks);
3427
3428 /* Compute probabilities, dominators, split_edges. */
3429 for (bb = 0; bb < current_nr_blocks; bb++)
3430 compute_dom_prob_ps (bb);
3431
3432 /* Cleanup ->aux used for EDGE_TO_BIT mapping. */
3433 /* We don't need them anymore. But we want to avoid duplication of
3434 aux fields in the newly created edges. */
3435 FOR_EACH_BB_FN (block, cfun)
3436 {
3437 if (CONTAINING_RGN (block->index) != rgn)
3438 continue;
3439 FOR_EACH_EDGE (e, ei, block->succs)
3440 e->aux = NULL;
3441 }
3442 }
3443 }
3444
3445 /* Free data computed for the finished region. */
3446 void
sched_rgn_local_free(void)3447 sched_rgn_local_free (void)
3448 {
3449 free (prob);
3450 sbitmap_vector_free (dom);
3451 sbitmap_vector_free (pot_split);
3452 sbitmap_vector_free (ancestor_edges);
3453 free (rgn_edges);
3454 }
3455
3456 /* Free data computed for the finished region. */
3457 void
sched_rgn_local_finish(void)3458 sched_rgn_local_finish (void)
3459 {
3460 if (current_nr_blocks > 1 && !sel_sched_p ())
3461 {
3462 sched_rgn_local_free ();
3463 }
3464 }
3465
3466 /* Setup scheduler infos. */
3467 void
rgn_setup_common_sched_info(void)3468 rgn_setup_common_sched_info (void)
3469 {
3470 memcpy (&rgn_common_sched_info, &haifa_common_sched_info,
3471 sizeof (rgn_common_sched_info));
3472
3473 rgn_common_sched_info.fix_recovery_cfg = rgn_fix_recovery_cfg;
3474 rgn_common_sched_info.add_block = rgn_add_block;
3475 rgn_common_sched_info.estimate_number_of_insns
3476 = rgn_estimate_number_of_insns;
3477 rgn_common_sched_info.sched_pass_id = SCHED_RGN_PASS;
3478
3479 common_sched_info = &rgn_common_sched_info;
3480 }
3481
3482 /* Setup all *_sched_info structures (for the Haifa frontend
3483 and for the dependence analysis) in the interblock scheduler. */
3484 void
rgn_setup_sched_infos(void)3485 rgn_setup_sched_infos (void)
3486 {
3487 if (!sel_sched_p ())
3488 memcpy (&rgn_sched_deps_info, &rgn_const_sched_deps_info,
3489 sizeof (rgn_sched_deps_info));
3490 else
3491 memcpy (&rgn_sched_deps_info, &rgn_const_sel_sched_deps_info,
3492 sizeof (rgn_sched_deps_info));
3493
3494 sched_deps_info = &rgn_sched_deps_info;
3495
3496 memcpy (&rgn_sched_info, &rgn_const_sched_info, sizeof (rgn_sched_info));
3497 current_sched_info = &rgn_sched_info;
3498 }
3499
3500 /* The one entry point in this file. */
3501 void
schedule_insns(void)3502 schedule_insns (void)
3503 {
3504 int rgn;
3505
3506 /* Taking care of this degenerate case makes the rest of
3507 this code simpler. */
3508 if (n_basic_blocks_for_fn (cfun) == NUM_FIXED_BLOCKS)
3509 return;
3510
3511 rgn_setup_common_sched_info ();
3512 rgn_setup_sched_infos ();
3513
3514 haifa_sched_init ();
3515 sched_rgn_init (reload_completed);
3516
3517 bitmap_initialize (¬_in_df, &bitmap_default_obstack);
3518
3519 /* Schedule every region in the subroutine. */
3520 for (rgn = 0; rgn < nr_regions; rgn++)
3521 if (dbg_cnt (sched_region))
3522 schedule_region (rgn);
3523
3524 /* Clean up. */
3525 sched_rgn_finish ();
3526 bitmap_release (¬_in_df);
3527
3528 haifa_sched_finish ();
3529 }
3530
3531 /* INSN has been added to/removed from current region. */
3532 static void
rgn_add_remove_insn(rtx_insn * insn,int remove_p)3533 rgn_add_remove_insn (rtx_insn *insn, int remove_p)
