1 | /* Basic block reordering routines for the GNU compiler. |
2 | Copyright (C) 2000-2023 Free Software Foundation, Inc. |
3 | |
4 | This file is part of GCC. |
5 | |
6 | GCC is free software; you can redistribute it and/or modify it |
7 | under the terms of the GNU General Public License as published by |
8 | the Free Software Foundation; either version 3, or (at your option) |
9 | any later version. |
10 | |
11 | GCC is distributed in the hope that it will be useful, but WITHOUT |
12 | ANY WARRANTY; without even the implied warranty of MERCHANTABILITY |
13 | or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public |
14 | License for more details. |
15 | |
16 | You should have received a copy of the GNU General Public License |
17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ |
19 | |
20 | /* This file contains the "reorder blocks" pass, which changes the control |
21 | flow of a function to encounter fewer branches; the "partition blocks" |
22 | pass, which divides the basic blocks into "hot" and "cold" partitions, |
23 | which are kept separate; and the "duplicate computed gotos" pass, which |
24 | duplicates blocks ending in an indirect jump. |
25 | |
26 | There are two algorithms for "reorder blocks": the "simple" algorithm, |
27 | which just rearranges blocks, trying to minimize the number of executed |
28 | unconditional branches; and the "software trace cache" algorithm, which |
29 | also copies code, and in general tries a lot harder to have long linear |
30 | pieces of machine code executed. This algorithm is described next. */ |
31 | |
32 | /* This (greedy) algorithm constructs traces in several rounds. |
33 | The construction starts from "seeds". The seed for the first round |
34 | is the entry point of the function. When there are more than one seed, |
35 | the one with the lowest key in the heap is selected first (see bb_to_key). |
36 | Then the algorithm repeatedly adds the most probable successor to the end |
37 | of a trace. Finally it connects the traces. |
38 | |
39 | There are two parameters: Branch Threshold and Exec Threshold. |
40 | If the probability of an edge to a successor of the current basic block is |
41 | lower than Branch Threshold or its count is lower than Exec Threshold, |
42 | then the successor will be the seed in one of the next rounds. |
43 | Each round has these parameters lower than the previous one. |
44 | The last round has to have these parameters set to zero so that the |
45 | remaining blocks are picked up. |
46 | |
47 | The algorithm selects the most probable successor from all unvisited |
48 | successors and successors that have been added to this trace. |
49 | The other successors (that has not been "sent" to the next round) will be |
50 | other seeds for this round and the secondary traces will start from them. |
51 | If the successor has not been visited in this trace, it is added to the |
52 | trace (however, there is some heuristic for simple branches). |
53 | If the successor has been visited in this trace, a loop has been found. |
54 | If the loop has many iterations, the loop is rotated so that the source |
55 | block of the most probable edge going out of the loop is the last block |
56 | of the trace. |
57 | If the loop has few iterations and there is no edge from the last block of |
58 | the loop going out of the loop, the loop header is duplicated. |
59 | |
60 | When connecting traces, the algorithm first checks whether there is an edge |
61 | from the last block of a trace to the first block of another trace. |
62 | When there are still some unconnected traces it checks whether there exists |
63 | a basic block BB such that BB is a successor of the last block of a trace |
64 | and BB is a predecessor of the first block of another trace. In this case, |
65 | BB is duplicated, added at the end of the first trace and the traces are |
66 | connected through it. |
67 | The rest of traces are simply connected so there will be a jump to the |
68 | beginning of the rest of traces. |
69 | |
70 | The above description is for the full algorithm, which is used when the |
71 | function is optimized for speed. When the function is optimized for size, |
72 | in order to reduce long jumps and connect more fallthru edges, the |
73 | algorithm is modified as follows: |
74 | (1) Break long traces to short ones. A trace is broken at a block that has |
75 | multiple predecessors/ successors during trace discovery. When connecting |
76 | traces, only connect Trace n with Trace n + 1. This change reduces most |
77 | long jumps compared with the above algorithm. |
78 | (2) Ignore the edge probability and count for fallthru edges. |
79 | (3) Keep the original order of blocks when there is no chance to fall |
80 | through. We rely on the results of cfg_cleanup. |
81 | |
82 | To implement the change for code size optimization, block's index is |
83 | selected as the key and all traces are found in one round. |
84 | |
85 | References: |
86 | |
87 | "Software Trace Cache" |
88 | A. Ramirez, J. Larriba-Pey, C. Navarro, J. Torrellas and M. Valero; 1999 |
89 | http://citeseer.nj.nec.com/15361.html |
90 | |
91 | */ |
92 | |
93 | #include "config.h" |
94 | #include "system.h" |
95 | #include "coretypes.h" |
96 | #include "backend.h" |
97 | #include "target.h" |
98 | #include "rtl.h" |
99 | #include "tree.h" |
100 | #include "cfghooks.h" |
101 | #include "df.h" |
102 | #include "memmodel.h" |
103 | #include "optabs.h" |
104 | #include "regs.h" |
105 | #include "emit-rtl.h" |
106 | #include "output.h" |
107 | #include "expr.h" |
108 | #include "tree-pass.h" |
109 | #include "cfgrtl.h" |
110 | #include "cfganal.h" |
111 | #include "cfgbuild.h" |
112 | #include "cfgcleanup.h" |
113 | #include "bb-reorder.h" |
114 | #include "except.h" |
115 | #include "alloc-pool.h" |
116 | #include "fibonacci_heap.h" |
117 | #include "stringpool.h" |
118 | #include "attribs.h" |
119 | #include "common/common-target.h" |
120 | |
121 | /* The number of rounds. In most cases there will only be 4 rounds, but |
122 | when partitioning hot and cold basic blocks into separate sections of |
123 | the object file there will be an extra round. */ |
124 | #define N_ROUNDS 5 |
125 | |
126 | struct target_bb_reorder default_target_bb_reorder; |
127 | #if SWITCHABLE_TARGET |
128 | struct target_bb_reorder *this_target_bb_reorder = &default_target_bb_reorder; |
129 | #endif |
130 | |
131 | #define uncond_jump_length \ |
132 | (this_target_bb_reorder->x_uncond_jump_length) |
133 | |
134 | /* Branch thresholds in thousandths (per mille) of the REG_BR_PROB_BASE. */ |
135 | static const int branch_threshold[N_ROUNDS] = {400, 200, 100, 0, 0}; |
136 | |
137 | /* Exec thresholds in thousandths (per mille) of the count of bb 0. */ |
138 | static const int exec_threshold[N_ROUNDS] = {500, 200, 50, 0, 0}; |
139 | |
140 | /* If edge count is lower than DUPLICATION_THRESHOLD per mille of entry |
141 | block the edge destination is not duplicated while connecting traces. */ |
142 | #define DUPLICATION_THRESHOLD 100 |
143 | |
144 | typedef fibonacci_heap <long, basic_block_def> bb_heap_t; |
145 | typedef fibonacci_node <long, basic_block_def> bb_heap_node_t; |
146 | |
147 | /* Structure to hold needed information for each basic block. */ |
148 | struct bbro_basic_block_data |
149 | { |
150 | /* Which trace is the bb start of (-1 means it is not a start of any). */ |
151 | int start_of_trace; |
152 | |
153 | /* Which trace is the bb end of (-1 means it is not an end of any). */ |
154 | int end_of_trace; |
155 | |
156 | /* Which trace is the bb in? */ |
157 | int in_trace; |
158 | |
159 | /* Which trace was this bb visited in? */ |
160 | int visited; |
161 | |
162 | /* Cached maximum frequency of interesting incoming edges. |
163 | Minus one means not yet computed. */ |
164 | int priority; |
165 | |
166 | /* Which heap is BB in (if any)? */ |
167 | bb_heap_t *heap; |
168 | |
169 | /* Which heap node is BB in (if any)? */ |
170 | bb_heap_node_t *node; |
171 | }; |
172 | |
173 | /* The current size of the following dynamic array. */ |
174 | static int array_size; |
175 | |
176 | /* The array which holds needed information for basic blocks. */ |
177 | static bbro_basic_block_data *bbd; |
178 | |
179 | /* To avoid frequent reallocation the size of arrays is greater than needed, |
180 | the number of elements is (not less than) 1.25 * size_wanted. */ |
181 | #define GET_ARRAY_SIZE(X) ((((X) / 4) + 1) * 5) |
182 | |
183 | /* Free the memory and set the pointer to NULL. */ |
184 | #define FREE(P) (gcc_assert (P), free (P), P = 0) |
185 | |
186 | /* Structure for holding information about a trace. */ |
187 | struct trace |
188 | { |
189 | /* First and last basic block of the trace. */ |
190 | basic_block first, last; |
191 | |
192 | /* The round of the STC creation which this trace was found in. */ |
193 | int round; |
194 | |
195 | /* The length (i.e. the number of basic blocks) of the trace. */ |
196 | int length; |
197 | }; |
198 | |
199 | /* Maximum count of one of the entry blocks. */ |
200 | static profile_count max_entry_count; |
201 | |
202 | /* Local function prototypes. */ |
203 | static void find_traces_1_round (int, profile_count, struct trace *, int *, |
204 | int, bb_heap_t **, int); |
205 | static basic_block copy_bb (basic_block, edge, basic_block, int); |
206 | static long bb_to_key (basic_block); |
207 | static bool better_edge_p (const_basic_block, const_edge, profile_probability, |
208 | profile_count, profile_probability, profile_count, |
209 | const_edge); |
210 | static bool copy_bb_p (const_basic_block, int); |
211 | |
212 | /* Return the trace number in which BB was visited. */ |
213 | |
214 | static int |
215 | bb_visited_trace (const_basic_block bb) |
216 | { |
217 | gcc_assert (bb->index < array_size); |
218 | return bbd[bb->index].visited; |
219 | } |
220 | |
221 | /* This function marks BB that it was visited in trace number TRACE. */ |
222 | |
223 | static void |
224 | mark_bb_visited (basic_block bb, int trace) |
225 | { |
226 | bbd[bb->index].visited = trace; |
227 | if (bbd[bb->index].heap) |
228 | { |
229 | bbd[bb->index].heap->delete_node (node: bbd[bb->index].node); |
230 | bbd[bb->index].heap = NULL; |
231 | bbd[bb->index].node = NULL; |
232 | } |
233 | } |
234 | |
235 | /* Check to see if bb should be pushed into the next round of trace |
236 | collections or not. Reasons for pushing the block forward are 1). |
237 | If the block is cold, we are doing partitioning, and there will be |
238 | another round (cold partition blocks are not supposed to be |
239 | collected into traces until the very last round); or 2). There will |
240 | be another round, and the basic block is not "hot enough" for the |
241 | current round of trace collection. */ |
242 | |
243 | static bool |
244 | push_to_next_round_p (const_basic_block bb, int round, int number_of_rounds, |
245 | profile_count count_th) |
246 | { |
247 | bool there_exists_another_round; |
248 | bool block_not_hot_enough; |
249 | |
250 | there_exists_another_round = round < number_of_rounds - 1; |
251 | |
252 | block_not_hot_enough = (bb->count < count_th |
253 | || probably_never_executed_bb_p (cfun, bb)); |
254 | |
255 | if (there_exists_another_round |
256 | && block_not_hot_enough) |
257 | return true; |
258 | else |
259 | return false; |
260 | } |
261 | |
262 | /* Find the traces for Software Trace Cache. Chain each trace through |
263 | RBI()->next. Store the number of traces to N_TRACES and description of |
264 | traces to TRACES. */ |
265 | |
266 | static void |
267 | find_traces (int *n_traces, struct trace *traces) |
268 | { |
269 | int i; |
270 | int number_of_rounds; |
271 | edge e; |
272 | edge_iterator ei; |
273 | bb_heap_t *heap = new bb_heap_t (LONG_MIN); |
274 | |
275 | /* Add one extra round of trace collection when partitioning hot/cold |
276 | basic blocks into separate sections. The last round is for all the |
277 | cold blocks (and ONLY the cold blocks). */ |
278 | |
279 | number_of_rounds = N_ROUNDS - 1; |
280 | |
281 | /* Insert entry points of function into heap. */ |
282 | max_entry_count = profile_count::zero (); |
283 | FOR_EACH_EDGE (e, ei, ENTRY_BLOCK_PTR_FOR_FN (cfun)->succs) |
284 | { |
285 | bbd[e->dest->index].heap = heap; |
286 | bbd[e->dest->index].node = heap->insert (key: bb_to_key (e->dest), data: e->dest); |
287 | if (e->dest->count > max_entry_count) |
288 | max_entry_count = e->dest->count; |
289 | } |
290 | |
291 | /* Find the traces. */ |
292 | for (i = 0; i < number_of_rounds; i++) |
293 | { |
294 | profile_count count_threshold; |
295 | |
296 | if (dump_file) |
297 | fprintf (stream: dump_file, format: "STC - round %d\n" , i + 1); |
298 | |
299 | count_threshold = max_entry_count.apply_scale (num: exec_threshold[i], den: 1000); |
300 | |
301 | find_traces_1_round (REG_BR_PROB_BASE * branch_threshold[i] / 1000, |
302 | count_threshold, traces, n_traces, i, &heap, |
303 | number_of_rounds); |
304 | } |
305 | delete heap; |
306 | |
307 | if (dump_file) |
308 | { |
309 | for (i = 0; i < *n_traces; i++) |
310 | { |
311 | basic_block bb; |
312 | fprintf (stream: dump_file, format: "Trace %d (round %d): " , i + 1, |
313 | traces[i].round + 1); |
314 | for (bb = traces[i].first; |
315 | bb != traces[i].last; |
316 | bb = (basic_block) bb->aux) |
317 | { |
318 | fprintf (stream: dump_file, format: "%d [" , bb->index); |
319 | bb->count.dump (f: dump_file); |
320 | fprintf (stream: dump_file, format: "] " ); |
321 | } |
322 | fprintf (stream: dump_file, format: "%d [" , bb->index); |
323 | bb->count.dump (f: dump_file); |
324 | fprintf (stream: dump_file, format: "]\n" ); |
325 | } |
326 | fflush (stream: dump_file); |
327 | } |
328 | } |
329 | |
330 | /* Rotate loop whose back edge is BACK_EDGE in the tail of trace TRACE |
331 | (with sequential number TRACE_N). */ |
332 | |
333 | static basic_block |
334 | rotate_loop (edge back_edge, struct trace *trace, int trace_n) |
335 | { |
336 | basic_block bb; |
337 | |
