]>
Commit | Line | Data |
---|---|---|
295ae817 | 1 | /* Basic block reordering routines for the GNU compiler. |
8d9254fc | 2 | Copyright (C) 2000-2020 Free Software Foundation, Inc. |
295ae817 | 3 | |
1322177d | 4 | This file is part of GCC. |
295ae817 | 5 | |
1322177d LB |
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 | |
9dcd6f09 | 8 | the Free Software Foundation; either version 3, or (at your option) |
295ae817 JE |
9 | any later version. |
10 | ||
1322177d LB |
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. | |
295ae817 JE |
15 | |
16 | You should have received a copy of the GNU General Public License | |
9dcd6f09 NC |
17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ | |
295ae817 | 19 | |
248c16c4 SB |
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 | ||
aa634f11 JZ |
32 | /* This (greedy) algorithm constructs traces in several rounds. |
33 | The construction starts from "seeds". The seed for the first round | |
4700dd70 EB |
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. | |
aa634f11 JZ |
38 | |
39 | There are two parameters: Branch Threshold and Exec Threshold. | |
4700dd70 | 40 | If the probability of an edge to a successor of the current basic block is |
79a6b3de | 41 | lower than Branch Threshold or its count is lower than Exec Threshold, |
4700dd70 | 42 | then the successor will be the seed in one of the next rounds. |
aa634f11 | 43 | Each round has these parameters lower than the previous one. |
4700dd70 EB |
44 | The last round has to have these parameters set to zero so that the |
45 | remaining blocks are picked up. | |
aa634f11 JZ |
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 | |
4700dd70 EB |
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. | |
aa634f11 | 57 | If the loop has few iterations and there is no edge from the last block of |
4700dd70 | 58 | the loop going out of the loop, the loop header is duplicated. |
aa634f11 | 59 | |
4700dd70 EB |
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. | |
aa634f11 | 62 | When there are still some unconnected traces it checks whether there exists |
4700dd70 EB |
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. | |
aa634f11 | 67 | The rest of traces are simply connected so there will be a jump to the |
4700dd70 | 68 | beginning of the rest of traces. |
aa634f11 | 69 | |
f7d0c571 ZC |
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. | |
79a6b3de | 78 | (2) Ignore the edge probability and count for fallthru edges. |
f7d0c571 ZC |
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. | |
aa634f11 JZ |
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 | ||
295ae817 JE |
91 | */ |
92 | ||
93 | #include "config.h" | |
94 | #include "system.h" | |
4977bab6 | 95 | #include "coretypes.h" |
c7131fb2 | 96 | #include "backend.h" |
957060b5 | 97 | #include "target.h" |
295ae817 | 98 | #include "rtl.h" |
957060b5 AM |
99 | #include "tree.h" |
100 | #include "cfghooks.h" | |
c7131fb2 | 101 | #include "df.h" |
4d0cdd0c | 102 | #include "memmodel.h" |
957060b5 | 103 | #include "optabs.h" |
7932a3db | 104 | #include "regs.h" |
957060b5 | 105 | #include "emit-rtl.h" |
295ae817 | 106 | #include "output.h" |
750054a2 | 107 | #include "expr.h" |
ef330312 | 108 | #include "tree-pass.h" |
60393bbc AM |
109 | #include "cfgrtl.h" |
110 | #include "cfganal.h" | |
111 | #include "cfgbuild.h" | |
112 | #include "cfgcleanup.h" | |
76ee381a | 113 | #include "bb-reorder.h" |
0be7e7a6 | 114 | #include "except.h" |
516fd7ce | 115 | #include "alloc-pool.h" |
8d261514 | 116 | #include "fibonacci_heap.h" |
314e6352 ML |
117 | #include "stringpool.h" |
118 | #include "attribs.h" | |
40305631 | 119 | #include "common/common-target.h" |
ef330312 | 120 | |
750054a2 CT |
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 | |
4700dd70 | 123 | the object file there will be an extra round. */ |
750054a2 | 124 | #define N_ROUNDS 5 |
aa634f11 | 125 | |
76ee381a RS |
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 | ||
aa634f11 | 134 | /* Branch thresholds in thousandths (per mille) of the REG_BR_PROB_BASE. */ |
cfb00b41 | 135 | static const int branch_threshold[N_ROUNDS] = {400, 200, 100, 0, 0}; |
aa634f11 | 136 | |
79a6b3de | 137 | /* Exec thresholds in thousandths (per mille) of the count of bb 0. */ |
cfb00b41 | 138 | static const int exec_threshold[N_ROUNDS] = {500, 200, 50, 0, 0}; |
aa634f11 | 139 | |
79a6b3de | 140 | /* If edge count is lower than DUPLICATION_THRESHOLD per mille of entry |
aa634f11 JZ |
141 | block the edge destination is not duplicated while connecting traces. */ |
142 | #define DUPLICATION_THRESHOLD 100 | |
143 | ||
8d261514 ML |
144 | typedef fibonacci_heap <long, basic_block_def> bb_heap_t; |
145 | typedef fibonacci_node <long, basic_block_def> bb_heap_node_t; | |
146 | ||
aa634f11 | 147 | /* Structure to hold needed information for each basic block. */ |
a79683d5 | 148 | struct bbro_basic_block_data |
aa634f11 | 149 | { |
4700dd70 | 150 | /* Which trace is the bb start of (-1 means it is not a start of any). */ |
aa634f11 JZ |
151 | int start_of_trace; |
152 | ||
4700dd70 | 153 | /* Which trace is the bb end of (-1 means it is not an end of any). */ |
aa634f11 JZ |
154 | int end_of_trace; |
155 | ||
87c8b4be CT |
156 | /* Which trace is the bb in? */ |
157 | int in_trace; | |
158 | ||
bcc708fc MM |
159 | /* Which trace was this bb visited in? */ |
160 | int visited; | |
161 | ||
375374ad RB |
162 | /* Cached maximum frequency of interesting incoming edges. |
163 | Minus one means not yet computed. */ | |
164 | int priority; | |
165 | ||
aa634f11 | 166 | /* Which heap is BB in (if any)? */ |
8d261514 | 167 | bb_heap_t *heap; |
aa634f11 JZ |
168 | |
169 | /* Which heap node is BB in (if any)? */ | |
8d261514 | 170 | bb_heap_node_t *node; |
a79683d5 | 171 | }; |
aa634f11 JZ |
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. */ | |
298e6adc | 184 | #define FREE(P) (gcc_assert (P), free (P), P = 0) |
aa634f11 JZ |
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 | ||
79a6b3de | 199 | /* Maximum count of one of the entry blocks. */ |
3995f3a2 | 200 | static profile_count max_entry_count; |
aa634f11 | 201 | |
295ae817 | 202 | /* Local function prototypes. */ |
370df3ce | 203 | static void find_traces_1_round (int, profile_count, struct trace *, int *, |
8d261514 | 204 | int, bb_heap_t **, int); |
4682ae04 | 205 | static basic_block copy_bb (basic_block, edge, basic_block, int); |
8d261514 | 206 | static long bb_to_key (basic_block); |
357067f2 | 207 | static bool better_edge_p (const_basic_block, const_edge, profile_probability, |
79a6b3de JH |
208 | profile_count, profile_probability, profile_count, |
209 | const_edge); | |
9678086d | 210 | static bool copy_bb_p (const_basic_block, int); |
f008a564 | 211 | \f |
bcc708fc MM |
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 | { | |
8d261514 | 229 | bbd[bb->index].heap->delete_node (bbd[bb->index].node); |
bcc708fc MM |
230 | bbd[bb->index].heap = NULL; |
231 | bbd[bb->index].node = NULL; | |
232 | } | |
233 | } | |
234 | ||
750054a2 CT |
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 | |
ed7a4b4b | 244 | push_to_next_round_p (const_basic_block bb, int round, int number_of_rounds, |
370df3ce | 245 | profile_count count_th) |
750054a2 CT |
246 | { |
247 | bool there_exists_another_round; | |
750054a2 CT |
248 | bool block_not_hot_enough; |
249 | ||
250 | there_exists_another_round = round < number_of_rounds - 1; | |
750054a2 | 251 | |
370df3ce | 252 | block_not_hot_enough = (bb->count < count_th |
2eb712b4 | 253 | || probably_never_executed_bb_p (cfun, bb)); |
750054a2 | 254 | |
87c8b4be CT |
255 | if (there_exists_another_round |
256 | && block_not_hot_enough) | |
750054a2 | 257 | return true; |
c22cacf3 | 258 | else |
750054a2 CT |
259 | return false; |
260 | } | |
261 | ||
aa634f11 JZ |
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. */ | |
295ae817 | 265 | |
f008a564 | 266 | static void |
4682ae04 | 267 | find_traces (int *n_traces, struct trace *traces) |
295ae817 | 268 | { |
aa634f11 | 269 | int i; |
750054a2 | 270 | int number_of_rounds; |
aa634f11 | 271 | edge e; |
628f6a4e | 272 | edge_iterator ei; |
8d261514 | 273 | bb_heap_t *heap = new bb_heap_t (LONG_MIN); |
aa634f11 | 274 | |
750054a2 CT |
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; | |
750054a2 | 280 | |
aa634f11 | 281 | /* Insert entry points of function into heap. */ |
3995f3a2 | 282 | max_entry_count = profile_count::zero (); |
fefa31b5 | 283 | FOR_EACH_EDGE (e, ei, ENTRY_BLOCK_PTR_FOR_FN (cfun)->succs) |
aa634f11 JZ |
284 | { |
285 | bbd[e->dest->index].heap = heap; | |
8d261514 | 286 | bbd[e->dest->index].node = heap->insert (bb_to_key (e->dest), e->dest); |
370df3ce | 287 | if (e->dest->count > max_entry_count) |
aa634f11 JZ |
288 | max_entry_count = e->dest->count; |
289 | } | |
290 | ||
291 | /* Find the traces. */ | |
750054a2 | 292 | for (i = 0; i < number_of_rounds; i++) |
aa634f11 | 293 | { |
370df3ce | 294 | profile_count count_threshold; |
f008a564 | 295 | |
c263766c RH |
296 | if (dump_file) |
297 | fprintf (dump_file, "STC - round %d\n", i + 1); | |
aa634f11 | 298 | |
370df3ce | 299 | count_threshold = max_entry_count.apply_scale (exec_threshold[i], 1000); |
aa634f11 JZ |
300 | |
301 | find_traces_1_round (REG_BR_PROB_BASE * branch_threshold[i] / 1000, | |
750054a2 CT |
302 | count_threshold, traces, n_traces, i, &heap, |
303 | number_of_rounds); | |
aa634f11 | 304 | } |
8d261514 | 305 | delete heap; |
aa634f11 | 306 | |
c263766c | 307 | if (dump_file) |
aa634f11 JZ |
308 | { |
309 | for (i = 0; i < *n_traces; i++) | |
310 | { | |
311 | basic_block bb; | |
c263766c | 312 | fprintf (dump_file, "Trace %d (round %d): ", i + 1, |
aa634f11 | 313 | traces[i].round + 1); |
4700dd70 EB |
314 | for (bb = traces[i].first; |
315 | bb != traces[i].last; | |
316 | bb = (basic_block) bb->aux) | |
79a6b3de JH |
317 | { |
318 | fprintf (dump_file, "%d [", bb->index); | |
319 | bb->count.dump (dump_file); | |
320 | fprintf (dump_file, "] "); | |
321 | } | |
322 | fprintf (dump_file, "%d [", bb->index); | |
323 | bb->count.dump (dump_file); | |
324 | fprintf (dump_file, "]\n"); | |
aa634f11 | 325 | } |
c263766c | 326 | fflush (dump_file); |
aa634f11 JZ |
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 | |
4682ae04 | 334 | rotate_loop (edge back_edge, struct trace *trace, int trace_n) |
aa634f11 JZ |
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; | |
3995f3a2 | 341 | profile_count best_count = profile_count::uninitialized (); |
aa634f11 JZ |
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; | |
f008a564 RH |
348 | do |
349 | { | |
aa634f11 | 350 | edge e; |
628f6a4e BE |
351 | edge_iterator ei; |
352 | ||
353 | FOR_EACH_EDGE (e, ei, bb->succs) | |
fefa31b5 | 354 | if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
bcc708fc | 355 | && bb_visited_trace (e->dest) != trace_n |
aa634f11 JZ |
356 | && (e->flags & EDGE_CAN_FALLTHRU) |
357 | && !(e->flags & EDGE_COMPLEX)) | |
358 | { | |
359 | if (is_preferred) | |
360 | { | |
361 | /* The best edge is preferred. */ | |
bcc708fc | 362 | if (!bb_visited_trace (e->dest) |
aa634f11 JZ |
363 | || bbd[e->dest->index].start_of_trace >= 0) |
364 | { | |
365 | /* The current edge E is also preferred. */ | |
370df3ce | 366 | if (e->count () > best_count) |
aa634f11 | 367 | { |
370df3ce | 368 | best_count = e->count (); |
aa634f11 JZ |
369 | best_edge = e; |
370 | best_bb = bb; | |
371 | } | |
372 | } | |
373 | } | |
374 | else | |
375 | { | |
bcc708fc | 376 | if (!bb_visited_trace (e->dest) |
aa634f11 JZ |
377 | || bbd[e->dest->index].start_of_trace >= 0) |
378 | { | |
379 | /* The current edge E is preferred. */ | |
380 | is_preferred = true; | |
ef30ab83 | 381 | best_count = e->count (); |
aa634f11 JZ |
382 | best_edge = e; |
383 | best_bb = bb; | |
384 | } | |
385 | else | |
386 | { | |
370df3ce | 387 | if (!best_edge || e->count () > best_count) |
aa634f11 | 388 | { |
ef30ab83 | 389 | best_count = e->count (); |
aa634f11 JZ |
390 | best_edge = e; |
391 | best_bb = bb; | |
392 | } | |
393 | } | |
394 | } | |
395 | } | |
f883e0a7 | 396 | bb = (basic_block) bb->aux; |
aa634f11 JZ |
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 | { | |
f883e0a7 | 406 | trace->first = (basic_block) best_bb->aux; |
aa634f11 JZ |