3534 {
3535 if (!remove_p)
3536 rgn_n_insns++;
3537 else
3538 rgn_n_insns--;
3539
3540 if (INSN_BB (insn) == target_bb)
3541 {
3542 if (!remove_p)
3543 target_n_insns++;
3544 else
3545 target_n_insns--;
3546 }
3547 }
3548
3549 /* Extend internal data structures. */
3550 void
extend_regions(void)3551 extend_regions (void)
3552 {
3553 rgn_table = XRESIZEVEC (region, rgn_table, n_basic_blocks_for_fn (cfun));
3554 rgn_bb_table = XRESIZEVEC (int, rgn_bb_table,
3555 n_basic_blocks_for_fn (cfun));
3556 block_to_bb = XRESIZEVEC (int, block_to_bb,
3557 last_basic_block_for_fn (cfun));
3558 containing_rgn = XRESIZEVEC (int, containing_rgn,
3559 last_basic_block_for_fn (cfun));
3560 }
3561
3562 void
rgn_make_new_region_out_of_new_block(basic_block bb)3563 rgn_make_new_region_out_of_new_block (basic_block bb)
3564 {
3565 int i;
3566
3567 i = RGN_BLOCKS (nr_regions);
3568 /* I - first free position in rgn_bb_table. */
3569
3570 rgn_bb_table[i] = bb->index;
3571 RGN_NR_BLOCKS (nr_regions) = 1;
3572 RGN_HAS_REAL_EBB (nr_regions) = 0;
3573 RGN_DONT_CALC_DEPS (nr_regions) = 0;
3574 CONTAINING_RGN (bb->index) = nr_regions;
3575 BLOCK_TO_BB (bb->index) = 0;
3576
3577 nr_regions++;
3578
3579 RGN_BLOCKS (nr_regions) = i + 1;
3580 }
3581
3582 /* BB was added to ebb after AFTER. */
3583 static void
rgn_add_block(basic_block bb,basic_block after)3584 rgn_add_block (basic_block bb, basic_block after)
3585 {
3586 extend_regions ();
3587 bitmap_set_bit (¬_in_df, bb->index);
3588
3589 if (after == 0 || after == EXIT_BLOCK_PTR_FOR_FN (cfun))
3590 {
3591 rgn_make_new_region_out_of_new_block (bb);
3592 RGN_DONT_CALC_DEPS (nr_regions - 1) = (after
3593 == EXIT_BLOCK_PTR_FOR_FN (cfun));
3594 }
3595 else
3596 {
3597 int i, pos;
3598
3599 /* We need to fix rgn_table, block_to_bb, containing_rgn
3600 and ebb_head. */
3601
3602 BLOCK_TO_BB (bb->index) = BLOCK_TO_BB (after->index);
3603
3604 /* We extend ebb_head to one more position to
3605 easily find the last position of the last ebb in
3606 the current region. Thus, ebb_head[BLOCK_TO_BB (after) + 1]
3607 is _always_ valid for access. */
3608
3609 i = BLOCK_TO_BB (after->index) + 1;
3610 pos = ebb_head[i] - 1;
3611 /* Now POS is the index of the last block in the region. */
3612
3613 /* Find index of basic block AFTER. */
3614 for (; rgn_bb_table[pos] != after->index; pos--)
3615 ;
3616
3617 pos++;
3618 gcc_assert (pos > ebb_head[i - 1]);
3619
3620 /* i - ebb right after "AFTER". */
3621 /* ebb_head[i] - VALID. */
3622
3623 /* Source position: ebb_head[i]
3624 Destination position: ebb_head[i] + 1
3625 Last position:
3626 RGN_BLOCKS (nr_regions) - 1
3627 Number of elements to copy: (last_position) - (source_position) + 1
3628 */
3629
3630 memmove (rgn_bb_table + pos + 1,
3631 rgn_bb_table + pos,
3632 ((RGN_BLOCKS (nr_regions) - 1) - (pos) + 1)
3633 * sizeof (*rgn_bb_table));
3634
3635 rgn_bb_table[pos] = bb->index;
3636
3637 for (; i <= current_nr_blocks; i++)
3638 ebb_head [i]++;
3639
3640 i = CONTAINING_RGN (after->index);
3641 CONTAINING_RGN (bb->index) = i;
3642
3643 RGN_HAS_REAL_EBB (i) = 1;
3644
3645 for (++i; i <= nr_regions; i++)
3646 RGN_BLOCKS (i)++;
3647 }
3648 }
3649
3650 /* Fix internal data after interblock movement of jump instruction.