338 | /* Information about the best end (end after rotation) of the loop. */ |
339 | basic_block best_bb = NULL; |
340 | edge best_edge = NULL; |
341 | profile_count best_count = profile_count::uninitialized (); |
342 | /* The best edge is preferred when its destination is not visited yet |
343 | or is a start block of some trace. */ |
344 | bool is_preferred = false; |
345 | |
346 | /* Find the most frequent edge that goes out from current trace. */ |
347 | bb = back_edge->dest; |
348 | do |
349 | { |
350 | edge e; |
351 | edge_iterator ei; |
352 | |
353 | FOR_EACH_EDGE (e, ei, bb->succs) |
354 | if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
355 | && bb_visited_trace (bb: e->dest) != trace_n |
356 | && (e->flags & EDGE_CAN_FALLTHRU) |
357 | && !(e->flags & EDGE_COMPLEX)) |
358 | { |
359 | if (is_preferred) |
360 | { |
361 | /* The best edge is preferred. */ |
362 | if (!bb_visited_trace (bb: e->dest) |
363 | || bbd[e->dest->index].start_of_trace >= 0) |
364 | { |
365 | /* The current edge E is also preferred. */ |
366 | if (e->count () > best_count) |
367 | { |
368 | best_count = e->count (); |
369 | best_edge = e; |
370 | best_bb = bb; |
371 | } |
372 | } |
373 | } |
374 | else |
375 | { |
376 | if (!bb_visited_trace (bb: e->dest) |
377 | || bbd[e->dest->index].start_of_trace >= 0) |
378 | { |
379 | /* The current edge E is preferred. */ |
380 | is_preferred = true; |
381 | best_count = e->count (); |
382 | best_edge = e; |
383 | best_bb = bb; |
384 | } |
385 | else |
386 | { |
387 | if (!best_edge || e->count () > best_count) |
388 | { |
389 | best_count = e->count (); |
390 | best_edge = e; |
391 | best_bb = bb; |
392 | } |
393 | } |
394 | } |
395 | } |
396 | bb = (basic_block) bb->aux; |
397 | } |
398 | while (bb != back_edge->dest); |
399 | |
400 | if (best_bb) |
401 | { |
402 | /* Rotate the loop so that the BEST_EDGE goes out from the last block of |
403 | the trace. */ |
404 | if (back_edge->dest == trace->first) |
405 | { |
406 | trace->first = (basic_block) best_bb->aux; |
407 | } |
408 | else |
409 | { |
410 | basic_block prev_bb; |
411 | |
412 | for (prev_bb = trace->first; |
413 | prev_bb->aux != back_edge->dest; |
414 | prev_bb = (basic_block) prev_bb->aux) |
415 | ; |
416 | prev_bb->aux = best_bb->aux; |
417 | |
418 | /* Try to get rid of uncond jump to cond jump. */ |
419 | if (single_succ_p (bb: prev_bb)) |
420 | { |
421 | basic_block = single_succ (bb: prev_bb); |
422 | |
423 | /* Duplicate HEADER if it is a small block containing cond jump |
424 | in the end. */ |
425 | if (any_condjump_p (BB_END (header)) && copy_bb_p (header, 0) |
426 | && !CROSSING_JUMP_P (BB_END (header))) |
427 | copy_bb (header, single_succ_edge (bb: prev_bb), prev_bb, trace_n); |
428 | } |
429 | } |
430 | } |
431 | else |
432 | { |
433 | /* We have not found suitable loop tail so do no rotation. */ |
434 | best_bb = back_edge->src; |
435 | } |
436 | best_bb->aux = NULL; |
437 | return best_bb; |
438 | } |
439 | |
440 | /* One round of finding traces. Find traces for BRANCH_TH and EXEC_TH i.e. do |
441 | not include basic blocks whose probability is lower than BRANCH_TH or whose |
442 | count is lower than EXEC_TH into traces (or whose count is lower than |
443 | COUNT_TH). Store the new traces into TRACES and modify the number of |
444 | traces *N_TRACES. Set the round (which the trace belongs to) to ROUND. |
445 | The function expects starting basic blocks to be in *HEAP and will delete |
446 | *HEAP and store starting points for the next round into new *HEAP. */ |
447 | |
448 | static void |
449 | find_traces_1_round (int branch_th, profile_count count_th, |
450 | struct trace *traces, int *n_traces, int round, |
451 | bb_heap_t **heap, int number_of_rounds) |
452 | { |
453 | /* Heap for discarded basic blocks which are possible starting points for |
454 | the next round. */ |
455 | bb_heap_t *new_heap = new bb_heap_t (LONG_MIN); |
456 | bool for_size = optimize_function_for_size_p (cfun); |
457 | |
458 | while (!(*heap)->empty ()) |
459 | { |
460 | basic_block bb; |
461 | struct trace *trace; |
462 | edge best_edge, e; |
463 | long key; |
464 | edge_iterator ei; |
465 | |
466 | bb = (*heap)->extract_min (); |
467 | bbd[bb->index].heap = NULL; |
468 | bbd[bb->index].node = NULL; |
469 | |
470 | if (dump_file) |
471 | fprintf (stream: dump_file, format: "Getting bb %d\n" , bb->index); |
472 | |
473 | /* If the BB's count is too low, send BB to the next round. When |
474 | partitioning hot/cold blocks into separate sections, make sure all |
475 | the cold blocks (and ONLY the cold blocks) go into the (extra) final |
476 | round. When optimizing for size, do not push to next round. */ |
477 | |
478 | if (!for_size |
479 | && push_to_next_round_p (bb, round, number_of_rounds, |
480 | count_th)) |
481 | { |
482 | int key = bb_to_key (bb); |
483 | bbd[bb->index].heap = new_heap; |
484 | bbd[bb->index].node = new_heap->insert (key, data: bb); |
485 | |
486 | if (dump_file) |
487 | fprintf (stream: dump_file, |
488 | format: " Possible start point of next round: %d (key: %d)\n" , |
489 | bb->index, key); |
490 | continue; |
491 | } |
492 | |
493 | trace = traces + *n_traces; |
494 | trace->first = bb; |
495 | trace->round = round; |
496 | trace->length = 0; |
497 | bbd[bb->index].in_trace = *n_traces; |
498 | (*n_traces)++; |
499 | |
500 | do |
501 | { |
502 | bool ends_in_call; |
503 | |
504 | /* The probability and count of the best edge. */ |
505 | profile_probability best_prob = profile_probability::uninitialized (); |
506 | profile_count best_count = profile_count::uninitialized (); |
507 | |
508 | best_edge = NULL; |
509 | mark_bb_visited (bb, trace: *n_traces); |
510 | trace->length++; |
511 | |
512 | if (dump_file) |
513 | fprintf (stream: dump_file, format: "Basic block %d was visited in trace %d\n" , |
514 | bb->index, *n_traces); |
515 | |
516 | ends_in_call = block_ends_with_call_p (bb); |
517 | |
518 | /* Select the successor that will be placed after BB. */ |
519 | FOR_EACH_EDGE (e, ei, bb->succs) |
520 | { |
521 | gcc_assert (!(e->flags & EDGE_FAKE)); |
522 | |
523 | if (e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
524 | continue; |
525 | |
526 | if (bb_visited_trace (bb: e->dest) |
527 | && bb_visited_trace (bb: e->dest) != *n_traces) |
528 | continue; |
529 | |
530 | /* If partitioning hot/cold basic blocks, don't consider edges |
531 | that cross section boundaries. */ |
532 | if (BB_PARTITION (e->dest) != BB_PARTITION (bb)) |
533 | continue; |
534 | |
535 | profile_probability prob = e->probability; |
536 | profile_count count = e->dest->count; |
537 | |
538 | /* The only sensible preference for a call instruction is the |
539 | fallthru edge. Don't bother selecting anything else. */ |
540 | if (ends_in_call) |
541 | { |
542 | if (e->flags & EDGE_CAN_FALLTHRU) |
543 | { |
544 | best_edge = e; |
545 | best_prob = prob; |
546 | best_count = count; |
547 | } |
548 | continue; |
549 | } |
550 | |
551 | /* Edge that cannot be fallthru or improbable or infrequent |
552 | successor (i.e. it is unsuitable successor). When optimizing |
553 | for size, ignore the probability and count. */ |
554 | if (!(e->flags & EDGE_CAN_FALLTHRU) || (e->flags & EDGE_COMPLEX) |
555 | || !prob.initialized_p () |
556 | || ((prob.to_reg_br_prob_base () < branch_th |
557 | || e->count () < count_th) && (!for_size))) |
558 | continue; |
559 | |
560 | if (better_edge_p (bb, e, prob, count, best_prob, best_count, |
561 | best_edge)) |
562 | { |
563 | best_edge = e; |
564 | best_prob = prob; |
565 | best_count = count; |
566 | } |
567 | } |
568 | |
569 | /* If the best destination has multiple predecessors and can be |
570 | duplicated cheaper than a jump, don't allow it to be added to |
571 | a trace; we'll duplicate it when connecting the traces later. |
572 | However, we need to check that this duplication wouldn't leave |
573 | the best destination with only crossing predecessors, because |
574 | this would change its effective partition from hot to cold. */ |
575 | if (best_edge |
576 | && EDGE_COUNT (best_edge->dest->preds) >= 2 |
577 | && copy_bb_p (best_edge->dest, 0)) |
578 | { |
579 | bool only_crossing_preds = true; |
580 | edge e; |
581 | edge_iterator ei; |
582 | FOR_EACH_EDGE (e, ei, best_edge->dest->preds) |
583 | if (e != best_edge && !(e->flags & EDGE_CROSSING)) |
584 | { |
585 | only_crossing_preds = false; |
586 | break; |
587 | } |
588 | if (!only_crossing_preds) |
589 | best_edge = NULL; |
590 | } |
591 | |
592 | /* If the best destination has multiple successors or predecessors, |
593 | don't allow it to be added when optimizing for size. This makes |
594 | sure predecessors with smaller index are handled before the best |
595 | destination. It breaks long trace and reduces long jumps. |
596 | |
597 | Take if-then-else as an example. |
598 | A |
599 | / \ |
600 | B C |
601 | \ / |
602 | D |
603 | If we do not remove the best edge B->D/C->D, the final order might |
604 | be A B D ... C. C is at the end of the program. If D's successors |
605 | and D are complicated, might need long jumps for A->C and C->D. |
606 | Similar issue for order: A C D ... B. |
607 | |
608 | After removing the best edge, the final result will be ABCD/ ACBD. |
609 | It does not add jump compared with the previous order. But it |
610 | reduces the possibility of long jumps. */ |
611 | if (best_edge && for_size |
612 | && (EDGE_COUNT (best_edge->dest->succs) > 1 |
613 | || EDGE_COUNT (best_edge->dest->preds) > 1)) |
614 | best_edge = NULL; |
615 | |
616 | /* Add all non-selected successors to the heaps. */ |
617 | FOR_EACH_EDGE (e, ei, bb->succs) |
618 | { |
619 | if (e == best_edge |
620 | || e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun) |
621 | || bb_visited_trace (bb: e->dest)) |
622 | continue; |
623 | |
624 | key = bb_to_key (e->dest); |
625 | |
626 | if (bbd[e->dest->index].heap) |
627 | { |
628 | /* E->DEST is already in some heap. */ |
629 | if (key != bbd[e->dest->index].node->get_key ()) |
630 | { |
631 | if (dump_file) |
632 | { |
633 | fprintf (stream: dump_file, |
634 | format: "Changing key for bb %d from %ld to %ld.\n" , |
635 | e->dest->index, |
636 | (long) bbd[e->dest->index].node->get_key (), |
637 | key); |
638 | } |
639 | bbd[e->dest->index].heap->replace_key |
640 | (node: bbd[e->dest->index].node, key); |
641 | } |
642 | } |
643 | else |
644 | { |
645 | bb_heap_t *which_heap = *heap; |
646 | |
647 | profile_probability prob = e->probability; |
648 | |
649 | if (!(e->flags & EDGE_CAN_FALLTHRU) |
650 | || (e->flags & EDGE_COMPLEX) |
651 | || !prob.initialized_p () |
652 | || prob.to_reg_br_prob_base () < branch_th |
653 | || e->count () < count_th) |
654 | { |
655 | /* When partitioning hot/cold basic blocks, make sure |
656 | the cold blocks (and only the cold blocks) all get |
657 | pushed to the last round of trace collection. When |
658 | optimizing for size, do not push to next round. */ |
659 | |
660 | if (!for_size && push_to_next_round_p (bb: e->dest, round, |
661 | number_of_rounds, |
662 | count_th)) |
663 | which_heap = new_heap; |
664 | } |
665 | |
666 | bbd[e->dest->index].heap = which_heap; |
667 | bbd[e->dest->index].node = which_heap->insert (key, data: e->dest); |
668 | |
669 | if (dump_file) |
670 | { |
671 | fprintf (stream: dump_file, |
672 | format: " Possible start of %s round: %d (key: %ld)\n" , |
673 | (which_heap == new_heap) ? "next" : "this" , |
674 | e->dest->index, (long) key); |
675 | } |
676 | |
677 | } |
678 | } |
679 | |
680 | if (best_edge) /* Suitable successor was found. */ |
681 | { |
682 | if (bb_visited_trace (bb: best_edge->dest) == *n_traces) |
683 | { |
684 | /* We do nothing with one basic block loops. */ |
685 | if (best_edge->dest != bb) |
686 | { |
687 | if (best_edge->count () |
688 | > best_edge->dest->count.apply_scale (num: 4, den: 5)) |
689 | { |
690 | /* The loop has at least 4 iterations. If the loop |
691 | header is not the first block of the function |
692 | we can rotate the loop. */ |
693 | |
694 | if (best_edge->dest |
695 | != ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb) |
696 | { |
697 | if (dump_file) |
698 | { |
699 | fprintf (stream: dump_file, |
700 | format: "Rotating loop %d - %d\n" , |
701 | best_edge->dest->index, bb->index); |
702 | } |
703 | bb->aux = best_edge->dest; |
704 | bbd[best_edge->dest->index].in_trace = |
705 | (*n_traces) - 1; |
706 | bb = rotate_loop (back_edge: best_edge, trace, trace_n: *n_traces); |
707 | } |
708 | } |
709 | else |
710 | { |
711 | /* The loop has less than 4 iterations. */ |
712 | |
713 | if (single_succ_p (bb) |
714 | && copy_bb_p (best_edge->dest, |
715 | optimize_edge_for_speed_p |
716 | (best_edge))) |
717 | { |
718 | bb = copy_bb (best_edge->dest, best_edge, bb, |
719 | *n_traces); |
720 | trace->length++; |
721 | } |
722 | } |
723 | } |
724 | |
725 | /* Terminate the trace. */ |
726 | break; |
727 | } |
728 | else |
729 | { |
730 | /* Check for a situation |
731 | |
732 | A |
733 | /| |
734 | B | |
735 | \| |
736 | C |
737 | |
738 | where |
739 | AB->count () + BC->count () >= AC->count (). |
740 | (i.e. 2 * B->count >= AC->count ) |
741 | Best ordering is then A B C. |
742 | |
743 | When optimizing for size, A B C is always the best order. |
744 | |
745 | This situation is created for example by: |
746 | |
747 | if (A) B; |
748 | C; |
749 | |
750 | */ |
751 | |
752 | FOR_EACH_EDGE (e, ei, bb->succs) |
753 | if (e != best_edge |
754 | && (e->flags & EDGE_CAN_FALLTHRU) |
755 | && !(e->flags & EDGE_COMPLEX) |
756 | && !bb_visited_trace (bb: e->dest) |
757 | && single_pred_p (bb: e->dest) |
758 | && !(e->flags & EDGE_CROSSING) |
759 | && single_succ_p (bb: e->dest) |
760 | && (single_succ_edge (bb: e->dest)->flags |
761 | & EDGE_CAN_FALLTHRU) |
762 | && !(single_succ_edge (bb: e->dest)->flags & EDGE_COMPLEX) |
763 | && single_succ (bb: e->dest) == best_edge->dest |
764 | && (e->dest->count * 2 |
765 | >= best_edge->count () || for_size)) |
766 | { |
767 | best_edge = e; |
768 | if (dump_file) |