407 | } |
408 | else | |
409 | { | |
410 | basic_block prev_bb; | |
ca7fd9cd | 411 | |
aa634f11 | 412 | for (prev_bb = trace->first; |
370369e1 | 413 | prev_bb->aux != back_edge->dest; |
f883e0a7 | 414 | prev_bb = (basic_block) prev_bb->aux) |
aa634f11 | 415 | ; |
370369e1 | 416 | prev_bb->aux = best_bb->aux; |
aa634f11 JZ |
417 | |
418 | /* Try to get rid of uncond jump to cond jump. */ | |
c5cbcccf | 419 | if (single_succ_p (prev_bb)) |
aa634f11 | 420 | { |
c5cbcccf | 421 | basic_block header = single_succ (prev_bb); |
aa634f11 JZ |
422 | |
423 | /* Duplicate HEADER if it is a small block containing cond jump | |
424 | in the end. */ | |
9fb32434 | 425 | if (any_condjump_p (BB_END (header)) && copy_bb_p (header, 0) |
339ba33b | 426 | && !CROSSING_JUMP_P (BB_END (header))) |
c5cbcccf | 427 | copy_bb (header, single_succ_edge (prev_bb), prev_bb, trace_n); |
aa634f11 JZ |
428 | } |
429 | } | |
430 | } | |
431 | else | |
432 | { | |
433 | /* We have not found suitable loop tail so do no rotation. */ | |
434 | best_bb = back_edge->src; | |
295ae817 | 435 | } |
370369e1 | 436 | best_bb->aux = NULL; |
aa634f11 | 437 | return best_bb; |
f008a564 | 438 | } |
295ae817 | 439 | |
4700dd70 EB |
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 | |
79a6b3de | 442 | count is lower than EXEC_TH into traces (or whose count is lower than |
4700dd70 EB |
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. */ | |
aa634f11 JZ |
447 | |
448 | static void | |
370df3ce | 449 | find_traces_1_round (int branch_th, profile_count count_th, |
4682ae04 | 450 | struct trace *traces, int *n_traces, int round, |
8d261514 | 451 | bb_heap_t **heap, int number_of_rounds) |
aa634f11 JZ |
452 | { |
453 | /* Heap for discarded basic blocks which are possible starting points for | |
454 | the next round. */ | |
8d261514 | 455 | bb_heap_t *new_heap = new bb_heap_t (LONG_MIN); |
f7d0c571 | 456 | bool for_size = optimize_function_for_size_p (cfun); |
aa634f11 | 457 | |
8d261514 | 458 | while (!(*heap)->empty ()) |
aa634f11 JZ |
459 | { |
460 | basic_block bb; | |
461 | struct trace *trace; | |
462 | edge best_edge, e; | |
8d261514 | 463 | long key; |
628f6a4e | 464 | edge_iterator ei; |
aa634f11 | 465 | |
8d261514 | 466 | bb = (*heap)->extract_min (); |
aa634f11 JZ |
467 | bbd[bb->index].heap = NULL; |
468 | bbd[bb->index].node = NULL; | |
469 | ||
c263766c RH |
470 | if (dump_file) |
471 | fprintf (dump_file, "Getting bb %d\n", bb->index); | |
aa634f11 | 472 | |
79a6b3de | 473 | /* If the BB's count is too low, send BB to the next round. When |
c22cacf3 MS |
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 | |
f7d0c571 | 476 | round. When optimizing for size, do not push to next round. */ |
750054a2 | 477 | |
f7d0c571 | 478 | if (!for_size |
370df3ce | 479 | && push_to_next_round_p (bb, round, number_of_rounds, |
f7d0c571 | 480 | count_th)) |
aa634f11 JZ |
481 | { |
482 | int key = bb_to_key (bb); | |
483 | bbd[bb->index].heap = new_heap; | |
8d261514 | 484 | bbd[bb->index].node = new_heap->insert (key, bb); |
aa634f11 | 485 | |
c263766c RH |
486 | if (dump_file) |
487 | fprintf (dump_file, | |
aa634f11 JZ |
488 | " 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; | |
87c8b4be | 497 | bbd[bb->index].in_trace = *n_traces; |
aa634f11 JZ |
498 | (*n_traces)++; |
499 | ||
500 | do | |
501 | { | |
934677f9 | 502 | bool ends_in_call; |
aa634f11 | 503 | |
79a6b3de | 504 | /* The probability and count of the best edge. */ |
357067f2 | 505 | profile_probability best_prob = profile_probability::uninitialized (); |
79a6b3de | 506 | profile_count best_count = profile_count::uninitialized (); |
aa634f11 JZ |
507 | |
508 | best_edge = NULL; | |
509 | mark_bb_visited (bb, *n_traces); | |
510 | trace->length++; | |
511 | ||
c263766c RH |
512 | if (dump_file) |
513 | fprintf (dump_file, "Basic block %d was visited in trace %d\n", | |
eada55b9 | 514 | bb->index, *n_traces); |
aa634f11 | 515 | |
c22cacf3 | 516 | ends_in_call = block_ends_with_call_p (bb); |
934677f9 | 517 | |
aa634f11 | 518 | /* Select the successor that will be placed after BB. */ |
628f6a4e | 519 | FOR_EACH_EDGE (e, ei, bb->succs) |
aa634f11 | 520 | { |
298e6adc | 521 | gcc_assert (!(e->flags & EDGE_FAKE)); |
aa634f11 | 522 | |
fefa31b5 | 523 | if (e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
aa634f11 JZ |
524 | continue; |
525 | ||
bcc708fc MM |
526 | if (bb_visited_trace (e->dest) |
527 | && bb_visited_trace (e->dest) != *n_traces) | |
aa634f11 JZ |
528 | continue; |
529 | ||
eada55b9 EB |
530 | /* If partitioning hot/cold basic blocks, don't consider edges |
531 | that cross section boundaries. */ | |
87c8b4be | 532 | if (BB_PARTITION (e->dest) != BB_PARTITION (bb)) |
750054a2 CT |
533 | continue; |
534 | ||
79a6b3de JH |
535 | profile_probability prob = e->probability; |
536 | profile_count count = e->dest->count; | |
aa634f11 | 537 | |
934677f9 RH |
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; | |
79a6b3de | 546 | best_count = count; |
934677f9 RH |
547 | } |
548 | continue; | |
549 | } | |
550 | ||
aa634f11 | 551 | /* Edge that cannot be fallthru or improbable or infrequent |
f7d0c571 | 552 | successor (i.e. it is unsuitable successor). When optimizing |
79a6b3de | 553 | for size, ignore the probability and count. */ |
aa634f11 | 554 | if (!(e->flags & EDGE_CAN_FALLTHRU) || (e->flags & EDGE_COMPLEX) |
357067f2 JH |
555 | || !prob.initialized_p () |
556 | || ((prob.to_reg_br_prob_base () < branch_th | |
370df3ce | 557 | || e->count () < count_th) && (!for_size))) |
aa634f11 JZ |
558 | continue; |
559 | ||
79a6b3de | 560 | if (better_edge_p (bb, e, prob, count, best_prob, best_count, |
750054a2 | 561 | best_edge)) |
aa634f11 JZ |
562 | { |
563 | best_edge = e; | |
564 | best_prob = prob; | |
79a6b3de | 565 | best_count = count; |
aa634f11 JZ |
566 | } |
567 | } | |
568 | ||
eada55b9 EB |
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 | |
6d9cc15b | 577 | && copy_bb_p (best_edge->dest, 0)) |
eada55b9 EB |
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 | } | |
6d9cc15b | 591 | |
f7d0c571 ZC |
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 | |
9c3da8cc | 595 | destination. It breaks long trace and reduces long jumps. |
f7d0c571 ZC |
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 | |
688010ba | 610 | reduces the possibility of long jumps. */ |
f7d0c571 ZC |
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 | ||
aa634f11 | 616 | /* Add all non-selected successors to the heaps. */ |
628f6a4e | 617 | FOR_EACH_EDGE (e, ei, bb->succs) |
aa634f11 JZ |
618 | { |
619 | if (e == best_edge | |
fefa31b5 | 620 | || e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun) |
bcc708fc | 621 | || bb_visited_trace (e->dest)) |
aa634f11 JZ |
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. */ | |
8d261514 | 629 | if (key != bbd[e->dest->index].node->get_key ()) |
aa634f11 | 630 | { |
c263766c | 631 | if (dump_file) |
aa634f11 | 632 | { |
c263766c | 633 | fprintf (dump_file, |
aa634f11 JZ |
634 | "Changing key for bb %d from %ld to %ld.\n", |
635 | e->dest->index, | |
8d261514 | 636 | (long) bbd[e->dest->index].node->get_key (), |
aa634f11 JZ |
637 | key); |
638 | } | |
a3dc1a45 | 639 | bbd[e->dest->index].heap->replace_key |
8d261514 | 640 | (bbd[e->dest->index].node, key); |
aa634f11 JZ |
641 | } |
642 | } | |
643 | else | |
644 | { | |
8d261514 | 645 | bb_heap_t *which_heap = *heap; |
aa634f11 | 646 | |
79a6b3de | 647 | profile_probability prob = e->probability; |
aa634f11 JZ |
648 | |
649 | if (!(e->flags & EDGE_CAN_FALLTHRU) | |
650 | || (e->flags & EDGE_COMPLEX) | |
357067f2 JH |
651 | || !prob.initialized_p () |
652 | || prob.to_reg_br_prob_base () < branch_th | |
370df3ce | 653 | || e->count () < count_th) |
aa634f11 | 654 | { |
750054a2 CT |
655 | /* When partitioning hot/cold basic blocks, make sure |
656 | the cold blocks (and only the cold blocks) all get | |
f7d0c571 ZC |
657 | pushed to the last round of trace collection. When |
658 | optimizing for size, do not push to next round. */ | |
750054a2 | 659 | |
f7d0c571 ZC |
660 | if (!for_size && push_to_next_round_p (e->dest, round, |
661 | number_of_rounds, | |
370df3ce | 662 | count_th)) |
aa634f11 JZ |
663 | which_heap = new_heap; |
664 | } | |
295ae817 | 665 | |
aa634f11 | 666 | bbd[e->dest->index].heap = which_heap; |
8d261514 | 667 | bbd[e->dest->index].node = which_heap->insert (key, e->dest); |
295ae817 | 668 | |
c263766c | 669 | if (dump_file) |
aa634f11 | 670 | { |
c263766c | 671 | fprintf (dump_file, |
aa634f11 JZ |
672 | " 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 | { | |
bcc708fc | 682 | if (bb_visited_trace (best_edge->dest) == *n_traces) |
aa634f11 JZ |
683 | { |
684 | /* We do nothing with one basic block loops. */ | |
685 | if (best_edge->dest != bb) | |
686 | { | |
370df3ce JH |
687 | if (best_edge->count () |
688 | > best_edge->dest->count.apply_scale (4, 5)) | |
aa634f11 JZ |
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 | ||
fefa31b5 DM |
694 | if (best_edge->dest |
695 | != ENTRY_BLOCK_PTR_FOR_FN (cfun)->next_bb) | |
aa634f11 | 696 | { |
c263766c | 697 | if (dump_file) |
aa634f11 | 698 | { |
c263766c | 699 | fprintf (dump_file, |
aa634f11 JZ |
700 | "Rotating loop %d - %d\n", |
701 | best_edge->dest->index, bb->index); | |
702 | } | |
370369e1 | 703 | bb->aux = best_edge->dest; |
c22cacf3 MS |
704 | bbd[best_edge->dest->index].in_trace = |
705 | (*n_traces) - 1; | |
aa634f11 JZ |
706 | bb = rotate_loop (best_edge, trace, *n_traces); |
707 | } | |
708 | } | |
709 | else | |
710 | { | |
711 | /* The loop has less than 4 iterations. */ | |
712 | ||
c5cbcccf | 713 | if (single_succ_p (bb) |
efd8f750 | 714 | && copy_bb_p (best_edge->dest, |
4700dd70 EB |
715 | optimize_edge_for_speed_p |
716 | (best_edge))) | |
aa634f11 JZ |
717 | { |
718 | bb = copy_bb (best_edge->dest, best_edge, bb, | |
719 | *n_traces); | |
87c8b4be | 720 | trace->length++; |
aa634f11 JZ |
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 | |
370df3ce JH |
739 | AB->count () + BC->count () >= AC->count (). |
740 | (i.e. 2 * B->count >= AC->count ) | |
aa634f11 JZ |
741 | Best ordering is then A B C. |
742 | ||
f7d0c571 ZC |
743 | When optimizing for size, A B C is always the best order. |
744 | ||
aa634f11 JZ |
745 | This situation is created for example by: |
746 | ||
747 | if (A) B; | |
748 | C; | |
749 | ||
750 | */ | |
751 | ||
628f6a4e | 752 | FOR_EACH_EDGE (e, ei, bb->succs) |
aa634f11 JZ |
753 | if (e != best_edge |
754 | && (e->flags & EDGE_CAN_FALLTHRU) | |
755 | && !(e->flags & EDGE_COMPLEX) | |
bcc708fc | 756 | && !bb_visited_trace (e->dest) |
c5cbcccf | 757 | && single_pred_p (e->dest) |
bd454efd | 758 | && !(e->flags & EDGE_CROSSING) |
87c8b4be | 759 | && single_succ_p (e->dest) |
c5cbcccf ZD |
760 | && (single_succ_edge (e->dest)->flags |
761 | & EDGE_CAN_FALLTHRU) | |
762 | && !(single_succ_edge (e->dest)->flags & EDGE_COMPLEX) | |
763 | && single_succ (e->dest) == best_edge->dest | |
370df3ce JH |
764 | && (e->dest->count.apply_scale (2, 1) |
765 | >= best_edge->count () || for_size)) | |
aa634f11 JZ |
766 | { |
767 | best_edge = e; | |
c263766c RH |
768 | if (dump_file) |
769 | fprintf (dump_file, "Selecting BB %d\n", | |
aa634f11 JZ |
770 | best_edge->dest->index); |
771 | break; | |
772 | } | |
773 | ||
370369e1 | 774 | bb->aux = best_edge->dest; |
87c8b4be | 775 | bbd[best_edge->dest->index].in_trace = (*n_traces) - 1; |
aa634f11 JZ |
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; | |
375374ad RB |
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 | } | |
aa634f11 JZ |
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). */ | |
628f6a4e | 796 | FOR_EACH_EDGE (e, ei, bb->succs) |
aa634f11 | 797 | { |
fefa31b5 | 798 | if (e->dest == EXIT_BLOCK_PTR_FOR_FN (cfun) |
bcc708fc | 799 | || bb_visited_trace (e->dest)) |
aa634f11 JZ |
800 | continue; |
801 | ||
802 | if (bbd[e->dest->index].heap) | |
803 | { | |
804 | key = bb_to_key (e->dest); | |
8d261514 | 805 | if (key != bbd[e->dest->index].node->get_key ()) |
aa634f11 | 806 | { |
c263766c | 807 | if (dump_file) |
aa634f11 | 808 | { |
c263766c | 809 | fprintf (dump_file, |
aa634f11 JZ |
810 | "Changing key for bb %d from %ld to %ld.\n", |
811 | e->dest->index, | |
8d261514 | 812 | (long) bbd[e->dest->index].node->get_key (), key); |