3651 For parameter meaning please refer to
3652 sched-int.h: struct sched_info: fix_recovery_cfg. */
3653 static void
rgn_fix_recovery_cfg(int bbi,int check_bbi,int check_bb_nexti)3654 rgn_fix_recovery_cfg (int bbi, int check_bbi, int check_bb_nexti)
3655 {
3656 int old_pos, new_pos, i;
3657
3658 BLOCK_TO_BB (check_bb_nexti) = BLOCK_TO_BB (bbi);
3659
3660 for (old_pos = ebb_head[BLOCK_TO_BB (check_bbi) + 1] - 1;
3661 rgn_bb_table[old_pos] != check_bb_nexti;
3662 old_pos--)
3663 ;
3664 gcc_assert (old_pos > ebb_head[BLOCK_TO_BB (check_bbi)]);
3665
3666 for (new_pos = ebb_head[BLOCK_TO_BB (bbi) + 1] - 1;
3667 rgn_bb_table[new_pos] != bbi;
3668 new_pos--)
3669 ;
3670 new_pos++;
3671 gcc_assert (new_pos > ebb_head[BLOCK_TO_BB (bbi)]);
3672
3673 gcc_assert (new_pos < old_pos);
3674
3675 memmove (rgn_bb_table + new_pos + 1,
3676 rgn_bb_table + new_pos,
3677 (old_pos - new_pos) * sizeof (*rgn_bb_table));
3678
3679 rgn_bb_table[new_pos] = check_bb_nexti;
3680
3681 for (i = BLOCK_TO_BB (bbi) + 1; i <= BLOCK_TO_BB (check_bbi); i++)
3682 ebb_head[i]++;
3683 }
3684
3685 /* Return next block in ebb chain. For parameter meaning please refer to
3686 sched-int.h: struct sched_info: advance_target_bb. */
3687 static basic_block
advance_target_bb(basic_block bb,rtx_insn * insn)3688 advance_target_bb (basic_block bb, rtx_insn *insn)
3689 {
3690 if (insn)
3691 return 0;
3692
3693 gcc_assert (BLOCK_TO_BB (bb->index) == target_bb
3694 && BLOCK_TO_BB (bb->next_bb->index) == target_bb);
3695 return bb->next_bb;
3696 }
3697
3698 #endif
3699
3700 /* Run instruction scheduler. */
3701 static unsigned int
rest_of_handle_live_range_shrinkage(void)3702 rest_of_handle_live_range_shrinkage (void)
3703 {
3704 #ifdef INSN_SCHEDULING
3705 int saved;
3706
3707 initialize_live_range_shrinkage ();
3708 saved = flag_schedule_interblock;
3709 flag_schedule_interblock = false;
3710 schedule_insns ();
3711 flag_schedule_interblock = saved;
3712 finish_live_range_shrinkage ();
3713 #endif
3714 return 0;
3715 }
3716
3717 /* Run instruction scheduler. */
3718 static unsigned int
rest_of_handle_sched(void)3719 rest_of_handle_sched (void)
3720 {
3721 #ifdef INSN_SCHEDULING
3722 if (flag_selective_scheduling
3723 && ! maybe_skip_selective_scheduling ())
3724 run_selective_scheduling ();
3725 else
3726 schedule_insns ();
3727 #endif
3728 return 0;
3729 }
3730
3731 /* Run second scheduling pass after reload. */
3732 static unsigned int
rest_of_handle_sched2(void)3733 rest_of_handle_sched2 (void)
3734 {
3735 #ifdef INSN_SCHEDULING
3736 if (flag_selective_scheduling2
3737 && ! maybe_skip_selective_scheduling ())
3738 run_selective_scheduling ();
3739 else
3740 {
3741 /* Do control and data sched analysis again,
3742 and write some more of the results to dump file. */
3743 if (flag_sched2_use_superblocks)