769 | fprintf (stream: dump_file, format: "Selecting BB %d\n" , |
770 | best_edge->dest->index); |
771 | break; |
772 | } |
773 | |
774 | bb->aux = best_edge->dest; |
775 | bbd[best_edge->dest->index].in_trace = (*n_traces) - 1; |
776 | bb = best_edge->dest; |
777 | } |
778 | } |
779 | } |
780 | while (best_edge); |
781 | trace->last = bb; |
782 | bbd[trace->first->index].start_of_trace = *n_traces - 1; |
783 | if (bbd[trace->last->index].end_of_trace != *n_traces - 1) |
784 | { |
785 | bbd[trace->last->index].end_of_trace = *n_traces - 1; |
786 | /* Update the cached maximum frequency for interesting predecessor |
787 | edges for successors of the new trace end. */ |
788 | FOR_EACH_EDGE (e, ei, trace->last->succs) |
789 | if (EDGE_FREQUENCY (e) > bbd[e->dest->index].priority) |
790 | bbd[e->dest->index].priority = EDGE_FREQUENCY (e); |
791 | } |
792 | |
793 | /* The trace is terminated so we have to recount the keys in heap |
794 | (some block can have a lower key because now one of its predecessors |
795 | is an end of the trace). */ |
796 | FOR_EACH_EDGE (e, ei, bb->succs) |
797 | { |
798 | if (e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun) |
799 | || bb_visited_trace (bb: e->dest)) |
800 | continue; |
801 | |
802 | if (bbd[e->dest->index].heap) |
803 | { |
804 | key = bb_to_key (e->dest); |
805 | if (key != bbd[e->dest->index].node->get_key ()) |
806 | { |
807 | if (dump_file) |
808 | { |
809 | fprintf (stream: dump_file, |
810 | format: "Changing key for bb %d from %ld to %ld.\n" , |
811 | e->dest->index, |
812 | (long) bbd[e->dest->index].node->get_key (), key); |
813 | } |
814 | bbd[e->dest->index].heap->replace_key |
815 | (node: bbd[e->dest->index].node, key); |
816 | } |
817 | } |
818 | } |
819 | } |
820 | |
821 | delete (*heap); |
822 | |
823 | /* "Return" the new heap. */ |
824 | *heap = new_heap; |
825 | } |
826 | |
827 | /* Create a duplicate of the basic block OLD_BB and redirect edge E to it, add |
828 | it to trace after BB, mark OLD_BB visited and update pass' data structures |
829 | (TRACE is a number of trace which OLD_BB is duplicated to). */ |
830 | |
831 | static basic_block |
832 | copy_bb (basic_block old_bb, edge e, basic_block bb, int trace) |
833 | { |
834 | basic_block new_bb; |
835 | |
836 | new_bb = duplicate_block (old_bb, e, bb); |
837 | BB_COPY_PARTITION (new_bb, old_bb); |
838 | |
839 | gcc_assert (e->dest == new_bb); |
840 | |
841 | if (dump_file) |
842 | fprintf (stream: dump_file, |
843 | format: "Duplicated bb %d (created bb %d)\n" , |
844 | old_bb->index, new_bb->index); |
845 | |
846 | if (new_bb->index >= array_size |
847 | || last_basic_block_for_fn (cfun) > array_size) |
848 | { |
849 | int i; |
850 | int new_size; |
851 | |
852 | new_size = MAX (last_basic_block_for_fn (cfun), new_bb->index + 1); |
853 | new_size = GET_ARRAY_SIZE (new_size); |
854 | bbd = XRESIZEVEC (bbro_basic_block_data, bbd, new_size); |
855 | for (i = array_size; i < new_size; i++) |
856 | { |
857 | bbd[i].start_of_trace = -1; |
858 | bbd[i].end_of_trace = -1; |
859 | bbd[i].in_trace = -1; |
860 | bbd[i].visited = 0; |
861 | bbd[i].priority = -1; |
862 | bbd[i].heap = NULL; |
863 | bbd[i].node = NULL; |
864 | } |
865 | array_size = new_size; |
866 | |
867 | if (dump_file) |
868 | { |
869 | fprintf (stream: dump_file, |
870 | format: "Growing the dynamic array to %d elements.\n" , |
871 | array_size); |
872 | } |
873 | } |
874 | |
875 | gcc_assert (!bb_visited_trace (e->dest)); |
876 | mark_bb_visited (bb: new_bb, trace); |
877 | new_bb->aux = bb->aux; |
878 | bb->aux = new_bb; |
879 | |
880 | bbd[new_bb->index].in_trace = trace; |
881 | |
882 | return new_bb; |
883 | } |
884 | |
885 | /* Compute and return the key (for the heap) of the basic block BB. */ |
886 | |
887 | static long |
888 | bb_to_key (basic_block bb) |
889 | { |
890 | edge e; |
891 | edge_iterator ei; |
892 | |
893 | /* Use index as key to align with its original order. */ |
894 | if (optimize_function_for_size_p (cfun)) |
895 | return bb->index; |
896 | |
897 | /* Do not start in probably never executed blocks. */ |
898 | |
899 | if (BB_PARTITION (bb) == BB_COLD_PARTITION |
900 | || probably_never_executed_bb_p (cfun, bb)) |
901 | return BB_FREQ_MAX; |
902 | |
903 | /* Prefer blocks whose predecessor is an end of some trace |
904 | or whose predecessor edge is EDGE_DFS_BACK. */ |
905 | int priority = bbd[bb->index].priority; |
906 | if (priority == -1) |
907 | { |
908 | priority = 0; |
909 | FOR_EACH_EDGE (e, ei, bb->preds) |
910 | { |
911 | if ((e->src != ENTRY_BLOCK_PTR_FOR_FN (cfun) |
912 | && bbd[e->src->index].end_of_trace >= 0) |
913 | || (e->flags & EDGE_DFS_BACK)) |
914 | { |
915 | int edge_freq = EDGE_FREQUENCY (e); |
916 | |
917 | if (edge_freq > priority) |
918 | priority = edge_freq; |
919 | } |
920 | } |
921 | bbd[bb->index].priority = priority; |
922 | } |
923 | |
924 | if (priority) |
925 | /* The block with priority should have significantly lower key. */ |
926 | return -(100 * BB_FREQ_MAX + 100 * priority + bb->count.to_frequency (cfun)); |
927 | |
928 | return -bb->count.to_frequency (cfun); |
929 | } |
930 | |
931 | /* Return true when the edge E from basic block BB is better than the temporary |
932 | best edge (details are in function). The probability of edge E is PROB. The |
933 | count of the successor is COUNT. The current best probability is |
934 | BEST_PROB, the best count is BEST_COUNT. |
935 | The edge is considered to be equivalent when PROB does not differ much from |
936 | BEST_PROB; similarly for count. */ |
937 | |
938 | static bool |
939 | better_edge_p (const_basic_block bb, const_edge e, profile_probability prob, |
940 | profile_count count, profile_probability best_prob, |
941 | profile_count best_count, const_edge cur_best_edge) |
942 | { |
943 | bool is_better_edge; |
944 | |
945 | /* The BEST_* values do not have to be best, but can be a bit smaller than |
946 | maximum values. */ |
947 | profile_probability diff_prob = best_prob / 10; |
948 | |
949 | /* The smaller one is better to keep the original order. */ |
950 | if (optimize_function_for_size_p (cfun)) |
951 | return !cur_best_edge |
952 | || cur_best_edge->dest->index > e->dest->index; |
953 | |
954 | /* Those edges are so expensive that continuing a trace is not useful |
955 | performance wise. */ |
956 | if (e->flags & (EDGE_ABNORMAL | EDGE_EH)) |
957 | return false; |
958 | |
959 | if (prob > best_prob + diff_prob |
960 | || (!best_prob.initialized_p () |
961 | && prob > profile_probability::guessed_never ())) |
962 | /* The edge has higher probability than the temporary best edge. */ |
963 | is_better_edge = true; |
964 | else if (prob < best_prob - diff_prob) |
965 | /* The edge has lower probability than the temporary best edge. */ |
966 | is_better_edge = false; |
967 | else |
968 | { |
969 | profile_count diff_count = best_count / 10; |
970 | if (count < best_count - diff_count |
971 | || (!best_count.initialized_p () |
972 | && count.nonzero_p ())) |
973 | /* The edge and the temporary best edge have almost equivalent |
974 | probabilities. The higher countuency of a successor now means |
975 | that there is another edge going into that successor. |
976 | This successor has lower countuency so it is better. */ |
977 | is_better_edge = true; |
978 | else if (count > best_count + diff_count) |
979 | /* This successor has higher countuency so it is worse. */ |
980 | is_better_edge = false; |
981 | else if (e->dest->prev_bb == bb) |
982 | /* The edges have equivalent probabilities and the successors |
983 | have equivalent frequencies. Select the previous successor. */ |
984 | is_better_edge = true; |
985 | else |
986 | is_better_edge = false; |
987 | } |
988 | |
989 | return is_better_edge; |
990 | } |
991 | |
992 | /* Return true when the edge E is better than the temporary best edge |
993 | CUR_BEST_EDGE. If SRC_INDEX_P is true, the function compares the src bb of |
994 | E and CUR_BEST_EDGE; otherwise it will compare the dest bb. |
995 | BEST_LEN is the trace length of src (or dest) bb in CUR_BEST_EDGE. |
996 | TRACES record the information about traces. |
997 | When optimizing for size, the edge with smaller index is better. |
998 | When optimizing for speed, the edge with bigger probability or longer trace |
999 | is better. */ |
1000 | |
1001 | static bool |
1002 | connect_better_edge_p (const_edge e, bool src_index_p, int best_len, |
1003 | const_edge cur_best_edge, struct trace *traces) |
1004 | { |
1005 | int e_index; |
1006 | int b_index; |
1007 | bool is_better_edge; |
1008 | |
1009 | if (!cur_best_edge) |
1010 | return true; |
1011 | |
1012 | if (optimize_function_for_size_p (cfun)) |
1013 | { |
1014 | e_index = src_index_p ? e->src->index : e->dest->index; |
1015 | b_index = src_index_p ? cur_best_edge->src->index |
1016 | : cur_best_edge->dest->index; |
1017 | /* The smaller one is better to keep the original order. */ |
1018 | return b_index > e_index; |
1019 | } |
1020 | |
1021 | if (src_index_p) |
1022 | { |
1023 | e_index = e->src->index; |
1024 | |
1025 | /* We are looking for predecessor, so probabilities are not that |
1026 | informative. We do not want to connect A to B because A has |
1027 | only one successor (probability is 100%) while there is edge |
1028 | A' to B where probability is 90% but which is much more frequent. */ |
1029 | if (e->count () > cur_best_edge->count ()) |
1030 | /* The edge has higher probability than the temporary best edge. */ |
1031 | is_better_edge = true; |
1032 | else if (e->count () < cur_best_edge->count ()) |
1033 | /* The edge has lower probability than the temporary best edge. */ |
1034 | is_better_edge = false; |
1035 | else if (e->probability > cur_best_edge->probability) |
1036 | /* The edge has higher probability than the temporary best edge. */ |
1037 | is_better_edge = true; |
1038 | else if (e->probability < cur_best_edge->probability) |
1039 | /* The edge has lower probability than the temporary best edge. */ |
1040 | is_better_edge = false; |
1041 | else if (traces[bbd[e_index].end_of_trace].length > best_len) |
1042 | /* The edge and the temporary best edge have equivalent probabilities. |
1043 | The edge with longer trace is better. */ |
1044 | is_better_edge = true; |
1045 | else |
1046 | is_better_edge = false; |
1047 | } |
1048 | else |
1049 | { |
1050 | e_index = e->dest->index; |
1051 | |
1052 | if (e->probability > cur_best_edge->probability) |
1053 | /* The edge has higher probability than the temporary best edge. */ |
1054 | is_better_edge = true; |
1055 | else if (e->probability < cur_best_edge->probability) |
1056 | /* The edge has lower probability than the temporary best edge. */ |
1057 | is_better_edge = false; |
1058 | else if (traces[bbd[e_index].start_of_trace].length > best_len) |
1059 | /* The edge and the temporary best edge have equivalent probabilities. |
1060 | The edge with longer trace is better. */ |
1061 | is_better_edge = true; |
1062 | else |
1063 | is_better_edge = false; |
1064 | } |
1065 | |
1066 | return is_better_edge; |
1067 | } |
1068 | |
1069 | /* Connect traces in array TRACES, N_TRACES is the count of traces. */ |
1070 | |
1071 | static void |
1072 | connect_traces (int n_traces, struct trace *traces) |
1073 | { |
1074 | int i; |
1075 | bool *connected; |
1076 | bool two_passes; |
1077 | int last_trace; |
1078 | int current_pass; |
1079 | int current_partition; |
1080 | profile_count count_threshold; |
1081 | bool for_size = optimize_function_for_size_p (cfun); |
1082 | |
1083 | count_threshold = max_entry_count.apply_scale (DUPLICATION_THRESHOLD, den: 1000); |
1084 | |
1085 | connected = XCNEWVEC (bool, n_traces); |
1086 | last_trace = -1; |
1087 | current_pass = 1; |
1088 | current_partition = BB_PARTITION (traces[0].first); |
1089 | two_passes = false; |
1090 | |
1091 | if (crtl->has_bb_partition) |
1092 | for (i = 0; i < n_traces && !two_passes; i++) |
1093 | if (BB_PARTITION (traces[0].first) |
1094 | != BB_PARTITION (traces[i].first)) |
1095 | two_passes = true; |
1096 | |
1097 | for (i = 0; i < n_traces || (two_passes && current_pass == 1) ; i++) |
1098 | { |
1099 | int t = i; |
1100 | int t2; |
1101 | edge e, best; |
1102 | int best_len; |
1103 | |
1104 | if (i >= n_traces) |
1105 | { |
1106 | gcc_assert (two_passes && current_pass == 1); |
1107 | i = 0; |
1108 | t = i; |
1109 | current_pass = 2; |
1110 | if (current_partition == BB_HOT_PARTITION) |
1111 | current_partition = BB_COLD_PARTITION; |
1112 | else |
1113 | current_partition = BB_HOT_PARTITION; |
1114 | } |
1115 | |
1116 | if (connected[t]) |
1117 | continue; |
1118 | |
1119 | if (two_passes |
1120 | && BB_PARTITION (traces[t].first) != current_partition) |
1121 | continue; |
1122 | |
1123 | connected[t] = true; |
1124 | |
1125 | /* Find the predecessor traces. */ |
1126 | for (t2 = t; t2 > 0;) |
1127 | { |
1128 | edge_iterator ei; |
1129 | best = NULL; |
1130 | best_len = 0; |
1131 | FOR_EACH_EDGE (e, ei, traces[t2].first->preds) |
1132 | { |
1133 | int si = e->src->index; |
1134 | |
1135 | if (e->src != ENTRY_BLOCK_PTR_FOR_FN (cfun) |
1136 | && (e->flags & EDGE_CAN_FALLTHRU) |
1137 | && !(e->flags & EDGE_COMPLEX) |
1138 | && bbd[si].end_of_trace >= 0 |
1139 | && !connected[bbd[si].end_of_trace] |
1140 | && (BB_PARTITION (e->src) == current_partition) |
1141 | && connect_better_edge_p (e, src_index_p: true, best_len, cur_best_edge: best, traces)) |
1142 | { |
1143 | best = e; |
1144 | best_len = traces[bbd[si].end_of_trace].length; |
1145 | } |
1146 | } |
1147 | if (best) |
1148 | { |
1149 | best->src->aux = best->dest; |
1150 | t2 = bbd[best->src->index].end_of_trace; |
1151 | connected[t2] = true; |
1152 | |
1153 | if (dump_file) |
1154 | { |
1155 | fprintf (stream: dump_file, format: "Connection: %d %d\n" , |
1156 | best->src->index, best->dest->index); |
1157 | } |
1158 | } |
1159 | else |
1160 | break; |
1161 | } |
1162 | |
1163 | if (last_trace >= 0) |
1164 | traces[last_trace].last->aux = traces[t2].first; |
1165 | last_trace = t; |
1166 | |
1167 | /* Find the successor traces. */ |