aa634f11 | 813 | } |
a3dc1a45 | 814 | bbd[e->dest->index].heap->replace_key |
8d261514 | 815 | (bbd[e->dest->index].node, key); |
aa634f11 JZ |
816 | } |
817 | } | |
818 | } | |
819 | } | |
820 | ||
8d261514 | 821 | delete (*heap); |
aa634f11 JZ |
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). */ | |
295ae817 | 830 | |
f008a564 | 831 | static basic_block |
4682ae04 | 832 | copy_bb (basic_block old_bb, edge e, basic_block bb, int trace) |
f008a564 | 833 | { |
aa634f11 JZ |
834 | basic_block new_bb; |
835 | ||
b9a66240 | 836 | new_bb = duplicate_block (old_bb, e, bb); |
076c7ab8 | 837 | BB_COPY_PARTITION (new_bb, old_bb); |
9fb32434 | 838 | |
298e6adc | 839 | gcc_assert (e->dest == new_bb); |
298e6adc | 840 | |
c263766c RH |
841 | if (dump_file) |
842 | fprintf (dump_file, | |
aa634f11 JZ |
843 | "Duplicated bb %d (created bb %d)\n", |
844 | old_bb->index, new_bb->index); | |
f008a564 | 845 | |
8b1c6fd7 DM |
846 | if (new_bb->index >= array_size |
847 | || last_basic_block_for_fn (cfun) > array_size) | |
295ae817 | 848 | { |
aa634f11 JZ |
849 | int i; |
850 | int new_size; | |
851 | ||
8b1c6fd7 | 852 | new_size = MAX (last_basic_block_for_fn (cfun), new_bb->index + 1); |
aa634f11 | 853 | new_size = GET_ARRAY_SIZE (new_size); |
f883e0a7 | 854 | bbd = XRESIZEVEC (bbro_basic_block_data, bbd, new_size); |
aa634f11 JZ |
855 | for (i = array_size; i < new_size; i++) |
856 | { | |
857 | bbd[i].start_of_trace = -1; | |
858 | bbd[i].end_of_trace = -1; | |
bcc708fc MM |
859 | bbd[i].in_trace = -1; |
860 | bbd[i].visited = 0; | |
375374ad | 861 | bbd[i].priority = -1; |
aa634f11 JZ |
862 | bbd[i].heap = NULL; |
863 | bbd[i].node = NULL; | |
864 | } | |
865 | array_size = new_size; | |
295ae817 | 866 | |
c263766c | 867 | if (dump_file) |
aa634f11 | 868 | { |
c263766c | 869 | fprintf (dump_file, |
aa634f11 JZ |
870 | "Growing the dynamic array to %d elements.\n", |
871 | array_size); | |
872 | } | |
295ae817 | 873 | } |
aa634f11 | 874 | |
bcc708fc MM |
875 | gcc_assert (!bb_visited_trace (e->dest)); |
876 | mark_bb_visited (new_bb, trace); | |
877 | new_bb->aux = bb->aux; | |
878 | bb->aux = new_bb; | |
879 | ||
87c8b4be CT |
880 | bbd[new_bb->index].in_trace = trace; |
881 | ||
aa634f11 JZ |
882 | return new_bb; |
883 | } | |
884 | ||
885 | /* Compute and return the key (for the heap) of the basic block BB. */ | |
886 | ||
8d261514 | 887 | static long |
4682ae04 | 888 | bb_to_key (basic_block bb) |
aa634f11 JZ |
889 | { |
890 | edge e; | |
628f6a4e | 891 | edge_iterator ei; |
aa634f11 | 892 | |
f7d0c571 ZC |
893 | /* Use index as key to align with its original order. */ |
894 | if (optimize_function_for_size_p (cfun)) | |
895 | return bb->index; | |
896 | ||
aa634f11 | 897 | /* Do not start in probably never executed blocks. */ |
750054a2 | 898 | |
076c7ab8 | 899 | if (BB_PARTITION (bb) == BB_COLD_PARTITION |
2eb712b4 | 900 | || probably_never_executed_bb_p (cfun, bb)) |
aa634f11 JZ |
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. */ | |
375374ad RB |
905 | int priority = bbd[bb->index].priority; |
906 | if (priority == -1) | |
402209ff | 907 | { |
375374ad RB |
908 | priority = 0; |
909 | FOR_EACH_EDGE (e, ei, bb->preds) | |
aa634f11 | 910 | { |
375374ad RB |
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); | |
aa634f11 | 916 | |
375374ad RB |
917 | if (edge_freq > priority) |
918 | priority = edge_freq; | |
919 | } | |
aa634f11 | 920 | } |
375374ad | 921 | bbd[bb->index].priority = priority; |
402209ff | 922 | } |
295ae817 | 923 | |
aa634f11 JZ |
924 | if (priority) |
925 | /* The block with priority should have significantly lower key. */ | |
e7a74006 | 926 | return -(100 * BB_FREQ_MAX + 100 * priority + bb->count.to_frequency (cfun)); |
4700dd70 | 927 | |
e7a74006 | 928 | return -bb->count.to_frequency (cfun); |
aa634f11 JZ |
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 | |
79a6b3de JH |
933 | count of the successor is COUNT. The current best probability is |
934 | BEST_PROB, the best count is BEST_COUNT. | |
aa634f11 | 935 | The edge is considered to be equivalent when PROB does not differ much from |
79a6b3de | 936 | BEST_PROB; similarly for count. */ |
aa634f11 JZ |
937 | |
938 | static bool | |
357067f2 | 939 | better_edge_p (const_basic_block bb, const_edge e, profile_probability prob, |
79a6b3de JH |
940 | profile_count count, profile_probability best_prob, |
941 | profile_count best_count, const_edge cur_best_edge) | |
aa634f11 JZ |
942 | { |
943 | bool is_better_edge; | |
f008a564 | 944 | |
aa634f11 JZ |
945 | /* The BEST_* values do not have to be best, but can be a bit smaller than |
946 | maximum values. */ | |
357067f2 | 947 | profile_probability diff_prob = best_prob.apply_scale (1, 10); |
f008a564 | 948 | |
f7d0c571 ZC |
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 | ||
b0a12b5e JH |
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 ())) | |
aa634f11 JZ |
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; | |
aa634f11 | 967 | else |
79a6b3de JH |
968 | { |
969 | profile_count diff_count = best_count.apply_scale (1, 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 | } | |
aa634f11 JZ |
988 | |
989 | return is_better_edge; | |
990 | } | |
991 | ||
f7d0c571 ZC |
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 | ||
79a6b3de | 1025 | /* We are looking for predecessor, so probabilities are not that |
9c3da8cc JJ |
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 | |
79a6b3de JH |
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; | |
02e18a0e | 1035 | else if (e->probability > cur_best_edge->probability) |
f7d0c571 ZC |
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 | ||
aa634f11 JZ |
1069 | /* Connect traces in array TRACES, N_TRACES is the count of traces. */ |
1070 | ||
1071 | static void | |
4682ae04 | 1072 | connect_traces (int n_traces, struct trace *traces) |
aa634f11 JZ |
1073 | { |
1074 | int i; | |
1075 | bool *connected; | |
87c8b4be | 1076 | bool two_passes; |
aa634f11 | 1077 | int last_trace; |
87c8b4be CT |
1078 | int current_pass; |
1079 | int current_partition; | |
370df3ce | 1080 | profile_count count_threshold; |
f7d0c571 | 1081 | bool for_size = optimize_function_for_size_p (cfun); |
aa634f11 | 1082 | |
370df3ce | 1083 | count_threshold = max_entry_count.apply_scale (DUPLICATION_THRESHOLD, 1000); |
aa634f11 | 1084 | |
5ed6ace5 | 1085 | connected = XCNEWVEC (bool, n_traces); |
aa634f11 | 1086 | last_trace = -1; |
87c8b4be CT |
1087 | current_pass = 1; |
1088 | current_partition = BB_PARTITION (traces[0].first); | |
1089 | two_passes = false; | |
750054a2 | 1090 | |
af205f67 | 1091 | if (crtl->has_bb_partition) |
87c8b4be | 1092 | for (i = 0; i < n_traces && !two_passes; i++) |
c22cacf3 | 1093 | if (BB_PARTITION (traces[0].first) |
87c8b4be CT |
1094 | != BB_PARTITION (traces[i].first)) |
1095 | two_passes = true; | |
1096 | ||
1097 | for (i = 0; i < n_traces || (two_passes && current_pass == 1) ; i++) | |
295ae817 | 1098 | { |
aa634f11 JZ |
1099 | int t = i; |
1100 | int t2; | |
1101 | edge e, best; | |
1102 | int best_len; | |
295ae817 | 1103 | |
87c8b4be | 1104 | if (i >= n_traces) |
750054a2 | 1105 | { |
41806d92 NS |
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; | |
87c8b4be | 1112 | else |
41806d92 | 1113 | current_partition = BB_HOT_PARTITION; |
750054a2 | 1114 | } |
c22cacf3 | 1115 | |
aa634f11 JZ |
1116 | if (connected[t]) |
1117 | continue; | |
f008a564 | 1118 | |
c22cacf3 | 1119 | if (two_passes |
87c8b4be CT |
1120 | && BB_PARTITION (traces[t].first) != current_partition) |
1121 | continue; | |
1122 | ||
aa634f11 | 1123 | connected[t] = true; |
b0cc7919 | 1124 | |
aa634f11 JZ |
1125 | /* Find the predecessor traces. */ |
1126 | for (t2 = t; t2 > 0;) | |
1127 | { | |
628f6a4e | 1128 | edge_iterator ei; |
aa634f11 JZ |
1129 | best = NULL; |
1130 | best_len = 0; | |
628f6a4e | 1131 | FOR_EACH_EDGE (e, ei, traces[t2].first->preds) |
aa634f11 JZ |
1132 | { |
1133 | int si = e->src->index; | |
b0cc7919 | 1134 | |
fefa31b5 | 1135 | if (e->src != ENTRY_BLOCK_PTR_FOR_FN (cfun) |
aa634f11 JZ |
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] | |
87c8b4be | 1140 | && (BB_PARTITION (e->src) == current_partition) |
f7d0c571 | 1141 | && connect_better_edge_p (e, true, best_len, best, traces)) |
aa634f11 JZ |
1142 | { |
1143 | best = e; | |
1144 | best_len = traces[bbd[si].end_of_trace].length; | |
1145 | } | |
1146 | } | |
1147 | if (best) | |
1148 | { | |
370369e1 | 1149 | best->src->aux = best->dest; |
aa634f11 JZ |
1150 | t2 = bbd[best->src->index].end_of_trace; |
1151 | connected[t2] = true; | |
750054a2 | 1152 | |
c263766c | 1153 | if (dump_file) |
aa634f11 | 1154 | { |
c263766c | 1155 | fprintf (dump_file, "Connection: %d %d\n", |
aa634f11 JZ |
1156 | best->src->index, best->dest->index); |
1157 | } | |
1158 | } | |
1159 | else | |
1160 | break; | |
1161 | } | |
f008a564 | 1162 | |
aa634f11 | 1163 | if (last_trace >= 0) |
370369e1 | 1164 | traces[last_trace].last->aux = traces[t2].first; |
aa634f11 JZ |
1165 | last_trace = t; |
1166 | ||
1167 | /* Find the successor traces. */ | |
1168 | while (1) | |
b0cc7919 | 1169 | { |
aa634f11 | 1170 | /* Find the continuation of the chain. */ |
628f6a4e | 1171 | edge_iterator ei; |
aa634f11 JZ |
1172 | best = NULL; |
1173 | best_len = 0; | |
628f6a4e | 1174 | FOR_EACH_EDGE (e, ei, traces[t].last->succs) |
aa634f11 JZ |
1175 | { |
1176 | int di = e->dest->index; | |
1177 | ||
fefa31b5 | 1178 | if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
aa634f11 JZ |
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] | |
87c8b4be | 1183 | && (BB_PARTITION (e->dest) == current_partition) |
f7d0c571 | 1184 | && connect_better_edge_p (e, false, best_len, best, traces)) |
aa634f11 JZ |
1185 | { |
1186 | best = e; | |
1187 | best_len = traces[bbd[di].start_of_trace].length; | |
1188 | } | |
1189 | } | |
1190 | ||
f7d0c571 ZC |
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 (dump_file, "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) | |
aa634f11 | 1230 | { |
c263766c | 1231 | if (dump_file) |
aa634f11 | 1232 | { |
c263766c | 1233 | fprintf (dump_file, "Connection: %d %d\n", |
aa634f11 JZ |
1234 | best->src->index, best->dest->index); |
1235 | } | |
1236 | t = bbd[best->dest->index].start_of_trace; | |
370369e1 | 1237 | traces[last_trace].last->aux = traces[t].first; |
aa634f11 JZ |
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; | |
99dc7277 | 1246 | bool try_copy = false; |
aa634f11 | 1247 | |
628f6a4e | 1248 | FOR_EACH_EDGE (e, ei, traces[t].last->succs) |
fefa31b5 | 1249 | if (e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) |
aa634f11 JZ |
1250 | && (e->flags & EDGE_CAN_FALLTHRU) |
1251 | && !(e->flags & EDGE_COMPLEX) | |
99dc7277 | 1252 | && (!best || e->probability > best->probability)) |
aa634f11 | 1253 | { |
628f6a4e | 1254 | edge_iterator ei; |
aa634f11 JZ |
1255 | edge best2 = NULL; |
1256 | int best2_len = 0; | |
1257 | ||
6d9cc15b JZ |
1258 | /* If the destination is a start of a trace which is only |
1259 | one block long, then no need to search the successor | |
99dc7277 | 1260 | blocks of the trace. Accept it. */ |
6d9cc15b JZ |
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 | } | |
99dc7277 | 1269 | |
628f6a4e | 1270 | FOR_EACH_EDGE (e2, ei, e->dest->succs) |
aa634f11 JZ |
1271 | { |
1272 | int di = e2->dest->index; | |
1273 | ||
fefa31b5 | 1274 | if (e2->dest == EXIT_BLOCK_PTR_FOR_FN (cfun) |
aa634f11 JZ |
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] | |
4700dd70 | 1279 | && BB_PARTITION (e2->dest) == current_partition |
370df3ce | 1280 | && e2->count () >= count_threshold |
aa634f11 JZ |
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; | |
fefa31b5 | 1289 | if (e2->dest != EXIT_BLOCK_PTR_FOR_FN (cfun)) |
aa634f11 JZ |
1290 | best2_len = traces[bbd[di].start_of_trace].length; |
1291 | else | |
1292 | best2_len = INT_MAX; | |
1293 | next_bb = e2->dest; | |
99dc7277 | 1294 | try_copy = true; |
aa634f11 JZ |
1295 | } |
1296 | } | |
1297 | } | |
99dc7277 RH |
1298 | |
1299 | /* Copy tiny blocks always; copy larger blocks only when the | |
1300 | edge is traversed frequently enough. */ | |
1301 | if (try_copy | |
584bafb6 | 1302 | && BB_PARTITION (best->src) == BB_PARTITION (best->dest) |
99dc7277 | 1303 | && copy_bb_p (best->dest, |