3744 schedule_ebbs ();
3745 else
3746 schedule_insns ();
3747 }
3748 #endif
3749 return 0;
3750 }
3751
3752 static unsigned int
rest_of_handle_sched_fusion(void)3753 rest_of_handle_sched_fusion (void)
3754 {
3755 #ifdef INSN_SCHEDULING
3756 sched_fusion = true;
3757 schedule_insns ();
3758 sched_fusion = false;
3759 #endif
3760 return 0;
3761 }
3762
3763 namespace {
3764
3765 const pass_data pass_data_live_range_shrinkage =
3766 {
3767 RTL_PASS, /* type */
3768 "lr_shrinkage", /* name */
3769 OPTGROUP_NONE, /* optinfo_flags */
3770 TV_LIVE_RANGE_SHRINKAGE, /* tv_id */
3771 0, /* properties_required */
3772 0, /* properties_provided */
3773 0, /* properties_destroyed */
3774 0, /* todo_flags_start */
3775 TODO_df_finish, /* todo_flags_finish */
3776 };
3777
3778 class pass_live_range_shrinkage : public rtl_opt_pass
3779 {
3780 public:
pass_live_range_shrinkage(gcc::context * ctxt)3781 pass_live_range_shrinkage(gcc::context *ctxt)
3782 : rtl_opt_pass(pass_data_live_range_shrinkage, ctxt)
3783 {}
3784
3785 /* opt_pass methods: */
gate(function *)3786 virtual bool gate (function *)
3787 {
3788 #ifdef INSN_SCHEDULING
3789 return flag_live_range_shrinkage;
3790 #else
3791 return 0;
3792 #endif
3793 }
3794
execute(function *)3795 virtual unsigned int execute (function *)
3796 {
3797 return rest_of_handle_live_range_shrinkage ();
3798 }
3799
3800 }; // class pass_live_range_shrinkage
3801
3802 } // anon namespace
3803
3804 rtl_opt_pass *
make_pass_live_range_shrinkage(gcc::context * ctxt)3805 make_pass_live_range_shrinkage (gcc::context *ctxt)
3806 {
3807 return new pass_live_range_shrinkage (ctxt);
3808 }
3809
3810 namespace {
3811
3812 const pass_data pass_data_sched =
3813 {
3814 RTL_PASS, /* type */
3815 "sched1", /* name */
3816 OPTGROUP_NONE, /* optinfo_flags */
3817 TV_SCHED, /* tv_id */
3818 0, /* properties_required */
3819 0, /* properties_provided */
3820 0, /* properties_destroyed */
3821 0, /* todo_flags_start */
3822 TODO_df_finish, /* todo_flags_finish */
3823 };
3824
3825 class pass_sched : public rtl_opt_pass
3826 {
3827 public:
pass_sched(gcc::context * ctxt)3828 pass_sched (gcc::context *ctxt)
3829 : rtl_opt_pass (pass_data_sched, ctxt)
3830 {}
3831
3832 /* opt_pass methods: */
3833 virtual bool gate (function *);
execute(function *)3834 virtual unsigned int execute (function *) { return rest_of_handle_sched (); }
3835
3836 }; // class pass_sched
3837
3838 bool
gate(function *)3839 pass_sched::gate (function *)
3840 {
3841 #ifdef INSN_SCHEDULING
3842 return optimize > 0 && flag_schedule_insns && dbg_cnt (sched_func);
3843 #else
3844 return 0;
3845 #endif
3846 }
3847
3848 } // anon namespace
3849
3850 rtl_opt_pass *
make_pass_sched(gcc::context * ctxt)3851 make_pass_sched (gcc::context *ctxt)
3852 {