1168 | while (1) |
1169 | { |
1170 | /* Find the continuation of the chain. */ |
1171 | edge_iterator ei; |
1172 | best = NULL; |
1173 | best_len = 0; |
1174 | FOR_EACH_EDGE (e, ei, traces[t].last->succs) |
1175 | { |
1176 | int di = e->dest->index; |
1177 | |
1178 | if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
1179 | && (e->flags & EDGE_CAN_FALLTHRU) |
1180 | && !(e->flags & EDGE_COMPLEX) |
1181 | && bbd[di].start_of_trace >= 0 |
1182 | && !connected[bbd[di].start_of_trace] |
1183 | && (BB_PARTITION (e->dest) == current_partition) |
1184 | && connect_better_edge_p (e, src_index_p: false, best_len, cur_best_edge: best, traces)) |
1185 | { |
1186 | best = e; |
1187 | best_len = traces[bbd[di].start_of_trace].length; |
1188 | } |
1189 | } |
1190 | |
1191 | if (for_size) |
1192 | { |
1193 | if (!best) |
1194 | /* Stop finding the successor traces. */ |
1195 | break; |
1196 | |
1197 | /* It is OK to connect block n with block n + 1 or a block |
1198 | before n. For others, only connect to the loop header. */ |
1199 | if (best->dest->index > (traces[t].last->index + 1)) |
1200 | { |
1201 | int count = EDGE_COUNT (best->dest->preds); |
1202 | |
1203 | FOR_EACH_EDGE (e, ei, best->dest->preds) |
1204 | if (e->flags & EDGE_DFS_BACK) |
1205 | count--; |
1206 | |
1207 | /* If dest has multiple predecessors, skip it. We expect |
1208 | that one predecessor with smaller index connects with it |
1209 | later. */ |
1210 | if (count != 1) |
1211 | break; |
1212 | } |
1213 | |
1214 | /* Only connect Trace n with Trace n + 1. It is conservative |
1215 | to keep the order as close as possible to the original order. |
1216 | It also helps to reduce long jumps. */ |
1217 | if (last_trace != bbd[best->dest->index].start_of_trace - 1) |
1218 | break; |
1219 | |
1220 | if (dump_file) |
1221 | fprintf (stream: dump_file, format: "Connection: %d %d\n" , |
1222 | best->src->index, best->dest->index); |
1223 | |
1224 | t = bbd[best->dest->index].start_of_trace; |
1225 | traces[last_trace].last->aux = traces[t].first; |
1226 | connected[t] = true; |
1227 | last_trace = t; |
1228 | } |
1229 | else if (best) |
1230 | { |
1231 | if (dump_file) |
1232 | { |
1233 | fprintf (stream: dump_file, format: "Connection: %d %d\n" , |
1234 | best->src->index, best->dest->index); |
1235 | } |
1236 | t = bbd[best->dest->index].start_of_trace; |
1237 | traces[last_trace].last->aux = traces[t].first; |
1238 | connected[t] = true; |
1239 | last_trace = t; |
1240 | } |
1241 | else |
1242 | { |
1243 | /* Try to connect the traces by duplication of 1 block. */ |
1244 | edge e2; |
1245 | basic_block next_bb = NULL; |
1246 | bool try_copy = false; |
1247 | |
1248 | FOR_EACH_EDGE (e, ei, traces[t].last->succs) |
1249 | if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
1250 | && (e->flags & EDGE_CAN_FALLTHRU) |
1251 | && !(e->flags & EDGE_COMPLEX) |
1252 | && (!best || e->probability > best->probability)) |
1253 | { |
1254 | edge_iterator ei; |
1255 | edge best2 = NULL; |
1256 | int best2_len = 0; |
1257 | |
1258 | /* If the destination is a start of a trace which is only |
1259 | one block long, then no need to search the successor |
1260 | blocks of the trace. Accept it. */ |
1261 | if (bbd[e->dest->index].start_of_trace >= 0 |
1262 | && traces[bbd[e->dest->index].start_of_trace].length |
1263 | == 1) |
1264 | { |
1265 | best = e; |
1266 | try_copy = true; |
1267 | continue; |
1268 | } |
1269 | |
1270 | FOR_EACH_EDGE (e2, ei, e->dest->succs) |
1271 | { |
1272 | int di = e2->dest->index; |
1273 | |
1274 | if (e2->dest == EXIT_BLOCK_PTR_FOR_FN (cfun) |
1275 | || ((e2->flags & EDGE_CAN_FALLTHRU) |
1276 | && !(e2->flags & EDGE_COMPLEX) |
1277 | && bbd[di].start_of_trace >= 0 |
1278 | && !connected[bbd[di].start_of_trace] |
1279 | && BB_PARTITION (e2->dest) == current_partition |
1280 | && e2->count () >= count_threshold |
1281 | && (!best2 |
1282 | || e2->probability > best2->probability |
1283 | || (e2->probability == best2->probability |
1284 | && traces[bbd[di].start_of_trace].length |
1285 | > best2_len)))) |
1286 | { |
1287 | best = e; |
1288 | best2 = e2; |
1289 | if (e2->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)) |
1290 | best2_len = traces[bbd[di].start_of_trace].length; |
1291 | else |
1292 | best2_len = INT_MAX; |
1293 | next_bb = e2->dest; |
1294 | try_copy = true; |
1295 | } |
1296 | } |
1297 | } |
1298 | |
1299 | /* Copy tiny blocks always; copy larger blocks only when the |
1300 | edge is traversed frequently enough. */ |
1301 | if (try_copy |
1302 | && BB_PARTITION (best->src) == BB_PARTITION (best->dest) |
1303 | && copy_bb_p (best->dest, |
1304 | optimize_edge_for_speed_p (best) |
1305 | && (!best->count ().initialized_p () |
1306 | || best->count () >= count_threshold))) |
1307 | { |
1308 | basic_block new_bb; |
1309 | |
1310 | if (dump_file) |
1311 | { |
1312 | fprintf (stream: dump_file, format: "Connection: %d %d " , |
1313 | traces[t].last->index, best->dest->index); |
1314 | if (!next_bb) |
1315 | fputc (c: '\n', stream: dump_file); |
1316 | else if (next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
1317 | fprintf (stream: dump_file, format: "exit\n" ); |
1318 | else |
1319 | fprintf (stream: dump_file, format: "%d\n" , next_bb->index); |
1320 | } |
1321 | |
1322 | new_bb = copy_bb (old_bb: best->dest, e: best, bb: traces[t].last, trace: t); |
1323 | traces[t].last = new_bb; |
1324 | if (next_bb && next_bb != EXIT_BLOCK_PTR_FOR_FN (cfun)) |
1325 | { |
1326 | t = bbd[next_bb->index].start_of_trace; |
1327 | traces[last_trace].last->aux = traces[t].first; |
1328 | connected[t] = true; |
1329 | last_trace = t; |
1330 | } |
1331 | else |
1332 | break; /* Stop finding the successor traces. */ |
1333 | } |
1334 | else |
1335 | break; /* Stop finding the successor traces. */ |
1336 | } |
1337 | } |
1338 | } |
1339 | |
1340 | if (dump_file) |
1341 | { |
1342 | basic_block bb; |
1343 | |
1344 | fprintf (stream: dump_file, format: "Final order:\n" ); |
1345 | for (bb = traces[0].first; bb; bb = (basic_block) bb->aux) |
1346 | fprintf (stream: dump_file, format: "%d " , bb->index); |
1347 | fprintf (stream: dump_file, format: "\n" ); |
1348 | fflush (stream: dump_file); |
1349 | } |
1350 | |
1351 | FREE (connected); |
1352 | } |
1353 | |
1354 | /* Return true when BB can and should be copied. CODE_MAY_GROW is true |
1355 | when code size is allowed to grow by duplication. */ |
1356 | |
1357 | static bool |
1358 | copy_bb_p (const_basic_block bb, int code_may_grow) |
1359 | { |
1360 | unsigned int size = 0; |
1361 | unsigned int max_size = uncond_jump_length; |
1362 | rtx_insn *insn; |
1363 | |
1364 | if (EDGE_COUNT (bb->preds) < 2) |
1365 | return false; |
1366 | if (!can_duplicate_block_p (bb)) |
1367 | return false; |
1368 | |
1369 | /* Avoid duplicating blocks which have many successors (PR/13430). */ |
1370 | if (EDGE_COUNT (bb->succs) > 8) |
1371 | return false; |
1372 | |
1373 | if (code_may_grow && optimize_bb_for_speed_p (bb)) |
1374 | max_size *= param_max_grow_copy_bb_insns; |
1375 | |
1376 | FOR_BB_INSNS (bb, insn) |
1377 | { |
1378 | if (INSN_P (insn)) |
1379 | { |
1380 | size += get_attr_min_length (insn); |
1381 | if (size > max_size) |
1382 | break; |
1383 | } |
1384 | } |
1385 | |
1386 | if (size <= max_size) |
1387 | return true; |
1388 | |
1389 | if (dump_file) |
1390 | { |
1391 | fprintf (stream: dump_file, |
1392 | format: "Block %d can't be copied because its size = %u.\n" , |
1393 | bb->index, size); |
1394 | } |
1395 | |
1396 | return false; |
1397 | } |
1398 | |
1399 | /* Return the length of unconditional jump instruction. */ |
1400 | |
1401 | int |
1402 | get_uncond_jump_length (void) |
1403 | { |
1404 | unsigned int length; |
1405 | |
1406 | start_sequence (); |
1407 | rtx_code_label *label = emit_label (gen_label_rtx ()); |
1408 | rtx_insn *jump = emit_jump_insn (targetm.gen_jump (label)); |
1409 | length = get_attr_min_length (jump); |
1410 | end_sequence (); |
1411 | |
1412 | gcc_assert (length < INT_MAX); |
1413 | return length; |
1414 | } |
1415 | |
1416 | /* Create a forwarder block to OLD_BB starting with NEW_LABEL and in the |
1417 | other partition wrt OLD_BB. */ |
1418 | |
1419 | static basic_block |
1420 | create_eh_forwarder_block (rtx_code_label *new_label, basic_block old_bb) |
1421 | { |
1422 | /* Split OLD_BB, so that EH pads have always only incoming EH edges, |
1423 | bb_has_eh_pred bbs are treated specially by DF infrastructure. */ |
1424 | old_bb = split_block_after_labels (old_bb)->dest; |
1425 | |
1426 | /* Put the new label and a jump in the new basic block. */ |
1427 | rtx_insn *label = emit_label (new_label); |
1428 | rtx_code_label *old_label = block_label (old_bb); |
1429 | rtx_insn *jump = emit_jump_insn (targetm.gen_jump (old_label)); |
1430 | JUMP_LABEL (jump) = old_label; |
1431 | |
1432 | /* Create the new basic block and put it in last position. */ |
1433 | basic_block last_bb = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb; |
1434 | basic_block new_bb = create_basic_block (label, jump, last_bb); |
1435 | new_bb->aux = last_bb->aux; |
1436 | new_bb->count = old_bb->count; |
1437 | last_bb->aux = new_bb; |
1438 | |
1439 | emit_barrier_after_bb (bb: new_bb); |
1440 | |
1441 | make_single_succ_edge (new_bb, old_bb, 0); |
1442 | |
1443 | /* Make sure the new basic block is in the other partition. */ |
1444 | unsigned new_partition = BB_PARTITION (old_bb); |
1445 | new_partition ^= BB_HOT_PARTITION | BB_COLD_PARTITION; |
1446 | BB_SET_PARTITION (new_bb, new_partition); |
1447 | |
1448 | return new_bb; |
1449 | } |
1450 | |
1451 | /* The common landing pad in block OLD_BB has edges from both partitions. |
1452 | Add a new landing pad that will just jump to the old one and split the |
1453 | edges so that no EH edge crosses partitions. */ |
1454 | |
1455 | static void |
1456 | sjlj_fix_up_crossing_landing_pad (basic_block old_bb) |
1457 | { |
1458 | const unsigned lp_len = cfun->eh->lp_array->length (); |
1459 | edge_iterator ei; |
1460 | edge e; |
1461 | |
1462 | /* Generate the new common landing-pad label. */ |
1463 | rtx_code_label *new_label = gen_label_rtx (); |
1464 | LABEL_PRESERVE_P (new_label) = 1; |
1465 | |
1466 | /* Create the forwarder block. */ |
1467 | basic_block new_bb = create_eh_forwarder_block (new_label, old_bb); |
1468 | |
1469 | /* Create the map from old to new lp index and initialize it. */ |
1470 | unsigned *index_map = (unsigned *) alloca (lp_len * sizeof (unsigned)); |
1471 | memset (s: index_map, c: 0, n: lp_len * sizeof (unsigned)); |
1472 | |
1473 | /* Fix up the edges. */ |
1474 | for (ei = ei_start (old_bb->preds); (e = ei_safe_edge (i: ei)) != NULL; ) |
1475 | if (e->src != new_bb && BB_PARTITION (e->src) == BB_PARTITION (new_bb)) |
1476 | { |
1477 | rtx_insn *insn = BB_END (e->src); |
1478 | rtx note = find_reg_note (insn, REG_EH_REGION, NULL_RTX); |
1479 | |
1480 | gcc_assert (note != NULL); |
1481 | const unsigned old_index = INTVAL (XEXP (note, 0)); |
1482 | |
1483 | /* Generate the new landing-pad structure. */ |
1484 | if (index_map[old_index] == 0) |
1485 | { |
1486 | eh_landing_pad old_lp = (*cfun->eh->lp_array)[old_index]; |
1487 | eh_landing_pad new_lp = gen_eh_landing_pad (old_lp->region); |
1488 | new_lp->post_landing_pad = old_lp->post_landing_pad; |
1489 | new_lp->landing_pad = new_label; |
1490 | index_map[old_index] = new_lp->index; |
1491 | } |
1492 | XEXP (note, 0) = GEN_INT (index_map[old_index]); |
1493 | |
1494 | /* Adjust the edge to the new destination. */ |
1495 | redirect_edge_succ (e, new_bb); |
1496 | } |
1497 | else |
1498 | ei_next (i: &ei); |
1499 | } |
1500 | |
1501 | /* The landing pad OLD_LP, in block OLD_BB, has edges from both partitions. |
1502 | Add a new landing pad that will just jump to the old one and split the |
1503 | edges so that no EH edge crosses partitions. */ |
1504 | |
1505 | static void |
1506 | dw2_fix_up_crossing_landing_pad (eh_landing_pad old_lp, basic_block old_bb) |
1507 | { |
1508 | eh_landing_pad new_lp; |
1509 | edge_iterator ei; |
1510 | edge e; |
1511 | |
1512 | /* Generate the new landing-pad structure. */ |
1513 | new_lp = gen_eh_landing_pad (old_lp->region); |
1514 | new_lp->post_landing_pad = old_lp->post_landing_pad; |
1515 | new_lp->landing_pad = gen_label_rtx (); |
1516 | LABEL_PRESERVE_P (new_lp->landing_pad) = 1; |
1517 | |
1518 | /* Create the forwarder block. */ |
1519 | basic_block new_bb = create_eh_forwarder_block (new_label: new_lp->landing_pad, old_bb); |
1520 | |
1521 | /* Fix up the edges. */ |
1522 | for (ei = ei_start (old_bb->preds); (e = ei_safe_edge (i: ei)) != NULL; ) |
1523 | if (e->src != new_bb && BB_PARTITION (e->src) == BB_PARTITION (new_bb)) |
1524 | { |
1525 | rtx_insn *insn = BB_END (e->src); |
1526 | rtx note = find_reg_note (insn, REG_EH_REGION, NULL_RTX); |
1527 | |
1528 | gcc_assert (note != NULL); |
1529 | gcc_checking_assert (INTVAL (XEXP (note, 0)) == old_lp->index); |
1530 | XEXP (note, 0) = GEN_INT (new_lp->index); |
1531 | |
1532 | /* Adjust the edge to the new destination. */ |
1533 | redirect_edge_succ (e, new_bb); |
1534 | } |
1535 | else |
1536 | ei_next (i: &ei); |
1537 | } |
1538 | |
1539 | |
1540 | /* Ensure that all hot bbs are included in a hot path through the |
1541 | procedure. This is done by calling this function twice, once |
1542 | with WALK_UP true (to look for paths from the entry to hot bbs) and |
1543 | once with WALK_UP false (to look for paths from hot bbs to the exit). |
1544 | Returns the updated value of COLD_BB_COUNT and adds newly-hot bbs |
1545 | to BBS_IN_HOT_PARTITION. */ |
1546 | |
1547 | static unsigned int |
1548 | sanitize_hot_paths (bool walk_up, unsigned int cold_bb_count, |
1549 | vec<basic_block> *bbs_in_hot_partition) |
1550 | { |
1551 | /* Callers check this. */ |
1552 | gcc_checking_assert (cold_bb_count); |
1553 | |
1554 | /* Keep examining hot bbs while we still have some left to check |
1555 | and there are remaining cold bbs. */ |
1556 | vec<basic_block> hot_bbs_to_check = bbs_in_hot_partition->copy (); |