efd8f750 | 1304 | optimize_edge_for_speed_p (best) |
ef30ab83 | 1305 | && (!best->count ().initialized_p () |
370df3ce | 1306 | || best->count () >= count_threshold))) |
aa634f11 JZ |
1307 | { |
1308 | basic_block new_bb; | |
1309 | ||
c263766c | 1310 | if (dump_file) |
aa634f11 | 1311 | { |
c263766c | 1312 | fprintf (dump_file, "Connection: %d %d ", |
aa634f11 | 1313 | traces[t].last->index, best->dest->index); |
99dc7277 | 1314 | if (!next_bb) |
c263766c | 1315 | fputc ('\n', dump_file); |
fefa31b5 | 1316 | else if (next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
c263766c | 1317 | fprintf (dump_file, "exit\n"); |
aa634f11 | 1318 | else |
c263766c | 1319 | fprintf (dump_file, "%d\n", next_bb->index); |
aa634f11 JZ |
1320 | } |
1321 | ||
1322 | new_bb = copy_bb (best->dest, best, traces[t].last, t); | |
1323 | traces[t].last = new_bb; | |
fefa31b5 | 1324 | if (next_bb && next_bb != EXIT_BLOCK_PTR_FOR_FN (cfun)) |
aa634f11 JZ |
1325 | { |
1326 | t = bbd[next_bb->index].start_of_trace; | |
370369e1 | 1327 | traces[last_trace].last->aux = traces[t].first; |
aa634f11 JZ |
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 | } | |
b0cc7919 | 1337 | } |
aa634f11 JZ |
1338 | } |
1339 | ||
c263766c | 1340 | if (dump_file) |
aa634f11 JZ |
1341 | { |
1342 | basic_block bb; | |
f008a564 | 1343 | |
c263766c | 1344 | fprintf (dump_file, "Final order:\n"); |
f883e0a7 | 1345 | for (bb = traces[0].first; bb; bb = (basic_block) bb->aux) |
c263766c RH |
1346 | fprintf (dump_file, "%d ", bb->index); |
1347 | fprintf (dump_file, "\n"); | |
1348 | fflush (dump_file); | |
295ae817 JE |
1349 | } |
1350 | ||
aa634f11 JZ |
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 | |
9678086d | 1358 | copy_bb_p (const_basic_block bb, int code_may_grow) |
aa634f11 | 1359 | { |
a0d87c29 AM |
1360 | unsigned int size = 0; |
1361 | unsigned int max_size = uncond_jump_length; | |
e93768e4 | 1362 | rtx_insn *insn; |
aa634f11 | 1363 | |
628f6a4e | 1364 | if (EDGE_COUNT (bb->preds) < 2) |
aa634f11 | 1365 | return false; |
6de9cd9a | 1366 | if (!can_duplicate_block_p (bb)) |
aa634f11 JZ |
1367 | return false; |
1368 | ||
6ae533cf | 1369 | /* Avoid duplicating blocks which have many successors (PR/13430). */ |
628f6a4e BE |
1370 | if (EDGE_COUNT (bb->succs) > 8) |
1371 | return false; | |
6ae533cf | 1372 | |
efd8f750 | 1373 | if (code_may_grow && optimize_bb_for_speed_p (bb)) |
028d4092 | 1374 | max_size *= param_max_grow_copy_bb_insns; |
aa634f11 | 1375 | |
bbcb0c05 | 1376 | FOR_BB_INSNS (bb, insn) |
f008a564 | 1377 | { |
aa634f11 | 1378 | if (INSN_P (insn)) |
a0d87c29 AM |
1379 | { |
1380 | size += get_attr_min_length (insn); | |
1381 | if (size > max_size) | |
1382 | break; | |
1383 | } | |
f008a564 | 1384 | } |
295ae817 | 1385 | |
aa634f11 JZ |
1386 | if (size <= max_size) |
1387 | return true; | |
1388 | ||
c263766c | 1389 | if (dump_file) |
f008a564 | 1390 | { |
c263766c | 1391 | fprintf (dump_file, |
a0d87c29 | 1392 | "Block %d can't be copied because its size = %u.\n", |
aa634f11 | 1393 | bb->index, size); |
f008a564 | 1394 | } |
295ae817 | 1395 | |
aa634f11 JZ |
1396 | return false; |
1397 | } | |
1398 | ||
1399 | /* Return the length of unconditional jump instruction. */ | |
1400 | ||
ffe14686 | 1401 | int |
4682ae04 | 1402 | get_uncond_jump_length (void) |
aa634f11 | 1403 | { |
a0d87c29 | 1404 | unsigned int length; |
aa634f11 | 1405 | |
ed2b2162 | 1406 | start_sequence (); |
e67d1102 | 1407 | rtx_code_label *label = emit_label (gen_label_rtx ()); |
ec4a505f | 1408 | rtx_insn *jump = emit_jump_insn (targetm.gen_jump (label)); |
070a7956 | 1409 | length = get_attr_min_length (jump); |
ed2b2162 | 1410 | end_sequence (); |
aa634f11 | 1411 | |
a0d87c29 | 1412 | gcc_assert (length < INT_MAX); |
aa634f11 | 1413 | return length; |
295ae817 JE |
1414 | } |
1415 | ||
40305631 EB |
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 | |
81e99446 | 1420 | create_eh_forwarder_block (rtx_code_label *new_label, basic_block old_bb) |
40305631 | 1421 | { |
81e99446 JJ |
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 | ||
40305631 EB |
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 (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. */ | |
81e99446 | 1467 | basic_block new_bb = create_eh_forwarder_block (new_label, old_bb); |
40305631 EB |
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 (index_map, 0, lp_len * sizeof (unsigned)); | |
1472 | ||
1473 | /* Fix up the edges. */ | |
1474 | for (ei = ei_start (old_bb->preds); (e = ei_safe_edge (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 (&ei); | |
1499 | } | |
1500 | ||
0be7e7a6 | 1501 | /* The landing pad OLD_LP, in block OLD_BB, has edges from both partitions. |
465d0087 JJ |
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. */ | |
0be7e7a6 RH |
1504 | |
1505 | static void | |
40305631 | 1506 | dw2_fix_up_crossing_landing_pad (eh_landing_pad old_lp, basic_block old_bb) |
0be7e7a6 RH |
1507 | { |
1508 | eh_landing_pad new_lp; | |
0be7e7a6 RH |
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 | ||
40305631 | 1518 | /* Create the forwarder block. */ |
81e99446 | 1519 | basic_block new_bb = create_eh_forwarder_block (new_lp->landing_pad, old_bb); |
0be7e7a6 RH |
1520 | |
1521 | /* Fix up the edges. */ | |
1522 | for (ei = ei_start (old_bb->preds); (e = ei_safe_edge (ei)) != NULL; ) | |
40305631 | 1523 | if (e->src != new_bb && BB_PARTITION (e->src) == BB_PARTITION (new_bb)) |
0be7e7a6 | 1524 | { |
e93768e4 | 1525 | rtx_insn *insn = BB_END (e->src); |
0be7e7a6 RH |
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 (&ei); | |
1537 | } | |
1538 | ||
600b5b1d TJ |
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; | |
357067f2 JH |
1564 | profile_probability highest_probability |
1565 | = profile_probability::uninitialized (); | |
3995f3a2 | 1566 | profile_count highest_count = profile_count::uninitialized (); |
600b5b1d TJ |
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 | ||
8b5d71cd JH |
1579 | /* Do not expect profile insanities when profile was not adjusted. */ |
1580 | if (e->probability == profile_probability::never () | |
ef30ab83 | 1581 | || e->count () == profile_count::zero ()) |
8b5d71cd JH |
1582 | continue; |
1583 | ||
600b5b1d TJ |
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 | |
79a6b3de JH |
1590 | the edge count, if it is non-zero, then fallback to |
1591 | the edge probability. */ | |
e7a74006 | 1592 | if (!(e->count () > highest_count)) |
ef30ab83 | 1593 | highest_count = e->count (); |
357067f2 JH |
1594 | if (!highest_probability.initialized_p () |
1595 | || e->probability > highest_probability) | |
600b5b1d TJ |
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; | |
8b5d71cd JH |
1611 | /* Do not expect profile insanities when profile was not adjusted. */ |
1612 | if (e->probability == profile_probability::never () | |
ef30ab83 | 1613 | || e->count () == profile_count::zero ()) |
8b5d71cd | 1614 | continue; |
600b5b1d | 1615 | /* Select the hottest edge using the edge count, if it is non-zero, |
79a6b3de | 1616 | then fallback to the edge probability. */ |
e7a74006 | 1617 | if (highest_count.initialized_p ()) |
600b5b1d | 1618 | { |
e7a74006 | 1619 | if (!(e->count () >= highest_count)) |
600b5b1d TJ |
1620 | continue; |
1621 | } | |
e7a74006 | 1622 | else if (!(e->probability >= highest_probability)) |
600b5b1d TJ |
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); | |
8b5d71cd JH |
1630 | if (dump_file) |
1631 | fprintf (dump_file, "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"); | |
600b5b1d TJ |
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 (reach_bb); | |
1639 | hot_bbs_to_check.safe_push (reach_bb); | |
1640 | } | |
1641 | } | |
dd172744 | 1642 | hot_bbs_to_check.release (); |
600b5b1d TJ |
1643 | |
1644 | return cold_bb_count; | |
1645 | } | |
1646 | ||
1647 | ||
750054a2 CT |
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 | |
ea6136a2 | 1650 | cache locality). Return a vector of all edges that cross. */ |
750054a2 | 1651 | |
600b5b1d | 1652 | static vec<edge> |
ea6136a2 | 1653 | find_rarely_executed_basic_blocks_and_crossing_edges (void) |
750054a2 | 1654 | { |
6e1aa848 | 1655 | vec<edge> crossing_edges = vNULL; |
750054a2 CT |
1656 | basic_block bb; |
1657 | edge e; | |
628f6a4e | 1658 | edge_iterator ei; |
600b5b1d | 1659 | unsigned int cold_bb_count = 0; |
548296b0 | 1660 | auto_vec<basic_block> bbs_in_hot_partition; |
750054a2 | 1661 | |
8b5d71cd JH |
1662 | propagate_unlikely_bbs_forward (); |
1663 | ||
750054a2 | 1664 | /* Mark which partition (hot/cold) each basic block belongs in. */ |
11cd3bed | 1665 | FOR_EACH_BB_FN (bb, cfun) |
750054a2 | 1666 | { |
79221839 TJ |
1667 | bool cold_bb = false; |
1668 | ||
2eb712b4 | 1669 | if (probably_never_executed_bb_p (cfun, bb)) |
79221839 | 1670 | { |
3b131f6a JH |
1671 | cold_bb = true; |
1672 | ||
79221839 TJ |
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. */ | |
3b131f6a JH |
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 | } | |
79221839 TJ |
1684 | } |
1685 | if (cold_bb) | |
600b5b1d TJ |
1686 | { |
1687 | BB_SET_PARTITION (bb, BB_COLD_PARTITION); | |
1688 | cold_bb_count++; | |
1689 | } | |
750054a2 | 1690 | else |
600b5b1d TJ |
1691 | { |
1692 | BB_SET_PARTITION (bb, BB_HOT_PARTITION); | |
1693 | bbs_in_hot_partition.safe_push (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 (true, cold_bb_count, | |
1711 | &bbs_in_hot_partition); | |
1712 | if (cold_bb_count) | |
1713 | sanitize_hot_paths (false, cold_bb_count, &bbs_in_hot_partition); | |
d9af4fea JH |
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 (bb)) | |
1719 | BB_SET_PARTITION (bb, BB_COLD_PARTITION); | |
750054a2 CT |
1720 | } |
1721 | ||
0be7e7a6 RH |
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 | |
40305631 | 1724 | moving the landing pads or inserting forwarder landing pads. */ |
0be7e7a6 | 1725 | if (cfun->eh->lp_array) |
9fb32434 | 1726 | { |
40305631 EB |
1727 | const bool sjlj |
1728 | = (targetm_common.except_unwind_info (&global_options) == UI_SJLJ); | |
0be7e7a6 RH |
1729 | unsigned i; |
1730 | eh_landing_pad lp; | |
1731 | ||
9771b263 | 1732 | FOR_EACH_VEC_ELT (*cfun->eh->lp_array, i, lp) |
c7466dee | 1733 | { |
0be7e7a6 RH |
1734 | bool all_same, all_diff; |
1735 | ||
b58d3391 JJ |
1736 | if (lp == NULL |
1737 | || lp->landing_pad == NULL_RTX | |
1738 | || !LABEL_P (lp->landing_pad)) | |
0be7e7a6 RH |
1739 | continue; |
1740 | ||
1741 | all_same = all_diff = true; | |
1742 | bb = BLOCK_FOR_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 | } | |
40305631 EB |
1760 | else if (sjlj) |
1761 | sjlj_fix_up_crossing_landing_pad (bb); | |
0be7e7a6 | 1762 | else |
40305631 EB |
1763 | dw2_fix_up_crossing_landing_pad (lp, bb); |
1764 | ||
1765 | /* There is a single, common landing pad in SJLJ mode. */ | |
1766 | if (sjlj) | |
1767 | break; | |
c7466dee | 1768 | } |
9fb32434 | 1769 | } |
ea6136a2 | 1770 | |
0be7e7a6 | 1771 | /* Mark every edge that crosses between sections. */ |
11cd3bed | 1772 | FOR_EACH_BB_FN (bb, cfun) |
0be7e7a6 RH |
1773 | FOR_EACH_EDGE (e, ei, bb->succs) |
1774 | { | |
1775 | unsigned int flags = e->flags; | |
4700dd70 | 1776 | |
0be7e7a6 RH |
1777 | /* We should never have EDGE_CROSSING set yet. */ |
1778 | gcc_checking_assert ((flags & EDGE_CROSSING) == 0); | |
1779 | ||
fefa31b5 DM |
1780 | if (e->src != ENTRY_BLOCK_PTR_FOR_FN (cfun) |
1781 | && e->dest != EXIT_BLOCK_PTR_FOR_FN (cfun) | |
0be7e7a6 RH |
1782 | && BB_PARTITION (e->src) != BB_PARTITION (e->dest)) |
1783 | { | |
9771b263 | 1784 | crossing_edges.safe_push (e); |
0be7e7a6 RH |
1785 | flags |= EDGE_CROSSING; |
1786 | } | |
9f68560b | 1787 | |
0be7e7a6 RH |
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; | |
9f68560b RH |
1796 | } |
1797 | ||
532aafad SB |
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 | ||
11cd3bed | 1805 | FOR_EACH_BB_FN (bb, cfun) |
532aafad SB |
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; | |
1476d1bd MM |
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)) | |
532aafad | 1828 | continue; |
1476d1bd | 1829 | invert_jump (bb_end_jump, JUMP_LABEL (bb_end_jump), 0); |
532aafad SB |
1830 | EDGE_SUCC (bb, 0)->flags |= EDGE_CAN_FALLTHRU; |
1831 | EDGE_SUCC (bb, 1)->flags |= EDGE_CAN_FALLTHRU; | |
1832 | } | |