3853 return new pass_sched (ctxt);
3854 }
3855
3856 namespace {
3857
3858 const pass_data pass_data_sched2 =
3859 {
3860 RTL_PASS, /* type */
3861 "sched2", /* name */
3862 OPTGROUP_NONE, /* optinfo_flags */
3863 TV_SCHED2, /* tv_id */
3864 0, /* properties_required */
3865 0, /* properties_provided */
3866 0, /* properties_destroyed */
3867 0, /* todo_flags_start */
3868 TODO_df_finish, /* todo_flags_finish */
3869 };
3870
3871 class pass_sched2 : public rtl_opt_pass
3872 {
3873 public:
pass_sched2(gcc::context * ctxt)3874 pass_sched2 (gcc::context *ctxt)
3875 : rtl_opt_pass (pass_data_sched2, ctxt)
3876 {}
3877
3878 /* opt_pass methods: */
3879 virtual bool gate (function *);
execute(function *)3880 virtual unsigned int execute (function *)
3881 {
3882 return rest_of_handle_sched2 ();
3883 }
3884
3885 }; // class pass_sched2
3886
3887 bool
gate(function *)3888 pass_sched2::gate (function *)
3889 {
3890 #ifdef INSN_SCHEDULING
3891 return optimize > 0 && flag_schedule_insns_after_reload
3892 && !targetm.delay_sched2 && dbg_cnt (sched2_func);
3893 #else
3894 return 0;
3895 #endif
3896 }
3897
3898 } // anon namespace
3899
3900 rtl_opt_pass *
make_pass_sched2(gcc::context * ctxt)3901 make_pass_sched2 (gcc::context *ctxt)
3902 {
3903 return new pass_sched2 (ctxt);
3904 }
3905
3906 namespace {
3907
3908 const pass_data pass_data_sched_fusion =
3909 {
3910 RTL_PASS, /* type */
3911 "sched_fusion", /* name */
3912 OPTGROUP_NONE, /* optinfo_flags */
3913 TV_SCHED_FUSION, /* tv_id */
3914 0, /* properties_required */
3915 0, /* properties_provided */
3916 0, /* properties_destroyed */
3917 0, /* todo_flags_start */
3918 TODO_df_finish, /* todo_flags_finish */
3919 };
3920
3921 class pass_sched_fusion : public rtl_opt_pass
3922 {
3923 public:
pass_sched_fusion(gcc::context * ctxt)3924 pass_sched_fusion (gcc::context *ctxt)
3925 : rtl_opt_pass (pass_data_sched_fusion, ctxt)
3926 {}
3927
3928 /* opt_pass methods: */
3929 virtual bool gate (function *);
execute(function *)3930 virtual unsigned int execute (function *)
3931 {
3932 return rest_of_handle_sched_fusion ();
3933 }
3934
3935 }; // class pass_sched2
3936
3937 bool
gate(function *)3938 pass_sched_fusion::gate (function *)
3939 {
3940 #ifdef INSN_SCHEDULING
3941 /* Scheduling fusion relies on peephole2 to do real fusion work,
3942 so only enable it if peephole2 is in effect. */
3943 return (optimize > 0 && flag_peephole2
3944 && flag_schedule_fusion && targetm.sched.fusion_priority != NULL);
3945 #else
3946 return 0;
3947 #endif
3948 }
3949
3950 } // anon namespace
3951
3952 rtl_opt_pass *
make_pass_sched_fusion(gcc::context * ctxt)3953 make_pass_sched_fusion (gcc::context *ctxt)
3954 {
3955 return new pass_sched_fusion (ctxt);
3956 }
3957
3958 #if __GNUC__ >= 10
3959 # pragma GCC diagnostic pop
3960 #endif
3961