1557 | while (! hot_bbs_to_check.is_empty () |
1558 | && cold_bb_count) |
1559 | { |
1560 | basic_block bb = hot_bbs_to_check.pop (); |
1561 | vec<edge, va_gc> *edges = walk_up ? bb->preds : bb->succs; |
1562 | edge e; |
1563 | edge_iterator ei; |
1564 | profile_probability highest_probability |
1565 | = profile_probability::uninitialized (); |
1566 | profile_count highest_count = profile_count::uninitialized (); |
1567 | bool found = false; |
1568 | |
1569 | /* Walk the preds/succs and check if there is at least one already |
1570 | marked hot. Keep track of the most frequent pred/succ so that we |
1571 | can mark it hot if we don't find one. */ |
1572 | FOR_EACH_EDGE (e, ei, edges) |
1573 | { |
1574 | basic_block reach_bb = walk_up ? e->src : e->dest; |
1575 | |
1576 | if (e->flags & EDGE_DFS_BACK) |
1577 | continue; |
1578 | |
1579 | /* Do not expect profile insanities when profile was not adjusted. */ |
1580 | if (e->probability == profile_probability::never () |
1581 | || e->count () == profile_count::zero ()) |
1582 | continue; |
1583 | |
1584 | if (BB_PARTITION (reach_bb) != BB_COLD_PARTITION) |
1585 | { |
1586 | found = true; |
1587 | break; |
1588 | } |
1589 | /* The following loop will look for the hottest edge via |
1590 | the edge count, if it is non-zero, then fallback to |
1591 | the edge probability. */ |
1592 | if (!(e->count () > highest_count)) |
1593 | highest_count = e->count (); |
1594 | if (!highest_probability.initialized_p () |
1595 | || e->probability > highest_probability) |
1596 | highest_probability = e->probability; |
1597 | } |
1598 | |
1599 | /* If bb is reached by (or reaches, in the case of !WALK_UP) another hot |
1600 | block (or unpartitioned, e.g. the entry block) then it is ok. If not, |
1601 | then the most frequent pred (or succ) needs to be adjusted. In the |
1602 | case where multiple preds/succs have the same frequency (e.g. a |
1603 | 50-50 branch), then both will be adjusted. */ |
1604 | if (found) |
1605 | continue; |
1606 | |
1607 | FOR_EACH_EDGE (e, ei, edges) |
1608 | { |
1609 | if (e->flags & EDGE_DFS_BACK) |
1610 | continue; |
1611 | /* Do not expect profile insanities when profile was not adjusted. */ |
1612 | if (e->probability == profile_probability::never () |
1613 | || e->count () == profile_count::zero ()) |
1614 | continue; |
1615 | /* Select the hottest edge using the edge count, if it is non-zero, |
1616 | then fallback to the edge probability. */ |
1617 | if (highest_count.initialized_p ()) |
1618 | { |
1619 | if (!(e->count () >= highest_count)) |
1620 | continue; |
1621 | } |
1622 | else if (!(e->probability >= highest_probability)) |
1623 | continue; |
1624 | |
1625 | basic_block reach_bb = walk_up ? e->src : e->dest; |
1626 | |
1627 | /* We have a hot bb with an immediate dominator that is cold. |
1628 | The dominator needs to be re-marked hot. */ |
1629 | BB_SET_PARTITION (reach_bb, BB_HOT_PARTITION); |
1630 | if (dump_file) |
1631 | fprintf (stream: dump_file, format: "Promoting bb %i to hot partition to sanitize " |
1632 | "profile of bb %i in %s walk\n" , reach_bb->index, |
1633 | bb->index, walk_up ? "backward" : "forward" ); |
1634 | cold_bb_count--; |
1635 | |
1636 | /* Now we need to examine newly-hot reach_bb to see if it is also |
1637 | dominated by a cold bb. */ |
1638 | bbs_in_hot_partition->safe_push (obj: reach_bb); |
1639 | hot_bbs_to_check.safe_push (obj: reach_bb); |
1640 | } |
1641 | } |
1642 | hot_bbs_to_check.release (); |
1643 | |
1644 | return cold_bb_count; |
1645 | } |
1646 | |
1647 | |
1648 | /* Find the basic blocks that are rarely executed and need to be moved to |
1649 | a separate section of the .o file (to cut down on paging and improve |
1650 | cache locality). Return a vector of all edges that cross. */ |
1651 | |
1652 | static vec<edge> |
1653 | find_rarely_executed_basic_blocks_and_crossing_edges (void) |
1654 | { |
1655 | vec<edge> crossing_edges = vNULL; |
1656 | basic_block bb; |
1657 | edge e; |
1658 | edge_iterator ei; |
1659 | unsigned int cold_bb_count = 0; |
1660 | auto_vec<basic_block> bbs_in_hot_partition; |
1661 | |
1662 | propagate_unlikely_bbs_forward (); |
1663 | |
1664 | /* Mark which partition (hot/cold) each basic block belongs in. */ |
1665 | FOR_EACH_BB_FN (bb, cfun) |
1666 | { |
1667 | bool cold_bb = false; |
1668 | |
1669 | if (probably_never_executed_bb_p (cfun, bb)) |
1670 | { |
1671 | cold_bb = true; |
1672 | |
1673 | /* Handle profile insanities created by upstream optimizations |
1674 | by also checking the incoming edge weights. If there is a non-cold |
1675 | incoming edge, conservatively prevent this block from being split |
1676 | into the cold section. */ |
1677 | if (!bb->count.precise_p ()) |
1678 | FOR_EACH_EDGE (e, ei, bb->preds) |
1679 | if (!probably_never_executed_edge_p (cfun, e)) |
1680 | { |
1681 | cold_bb = false; |
1682 | break; |
1683 | } |
1684 | } |
1685 | if (cold_bb) |
1686 | { |
1687 | BB_SET_PARTITION (bb, BB_COLD_PARTITION); |
1688 | cold_bb_count++; |
1689 | } |
1690 | else |
1691 | { |
1692 | BB_SET_PARTITION (bb, BB_HOT_PARTITION); |
1693 | bbs_in_hot_partition.safe_push (obj: bb); |
1694 | } |
1695 | } |
1696 | |
1697 | /* Ensure that hot bbs are included along a hot path from the entry to exit. |
1698 | Several different possibilities may include cold bbs along all paths |
1699 | to/from a hot bb. One is that there are edge weight insanities |
1700 | due to optimization phases that do not properly update basic block profile |
1701 | counts. The second is that the entry of the function may not be hot, because |
1702 | it is entered fewer times than the number of profile training runs, but there |
1703 | is a loop inside the function that causes blocks within the function to be |
1704 | above the threshold for hotness. This is fixed by walking up from hot bbs |
1705 | to the entry block, and then down from hot bbs to the exit, performing |
1706 | partitioning fixups as necessary. */ |
1707 | if (cold_bb_count) |
1708 | { |
1709 | mark_dfs_back_edges (); |
1710 | cold_bb_count = sanitize_hot_paths (walk_up: true, cold_bb_count, |
1711 | bbs_in_hot_partition: &bbs_in_hot_partition); |
1712 | if (cold_bb_count) |
1713 | sanitize_hot_paths (walk_up: false, cold_bb_count, bbs_in_hot_partition: &bbs_in_hot_partition); |
1714 | |
1715 | hash_set <basic_block> set; |
1716 | find_bbs_reachable_by_hot_paths (&set); |
1717 | FOR_EACH_BB_FN (bb, cfun) |
1718 | if (!set.contains (k: bb)) |
1719 | BB_SET_PARTITION (bb, BB_COLD_PARTITION); |
1720 | } |
1721 | |
1722 | /* The format of .gcc_except_table does not allow landing pads to |
1723 | be in a different partition as the throw. Fix this by either |
1724 | moving the landing pads or inserting forwarder landing pads. */ |
1725 | if (cfun->eh->lp_array) |
1726 | { |
1727 | const bool sjlj |
1728 | = (targetm_common.except_unwind_info (&global_options) == UI_SJLJ); |
1729 | unsigned i; |
1730 | eh_landing_pad lp; |
1731 | |
1732 | FOR_EACH_VEC_ELT (*cfun->eh->lp_array, i, lp) |
1733 | { |
1734 | bool all_same, all_diff; |
1735 | |
1736 | if (lp == NULL |
1737 | || lp->landing_pad == NULL_RTX |
1738 | || !LABEL_P (lp->landing_pad)) |
1739 | continue; |
1740 | |
1741 | all_same = all_diff = true; |
1742 | bb = BLOCK_FOR_INSN (insn: lp->landing_pad); |
1743 | FOR_EACH_EDGE (e, ei, bb->preds) |
1744 | { |
1745 | gcc_assert (e->flags & EDGE_EH); |
1746 | if (BB_PARTITION (bb) == BB_PARTITION (e->src)) |
1747 | all_diff = false; |
1748 | else |
1749 | all_same = false; |
1750 | } |
1751 | |
1752 | if (all_same) |
1753 | ; |
1754 | else if (all_diff) |
1755 | { |
1756 | int which = BB_PARTITION (bb); |
1757 | which ^= BB_HOT_PARTITION | BB_COLD_PARTITION; |
1758 | BB_SET_PARTITION (bb, which); |
1759 | } |
1760 | else if (sjlj) |
1761 | sjlj_fix_up_crossing_landing_pad (old_bb: bb); |
1762 | else |
1763 | dw2_fix_up_crossing_landing_pad (old_lp: lp, old_bb: bb); |
1764 | |
1765 | /* There is a single, common landing pad in SJLJ mode. */ |
1766 | if (sjlj) |
1767 | break; |
1768 | } |
1769 | } |
1770 | |
1771 | /* Mark every edge that crosses between sections. */ |
1772 | FOR_EACH_BB_FN (bb, cfun) |
1773 | FOR_EACH_EDGE (e, ei, bb->succs) |
1774 | { |
1775 | unsigned int flags = e->flags; |
1776 | |
1777 | /* We should never have EDGE_CROSSING set yet. */ |
1778 | gcc_checking_assert ((flags & EDGE_CROSSING) == 0); |
1779 | |
1780 | if (e->src != ENTRY_BLOCK_PTR_FOR_FN (cfun) |
1781 | && e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
1782 | && BB_PARTITION (e->src) != BB_PARTITION (e->dest)) |
1783 | { |
1784 | crossing_edges.safe_push (obj: e); |
1785 | flags |= EDGE_CROSSING; |
1786 | } |
1787 | |
1788 | /* Now that we've split eh edges as appropriate, allow landing pads |
1789 | to be merged with the post-landing pads. */ |
1790 | flags &= ~EDGE_PRESERVE; |
1791 | |
1792 | e->flags = flags; |
1793 | } |
1794 | |
1795 | return crossing_edges; |
1796 | } |
1797 | |
1798 | /* Set the flag EDGE_CAN_FALLTHRU for edges that can be fallthru. */ |
1799 | |
1800 | static void |
1801 | set_edge_can_fallthru_flag (void) |
1802 | { |
1803 | basic_block bb; |
1804 | |
1805 | FOR_EACH_BB_FN (bb, cfun) |
1806 | { |
1807 | edge e; |
1808 | edge_iterator ei; |
1809 | |
1810 | FOR_EACH_EDGE (e, ei, bb->succs) |
1811 | { |
1812 | e->flags &= ~EDGE_CAN_FALLTHRU; |
1813 | |
1814 | /* The FALLTHRU edge is also CAN_FALLTHRU edge. */ |
1815 | if (e->flags & EDGE_FALLTHRU) |
1816 | e->flags |= EDGE_CAN_FALLTHRU; |
1817 | } |
1818 | |
1819 | /* If the BB ends with an invertible condjump all (2) edges are |
1820 | CAN_FALLTHRU edges. */ |
1821 | if (EDGE_COUNT (bb->succs) != 2) |
1822 | continue; |
1823 | if (!any_condjump_p (BB_END (bb))) |
1824 | continue; |
1825 | |
1826 | rtx_jump_insn *bb_end_jump = as_a <rtx_jump_insn *> (BB_END (bb)); |
1827 | if (!invert_jump (bb_end_jump, JUMP_LABEL (bb_end_jump), 0)) |
1828 | continue; |
1829 | invert_jump (bb_end_jump, JUMP_LABEL (bb_end_jump), 0); |
1830 | EDGE_SUCC (bb, 0)->flags |= EDGE_CAN_FALLTHRU; |
1831 | EDGE_SUCC (bb, 1)->flags |= EDGE_CAN_FALLTHRU; |
1832 | } |
1833 | } |
1834 | |
1835 | /* If any destination of a crossing edge does not have a label, add label; |
1836 | Convert any easy fall-through crossing edges to unconditional jumps. */ |
1837 | |
1838 | static void |
1839 | add_labels_and_missing_jumps (vec<edge> crossing_edges) |
1840 | { |
1841 | size_t i; |
1842 | edge e; |
1843 | |
1844 | FOR_EACH_VEC_ELT (crossing_edges, i, e) |
1845 | { |
1846 | basic_block src = e->src; |
1847 | basic_block dest = e->dest; |
1848 | rtx_jump_insn *new_jump; |
1849 | |
1850 | if (dest == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
1851 | continue; |
1852 | |
1853 | /* Make sure dest has a label. */ |
1854 | rtx_code_label *label = block_label (dest); |
1855 | |
1856 | /* Nothing to do for non-fallthru edges. */ |
1857 | if (src == ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
1858 | continue; |
1859 | if ((e->flags & EDGE_FALLTHRU) == 0) |
1860 | continue; |
1861 | |
1862 | /* If the block does not end with a control flow insn, then we |
1863 | can trivially add a jump to the end to fixup the crossing. |
1864 | Otherwise the jump will have to go in a new bb, which will |
1865 | be handled by fix_up_fall_thru_edges function. */ |
1866 | if (control_flow_insn_p (BB_END (src))) |
1867 | continue; |
1868 | |
1869 | /* Make sure there's only one successor. */ |
1870 | gcc_assert (single_succ_p (src)); |
1871 | |
1872 | new_jump = emit_jump_insn_after (targetm.gen_jump (label), BB_END (src)); |
1873 | BB_END (src) = new_jump; |
1874 | JUMP_LABEL (new_jump) = label; |
1875 | LABEL_NUSES (label) += 1; |
1876 | |
1877 | emit_barrier_after_bb (bb: src); |
1878 | |
1879 | /* Mark edge as non-fallthru. */ |
1880 | e->flags &= ~EDGE_FALLTHRU; |
1881 | } |
1882 | } |
1883 | |
1884 | /* Find any bb's where the fall-through edge is a crossing edge (note that |
1885 | these bb's must also contain a conditional jump or end with a call |
1886 | instruction; we've already dealt with fall-through edges for blocks |
1887 | that didn't have a conditional jump or didn't end with call instruction |
1888 | in the call to add_labels_and_missing_jumps). Convert the fall-through |
1889 | edge to non-crossing edge by inserting a new bb to fall-through into. |
1890 | The new bb will contain an unconditional jump (crossing edge) to the |
1891 | original fall through destination. */ |
1892 | |
1893 | static void |
1894 | fix_up_fall_thru_edges (void) |
1895 | { |
1896 | basic_block cur_bb; |
1897 | |
1898 | FOR_EACH_BB_FN (cur_bb, cfun) |
1899 | { |
1900 | edge succ1; |
1901 | edge succ2; |
1902 | edge fall_thru = NULL; |
1903 | edge cond_jump = NULL; |
1904 | |
1905 | fall_thru = NULL; |
1906 | if (EDGE_COUNT (cur_bb->succs) > 0) |
1907 | succ1 = EDGE_SUCC (cur_bb, 0); |
1908 | else |
1909 | succ1 = NULL; |
1910 | |
1911 | if (EDGE_COUNT (cur_bb->succs) > 1) |
1912 | succ2 = EDGE_SUCC (cur_bb, 1); |
1913 | else |
1914 | succ2 = NULL; |
1915 | |
1916 | /* Find the fall-through edge. */ |
1917 | |
1918 | if (succ1 |
1919 | && (succ1->flags & EDGE_FALLTHRU)) |
1920 | { |
1921 | fall_thru = succ1; |
1922 | cond_jump = succ2; |
1923 | } |
1924 | else if (succ2 |
1925 | && (succ2->flags & EDGE_FALLTHRU)) |
1926 | { |
1927 | fall_thru = succ2; |
1928 | cond_jump = succ1; |
1929 | } |
1930 | else if (succ2 && EDGE_COUNT (cur_bb->succs) > 2) |
1931 | fall_thru = find_fallthru_edge (edges: cur_bb->succs); |
1932 | |
1933 | if (fall_thru && (fall_thru->dest != EXIT_BLOCK_PTR_FOR_FN (cfun))) |
1934 | { |
1935 | /* Check to see if the fall-thru edge is a crossing edge. */ |
1936 | |
1937 | if (fall_thru->flags & EDGE_CROSSING) |
1938 | { |
1939 | /* The fall_thru edge crosses; now check the cond jump edge, if |
1940 | it exists. */ |
1941 | |
1942 | bool cond_jump_crosses = true; |
1943 | int invert_worked = 0; |
1944 | rtx_insn *old_jump = BB_END (cur_bb); |
1945 | |
1946 | /* Find the jump instruction, if there is one. */ |
1947 | |
1948 | if (cond_jump) |