1833 | } | |
1834 | ||
750054a2 | 1835 | /* If any destination of a crossing edge does not have a label, add label; |
ea6136a2 | 1836 | Convert any easy fall-through crossing edges to unconditional jumps. */ |
750054a2 | 1837 | |
c22cacf3 | 1838 | static void |
9771b263 | 1839 | add_labels_and_missing_jumps (vec<edge> crossing_edges) |
750054a2 | 1840 | { |
ea6136a2 RH |
1841 | size_t i; |
1842 | edge e; | |
c22cacf3 | 1843 | |
9771b263 | 1844 | FOR_EACH_VEC_ELT (crossing_edges, i, e) |
750054a2 | 1845 | { |
ea6136a2 RH |
1846 | basic_block src = e->src; |
1847 | basic_block dest = e->dest; | |
e67d1102 | 1848 | rtx_jump_insn *new_jump; |
c22cacf3 | 1849 | |
fefa31b5 | 1850 | if (dest == EXIT_BLOCK_PTR_FOR_FN (cfun)) |
ea6136a2 | 1851 | continue; |
c22cacf3 | 1852 | |
ea6136a2 | 1853 | /* Make sure dest has a label. */ |
e67d1102 | 1854 | rtx_code_label *label = block_label (dest); |
c22cacf3 | 1855 | |
ea6136a2 | 1856 | /* Nothing to do for non-fallthru edges. */ |
fefa31b5 | 1857 | if (src == ENTRY_BLOCK_PTR_FOR_FN (cfun)) |
ea6136a2 RH |
1858 | continue; |
1859 | if ((e->flags & EDGE_FALLTHRU) == 0) | |
1860 | continue; | |
c22cacf3 | 1861 | |
ea6136a2 RH |
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 | ||
ec4a505f | 1872 | new_jump = emit_jump_insn_after (targetm.gen_jump (label), BB_END (src)); |
1130d5e3 | 1873 | BB_END (src) = new_jump; |
ea6136a2 RH |
1874 | JUMP_LABEL (new_jump) = label; |
1875 | LABEL_NUSES (label) += 1; | |
9f68560b RH |
1876 | |
1877 | emit_barrier_after_bb (src); | |
ea6136a2 RH |
1878 | |
1879 | /* Mark edge as non-fallthru. */ | |
1880 | e->flags &= ~EDGE_FALLTHRU; | |
1881 | } | |
750054a2 CT |
1882 | } |
1883 | ||
1884 | /* Find any bb's where the fall-through edge is a crossing edge (note that | |
8bf6e270 RE |
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. */ | |
750054a2 | 1892 | |
c22cacf3 | 1893 | static void |
750054a2 CT |
1894 | fix_up_fall_thru_edges (void) |
1895 | { | |
1896 | basic_block cur_bb; | |
c22cacf3 | 1897 | |
11cd3bed | 1898 | FOR_EACH_BB_FN (cur_bb, cfun) |
750054a2 | 1899 | { |
c52d6043 JJ |
1900 | edge succ1; |
1901 | edge succ2; | |
1902 | edge fall_thru = NULL; | |
1903 | edge cond_jump = NULL; | |
c52d6043 | 1904 | |
750054a2 | 1905 | fall_thru = NULL; |
628f6a4e BE |
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) | |
c22cacf3 | 1912 | succ2 = EDGE_SUCC (cur_bb, 1); |
750054a2 | 1913 | else |
c22cacf3 MS |
1914 | succ2 = NULL; |
1915 | ||
750054a2 | 1916 | /* Find the fall-through edge. */ |
c22cacf3 MS |
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 | } | |
c52d6043 JJ |
1930 | else if (succ2 && EDGE_COUNT (cur_bb->succs) > 2) |
1931 | fall_thru = find_fallthru_edge (cur_bb->succs); | |
c22cacf3 | 1932 | |
fefa31b5 | 1933 | if (fall_thru && (fall_thru->dest != EXIT_BLOCK_PTR_FOR_FN (cfun))) |
c22cacf3 MS |
1934 | { |
1935 | /* Check to see if the fall-thru edge is a crossing edge. */ | |
1936 | ||
bd454efd | 1937 | if (fall_thru->flags & EDGE_CROSSING) |
c22cacf3 | 1938 | { |
750054a2 | 1939 | /* The fall_thru edge crosses; now check the cond jump edge, if |
c22cacf3 MS |
1940 | it exists. */ |
1941 | ||
c52d6043 JJ |
1942 | bool cond_jump_crosses = true; |
1943 | int invert_worked = 0; | |
1944 | rtx_insn *old_jump = BB_END (cur_bb); | |
c22cacf3 MS |
1945 | |
1946 | /* Find the jump instruction, if there is one. */ | |
1947 | ||
1948 | if (cond_jump) | |
1949 | { | |
bd454efd | 1950 | if (!(cond_jump->flags & EDGE_CROSSING)) |
c22cacf3 MS |
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 | |
750054a2 | 1955 | next block in the bb order, invert the jump |
073a8998 | 1956 | (i.e. fix it so the fall through does not cross and |
c22cacf3 MS |
1957 | the cond jump does). */ |
1958 | ||
3371a64f | 1959 | if (!cond_jump_crosses) |
c22cacf3 MS |
1960 | { |
1961 | /* Find label in fall_thru block. We've already added | |
1962 | any missing labels, so there must be one. */ | |
1963 | ||
c52d6043 JJ |
1964 | rtx_code_label *fall_thru_label |
1965 | = block_label (fall_thru->dest); | |
c22cacf3 | 1966 | |
1476d1bd MM |
1967 | if (old_jump && fall_thru_label) |
1968 | { | |
c52d6043 JJ |
1969 | rtx_jump_insn *old_jump_insn |
1970 | = dyn_cast <rtx_jump_insn *> (old_jump); | |
1476d1bd MM |
1971 | if (old_jump_insn) |
1972 | invert_worked = invert_jump (old_jump_insn, | |
1973 | fall_thru_label, 0); | |
1974 | } | |
1975 | ||
c22cacf3 MS |
1976 | if (invert_worked) |
1977 | { | |
1978 | fall_thru->flags &= ~EDGE_FALLTHRU; | |
1979 | cond_jump->flags |= EDGE_FALLTHRU; | |
1980 | update_br_prob_note (cur_bb); | |
fab27f52 | 1981 | std::swap (fall_thru, cond_jump); |
bd454efd SB |
1982 | cond_jump->flags |= EDGE_CROSSING; |
1983 | fall_thru->flags &= ~EDGE_CROSSING; | |
c22cacf3 MS |
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 | |
750054a2 | 1992 | fall through edge. The jump edge will be taken care |
b8698a0f | 1993 | of later. The EDGE_CROSSING flag of fall_thru edge |
4700dd70 EB |
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. */ | |
8bf6e270 | 1999 | |
4700dd70 | 2000 | fall_thru->flags &= ~EDGE_CROSSING; |
c52d6043 | 2001 | basic_block new_bb = force_nonfallthru (fall_thru); |
c22cacf3 MS |
2002 | |
2003 | if (new_bb) | |
2004 | { | |
2005 | new_bb->aux = cur_bb->aux; | |
2006 | cur_bb->aux = new_bb; | |
2007 | ||
3371a64f TJ |
2008 | /* This is done by force_nonfallthru_and_redirect. */ |
2009 | gcc_assert (BB_PARTITION (new_bb) | |
2010 | == BB_PARTITION (cur_bb)); | |
076c7ab8 | 2011 | |
c5cbcccf | 2012 | single_succ_edge (new_bb)->flags |= EDGE_CROSSING; |
c22cacf3 | 2013 | } |
4700dd70 EB |
2014 | else |
2015 | { | |
2016 | /* If a new basic-block was not created; restore | |
2017 | the EDGE_CROSSING flag. */ | |
2018 | fall_thru->flags |= EDGE_CROSSING; | |
2019 | } | |
c22cacf3 MS |
2020 | |
2021 | /* Add barrier after new jump */ | |
9f68560b | 2022 | emit_barrier_after_bb (new_bb ? new_bb : cur_bb); |
c22cacf3 MS |
2023 | } |
2024 | } | |
2025 | } | |
750054a2 CT |
2026 | } |
2027 | } | |
2028 | ||
fa10beec | 2029 | /* This function checks the destination block of a "crossing jump" to |
750054a2 CT |
2030 | see if it has any crossing predecessors that begin with a code label |
2031 | and end with an unconditional jump. If so, it returns that predecessor | |
2032 | block. (This is to avoid creating lots of new basic blocks that all | |
2033 | contain unconditional jumps to the same destination). */ | |
2034 | ||
2035 | static basic_block | |
c22cacf3 MS |
2036 | find_jump_block (basic_block jump_dest) |
2037 | { | |
2038 | basic_block source_bb = NULL; | |
750054a2 | 2039 | edge e; |
e93768e4 | 2040 | rtx_insn *insn; |
628f6a4e | 2041 | edge_iterator ei; |
750054a2 | 2042 | |
628f6a4e | 2043 | FOR_EACH_EDGE (e, ei, jump_dest->preds) |
bd454efd | 2044 | if (e->flags & EDGE_CROSSING) |
750054a2 CT |
2045 | { |
2046 | basic_block src = e->src; | |
c22cacf3 | 2047 | |
750054a2 CT |
2048 | /* Check each predecessor to see if it has a label, and contains |
2049 | only one executable instruction, which is an unconditional jump. | |
9cf737f8 | 2050 | If so, we can use it. */ |
c22cacf3 | 2051 | |
4b4bf941 | 2052 | if (LABEL_P (BB_HEAD (src))) |
c22cacf3 | 2053 | for (insn = BB_HEAD (src); |
750054a2 CT |
2054 | !INSN_P (insn) && insn != NEXT_INSN (BB_END (src)); |
2055 | insn = NEXT_INSN (insn)) | |
2056 | { | |
2057 | if (INSN_P (insn) | |
2058 | && insn == BB_END (src) | |
4b4bf941 | 2059 | && JUMP_P (insn) |
750054a2 CT |
2060 | && !any_condjump_p (insn)) |
2061 | { | |
2062 | source_bb = src; | |
2063 | break; | |
2064 | } | |
2065 | } | |
c22cacf3 | 2066 | |
750054a2 CT |
2067 | if (source_bb) |
2068 | break; | |
2069 | } | |
2070 | ||
2071 | return source_bb; | |
2072 | } | |
2073 | ||
2074 | /* Find all BB's with conditional jumps that are crossing edges; | |
2075 | insert a new bb and make the conditional jump branch to the new | |
2076 | bb instead (make the new bb same color so conditional branch won't | |
2077 | be a 'crossing' edge). Insert an unconditional jump from the | |
2078 | new bb to the original destination of the conditional jump. */ | |
2079 | ||
2080 | static void | |
2081 | fix_crossing_conditional_branches (void) | |
2082 | { | |
2083 | basic_block cur_bb; | |
2084 | basic_block new_bb; | |
750054a2 | 2085 | basic_block dest; |
750054a2 CT |
2086 | edge succ1; |
2087 | edge succ2; | |
2088 | edge crossing_edge; | |
2089 | edge new_edge; | |
750054a2 CT |
2090 | rtx set_src; |
2091 | rtx old_label = NULL_RTX; | |
1476d1bd | 2092 | rtx_code_label *new_label; |
c22cacf3 | 2093 | |
11cd3bed | 2094 | FOR_EACH_BB_FN (cur_bb, cfun) |
750054a2 CT |
2095 | { |
2096 | crossing_edge = NULL; | |
628f6a4e BE |
2097 | if (EDGE_COUNT (cur_bb->succs) > 0) |
2098 | succ1 = EDGE_SUCC (cur_bb, 0); | |
2099 | else | |
2100 | succ1 = NULL; | |
c22cacf3 | 2101 | |
628f6a4e BE |
2102 | if (EDGE_COUNT (cur_bb->succs) > 1) |
2103 | succ2 = EDGE_SUCC (cur_bb, 1); | |
750054a2 | 2104 | else |
628f6a4e | 2105 | succ2 = NULL; |
c22cacf3 | 2106 | |
750054a2 CT |
2107 | /* We already took care of fall-through edges, so only one successor |
2108 | can be a crossing edge. */ | |
c22cacf3 | 2109 | |
bd454efd | 2110 | if (succ1 && (succ1->flags & EDGE_CROSSING)) |
750054a2 | 2111 | crossing_edge = succ1; |
bd454efd | 2112 | else if (succ2 && (succ2->flags & EDGE_CROSSING)) |
c22cacf3 MS |
2113 | crossing_edge = succ2; |
2114 | ||
2115 | if (crossing_edge) | |
2116 | { | |
d08093db | 2117 | rtx_insn *old_jump = BB_END (cur_bb); |
c22cacf3 | 2118 | |
750054a2 CT |
2119 | /* Check to make sure the jump instruction is a |
2120 | conditional jump. */ | |
c22cacf3 | 2121 | |
750054a2 CT |
2122 | set_src = NULL_RTX; |
2123 | ||
2124 | if (any_condjump_p (old_jump)) | |
2125 | { | |
2126 | if (GET_CODE (PATTERN (old_jump)) == SET) | |
2127 | set_src = SET_SRC (PATTERN (old_jump)); | |
2128 | else if (GET_CODE (PATTERN (old_jump)) == PARALLEL) | |
2129 | { | |
2130 | set_src = XVECEXP (PATTERN (old_jump), 0,0); | |
2131 | if (GET_CODE (set_src) == SET) | |
2132 | set_src = SET_SRC (set_src); | |
2133 | else | |
2134 | set_src = NULL_RTX; | |
2135 | } | |
2136 | } | |
2137 | ||
2138 | if (set_src && (GET_CODE (set_src) == IF_THEN_ELSE)) | |
2139 | { | |
d08093db MM |
2140 | rtx_jump_insn *old_jump_insn = |
2141 | as_a <rtx_jump_insn *> (old_jump); | |
2142 | ||
750054a2 CT |
2143 | if (GET_CODE (XEXP (set_src, 1)) == PC) |
2144 | old_label = XEXP (set_src, 2); | |
2145 | else if (GET_CODE (XEXP (set_src, 2)) == PC) | |
2146 | old_label = XEXP (set_src, 1); | |
c22cacf3 | 2147 | |
750054a2 CT |
2148 | /* Check to see if new bb for jumping to that dest has |
2149 | already been created; if so, use it; if not, create | |
2150 | a new one. */ | |
2151 | ||
2152 | new_bb = find_jump_block (crossing_edge->dest); | |
c22cacf3 | 2153 | |
750054a2 CT |
2154 | if (new_bb) |
2155 | new_label = block_label (new_bb); | |
2156 | else | |
2157 | { | |
6774a66f | 2158 | basic_block last_bb; |
1476d1bd MM |
2159 | rtx_code_label *old_jump_target; |
2160 | rtx_jump_insn *new_jump; | |
6774a66f | 2161 | |
750054a2 CT |
2162 | /* Create new basic block to be dest for |
2163 | conditional jump. */ | |
c22cacf3 | 2164 | |
750054a2 | 2165 | /* Put appropriate instructions in new bb. */ |
c22cacf3 | 2166 | |
750054a2 | 2167 | new_label = gen_label_rtx (); |
6774a66f | 2168 | emit_label (new_label); |
c22cacf3 | 2169 | |
6774a66f | 2170 | gcc_assert (GET_CODE (old_label) == LABEL_REF); |
d08093db | 2171 | old_jump_target = old_jump_insn->jump_target (); |
1476d1bd | 2172 | new_jump = as_a <rtx_jump_insn *> |
ec4a505f | 2173 | (emit_jump_insn (targetm.gen_jump (old_jump_target))); |
1476d1bd | 2174 | new_jump->set_jump_target (old_jump_target); |
6774a66f | 2175 | |
fefa31b5 | 2176 | last_bb = EXIT_BLOCK_PTR_FOR_FN (cfun)->prev_bb; |
6774a66f RH |
2177 | new_bb = create_basic_block (new_label, new_jump, last_bb); |
2178 | new_bb->aux = last_bb->aux; | |
2179 | last_bb->aux = new_bb; | |
9f68560b RH |
2180 | |
2181 | emit_barrier_after_bb (new_bb); | |
c22cacf3 | 2182 | |
750054a2 CT |
2183 | /* Make sure new bb is in same partition as source |
2184 | of conditional branch. */ | |
076c7ab8 | 2185 | BB_COPY_PARTITION (new_bb, cur_bb); |
750054a2 | 2186 | } |
c22cacf3 | 2187 | |