1949 | { |
1950 | if (!(cond_jump->flags & EDGE_CROSSING)) |
1951 | cond_jump_crosses = false; |
1952 | |
1953 | /* We know the fall-thru edge crosses; if the cond |
1954 | jump edge does NOT cross, and its destination is the |
1955 | next block in the bb order, invert the jump |
1956 | (i.e. fix it so the fall through does not cross and |
1957 | the cond jump does). */ |
1958 | |
1959 | if (!cond_jump_crosses) |
1960 | { |
1961 | /* Find label in fall_thru block. We've already added |
1962 | any missing labels, so there must be one. */ |
1963 | |
1964 | rtx_code_label *fall_thru_label |
1965 | = block_label (fall_thru->dest); |
1966 | |
1967 | if (old_jump && fall_thru_label) |
1968 | { |
1969 | rtx_jump_insn *old_jump_insn |
1970 | = dyn_cast <rtx_jump_insn *> (p: old_jump); |
1971 | if (old_jump_insn) |
1972 | invert_worked = invert_jump (old_jump_insn, |
1973 | fall_thru_label, 0); |
1974 | } |
1975 | |
1976 | if (invert_worked) |
1977 | { |
1978 | fall_thru->flags &= ~EDGE_FALLTHRU; |
1979 | cond_jump->flags |= EDGE_FALLTHRU; |
1980 | update_br_prob_note (cur_bb); |
1981 | std::swap (a&: fall_thru, b&: cond_jump); |
1982 | cond_jump->flags |= EDGE_CROSSING; |
1983 | fall_thru->flags &= ~EDGE_CROSSING; |
1984 | } |
1985 | } |
1986 | } |
1987 | |
1988 | if (cond_jump_crosses || !invert_worked) |
1989 | { |
1990 | /* This is the case where both edges out of the basic |
1991 | block are crossing edges. Here we will fix up the |
1992 | fall through edge. The jump edge will be taken care |
1993 | of later. The EDGE_CROSSING flag of fall_thru edge |
1994 | is unset before the call to force_nonfallthru |
1995 | function because if a new basic-block is created |
1996 | this edge remains in the current section boundary |
1997 | while the edge between new_bb and the fall_thru->dest |
1998 | becomes EDGE_CROSSING. */ |
1999 | |
2000 | fall_thru->flags &= ~EDGE_CROSSING; |
2001 | unsigned old_count = EDGE_COUNT (cur_bb->succs); |
2002 | basic_block new_bb = force_nonfallthru (fall_thru); |
2003 | |
2004 | if (new_bb) |
2005 | { |
2006 | new_bb->aux = cur_bb->aux; |
2007 | cur_bb->aux = new_bb; |
2008 | |
2009 | /* This is done by force_nonfallthru_and_redirect. */ |
2010 | gcc_assert (BB_PARTITION (new_bb) |
2011 | == BB_PARTITION (cur_bb)); |
2012 | |
2013 | edge e = single_succ_edge (bb: new_bb); |
2014 | e->flags |= EDGE_CROSSING; |
2015 | if (EDGE_COUNT (cur_bb->succs) > old_count) |
2016 | { |
2017 | /* If asm goto has a crossing fallthrough edge |
2018 | and at least one of the labels to the same bb, |
2019 | force_nonfallthru can result in the fallthrough |
2020 | edge being redirected and a new edge added for the |
2021 | label or more labels to e->dest. As we've |
2022 | temporarily cleared EDGE_CROSSING flag on the |
2023 | fallthrough edge, we need to restore it again. |
2024 | See PR108596. */ |
2025 | rtx_insn *j = BB_END (cur_bb); |
2026 | gcc_checking_assert (JUMP_P (j) |
2027 | && asm_noperands (PATTERN (j))); |
2028 | edge e2 = find_edge (cur_bb, e->dest); |
2029 | if (e2) |
2030 | e2->flags |= EDGE_CROSSING; |
2031 | } |
2032 | } |
2033 | else |
2034 | { |
2035 | /* If a new basic-block was not created; restore |
2036 | the EDGE_CROSSING flag. */ |
2037 | fall_thru->flags |= EDGE_CROSSING; |
2038 | } |
2039 | |
2040 | /* Add barrier after new jump */ |
2041 | emit_barrier_after_bb (bb: new_bb ? new_bb : cur_bb); |
2042 | } |
2043 | } |
2044 | } |
2045 | } |
2046 | } |
2047 | |
2048 | /* This function checks the destination block of a "crossing jump" to |
2049 | see if it has any crossing predecessors that begin with a code label |
2050 | and end with an unconditional jump. If so, it returns that predecessor |
2051 | block. (This is to avoid creating lots of new basic blocks that all |
2052 | contain unconditional jumps to the same destination). */ |
2053 | |
2054 | static basic_block |
2055 | find_jump_block (basic_block jump_dest) |
2056 | { |
2057 | basic_block source_bb = NULL; |
2058 | edge e; |
2059 | rtx_insn *insn; |
2060 | edge_iterator ei; |
2061 | |
2062 | FOR_EACH_EDGE (e, ei, jump_dest->preds) |
2063 | if (e->flags & EDGE_CROSSING) |
2064 | { |
2065 | basic_block src = e->src; |
2066 | |
2067 | /* Check each predecessor to see if it has a label, and contains |
2068 | only one executable instruction, which is an unconditional jump. |
2069 | If so, we can use it. */ |
2070 | |
2071 | if (LABEL_P (BB_HEAD (src))) |
2072 | for (insn = BB_HEAD (src); |
2073 | !INSN_P (insn) && insn != NEXT_INSN (BB_END (src)); |
2074 | insn = NEXT_INSN (insn)) |
2075 | { |
2076 | if (INSN_P (insn) |
2077 | && insn == BB_END (src) |
2078 | && JUMP_P (insn) |
2079 | && !any_condjump_p (insn)) |
2080 | { |
2081 | source_bb = src; |
2082 | break; |
2083 | } |
2084 | } |
2085 | |
2086 | if (source_bb) |
2087 | break; |
2088 | } |
2089 | |
2090 | return source_bb; |
2091 | } |
2092 | |
2093 | /* Find all BB's with conditional jumps that are crossing edges; |
2094 | insert a new bb and make the conditional jump branch to the new |
2095 | bb instead (make the new bb same color so conditional branch won't |
2096 | be a 'crossing' edge). Insert an unconditional jump from the |
2097 | new bb to the original destination of the conditional jump. */ |
2098 | |
2099 | static void |
2100 | fix_crossing_conditional_branches (void) |
2101 | { |
2102 | basic_block cur_bb; |
2103 | basic_block new_bb; |
2104 | basic_block dest; |
2105 | edge succ1; |
2106 | edge succ2; |
2107 | edge crossing_edge; |
2108 | edge new_edge; |
2109 | rtx set_src; |
2110 | rtx old_label = NULL_RTX; |
2111 | rtx_code_label *new_label; |
2112 | |
2113 | FOR_EACH_BB_FN (cur_bb, cfun) |
2114 | { |
2115 | crossing_edge = NULL; |
2116 | if (EDGE_COUNT (cur_bb->succs) > 0) |
2117 | succ1 = EDGE_SUCC (cur_bb, 0); |
2118 | else |
2119 | succ1 = NULL; |
2120 | |
2121 | if (EDGE_COUNT (cur_bb->succs) > 1) |
2122 | succ2 = EDGE_SUCC (cur_bb, 1); |
2123 | else |
2124 | succ2 = NULL; |
2125 | |
2126 | /* We already took care of fall-through edges, so only one successor |
2127 | can be a crossing edge. */ |
2128 | |
2129 | if (succ1 && (succ1->flags & EDGE_CROSSING)) |
2130 | crossing_edge = succ1; |
2131 | else if (succ2 && (succ2->flags & EDGE_CROSSING)) |
2132 | crossing_edge = succ2; |
2133 | |
2134 | if (crossing_edge) |
2135 | { |
2136 | rtx_insn *old_jump = BB_END (cur_bb); |
2137 | |
2138 | /* Check to make sure the jump instruction is a |
2139 | conditional jump. */ |
2140 | |
2141 | set_src = NULL_RTX; |
2142 | |
2143 | if (any_condjump_p (old_jump)) |
2144 | { |
2145 | if (GET_CODE (PATTERN (old_jump)) == SET) |
2146 | set_src = SET_SRC (PATTERN (old_jump)); |
2147 | else if (GET_CODE (PATTERN (old_jump)) == PARALLEL) |
2148 | { |
2149 | set_src = XVECEXP (PATTERN (old_jump), 0,0); |
2150 | if (GET_CODE (set_src) == SET) |
2151 | set_src = SET_SRC (set_src); |
2152 | else |
2153 | set_src = NULL_RTX; |
2154 | } |
2155 | } |
2156 | |
2157 | if (set_src && (GET_CODE (set_src) == IF_THEN_ELSE)) |
2158 | { |
2159 | rtx_jump_insn *old_jump_insn = |
2160 | as_a <rtx_jump_insn *> (p: old_jump); |
2161 | |
2162 | if (GET_CODE (XEXP (set_src, 1)) == PC) |
2163 | old_label = XEXP (set_src, 2); |
2164 | else if (GET_CODE (XEXP (set_src, 2)) == PC) |
2165 | old_label = XEXP (set_src, 1); |
2166 | |
2167 | /* Check to see if new bb for jumping to that dest has |
2168 | already been created; if so, use it; if not, create |
2169 | a new one. */ |
2170 | |
2171 | new_bb = find_jump_block (jump_dest: crossing_edge->dest); |
2172 | |
2173 | if (new_bb) |
2174 | new_label = block_label (new_bb); |
2175 | else |
2176 | { |
2177 | basic_block last_bb; |
2178 | rtx_code_label *old_jump_target; |
2179 | rtx_jump_insn *new_jump; |
2180 | |
2181 | /* Create new basic block to be dest for |
2182 | conditional jump. */ |
2183 | |
2184 | /* Put appropriate instructions in new bb. */ |
2185 | |
2186 | new_label = gen_label_rtx (); |
2187 | emit_label (new_label); |
2188 | |
2189 | gcc_assert (GET_CODE (old_label) == LABEL_REF); |
2190 | old_jump_target = old_jump_insn->jump_target (); |
2191 | new_jump = as_a <rtx_jump_insn *> |
2192 | (p: emit_jump_insn (targetm.gen_jump (old_jump_target))); |
2193 | new_jump->set_jump_target (old_jump_target); |
2194 | |
2195 | last_bb = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb; |
2196 | new_bb = create_basic_block (new_label, new_jump, last_bb); |
2197 | new_bb->aux = last_bb->aux; |
2198 | last_bb->aux = new_bb; |
2199 | |
2200 | emit_barrier_after_bb (bb: new_bb); |
2201 | |
2202 | /* Make sure new bb is in same partition as source |
2203 | of conditional branch. */ |
2204 | BB_COPY_PARTITION (new_bb, cur_bb); |
2205 | } |
2206 | |
2207 | /* Make old jump branch to new bb. */ |
2208 | |
2209 | redirect_jump (old_jump_insn, new_label, 0); |
2210 | |
2211 | /* Remove crossing_edge as predecessor of 'dest'. */ |
2212 | |
2213 | dest = crossing_edge->dest; |
2214 | |
2215 | redirect_edge_succ (crossing_edge, new_bb); |
2216 | |
2217 | /* Make a new edge from new_bb to old dest; new edge |
2218 | will be a successor for new_bb and a predecessor |
2219 | for 'dest'. */ |
2220 | |
2221 | if (EDGE_COUNT (new_bb->succs) == 0) |
2222 | new_edge = make_single_succ_edge (new_bb, dest, 0); |
2223 | else |
2224 | new_edge = EDGE_SUCC (new_bb, 0); |
2225 | |
2226 | crossing_edge->flags &= ~EDGE_CROSSING; |
2227 | new_edge->flags |= EDGE_CROSSING; |
2228 | } |
2229 | } |
2230 | } |
2231 | } |
2232 | |
2233 | /* Find any unconditional branches that cross between hot and cold |
2234 | sections. Convert them into indirect jumps instead. */ |
2235 | |
2236 | static void |
2237 | fix_crossing_unconditional_branches (void) |
2238 | { |
2239 | basic_block cur_bb; |
2240 | rtx_insn *last_insn; |
2241 | rtx label; |
2242 | rtx label_addr; |
2243 | rtx_insn *indirect_jump_sequence; |
2244 | rtx_insn *jump_insn = NULL; |
2245 | rtx new_reg; |
2246 | rtx_insn *cur_insn; |
2247 | edge succ; |
2248 | |
2249 | FOR_EACH_BB_FN (cur_bb, cfun) |
2250 | { |
2251 | last_insn = BB_END (cur_bb); |
2252 | |
2253 | if (EDGE_COUNT (cur_bb->succs) < 1) |
2254 | continue; |
2255 | |
2256 | succ = EDGE_SUCC (cur_bb, 0); |
2257 | |
2258 | /* Check to see if bb ends in a crossing (unconditional) jump. At |
2259 | this point, no crossing jumps should be conditional. */ |
2260 | |
2261 | if (JUMP_P (last_insn) |
2262 | && (succ->flags & EDGE_CROSSING)) |
2263 | { |
2264 | gcc_assert (!any_condjump_p (last_insn)); |
2265 | |
2266 | /* Make sure the jump is not already an indirect or table jump. */ |
2267 | |
2268 | if (!computed_jump_p (last_insn) |
2269 | && !tablejump_p (last_insn, NULL, NULL)) |
2270 | { |
2271 | /* We have found a "crossing" unconditional branch. Now |
2272 | we must convert it to an indirect jump. First create |
2273 | reference of label, as target for jump. */ |
2274 | |
2275 | label = JUMP_LABEL (last_insn); |
2276 | label_addr = gen_rtx_LABEL_REF (Pmode, label); |
2277 | LABEL_NUSES (label) += 1; |
2278 | |
2279 | /* Get a register to use for the indirect jump. */ |
2280 | |
2281 | new_reg = gen_reg_rtx (Pmode); |
2282 | |
2283 | /* Generate indirect the jump sequence. */ |
2284 | |
2285 | start_sequence (); |
2286 | emit_move_insn (new_reg, label_addr); |
2287 | emit_indirect_jump (new_reg); |
2288 | indirect_jump_sequence = get_insns (); |
2289 | end_sequence (); |
2290 | |
2291 | /* Make sure every instruction in the new jump sequence has |
2292 | its basic block set to be cur_bb. */ |
2293 | |
2294 | for (cur_insn = indirect_jump_sequence; cur_insn; |
2295 | cur_insn = NEXT_INSN (insn: cur_insn)) |
2296 | { |
2297 | if (!BARRIER_P (cur_insn)) |
2298 | BLOCK_FOR_INSN (insn: cur_insn) = cur_bb; |
2299 | if (JUMP_P (cur_insn)) |
2300 | jump_insn = cur_insn; |
2301 | } |
2302 | |
2303 | /* Insert the new (indirect) jump sequence immediately before |
2304 | the unconditional jump, then delete the unconditional jump. */ |
2305 | |
2306 | emit_insn_before (indirect_jump_sequence, last_insn); |
2307 | delete_insn (last_insn); |
2308 | |
2309 | JUMP_LABEL (jump_insn) = label; |
2310 | LABEL_NUSES (label)++; |
2311 | |
2312 | /* Make BB_END for cur_bb be the jump instruction (NOT the |
2313 | barrier instruction at the end of the sequence...). */ |
2314 | |
2315 | BB_END (cur_bb) = jump_insn; |
2316 | } |
2317 | } |
2318 | } |
2319 | } |
2320 | |
2321 | /* Update CROSSING_JUMP_P flags on all jump insns. */ |
2322 | |
2323 | static void |
2324 | update_crossing_jump_flags (void) |
2325 | { |
2326 | basic_block bb; |
2327 | edge e; |
2328 | edge_iterator ei; |
2329 | |
2330 | FOR_EACH_BB_FN (bb, cfun) |
2331 | FOR_EACH_EDGE (e, ei, bb->succs) |
2332 | if (e->flags & EDGE_CROSSING) |
2333 | { |
2334 | if (JUMP_P (BB_END (bb))) |
2335 | CROSSING_JUMP_P (BB_END (bb)) = 1; |
2336 | break; |
2337 | } |
2338 | } |
2339 | |
2340 | /* Reorder basic blocks using the software trace cache (STC) algorithm. */ |
2341 | |
2342 | static void |
2343 | reorder_basic_blocks_software_trace_cache (void) |
2344 | { |
2345 | if (dump_file) |
2346 | fprintf (stream: dump_file, format: "\nReordering with the STC algorithm.\n\n" ); |
2347 | |
2348 | int n_traces; |
2349 | int i; |
2350 | struct trace *traces; |
2351 | |
2352 | /* We are estimating the length of uncond jump insn only once since the code |
2353 | for getting the insn length always returns the minimal length now. */ |
2354 | if (uncond_jump_length == 0) |
2355 | uncond_jump_length = get_uncond_jump_length (); |
2356 | |
2357 | /* We need to know some information for each basic block. */ |
2358 | array_size = GET_ARRAY_SIZE (last_basic_block_for_fn (cfun)); |
2359 | bbd = XNEWVEC (bbro_basic_block_data, array_size); |
2360 | for (i = 0; i < array_size; i++) |
2361 | { |
2362 | bbd[i].start_of_trace = -1; |
2363 | bbd[i].end_of_trace = -1; |
2364 | bbd[i].in_trace = -1; |
2365 | bbd[i].visited = 0; |
2366 | bbd[i].priority = -1; |
2367 | bbd[i].heap = NULL; |
2368 | bbd[i].node = NULL; |
2369 | } |
2370 | |
2371 | traces = XNEWVEC (struct trace, n_basic_blocks_for_fn (cfun)); |
2372 | n_traces = 0; |
2373 | find_traces (n_traces: &n_traces, traces); |
2374 | connect_traces (n_traces, traces); |
2375 | FREE (traces); |
2376 | FREE (bbd); |
2377 | } |
2378 | |
2379 | /* Order edges by execution frequency, higher first. */ |
2380 | |
2381 | static int |
2382 | edge_order (const void *ve1, const void *ve2) |
2383 | { |
2384 | edge e1 = *(const edge *) ve1; |
2385 | edge e2 = *(const edge *) ve2; |
2386 | profile_count c1 = e1->count (); |
2387 | profile_count c2 = e2->count (); |
2388 | /* Since profile_count::operator< does not establish a strict weak order |
2389 | in presence of uninitialized counts, use 'max': this makes them appear |
2390 | as if having execution frequency less than any initialized count. */ |
2391 | profile_count m = c1.max (other: c2); |
2392 | return (m == c2) - (m == c1); |
2393 | } |
2394 | |
2395 | /* Reorder basic blocks using the "simple" algorithm. This tries to |
2396 | maximize the dynamic number of branches that are fallthrough, without |
2397 | copying instructions. The algorithm is greedy, looking at the most |
2398 | frequently executed branch first. */ |
2399 | |
2400 | static void |
2401 | reorder_basic_blocks_simple (void) |
2402 | { |
2403 | if (dump_file) |
2404 | fprintf (stream: dump_file, format: "\nReordering with the \"simple\" algorithm.\n\n" ); |
2405 | |
2406 | edge *edges = new edge[2 * n_basic_blocks_for_fn (cfun)]; |
2407 | |
2408 | /* First, collect all edges that can be optimized by reordering blocks: |
2409 | simple jumps and conditional jumps, as well as the function entry edge. */ |
2410 | |
2411 | int n = 0; |
2412 | edges[n++] = EDGE_SUCC (ENTRY_BLOCK_PTR_FOR_FN (cfun), 0); |
2413 | |
2414 | basic_block bb; |
2415 | FOR_EACH_BB_FN (bb, cfun) |
2416 | { |
2417 | rtx_insn *end = BB_END (bb); |
2418 | |
2419 | if (computed_jump_p (end) || tablejump_p (end, NULL, NULL)) |
2420 | continue; |
2421 | |
2422 | /* We cannot optimize asm goto. */ |
2423 | if (JUMP_P (end) && extract_asm_operands (end)) |
2424 | continue; |
2425 | |
2426 | if (single_succ_p (bb)) |
2427 | edges[n++] = EDGE_SUCC (bb, 0); |
2428 | else if (any_condjump_p (end)) |
2429 | { |
2430 | edge e0 = EDGE_SUCC (bb, 0); |
2431 | edge e1 = EDGE_SUCC (bb, 1); |
2432 | /* When optimizing for size it is best to keep the original |
2433 | fallthrough edges. */ |
2434 | if (e1->flags & EDGE_FALLTHRU) |
2435 | std::swap (a&: e0, b&: e1); |
2436 | edges[n++] = e0; |
2437 | edges[n++] = e1; |
2438 | } |
2439 | } |
2440 | |
2441 | /* Sort the edges, the most desirable first. When optimizing for size |
2442 | all edges are equally desirable. */ |
2443 | |
2444 | if (optimize_function_for_speed_p (cfun)) |
2445 | gcc_stablesort (edges, n, sizeof *edges, edge_order); |
2446 | |
2447 | /* Now decide which of those edges to make fallthrough edges. We set |
2448 | BB_VISITED if a block already has a fallthrough successor assigned |
2449 | to it. We make ->AUX of an endpoint point to the opposite endpoint |
2450 | of a sequence of blocks that fall through, and ->AUX will be NULL |
2451 | for a block that is in such a sequence but not an endpoint anymore. |
2452 | |
2453 | To start with, everything points to itself, nothing is assigned yet. */ |
2454 | |
2455 | FOR_ALL_BB_FN (bb, cfun) |
2456 | { |
2457 | bb->aux = bb; |
2458 | bb->flags &= ~BB_VISITED; |
2459 | } |
2460 | |
2461 | EXIT_BLOCK_PTR_FOR_FN (cfun)->aux = 0; |
2462 | |
2463 | /* Now for all edges, the most desirable first, see if that edge can |
2464 | connect two sequences. If it can, update AUX and BB_VISITED; if it |
2465 | cannot, zero out the edge in the table. */ |
2466 | |
2467 | for (int j = 0; j < n; j++) |
2468 | { |
2469 | edge e = edges[j]; |
2470 | |
2471 | basic_block tail_a = e->src; |
2472 | basic_block head_b = e->dest; |
2473 | basic_block head_a = (basic_block) tail_a->aux; |
2474 | basic_block tail_b = (basic_block) head_b->aux; |
2475 | |
2476 | /* An edge cannot connect two sequences if: |
2477 | - it crosses partitions; |
2478 | - its src is not a current endpoint; |
2479 | - its dest is not a current endpoint; |
2480 | - or, it would create a loop. */ |
2481 | |
2482 | if (e->flags & EDGE_CROSSING |
2483 | || tail_a->flags & BB_VISITED |
2484 | || !tail_b |
2485 | || (!(head_b->flags & BB_VISITED) && head_b != tail_b) |
2486 | || tail_a == tail_b) |
2487 | { |
2488 | edges[j] = 0; |
2489 | continue; |
2490 | } |
2491 | |
2492 | tail_a->aux = 0; |
2493 | head_b->aux = 0; |
2494 | head_a->aux = tail_b; |
2495 | tail_b->aux = head_a; |
2496 | tail_a->flags |= BB_VISITED; |
2497 | } |
2498 | |
2499 | /* Put the pieces together, in the same order that the start blocks of |
2500 | the sequences already had. The hot/cold partitioning gives a little |
2501 | complication: as a first pass only do this for blocks in the same |
2502 | partition as the start block, and (if there is anything left to do) |
2503 | in a second pass handle the other partition. */ |
2504 | |
2505 | basic_block last_tail = (basic_block) ENTRY_BLOCK_PTR_FOR_FN (cfun)->aux; |
2506 | |
2507 | int current_partition |
2508 | = BB_PARTITION (last_tail == ENTRY_BLOCK_PTR_FOR_FN (cfun) |
2509 | ? EDGE_SUCC (ENTRY_BLOCK_PTR_FOR_FN (cfun), 0)->dest |
2510 | : last_tail); |
2511 | bool need_another_pass = true; |
2512 | |
2513 | for (int pass = 0; pass < 2 && need_another_pass; pass++) |
2514 | { |
2515 | need_another_pass = false; |
2516 | |
2517 | FOR_EACH_BB_FN (bb, cfun) |
2518 | if ((bb->flags & BB_VISITED && bb->aux) || bb->aux == bb) |
2519 | { |
2520 | if (BB_PARTITION (bb) != current_partition) |
2521 | { |
2522 | need_another_pass = true; |
2523 | continue; |
2524 | } |
2525 | |
2526 | last_tail->aux = bb; |
2527 | last_tail = (basic_block) bb->aux; |
2528 | } |
2529 | |
2530 | current_partition ^= BB_HOT_PARTITION | BB_COLD_PARTITION; |
2531 | } |
2532 | |
2533 | last_tail->aux = 0; |
2534 | |
2535 | /* Finally, link all the chosen fallthrough edges. */ |
2536 | |
2537 | for (int j = 0; j < n; j++) |
2538 | if (edges[j]) |
2539 | edges[j]->src->aux = edges[j]->dest; |
2540 | |
2541 | delete[] edges; |
2542 | |
2543 | /* If the entry edge no longer falls through we have to make a new |
2544 | block so it can do so again. */ |
2545 | |
2546 | edge e = EDGE_SUCC (ENTRY_BLOCK_PTR_FOR_FN (cfun), 0); |
2547 | if (e->dest != ENTRY_BLOCK_PTR_FOR_FN (cfun)->aux) |
2548 | { |
2549 | force_nonfallthru (e); |
2550 | e->src->aux = ENTRY_BLOCK_PTR_FOR_FN (cfun)->aux; |
2551 | } |
2552 | } |
2553 | |
2554 | /* Reorder basic blocks. The main entry point to this file. */ |
2555 | |
2556 | static void |
2557 | reorder_basic_blocks (void) |
2558 | { |
2559 | gcc_assert (current_ir_type () == IR_RTL_CFGLAYOUT); |
2560 | |
2561 | if (n_basic_blocks_for_fn (cfun) <= NUM_FIXED_BLOCKS + 1) |
2562 | return; |
2563 | |
2564 | set_edge_can_fallthru_flag (); |
2565 | mark_dfs_back_edges (); |
2566 | |
2567 | switch (flag_reorder_blocks_algorithm) |
2568 | { |
2569 | case REORDER_BLOCKS_ALGORITHM_SIMPLE: |
2570 | reorder_basic_blocks_simple (); |
2571 | break; |
2572 | |
2573 | case REORDER_BLOCKS_ALGORITHM_STC: |
2574 | reorder_basic_blocks_software_trace_cache (); |
2575 | break; |
2576 | |
2577 | default: |
2578 | gcc_unreachable (); |
2579 | } |
2580 | |
2581 | relink_block_chain (/*stay_in_cfglayout_mode=*/true); |
2582 | |
2583 | if (dump_file) |
2584 | { |
2585 | if (dump_flags & TDF_DETAILS) |
2586 | dump_reg_info (dump_file); |
2587 | dump_flow_info (dump_file, dump_flags); |
2588 | } |
2589 | |
2590 | /* Signal that rtl_verify_flow_info_1 can now verify that there |
2591 | is at most one switch between hot/cold sections. */ |
2592 | crtl->bb_reorder_complete = true; |
2593 | } |
2594 | |
2595 | /* Determine which partition the first basic block in the function |
2596 | belongs to, then find the first basic block in the current function |
2597 | that belongs to a different section, and insert a |
2598 | NOTE_INSN_SWITCH_TEXT_SECTIONS note immediately before it in the |
2599 | instruction stream. When writing out the assembly code, |
2600 | encountering this note will make the compiler switch between the |
2601 | hot and cold text sections. */ |
2602 | |
2603 | void |
2604 | insert_section_boundary_note (void) |
2605 | { |
2606 | basic_block bb; |
2607 | bool switched_sections = false; |
2608 | int current_partition = 0; |
2609 | |
2610 | if (!crtl->has_bb_partition) |
2611 | return; |
2612 | |
2613 | FOR_EACH_BB_FN (bb, cfun) |
2614 | { |
2615 | if (!current_partition) |
2616 | current_partition = BB_PARTITION (bb); |
2617 | if (BB_PARTITION (bb) != current_partition) |
2618 | { |
2619 | gcc_assert (!switched_sections); |
2620 | switched_sections = true; |
2621 | emit_note_before (NOTE_INSN_SWITCH_TEXT_SECTIONS, BB_HEAD (bb)); |
2622 | current_partition = BB_PARTITION (bb); |
2623 | } |
2624 | } |
2625 | |
2626 | /* Make sure crtl->has_bb_partition matches reality even if bbpart finds |
2627 | some hot and some cold basic blocks, but later one of those kinds is |
2628 | optimized away. */ |
2629 | crtl->has_bb_partition = switched_sections; |
2630 | } |
2631 | |
2632 | namespace { |
2633 | |
2634 | const pass_data pass_data_reorder_blocks = |
2635 | { |
2636 | .type: RTL_PASS, /* type */ |
2637 | .name: "bbro" , /* name */ |
2638 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
2639 | .tv_id: TV_REORDER_BLOCKS, /* tv_id */ |
2640 | .properties_required: 0, /* properties_required */ |
2641 | .properties_provided: 0, /* properties_provided */ |
2642 | .properties_destroyed: 0, /* properties_destroyed */ |
2643 | .todo_flags_start: 0, /* todo_flags_start */ |
2644 | .todo_flags_finish: 0, /* todo_flags_finish */ |
2645 | }; |
2646 | |
2647 | class pass_reorder_blocks : public rtl_opt_pass |
2648 | { |
2649 | public: |
2650 | pass_reorder_blocks (gcc::context *ctxt) |
2651 | : rtl_opt_pass (pass_data_reorder_blocks, ctxt) |
2652 | {} |
2653 | |
2654 | /* opt_pass methods: */ |
2655 | bool gate (function *) final override |
2656 | { |
2657 | if (targetm.cannot_modify_jumps_p ()) |
2658 | return false; |
2659 | return (optimize > 0 |
2660 | && (flag_reorder_blocks || flag_reorder_blocks_and_partition)); |
2661 | } |
2662 | |
2663 | unsigned int execute (function *) final override; |
2664 | |
2665 | }; // class pass_reorder_blocks |
2666 | |
2667 | unsigned int |
2668 | pass_reorder_blocks::execute (function *fun) |
2669 | { |
2670 | basic_block bb; |
2671 | |
2672 | /* Last attempt to optimize CFG, as scheduling, peepholing and insn |
2673 | splitting possibly introduced more crossjumping opportunities. */ |
2674 | cfg_layout_initialize (CLEANUP_EXPENSIVE); |
2675 | |
2676 | reorder_basic_blocks (); |
2677 | cleanup_cfg (CLEANUP_EXPENSIVE | CLEANUP_NO_PARTITIONING); |
2678 | |
2679 | FOR_EACH_BB_FN (bb, fun) |
2680 | if (bb->next_bb != EXIT_BLOCK_PTR_FOR_FN (fun)) |
2681 | bb->aux = bb->next_bb; |
2682 | cfg_layout_finalize (); |
2683 | |
2684 | FOR_EACH_BB_FN (bb, fun) |
2685 | df_recompute_luids (bb); |
2686 | return 0; |
2687 | } |
2688 | |
2689 | } // anon namespace |
2690 | |
2691 | rtl_opt_pass * |
2692 | make_pass_reorder_blocks (gcc::context *ctxt) |
2693 | { |
2694 | return new pass_reorder_blocks (ctxt); |
2695 | } |
2696 | |
2697 | /* Duplicate a block (that we already know ends in a computed jump) into its |
2698 | predecessors, where possible. Return whether anything is changed. */ |
2699 | static bool |
2700 | maybe_duplicate_computed_goto (basic_block bb, int max_size) |
2701 | { |
2702 | /* Make sure that the block is small enough. */ |
2703 | rtx_insn *insn; |
2704 | FOR_BB_INSNS (bb, insn) |
2705 | if (INSN_P (insn)) |
2706 | { |
2707 | max_size -= get_attr_min_length (insn); |
2708 | if (max_size < 0) |
2709 | return false; |
2710 | } |
2711 | |
2712 | bool changed = false; |
2713 | edge e; |
2714 | edge_iterator ei; |
2715 | for (ei = ei_start (bb->preds); (e = ei_safe_edge (i: ei)); ) |
2716 | { |
2717 | basic_block pred = e->src; |
2718 | |
2719 | /* Do not duplicate BB into PRED if we cannot merge a copy of BB |
2720 | with PRED. */ |
2721 | if (!single_succ_p (bb: pred) |
2722 | || e->flags & EDGE_COMPLEX |
2723 | || pred->index < NUM_FIXED_BLOCKS |
2724 | || (JUMP_P (BB_END (pred)) && !simplejump_p (BB_END (pred))) |
2725 | || (JUMP_P (BB_END (pred)) && CROSSING_JUMP_P (BB_END (pred)))) |
2726 | { |
2727 | ei_next (i: &ei); |
2728 | continue; |
2729 | } |
2730 | |
2731 | if (dump_file) |
2732 | fprintf (stream: dump_file, format: "Duplicating computed goto bb %d into bb %d\n" , |
2733 | bb->index, e->src->index); |
2734 | |
2735 | /* Remember if PRED can be duplicated; if so, the copy of BB merged |
2736 | with PRED can be duplicated as well. */ |
2737 | bool can_dup_more = can_duplicate_block_p (pred); |
2738 | |
2739 | /* Make a copy of BB, merge it into PRED. */ |
2740 | basic_block copy = duplicate_block (bb, e, NULL); |
2741 | emit_barrier_after_bb (bb: copy); |
2742 | reorder_insns_nobb (BB_HEAD (copy), BB_END (copy), BB_END (pred)); |
2743 | merge_blocks (pred, copy); |
2744 | |
2745 | changed = true; |
2746 | |
2747 | /* Try to merge the resulting merged PRED into further predecessors. */ |
2748 | if (can_dup_more) |
2749 | maybe_duplicate_computed_goto (bb: pred, max_size); |
2750 | } |
2751 | |
2752 | return changed; |
2753 | } |
2754 | |
2755 | /* Duplicate the blocks containing computed gotos. This basically unfactors |
2756 | computed gotos that were factored early on in the compilation process to |
2757 | speed up edge based data flow. We used to not unfactor them again, which |
2758 | can seriously pessimize code with many computed jumps in the source code, |
2759 | such as interpreters. See e.g. PR15242. */ |
2760 | static void |
2761 | duplicate_computed_gotos (function *fun) |
2762 | { |
2763 | /* We are estimating the length of uncond jump insn only once |
2764 | since the code for getting the insn length always returns |
2765 | the minimal length now. */ |
2766 | if (uncond_jump_length == 0) |