750054a2 | 2188 | /* Make old jump branch to new bb. */ |
c22cacf3 | 2189 | |
d08093db | 2190 | redirect_jump (old_jump_insn, new_label, 0); |
c22cacf3 | 2191 | |
750054a2 | 2192 | /* Remove crossing_edge as predecessor of 'dest'. */ |
c22cacf3 | 2193 | |
750054a2 | 2194 | dest = crossing_edge->dest; |
c22cacf3 | 2195 | |
750054a2 | 2196 | redirect_edge_succ (crossing_edge, new_bb); |
c22cacf3 | 2197 | |
750054a2 CT |
2198 | /* Make a new edge from new_bb to old dest; new edge |
2199 | will be a successor for new_bb and a predecessor | |
2200 | for 'dest'. */ | |
c22cacf3 | 2201 | |
628f6a4e | 2202 | if (EDGE_COUNT (new_bb->succs) == 0) |
a4f20288 | 2203 | new_edge = make_single_succ_edge (new_bb, dest, 0); |
750054a2 | 2204 | else |
628f6a4e | 2205 | new_edge = EDGE_SUCC (new_bb, 0); |
c22cacf3 | 2206 | |
bd454efd SB |
2207 | crossing_edge->flags &= ~EDGE_CROSSING; |
2208 | new_edge->flags |= EDGE_CROSSING; | |
750054a2 | 2209 | } |
c22cacf3 | 2210 | } |
750054a2 CT |
2211 | } |
2212 | } | |
2213 | ||
2214 | /* Find any unconditional branches that cross between hot and cold | |
2215 | sections. Convert them into indirect jumps instead. */ | |
2216 | ||
2217 | static void | |
2218 | fix_crossing_unconditional_branches (void) | |
2219 | { | |
2220 | basic_block cur_bb; | |
e93768e4 | 2221 | rtx_insn *last_insn; |
750054a2 CT |
2222 | rtx label; |
2223 | rtx label_addr; | |
e93768e4 DM |
2224 | rtx_insn *indirect_jump_sequence; |
2225 | rtx_insn *jump_insn = NULL; | |
750054a2 | 2226 | rtx new_reg; |
e93768e4 | 2227 | rtx_insn *cur_insn; |
750054a2 | 2228 | edge succ; |
9fb32434 | 2229 | |
11cd3bed | 2230 | FOR_EACH_BB_FN (cur_bb, cfun) |
750054a2 CT |
2231 | { |
2232 | last_insn = BB_END (cur_bb); | |
87c8b4be CT |
2233 | |
2234 | if (EDGE_COUNT (cur_bb->succs) < 1) | |
2235 | continue; | |
2236 | ||
628f6a4e | 2237 | succ = EDGE_SUCC (cur_bb, 0); |
750054a2 CT |
2238 | |
2239 | /* Check to see if bb ends in a crossing (unconditional) jump. At | |
c22cacf3 | 2240 | this point, no crossing jumps should be conditional. */ |
750054a2 | 2241 | |
4b4bf941 | 2242 | if (JUMP_P (last_insn) |
bd454efd | 2243 | && (succ->flags & EDGE_CROSSING)) |
750054a2 | 2244 | { |
298e6adc | 2245 | gcc_assert (!any_condjump_p (last_insn)); |
750054a2 CT |
2246 | |
2247 | /* Make sure the jump is not already an indirect or table jump. */ | |
2248 | ||
298e6adc | 2249 | if (!computed_jump_p (last_insn) |
7bca81dc | 2250 | && !tablejump_p (last_insn, NULL, NULL)) |
750054a2 CT |
2251 | { |
2252 | /* We have found a "crossing" unconditional branch. Now | |
2253 | we must convert it to an indirect jump. First create | |
2254 | reference of label, as target for jump. */ | |
c22cacf3 | 2255 | |
750054a2 | 2256 | label = JUMP_LABEL (last_insn); |
b50b729d | 2257 | label_addr = gen_rtx_LABEL_REF (Pmode, label); |
750054a2 | 2258 | LABEL_NUSES (label) += 1; |
c22cacf3 | 2259 | |
750054a2 | 2260 | /* Get a register to use for the indirect jump. */ |
c22cacf3 | 2261 | |
750054a2 | 2262 | new_reg = gen_reg_rtx (Pmode); |
c22cacf3 | 2263 | |
750054a2 | 2264 | /* Generate indirect the jump sequence. */ |
c22cacf3 | 2265 | |
750054a2 CT |
2266 | start_sequence (); |
2267 | emit_move_insn (new_reg, label_addr); | |
2268 | emit_indirect_jump (new_reg); | |
2269 | indirect_jump_sequence = get_insns (); | |
2270 | end_sequence (); | |
c22cacf3 | 2271 | |
750054a2 CT |
2272 | /* Make sure every instruction in the new jump sequence has |
2273 | its basic block set to be cur_bb. */ | |
c22cacf3 | 2274 | |
750054a2 CT |
2275 | for (cur_insn = indirect_jump_sequence; cur_insn; |
2276 | cur_insn = NEXT_INSN (cur_insn)) | |
2277 | { | |
0e714a54 EB |
2278 | if (!BARRIER_P (cur_insn)) |
2279 | BLOCK_FOR_INSN (cur_insn) = cur_bb; | |
4b4bf941 | 2280 | if (JUMP_P (cur_insn)) |
750054a2 CT |
2281 | jump_insn = cur_insn; |
2282 | } | |
c22cacf3 | 2283 | |
750054a2 CT |
2284 | /* Insert the new (indirect) jump sequence immediately before |
2285 | the unconditional jump, then delete the unconditional jump. */ | |
c22cacf3 | 2286 | |
750054a2 CT |
2287 | emit_insn_before (indirect_jump_sequence, last_insn); |
2288 | delete_insn (last_insn); | |
c22cacf3 | 2289 | |
8fce217e JJ |
2290 | JUMP_LABEL (jump_insn) = label; |
2291 | LABEL_NUSES (label)++; | |
2292 | ||
750054a2 CT |
2293 | /* Make BB_END for cur_bb be the jump instruction (NOT the |
2294 | barrier instruction at the end of the sequence...). */ | |
c22cacf3 | 2295 | |
1130d5e3 | 2296 | BB_END (cur_bb) = jump_insn; |
750054a2 CT |
2297 | } |
2298 | } | |
2299 | } | |
2300 | } | |
2301 | ||
339ba33b | 2302 | /* Update CROSSING_JUMP_P flags on all jump insns. */ |
750054a2 CT |
2303 | |
2304 | static void | |
339ba33b | 2305 | update_crossing_jump_flags (void) |
750054a2 CT |
2306 | { |
2307 | basic_block bb; | |
2308 | edge e; | |
628f6a4e | 2309 | edge_iterator ei; |
750054a2 | 2310 | |
11cd3bed | 2311 | FOR_EACH_BB_FN (bb, cfun) |
628f6a4e | 2312 | FOR_EACH_EDGE (e, ei, bb->succs) |
339ba33b RS |
2313 | if (e->flags & EDGE_CROSSING) |
2314 | { | |
52af3804 | 2315 | if (JUMP_P (BB_END (bb))) |
339ba33b RS |
2316 | CROSSING_JUMP_P (BB_END (bb)) = 1; |
2317 | break; | |
2318 | } | |
750054a2 CT |
2319 | } |
2320 | ||
af71fa39 | 2321 | /* Reorder basic blocks using the software trace cache (STC) algorithm. */ |
295ae817 | 2322 | |
711417cd | 2323 | static void |
af71fa39 | 2324 | reorder_basic_blocks_software_trace_cache (void) |
295ae817 | 2325 | { |
248c16c4 SB |
2326 | if (dump_file) |
2327 | fprintf (dump_file, "\nReordering with the STC algorithm.\n\n"); | |
2328 | ||
aa634f11 JZ |
2329 | int n_traces; |
2330 | int i; | |
2331 | struct trace *traces; | |
2332 | ||
e0bb17a8 KH |
2333 | /* We are estimating the length of uncond jump insn only once since the code |
2334 | for getting the insn length always returns the minimal length now. */ | |
4682ae04 | 2335 | if (uncond_jump_length == 0) |
aa634f11 JZ |
2336 | uncond_jump_length = get_uncond_jump_length (); |
2337 | ||
2338 | /* We need to know some information for each basic block. */ | |
8b1c6fd7 | 2339 | array_size = GET_ARRAY_SIZE (last_basic_block_for_fn (cfun)); |
5ed6ace5 | 2340 | bbd = XNEWVEC (bbro_basic_block_data, array_size); |
aa634f11 JZ |
2341 | for (i = 0; i < array_size; i++) |
2342 | { | |
2343 | bbd[i].start_of_trace = -1; | |
2344 | bbd[i].end_of_trace = -1; | |
bcc708fc MM |
2345 | bbd[i].in_trace = -1; |
2346 | bbd[i].visited = 0; | |
375374ad | 2347 | bbd[i].priority = -1; |
aa634f11 JZ |
2348 | bbd[i].heap = NULL; |
2349 | bbd[i].node = NULL; | |
2350 | } | |
2351 | ||
0cae8d31 | 2352 | traces = XNEWVEC (struct trace, n_basic_blocks_for_fn (cfun)); |
aa634f11 JZ |
2353 | n_traces = 0; |
2354 | find_traces (&n_traces, traces); | |
2355 | connect_traces (n_traces, traces); | |
2356 | FREE (traces); | |
2357 | FREE (bbd); | |
af71fa39 SB |
2358 | } |
2359 | ||
64979e04 | 2360 | /* Order edges by execution frequency, higher first. */ |
248c16c4 | 2361 | |
64979e04 AM |
2362 | static int |
2363 | edge_order (const void *ve1, const void *ve2) | |
248c16c4 | 2364 | { |
64979e04 AM |
2365 | edge e1 = *(const edge *) ve1; |
2366 | edge e2 = *(const edge *) ve2; | |
2367 | profile_count c1 = e1->count (); | |
2368 | profile_count c2 = e2->count (); | |
2369 | /* Since profile_count::operator< does not establish a strict weak order | |
2370 | in presence of uninitialized counts, use 'max': this makes them appear | |
2371 | as if having execution frequency less than any initialized count. */ | |
2372 | profile_count m = c1.max (c2); | |
2373 | return (m == c2) - (m == c1); | |
248c16c4 SB |
2374 | } |
2375 | ||
2376 | /* Reorder basic blocks using the "simple" algorithm. This tries to | |
2377 | maximize the dynamic number of branches that are fallthrough, without | |
2378 | copying instructions. The algorithm is greedy, looking at the most | |
2379 | frequently executed branch first. */ | |
2380 | ||
2381 | static void | |
2382 | reorder_basic_blocks_simple (void) | |
2383 | { | |
2384 | if (dump_file) | |
2385 | fprintf (dump_file, "\nReordering with the \"simple\" algorithm.\n\n"); | |
2386 | ||
2387 | edge *edges = new edge[2 * n_basic_blocks_for_fn (cfun)]; | |
2388 | ||
2389 | /* First, collect all edges that can be optimized by reordering blocks: | |
2390 | simple jumps and conditional jumps, as well as the function entry edge. */ | |
2391 | ||
2392 | int n = 0; | |
2393 | edges[n++] = EDGE_SUCC (ENTRY_BLOCK_PTR_FOR_FN (cfun), 0); | |
2394 | ||
2395 | basic_block bb; | |
2396 | FOR_EACH_BB_FN (bb, cfun) | |
2397 | { | |
2398 | rtx_insn *end = BB_END (bb); | |
2399 | ||
2400 | if (computed_jump_p (end) || tablejump_p (end, NULL, NULL)) | |
2401 | continue; | |
2402 | ||
2403 | /* We cannot optimize asm goto. */ | |
2404 | if (JUMP_P (end) && extract_asm_operands (end)) | |
2405 | continue; | |
2406 | ||
d4c8d5ed SB |
2407 | if (single_succ_p (bb)) |
2408 | edges[n++] = EDGE_SUCC (bb, 0); | |
2409 | else if (any_condjump_p (end)) | |
248c16c4 | 2410 | { |
c70f0ca2 SB |
2411 | edge e0 = EDGE_SUCC (bb, 0); |
2412 | edge e1 = EDGE_SUCC (bb, 1); | |
2413 | /* When optimizing for size it is best to keep the original | |
2414 | fallthrough edges. */ | |
2415 | if (e1->flags & EDGE_FALLTHRU) | |
2416 | std::swap (e0, e1); | |
2417 | edges[n++] = e0; | |
2418 | edges[n++] = e1; | |
248c16c4 | 2419 | } |
248c16c4 SB |
2420 | } |
2421 | ||
c70f0ca2 SB |
2422 | /* Sort the edges, the most desirable first. When optimizing for size |
2423 | all edges are equally desirable. */ | |
248c16c4 | 2424 | |
c70f0ca2 | 2425 | if (optimize_function_for_speed_p (cfun)) |
64979e04 | 2426 | gcc_stablesort (edges, n, sizeof *edges, edge_order); |
248c16c4 SB |
2427 | |
2428 | /* Now decide which of those edges to make fallthrough edges. We set | |
2429 | BB_VISITED if a block already has a fallthrough successor assigned | |
2430 | to it. We make ->AUX of an endpoint point to the opposite endpoint | |
2431 | of a sequence of blocks that fall through, and ->AUX will be NULL | |
2432 | for a block that is in such a sequence but not an endpoint anymore. | |
2433 | ||
2434 | To start with, everything points to itself, nothing is assigned yet. */ | |
2435 | ||
2436 | FOR_ALL_BB_FN (bb, cfun) | |
f6a1ec62 RB |
2437 | { |
2438 | bb->aux = bb; | |
2439 | bb->flags &= ~BB_VISITED; | |
2440 | } | |
248c16c4 SB |
2441 | |
2442 | EXIT_BLOCK_PTR_FOR_FN (cfun)->aux = 0; | |
2443 | ||
2444 | /* Now for all edges, the most desirable first, see if that edge can | |
2445 | connect two sequences. If it can, update AUX and BB_VISITED; if it | |
2446 | cannot, zero out the edge in the table. */ | |
2447 | ||
2448 | for (int j = 0; j < n; j++) | |
2449 | { | |
2450 | edge e = edges[j]; | |
2451 | ||
2452 | basic_block tail_a = e->src; | |
2453 | basic_block head_b = e->dest; | |
2454 | basic_block head_a = (basic_block) tail_a->aux; | |
2455 | basic_block tail_b = (basic_block) head_b->aux; | |
2456 | ||
2457 | /* An edge cannot connect two sequences if: | |
2458 | - it crosses partitions; | |
2459 | - its src is not a current endpoint; | |
2460 | - its dest is not a current endpoint; | |
2461 | - or, it would create a loop. */ | |
2462 | ||
2463 | if (e->flags & EDGE_CROSSING | |
2464 | || tail_a->flags & BB_VISITED | |
2465 | || !tail_b | |
2466 | || (!(head_b->flags & BB_VISITED) && head_b != tail_b) | |
2467 | || tail_a == tail_b) | |
2468 | { | |
2469 | edges[j] = 0; | |
2470 | continue; | |
2471 | } | |
2472 | ||
2473 | tail_a->aux = 0; | |
2474 | head_b->aux = 0; | |
2475 | head_a->aux = tail_b; | |
2476 | tail_b->aux = head_a; | |
2477 | tail_a->flags |= BB_VISITED; | |
2478 | } | |
2479 | ||
2480 | /* Put the pieces together, in the same order that the start blocks of | |
2481 | the sequences already had. The hot/cold partitioning gives a little | |
2482 | complication: as a first pass only do this for blocks in the same | |
2483 | partition as the start block, and (if there is anything left to do) | |
2484 | in a second pass handle the other partition. */ | |
2485 | ||
2486 | basic_block last_tail = (basic_block) ENTRY_BLOCK_PTR_FOR_FN (cfun)->aux; | |
2487 | ||
63e1159c JJ |
2488 | int current_partition |
2489 | = BB_PARTITION (last_tail == ENTRY_BLOCK_PTR_FOR_FN (cfun) | |
2490 | ? EDGE_SUCC (ENTRY_BLOCK_PTR_FOR_FN (cfun), 0)->dest | |