2767 | uncond_jump_length = get_uncond_jump_length (); |
2768 | |
2769 | /* Never copy a block larger than this. */ |
2770 | int max_size |
2771 | = uncond_jump_length * param_max_goto_duplication_insns; |
2772 | |
2773 | bool changed = false; |
2774 | |
2775 | /* Try to duplicate all blocks that end in a computed jump and that |
2776 | can be duplicated at all. */ |
2777 | basic_block bb; |
2778 | FOR_EACH_BB_FN (bb, fun) |
2779 | if (computed_jump_p (BB_END (bb)) && can_duplicate_block_p (bb)) |
2780 | changed |= maybe_duplicate_computed_goto (bb, max_size); |
2781 | |
2782 | /* Some blocks may have become unreachable. */ |
2783 | if (changed) |
2784 | cleanup_cfg (0); |
2785 | |
2786 | /* Duplicating blocks will redirect edges and may cause hot blocks |
2787 | previously reached by both hot and cold blocks to become dominated |
2788 | only by cold blocks. */ |
2789 | if (changed) |
2790 | fixup_partitions (); |
2791 | } |
2792 | |
2793 | namespace { |
2794 | |
2795 | const pass_data pass_data_duplicate_computed_gotos = |
2796 | { |
2797 | .type: RTL_PASS, /* type */ |
2798 | .name: "compgotos" , /* name */ |
2799 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
2800 | .tv_id: TV_REORDER_BLOCKS, /* tv_id */ |
2801 | .properties_required: 0, /* properties_required */ |
2802 | .properties_provided: 0, /* properties_provided */ |
2803 | .properties_destroyed: 0, /* properties_destroyed */ |
2804 | .todo_flags_start: 0, /* todo_flags_start */ |
2805 | .todo_flags_finish: 0, /* todo_flags_finish */ |
2806 | }; |
2807 | |
2808 | class pass_duplicate_computed_gotos : public rtl_opt_pass |
2809 | { |
2810 | public: |
2811 | pass_duplicate_computed_gotos (gcc::context *ctxt) |
2812 | : rtl_opt_pass (pass_data_duplicate_computed_gotos, ctxt) |
2813 | {} |
2814 | |
2815 | /* opt_pass methods: */ |
2816 | bool gate (function *) final override; |
2817 | unsigned int execute (function *) final override; |
2818 | |
2819 | }; // class pass_duplicate_computed_gotos |
2820 | |
2821 | bool |
2822 | pass_duplicate_computed_gotos::gate (function *fun) |
2823 | { |
2824 | if (targetm.cannot_modify_jumps_p ()) |
2825 | return false; |
2826 | return (optimize > 0 |
2827 | && flag_expensive_optimizations |
2828 | && ! optimize_function_for_size_p (fun)); |
2829 | } |
2830 | |
2831 | unsigned int |
2832 | pass_duplicate_computed_gotos::execute (function *fun) |
2833 | { |
2834 | duplicate_computed_gotos (fun); |
2835 | |
2836 | return 0; |
2837 | } |
2838 | |
2839 | } // anon namespace |
2840 | |
2841 | rtl_opt_pass * |
2842 | make_pass_duplicate_computed_gotos (gcc::context *ctxt) |
2843 | { |
2844 | return new pass_duplicate_computed_gotos (ctxt); |
2845 | } |
2846 | |
2847 | /* This function is the main 'entrance' for the optimization that |
2848 | partitions hot and cold basic blocks into separate sections of the |
2849 | .o file (to improve performance and cache locality). Ideally it |
2850 | would be called after all optimizations that rearrange the CFG have |
2851 | been called. However part of this optimization may introduce new |
2852 | register usage, so it must be called before register allocation has |
2853 | occurred. This means that this optimization is actually called |
2854 | well before the optimization that reorders basic blocks (see |
2855 | function above). |
2856 | |
2857 | This optimization checks the feedback information to determine |
2858 | which basic blocks are hot/cold, updates flags on the basic blocks |
2859 | to indicate which section they belong in. This information is |
2860 | later used for writing out sections in the .o file. Because hot |
2861 | and cold sections can be arbitrarily large (within the bounds of |
2862 | memory), far beyond the size of a single function, it is necessary |
2863 | to fix up all edges that cross section boundaries, to make sure the |
2864 | instructions used can actually span the required distance. The |
2865 | fixes are described below. |
2866 | |
2867 | Fall-through edges must be changed into jumps; it is not safe or |
2868 | legal to fall through across a section boundary. Whenever a |
2869 | fall-through edge crossing a section boundary is encountered, a new |
2870 | basic block is inserted (in the same section as the fall-through |
2871 | source), and the fall through edge is redirected to the new basic |
2872 | block. The new basic block contains an unconditional jump to the |
2873 | original fall-through target. (If the unconditional jump is |
2874 | insufficient to cross section boundaries, that is dealt with a |
2875 | little later, see below). |
2876 | |
2877 | In order to deal with architectures that have short conditional |
2878 | branches (which cannot span all of memory) we take any conditional |
2879 | jump that attempts to cross a section boundary and add a level of |
2880 | indirection: it becomes a conditional jump to a new basic block, in |
2881 | the same section. The new basic block contains an unconditional |
2882 | jump to the original target, in the other section. |
2883 | |
2884 | For those architectures whose unconditional branch is also |
2885 | incapable of reaching all of memory, those unconditional jumps are |
2886 | converted into indirect jumps, through a register. |
2887 | |
2888 | IMPORTANT NOTE: This optimization causes some messy interactions |
2889 | with the cfg cleanup optimizations; those optimizations want to |
2890 | merge blocks wherever possible, and to collapse indirect jump |
2891 | sequences (change "A jumps to B jumps to C" directly into "A jumps |
2892 | to C"). Those optimizations can undo the jump fixes that |
2893 | partitioning is required to make (see above), in order to ensure |
2894 | that jumps attempting to cross section boundaries are really able |
2895 | to cover whatever distance the jump requires (on many architectures |
2896 | conditional or unconditional jumps are not able to reach all of |
2897 | memory). Therefore tests have to be inserted into each such |
2898 | optimization to make sure that it does not undo stuff necessary to |
2899 | cross partition boundaries. This would be much less of a problem |
2900 | if we could perform this optimization later in the compilation, but |
2901 | unfortunately the fact that we may need to create indirect jumps |
2902 | (through registers) requires that this optimization be performed |
2903 | before register allocation. |
2904 | |
2905 | Hot and cold basic blocks are partitioned and put in separate |
2906 | sections of the .o file, to reduce paging and improve cache |
2907 | performance (hopefully). This can result in bits of code from the |
2908 | same function being widely separated in the .o file. However this |
2909 | is not obvious to the current bb structure. Therefore we must take |
2910 | care to ensure that: 1). There are no fall_thru edges that cross |
2911 | between sections; 2). For those architectures which have "short" |
2912 | conditional branches, all conditional branches that attempt to |
2913 | cross between sections are converted to unconditional branches; |
2914 | and, 3). For those architectures which have "short" unconditional |
2915 | branches, all unconditional branches that attempt to cross between |
2916 | sections are converted to indirect jumps. |
2917 | |
2918 | The code for fixing up fall_thru edges that cross between hot and |
2919 | cold basic blocks does so by creating new basic blocks containing |
2920 | unconditional branches to the appropriate label in the "other" |
2921 | section. The new basic block is then put in the same (hot or cold) |
2922 | section as the original conditional branch, and the fall_thru edge |
2923 | is modified to fall into the new basic block instead. By adding |
2924 | this level of indirection we end up with only unconditional branches |
2925 | crossing between hot and cold sections. |
2926 | |
2927 | Conditional branches are dealt with by adding a level of indirection. |
2928 | A new basic block is added in the same (hot/cold) section as the |
2929 | conditional branch, and the conditional branch is retargeted to the |
2930 | new basic block. The new basic block contains an unconditional branch |
2931 | to the original target of the conditional branch (in the other section). |
2932 | |
2933 | Unconditional branches are dealt with by converting them into |
2934 | indirect jumps. */ |
2935 | |
2936 | namespace { |
2937 | |
2938 | const pass_data pass_data_partition_blocks = |
2939 | { |
2940 | .type: RTL_PASS, /* type */ |
2941 | .name: "bbpart" , /* name */ |
2942 | .optinfo_flags: OPTGROUP_NONE, /* optinfo_flags */ |
2943 | .tv_id: TV_REORDER_BLOCKS, /* tv_id */ |
2944 | PROP_cfglayout, /* properties_required */ |
2945 | .properties_provided: 0, /* properties_provided */ |
2946 | .properties_destroyed: 0, /* properties_destroyed */ |
2947 | .todo_flags_start: 0, /* todo_flags_start */ |
2948 | .todo_flags_finish: 0, /* todo_flags_finish */ |
2949 | }; |
2950 | |
2951 | class pass_partition_blocks : public rtl_opt_pass |
2952 | { |
2953 | public: |
2954 | pass_partition_blocks (gcc::context *ctxt) |
2955 | : rtl_opt_pass (pass_data_partition_blocks, ctxt) |
2956 | {} |
2957 | |
2958 | /* opt_pass methods: */ |
2959 | bool gate (function *) final override; |
2960 | unsigned int execute (function *) final override; |
2961 | |
2962 | }; // class pass_partition_blocks |
2963 | |
2964 | bool |
2965 | pass_partition_blocks::gate (function *fun) |
2966 | { |
2967 | /* The optimization to partition hot/cold basic blocks into separate |
2968 | sections of the .o file does not work well with linkonce or with |
2969 | user defined section attributes or with naked attribute. Don't call |
2970 | it if either case arises. */ |
2971 | return (flag_reorder_blocks_and_partition |
2972 | && optimize |
2973 | /* See pass_reorder_blocks::gate. We should not partition if |
2974 | we are going to omit the reordering. */ |
2975 | && optimize_function_for_speed_p (fun) |
2976 | && !DECL_COMDAT_GROUP (current_function_decl) |
2977 | && !lookup_attribute (attr_name: "section" , DECL_ATTRIBUTES (fun->decl)) |
2978 | && !lookup_attribute (attr_name: "naked" , DECL_ATTRIBUTES (fun->decl)) |
2979 | /* Workaround a bug in GDB where read_partial_die doesn't cope |
2980 | with DIEs with DW_AT_ranges, see PR81115. */ |
2981 | && !(in_lto_p && MAIN_NAME_P (DECL_NAME (fun->decl)))); |
2982 | } |
2983 | |
2984 | unsigned |
2985 | pass_partition_blocks::execute (function *fun) |
2986 | { |
2987 | vec<edge> crossing_edges; |
2988 | |
2989 | if (n_basic_blocks_for_fn (fun) <= NUM_FIXED_BLOCKS + 1) |
2990 | return 0; |
2991 | |
2992 | df_set_flags (DF_DEFER_INSN_RESCAN); |
2993 | |
2994 | crossing_edges = find_rarely_executed_basic_blocks_and_crossing_edges (); |
2995 | if (!crossing_edges.exists ()) |
2996 | /* Make sure to process deferred rescans and clear changeable df flags. */ |
2997 | return TODO_df_finish; |
2998 | |
2999 | crtl->has_bb_partition = true; |
3000 | |
3001 | /* Make sure the source of any crossing edge ends in a jump and the |
3002 | destination of any crossing edge has a label. */ |
3003 | add_labels_and_missing_jumps (crossing_edges); |
3004 | |
3005 | /* Convert all crossing fall_thru edges to non-crossing fall |
3006 | thrus to unconditional jumps (that jump to the original fall |
3007 | through dest). */ |
3008 | fix_up_fall_thru_edges (); |
3009 | |
3010 | /* If the architecture does not have conditional branches that can |
3011 | span all of memory, convert crossing conditional branches into |
3012 | crossing unconditional branches. */ |
3013 | if (!HAS_LONG_COND_BRANCH) |
3014 | fix_crossing_conditional_branches (); |
3015 | |
3016 | /* If the architecture does not have unconditional branches that |
3017 | can span all of memory, convert crossing unconditional branches |
3018 | into indirect jumps. Since adding an indirect jump also adds |
3019 | a new register usage, update the register usage information as |
3020 | well. */ |
3021 | if (!HAS_LONG_UNCOND_BRANCH) |
3022 | fix_crossing_unconditional_branches (); |
3023 | |
3024 | update_crossing_jump_flags (); |
3025 | |
3026 | /* Clear bb->aux fields that the above routines were using. */ |
3027 | clear_aux_for_blocks (); |
3028 | |
3029 | crossing_edges.release (); |
3030 | |
3031 | /* ??? FIXME: DF generates the bb info for a block immediately. |
3032 | And by immediately, I mean *during* creation of the block. |
3033 | |
3034 | #0 df_bb_refs_collect |
3035 | #1 in df_bb_refs_record |
3036 | #2 in create_basic_block_structure |
3037 | |
3038 | Which means that the bb_has_eh_pred test in df_bb_refs_collect |
3039 | will *always* fail, because no edges can have been added to the |
3040 | block yet. Which of course means we don't add the right |
3041 | artificial refs, which means we fail df_verify (much) later. |
3042 | |
3043 | Cleanest solution would seem to make DF_DEFER_INSN_RESCAN imply |
3044 | that we also shouldn't grab data from the new blocks those new |
3045 | insns are in either. In this way one can create the block, link |
3046 | it up properly, and have everything Just Work later, when deferred |
3047 | insns are processed. |
3048 | |
3049 | In the meantime, we have no other option but to throw away all |
3050 | of the DF data and recompute it all. */ |
3051 | if (fun->eh->lp_array) |
3052 | { |
3053 | df_finish_pass (true); |
3054 | df_scan_alloc (NULL); |
3055 | df_scan_blocks (); |
3056 | /* Not all post-landing pads use all of the EH_RETURN_DATA_REGNO |
3057 | data. We blindly generated all of them when creating the new |
3058 | landing pad. Delete those assignments we don't use. */ |
3059 | df_set_flags (DF_LR_RUN_DCE); |
3060 | df_analyze (); |
3061 | } |
3062 | |
3063 | /* Make sure to process deferred rescans and clear changeable df flags. */ |
3064 | return TODO_df_finish; |
3065 | } |
3066 | |
3067 | } // anon namespace |
3068 | |
3069 | rtl_opt_pass * |
3070 | make_pass_partition_blocks (gcc::context *ctxt) |
3071 | { |
3072 | return new pass_partition_blocks (ctxt); |
3073 | } |
3074 | |