2491 | : last_tail); | |
248c16c4 SB |
2492 | bool need_another_pass = true; |
2493 | ||
2494 | for (int pass = 0; pass < 2 && need_another_pass; pass++) | |
2495 | { | |
2496 | need_another_pass = false; | |
2497 | ||
2498 | FOR_EACH_BB_FN (bb, cfun) | |
2499 | if ((bb->flags & BB_VISITED && bb->aux) || bb->aux == bb) | |
2500 | { | |
2501 | if (BB_PARTITION (bb) != current_partition) | |
2502 | { | |
2503 | need_another_pass = true; | |
2504 | continue; | |
2505 | } | |
2506 | ||
2507 | last_tail->aux = bb; | |
2508 | last_tail = (basic_block) bb->aux; | |
2509 | } | |
2510 | ||
2511 | current_partition ^= BB_HOT_PARTITION | BB_COLD_PARTITION; | |
2512 | } | |
2513 | ||
2514 | last_tail->aux = 0; | |
2515 | ||
2516 | /* Finally, link all the chosen fallthrough edges. */ | |
2517 | ||
2518 | for (int j = 0; j < n; j++) | |
2519 | if (edges[j]) | |
2520 | edges[j]->src->aux = edges[j]->dest; | |
2521 | ||
2522 | delete[] edges; | |
2523 | ||
2524 | /* If the entry edge no longer falls through we have to make a new | |
2525 | block so it can do so again. */ | |
2526 | ||
2527 | edge e = EDGE_SUCC (ENTRY_BLOCK_PTR_FOR_FN (cfun), 0); | |
2528 | if (e->dest != ENTRY_BLOCK_PTR_FOR_FN (cfun)->aux) | |
2529 | { | |
2530 | force_nonfallthru (e); | |
2531 | e->src->aux = ENTRY_BLOCK_PTR_FOR_FN (cfun)->aux; | |
248c16c4 SB |
2532 | } |
2533 | } | |
2534 | ||
af71fa39 SB |
2535 | /* Reorder basic blocks. The main entry point to this file. */ |
2536 | ||
2537 | static void | |
2538 | reorder_basic_blocks (void) | |
2539 | { | |
2540 | gcc_assert (current_ir_type () == IR_RTL_CFGLAYOUT); | |
2541 | ||
2542 | if (n_basic_blocks_for_fn (cfun) <= NUM_FIXED_BLOCKS + 1) | |
2543 | return; | |
2544 | ||
2545 | set_edge_can_fallthru_flag (); | |
2546 | mark_dfs_back_edges (); | |
2547 | ||
59faab7c SB |
2548 | switch (flag_reorder_blocks_algorithm) |
2549 | { | |
2550 | case REORDER_BLOCKS_ALGORITHM_SIMPLE: | |
2551 | reorder_basic_blocks_simple (); | |
2552 | break; | |
2553 | ||
2554 | case REORDER_BLOCKS_ALGORITHM_STC: | |
2555 | reorder_basic_blocks_software_trace_cache (); | |
2556 | break; | |
2557 | ||
2558 | default: | |
2559 | gcc_unreachable (); | |
2560 | } | |
402209ff | 2561 | |
ad21dab7 SB |
2562 | relink_block_chain (/*stay_in_cfglayout_mode=*/true); |
2563 | ||
c263766c | 2564 | if (dump_file) |
532aafad SB |
2565 | { |
2566 | if (dump_flags & TDF_DETAILS) | |
2567 | dump_reg_info (dump_file); | |
2568 | dump_flow_info (dump_file, dump_flags); | |
2569 | } | |
402209ff | 2570 | |
af205f67 TJ |
2571 | /* Signal that rtl_verify_flow_info_1 can now verify that there |
2572 | is at most one switch between hot/cold sections. */ | |
2573 | crtl->bb_reorder_complete = true; | |
295ae817 | 2574 | } |
bbcb0c05 | 2575 | |
87c8b4be CT |
2576 | /* Determine which partition the first basic block in the function |
2577 | belongs to, then find the first basic block in the current function | |
2578 | that belongs to a different section, and insert a | |
2579 | NOTE_INSN_SWITCH_TEXT_SECTIONS note immediately before it in the | |
2580 | instruction stream. When writing out the assembly code, | |
2581 | encountering this note will make the compiler switch between the | |
2582 | hot and cold text sections. */ | |
2583 | ||
3371a64f | 2584 | void |
87c8b4be CT |
2585 | insert_section_boundary_note (void) |
2586 | { | |
2587 | basic_block bb; | |
3371a64f TJ |
2588 | bool switched_sections = false; |
2589 | int current_partition = 0; | |
c22cacf3 | 2590 | |
3371a64f | 2591 | if (!crtl->has_bb_partition) |
9645d434 BS |
2592 | return; |
2593 | ||
11cd3bed | 2594 | FOR_EACH_BB_FN (bb, cfun) |
87c8b4be | 2595 | { |
3371a64f TJ |
2596 | if (!current_partition) |
2597 | current_partition = BB_PARTITION (bb); | |
2598 | if (BB_PARTITION (bb) != current_partition) | |
87c8b4be | 2599 | { |
3371a64f TJ |
2600 | gcc_assert (!switched_sections); |
2601 | switched_sections = true; | |
2602 | emit_note_before (NOTE_INSN_SWITCH_TEXT_SECTIONS, BB_HEAD (bb)); | |
2603 | current_partition = BB_PARTITION (bb); | |
87c8b4be CT |
2604 | } |
2605 | } | |
64d620ef JJ |
2606 | |
2607 | /* Make sure crtl->has_bb_partition matches reality even if bbpart finds | |
2608 | some hot and some cold basic blocks, but later one of those kinds is | |
2609 | optimized away. */ | |
2610 | crtl->has_bb_partition = switched_sections; | |
87c8b4be CT |
2611 | } |
2612 | ||
27a4cd48 DM |
2613 | namespace { |
2614 | ||
2615 | const pass_data pass_data_reorder_blocks = | |
427b8bb8 | 2616 | { |
27a4cd48 DM |
2617 | RTL_PASS, /* type */ |
2618 | "bbro", /* name */ | |
2619 | OPTGROUP_NONE, /* optinfo_flags */ | |
27a4cd48 DM |
2620 | TV_REORDER_BLOCKS, /* tv_id */ |
2621 | 0, /* properties_required */ | |
2622 | 0, /* properties_provided */ | |
2623 | 0, /* properties_destroyed */ | |
2624 | 0, /* todo_flags_start */ | |
3bea341f | 2625 | 0, /* todo_flags_finish */ |
427b8bb8 EB |
2626 | }; |
2627 | ||
27a4cd48 DM |
2628 | class pass_reorder_blocks : public rtl_opt_pass |
2629 | { | |
2630 | public: | |
c3284718 RS |
2631 | pass_reorder_blocks (gcc::context *ctxt) |
2632 | : rtl_opt_pass (pass_data_reorder_blocks, ctxt) | |
27a4cd48 DM |
2633 | {} |
2634 | ||
2635 | /* opt_pass methods: */ | |
1a3d085c TS |
2636 | virtual bool gate (function *) |
2637 | { | |
2638 | if (targetm.cannot_modify_jumps_p ()) | |
2639 | return false; | |
2640 | return (optimize > 0 | |
2641 | && (flag_reorder_blocks || flag_reorder_blocks_and_partition)); | |
2642 | } | |
2643 | ||
be55bfe6 | 2644 | virtual unsigned int execute (function *); |
27a4cd48 DM |
2645 | |
2646 | }; // class pass_reorder_blocks | |
2647 | ||
be55bfe6 TS |
2648 | unsigned int |
2649 | pass_reorder_blocks::execute (function *fun) | |
2650 | { | |
2651 | basic_block bb; | |
2652 | ||
2653 | /* Last attempt to optimize CFG, as scheduling, peepholing and insn | |
2654 | splitting possibly introduced more crossjumping opportunities. */ | |
2655 | cfg_layout_initialize (CLEANUP_EXPENSIVE); | |
2656 | ||
2657 | reorder_basic_blocks (); | |
3ff0dc17 | 2658 | cleanup_cfg (CLEANUP_EXPENSIVE | CLEANUP_NO_PARTITIONING); |
be55bfe6 TS |
2659 | |
2660 | FOR_EACH_BB_FN (bb, fun) | |
2661 | if (bb->next_bb != EXIT_BLOCK_PTR_FOR_FN (fun)) | |
2662 | bb->aux = bb->next_bb; | |
2663 | cfg_layout_finalize (); | |
2664 | ||
fdc5359f MM |
2665 | FOR_EACH_BB_FN (bb, fun) |
2666 | df_recompute_luids (bb); | |
be55bfe6 TS |
2667 | return 0; |
2668 | } | |
2669 | ||
27a4cd48 DM |
2670 | } // anon namespace |
2671 | ||
2672 | rtl_opt_pass * | |
2673 | make_pass_reorder_blocks (gcc::context *ctxt) | |
2674 | { | |
2675 | return new pass_reorder_blocks (ctxt); | |
2676 | } | |
2677 | ||
cd36a451 SB |
2678 | /* Duplicate a block (that we already know ends in a computed jump) into its |
2679 | predecessors, where possible. Return whether anything is changed. */ | |
2680 | static bool | |
2681 | maybe_duplicate_computed_goto (basic_block bb, int max_size) | |
2682 | { | |
2683 | if (single_pred_p (bb)) | |
2684 | return false; | |
2685 | ||
2686 | /* Make sure that the block is small enough. */ | |
2687 | rtx_insn *insn; | |
2688 | FOR_BB_INSNS (bb, insn) | |
2689 | if (INSN_P (insn)) | |
2690 | { | |
2691 | max_size -= get_attr_min_length (insn); | |
2692 | if (max_size < 0) | |
2693 | return false; | |
2694 | } | |
2695 | ||
2696 | bool changed = false; | |
2697 | edge e; | |
2698 | edge_iterator ei; | |
2699 | for (ei = ei_start (bb->preds); (e = ei_safe_edge (ei)); ) | |
2700 | { | |
2701 | basic_block pred = e->src; | |
2702 | ||
2703 | /* Do not duplicate BB into PRED if that is the last predecessor, or if | |
2704 | we cannot merge a copy of BB with PRED. */ | |
2705 | if (single_pred_p (bb) | |
2706 | || !single_succ_p (pred) | |
2707 | || e->flags & EDGE_COMPLEX | |
2708 | || pred->index < NUM_FIXED_BLOCKS | |
2709 | || (JUMP_P (BB_END (pred)) && !simplejump_p (BB_END (pred))) | |
2710 | || (JUMP_P (BB_END (pred)) && CROSSING_JUMP_P (BB_END (pred)))) | |
2711 | { | |
2712 | ei_next (&ei); | |
2713 | continue; | |
2714 | } | |
2715 | ||
2716 | if (dump_file) | |
2717 | fprintf (dump_file, "Duplicating computed goto bb %d into bb %d\n", | |
2718 | bb->index, e->src->index); | |
2719 | ||
2720 | /* Remember if PRED can be duplicated; if so, the copy of BB merged | |
2721 | with PRED can be duplicated as well. */ | |
2722 | bool can_dup_more = can_duplicate_block_p (pred); | |
2723 | ||
2724 | /* Make a copy of BB, merge it into PRED. */ | |
2725 | basic_block copy = duplicate_block (bb, e, NULL); | |
2726 | emit_barrier_after_bb (copy); | |
2727 | reorder_insns_nobb (BB_HEAD (copy), BB_END (copy), BB_END (pred)); | |
2728 | merge_blocks (pred, copy); | |
2729 | ||
2730 | changed = true; | |
2731 | ||
2732 | /* Try to merge the resulting merged PRED into further predecessors. */ | |
2733 | if (can_dup_more) | |
2734 | maybe_duplicate_computed_goto (pred, max_size); | |
2735 | } | |
2736 | ||
2737 | return changed; | |
2738 | } | |
2739 | ||
bbcb0c05 SB |
2740 | /* Duplicate the blocks containing computed gotos. This basically unfactors |
2741 | computed gotos that were factored early on in the compilation process to | |
cd36a451 SB |
2742 | speed up edge based data flow. We used to not unfactor them again, which |
2743 | can seriously pessimize code with many computed jumps in the source code, | |
2744 | such as interpreters. See e.g. PR15242. */ | |
2745 | static void | |
2746 | duplicate_computed_gotos (function *fun) | |
2747 | { | |
2748 | /* We are estimating the length of uncond jump insn only once | |
2749 | since the code for getting the insn length always returns | |
2750 | the minimal length now. */ | |
2751 | if (uncond_jump_length == 0) | |
2752 | uncond_jump_length = get_uncond_jump_length (); | |
2753 | ||
2754 | /* Never copy a block larger than this. */ | |
2755 | int max_size | |
028d4092 | 2756 | = uncond_jump_length * param_max_goto_duplication_insns; |
cd36a451 SB |
2757 | |
2758 | bool changed = false; | |
2759 | ||
2760 | /* Try to duplicate all blocks that end in a computed jump and that | |
2761 | can be duplicated at all. */ | |
2762 | basic_block bb; | |
2763 | FOR_EACH_BB_FN (bb, fun) | |
2764 | if (computed_jump_p (BB_END (bb)) && can_duplicate_block_p (bb)) | |
2765 | changed |= maybe_duplicate_computed_goto (bb, max_size); | |
2766 | ||
2767 | /* Duplicating blocks will redirect edges and may cause hot blocks | |
2768 | previously reached by both hot and cold blocks to become dominated | |
2769 | only by cold blocks. */ | |
2770 | if (changed) | |
2771 | fixup_partitions (); | |
2772 | } | |
bbcb0c05 | 2773 | |
be55bfe6 | 2774 | namespace { |
ef330312 | 2775 | |
be55bfe6 TS |
2776 | const pass_data pass_data_duplicate_computed_gotos = |
2777 | { | |
2778 | RTL_PASS, /* type */ | |
2779 | "compgotos", /* name */ | |
2780 | OPTGROUP_NONE, /* optinfo_flags */ | |
be55bfe6 TS |
2781 | TV_REORDER_BLOCKS, /* tv_id */ |
2782 | 0, /* properties_required */ | |
2783 | 0, /* properties_provided */ | |
2784 | 0, /* properties_destroyed */ | |
2785 | 0, /* todo_flags_start */ | |
3bea341f | 2786 | 0, /* todo_flags_finish */ |
be55bfe6 TS |
2787 | }; |
2788 | ||
2789 | class pass_duplicate_computed_gotos : public rtl_opt_pass | |
2790 | { | |
2791 | public: | |
2792 | pass_duplicate_computed_gotos (gcc::context *ctxt) | |
2793 | : rtl_opt_pass (pass_data_duplicate_computed_gotos, ctxt) | |
2794 | {} | |
2795 | ||
2796 | /* opt_pass methods: */ | |
2797 | virtual bool gate (function *); | |
2798 | virtual unsigned int execute (function *); | |
2799 | ||
2800 | }; // class pass_duplicate_computed_gotos | |
2801 | ||
2802 | bool | |
2803 | pass_duplicate_computed_gotos::gate (function *fun) | |
2804 | { | |
2805 | if (targetm.cannot_modify_jumps_p ()) | |
2806 | return false; | |
2807 | return (optimize > 0 | |
2808 | && flag_expensive_optimizations | |
2809 | && ! optimize_function_for_size_p (fun)); | |
2810 | } | |
2811 | ||
2812 | unsigned int | |
2813 | pass_duplicate_computed_gotos::execute (function *fun) | |
bbcb0c05 | 2814 | { |
cd36a451 | 2815 | duplicate_computed_gotos (fun); |
bbcb0c05 | 2816 | |
c2924966 | 2817 | return 0; |
bbcb0c05 | 2818 | } |
750054a2 | 2819 | |
27a4cd48 DM |
2820 | } // anon namespace |
2821 | ||
2822 | rtl_opt_pass * | |
2823 | make_pass_duplicate_computed_gotos (gcc::context *ctxt) | |
2824 | { | |
2825 | return new pass_duplicate_computed_gotos (ctxt); | |
2826 | } | |
2827 | ||
750054a2 CT |
2828 | /* This function is the main 'entrance' for the optimization that |
2829 | partitions hot and cold basic blocks into separate sections of the | |
2830 | .o file (to improve performance and cache locality). Ideally it | |
2831 | would be called after all optimizations that rearrange the CFG have | |
2832 | been called. However part of this optimization may introduce new | |
2833 | register usage, so it must be called before register allocation has | |
2834 | occurred. This means that this optimization is actually called | |
8e8d5162 CT |
2835 | well before the optimization that reorders basic blocks (see |
2836 | function above). | |
750054a2 CT |
2837 | |
2838 | This optimization checks the feedback information to determine | |
87c8b4be CT |
2839 | which basic blocks are hot/cold, updates flags on the basic blocks |
2840 | to indicate which section they belong in. This information is | |
2841 | later used for writing out sections in the .o file. Because hot | |
2842 | and cold sections can be arbitrarily large (within the bounds of | |
2843 | memory), far beyond the size of a single function, it is necessary | |
2844 | to fix up all edges that cross section boundaries, to make sure the | |
2845 | instructions used can actually span the required distance. The | |
2846 | fixes are described below. | |
8e8d5162 CT |
2847 | |
2848 | Fall-through edges must be changed into jumps; it is not safe or | |
2849 | legal to fall through across a section boundary. Whenever a | |
2850 | fall-through edge crossing a section boundary is encountered, a new | |
2851 | basic block is inserted (in the same section as the fall-through | |
2852 | source), and the fall through edge is redirected to the new basic | |
2853 | block. The new basic block contains an unconditional jump to the | |
2854 | original fall-through target. (If the unconditional jump is | |
2855 | insufficient to cross section boundaries, that is dealt with a | |
2856 | little later, see below). | |
2857 | ||
2858 | In order to deal with architectures that have short conditional | |
2859 | branches (which cannot span all of memory) we take any conditional | |
2860 | jump that attempts to cross a section boundary and add a level of | |
2861 | indirection: it becomes a conditional jump to a new basic block, in | |
2862 | the same section. The new basic block contains an unconditional | |
2863 | jump to the original target, in the other section. | |
2864 | ||
2865 | For those architectures whose unconditional branch is also | |
2866 | incapable of reaching all of memory, those unconditional jumps are | |
2867 | converted into indirect jumps, through a register. | |
2868 | ||
2869 | IMPORTANT NOTE: This optimization causes some messy interactions | |
2870 | with the cfg cleanup optimizations; those optimizations want to | |
2871 | merge blocks wherever possible, and to collapse indirect jump | |
2872 | sequences (change "A jumps to B jumps to C" directly into "A jumps | |
2873 | to C"). Those optimizations can undo the jump fixes that | |
2874 | partitioning is required to make (see above), in order to ensure | |
2875 | that jumps attempting to cross section boundaries are really able | |
2876 | to cover whatever distance the jump requires (on many architectures | |
2877 | conditional or unconditional jumps are not able to reach all of | |
2878 | memory). Therefore tests have to be inserted into each such | |
2879 | optimization to make sure that it does not undo stuff necessary to | |
2880 | cross partition boundaries. This would be much less of a problem | |
2881 | if we could perform this optimization later in the compilation, but | |
2882 | unfortunately the fact that we may need to create indirect jumps | |
2883 | (through registers) requires that this optimization be performed | |
ea6136a2 | 2884 | before register allocation. |
750054a2 | 2885 | |
ea6136a2 RH |
2886 | Hot and cold basic blocks are partitioned and put in separate |
2887 | sections of the .o file, to reduce paging and improve cache | |
2888 | performance (hopefully). This can result in bits of code from the | |
2889 | same function being widely separated in the .o file. However this | |
2890 | is not obvious to the current bb structure. Therefore we must take | |
2891 | care to ensure that: 1). There are no fall_thru edges that cross | |
2892 | between sections; 2). For those architectures which have "short" | |
2893 | conditional branches, all conditional branches that attempt to | |
2894 | cross between sections are converted to unconditional branches; | |
2895 | and, 3). For those architectures which have "short" unconditional | |
2896 | branches, all unconditional branches that attempt to cross between | |
2897 | sections are converted to indirect jumps. | |
2898 | ||
2899 | The code for fixing up fall_thru edges that cross between hot and | |
2900 | cold basic blocks does so by creating new basic blocks containing | |
2901 | unconditional branches to the appropriate label in the "other" | |
2902 | section. The new basic block is then put in the same (hot or cold) | |
2903 | section as the original conditional branch, and the fall_thru edge | |
2904 | is modified to fall into the new basic block instead. By adding | |
2905 | this level of indirection we end up with only unconditional branches | |
2906 | crossing between hot and cold sections. | |
2907 | ||
2908 | Conditional branches are dealt with by adding a level of indirection. | |
2909 | A new basic block is added in the same (hot/cold) section as the | |
2910 | conditional branch, and the conditional branch is retargeted to the | |
2911 | new basic block. The new basic block contains an unconditional branch | |
2912 | to the original target of the conditional branch (in the other section). | |
2913 | ||
2914 | Unconditional branches are dealt with by converting them into | |
2915 | indirect jumps. */ | |
2916 | ||
be55bfe6 TS |
2917 | namespace { |
2918 | ||
2919 | const pass_data pass_data_partition_blocks = | |
2920 | { | |
2921 | RTL_PASS, /* type */ | |
2922 | "bbpart", /* name */ | |
2923 | OPTGROUP_NONE, /* optinfo_flags */ | |
be55bfe6 TS |
2924 | TV_REORDER_BLOCKS, /* tv_id */ |
2925 | PROP_cfglayout, /* properties_required */ | |
2926 | 0, /* properties_provided */ | |
2927 | 0, /* properties_destroyed */ | |
2928 | 0, /* todo_flags_start */ | |
2929 | 0, /* todo_flags_finish */ | |
2930 | }; | |
2931 | ||
2932 | class pass_partition_blocks : public rtl_opt_pass | |
2933 | { | |
2934 | public: | |
2935 | pass_partition_blocks (gcc::context *ctxt) | |
2936 | : rtl_opt_pass (pass_data_partition_blocks, ctxt) | |
2937 | {} | |
2938 | ||
2939 | /* opt_pass methods: */ | |
2940 | virtual bool gate (function *); | |
2941 | virtual unsigned int execute (function *); | |
2942 | ||
2943 | }; // class pass_partition_blocks | |
2944 | ||
2945 | bool | |
2946 | pass_partition_blocks::gate (function *fun) | |
2947 | { | |
2948 | /* The optimization to partition hot/cold basic blocks into separate | |
2949 | sections of the .o file does not work well with linkonce or with | |
28c84b62 JJ |
2950 | user defined section attributes or with naked attribute. Don't call |
2951 | it if either case arises. */ | |
be55bfe6 TS |
2952 | return (flag_reorder_blocks_and_partition |
2953 | && optimize | |
cd98b1cf | 2954 | /* See pass_reorder_blocks::gate. We should not partition if |
be55bfe6 TS |
2955 | we are going to omit the reordering. */ |
2956 | && optimize_function_for_speed_p (fun) | |
cf288ed3 | 2957 | && !DECL_COMDAT_GROUP (current_function_decl) |
8f72ce2c | 2958 | && !lookup_attribute ("section", DECL_ATTRIBUTES (fun->decl)) |
28c84b62 | 2959 | && !lookup_attribute ("naked", DECL_ATTRIBUTES (fun->decl)) |
8f72ce2c JJ |
2960 | /* Workaround a bug in GDB where read_partial_die doesn't cope |
2961 | with DIEs with DW_AT_ranges, see PR81115. */ | |
2962 | && !(in_lto_p && MAIN_NAME_P (DECL_NAME (fun->decl)))); | |
be55bfe6 TS |
2963 | } |
2964 | ||
2965 | unsigned | |
2966 | pass_partition_blocks::execute (function *fun) | |
750054a2 | 2967 | { |
9771b263 | 2968 | vec<edge> crossing_edges; |
c22cacf3 | 2969 | |
be55bfe6 | 2970 | if (n_basic_blocks_for_fn (fun) <= NUM_FIXED_BLOCKS + 1) |
ea6136a2 RH |
2971 | return 0; |
2972 | ||
0be7e7a6 RH |
2973 | df_set_flags (DF_DEFER_INSN_RESCAN); |
2974 | ||
ea6136a2 | 2975 | crossing_edges = find_rarely_executed_basic_blocks_and_crossing_edges (); |
9771b263 | 2976 | if (!crossing_edges.exists ()) |
2fdaed89 JJ |
2977 | /* Make sure to process deferred rescans and clear changeable df flags. */ |
2978 | return TODO_df_finish; | |
ea6136a2 | 2979 | |
af205f67 TJ |
2980 | crtl->has_bb_partition = true; |
2981 | ||
ea6136a2 RH |
2982 | /* Make sure the source of any crossing edge ends in a jump and the |
2983 | destination of any crossing edge has a label. */ | |
2984 | add_labels_and_missing_jumps (crossing_edges); | |
2985 | ||
2986 | /* Convert all crossing fall_thru edges to non-crossing fall | |
2987 | thrus to unconditional jumps (that jump to the original fall | |
073a8998 | 2988 | through dest). */ |
ea6136a2 RH |
2989 | fix_up_fall_thru_edges (); |
2990 | ||
2991 | /* If the architecture does not have conditional branches that can | |
2992 | span all of memory, convert crossing conditional branches into | |
2993 | crossing unconditional branches. */ | |
2994 | if (!HAS_LONG_COND_BRANCH) | |
2995 | fix_crossing_conditional_branches (); | |
c22cacf3 | 2996 | |
ea6136a2 RH |
2997 | /* If the architecture does not have unconditional branches that |
2998 | can span all of memory, convert crossing unconditional branches | |
2999 | into indirect jumps. Since adding an indirect jump also adds | |
3000 | a new register usage, update the register usage information as | |
3001 | well. */ | |
3002 | if (!HAS_LONG_UNCOND_BRANCH) | |
3003 | fix_crossing_unconditional_branches (); | |
750054a2 | 3004 | |
339ba33b | 3005 | update_crossing_jump_flags (); |
750054a2 | 3006 | |
b71d7f85 JJ |
3007 | /* Clear bb->aux fields that the above routines were using. */ |
3008 | clear_aux_for_blocks (); | |
3009 | ||
9771b263 | 3010 | crossing_edges.release (); |
c22cacf3 | 3011 | |
0be7e7a6 RH |
3012 | /* ??? FIXME: DF generates the bb info for a block immediately. |
3013 | And by immediately, I mean *during* creation of the block. | |
3014 | ||
3015 | #0 df_bb_refs_collect | |
3016 | #1 in df_bb_refs_record | |
3017 | #2 in create_basic_block_structure | |
3018 | ||
3019 | Which means that the bb_has_eh_pred test in df_bb_refs_collect | |
3020 | will *always* fail, because no edges can have been added to the | |
3021 | block yet. Which of course means we don't add the right | |
3022 | artificial refs, which means we fail df_verify (much) later. | |
3023 | ||
3024 | Cleanest solution would seem to make DF_DEFER_INSN_RESCAN imply | |
3025 | that we also shouldn't grab data from the new blocks those new | |
3026 | insns are in either. In this way one can create the block, link | |
3027 | it up properly, and have everything Just Work later, when deferred | |
3028 | insns are processed. | |
3029 | ||
3030 | In the meantime, we have no other option but to throw away all | |
3031 | of the DF data and recompute it all. */ | |
be55bfe6 | 3032 | if (fun->eh->lp_array) |
0be7e7a6 RH |
3033 | { |
3034 | df_finish_pass (true); | |
3035 | df_scan_alloc (NULL); | |
3036 | df_scan_blocks (); | |
3037 | /* Not all post-landing pads use all of the EH_RETURN_DATA_REGNO | |
3038 | data. We blindly generated all of them when creating the new | |
3039 | landing pad. Delete those assignments we don't use. */ | |
3040 | df_set_flags (DF_LR_RUN_DCE); | |
3041 | df_analyze (); | |
3042 | } | |
3043 | ||
2fdaed89 JJ |
3044 | /* Make sure to process deferred rescans and clear changeable df flags. */ |
3045 | return TODO_df_finish; | |
750054a2 | 3046 | } |
ef330312 | 3047 | |
27a4cd48 DM |
3048 | } // anon namespace |
3049 | ||
3050 | rtl_opt_pass * | |
3051 | make_pass_partition_blocks (gcc::context *ctxt) | |
3052 | { | |
3053 | return new pass_partition_blocks (ctxt); | |
3054 | } |