]>
Commit | Line | Data |
---|---|---|
7a31a7bd | 1 | /* Instruction scheduling pass. |
711789cc | 2 | Copyright (C) 1992-2013 Free Software Foundation, Inc. |
7a31a7bd | 3 | Contributed by Michael Tiemann (tiemann@cygnus.com) Enhanced by, |
4 | and currently maintained by, Jim Wilson (wilson@cygnus.com) | |
5 | ||
f12b58b3 | 6 | This file is part of GCC. |
7a31a7bd | 7 | |
f12b58b3 | 8 | GCC is free software; you can redistribute it and/or modify it under |
9 | the terms of the GNU General Public License as published by the Free | |
8c4c00c1 | 10 | Software Foundation; either version 3, or (at your option) any later |
f12b58b3 | 11 | version. |
7a31a7bd | 12 | |
f12b58b3 | 13 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
14 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
7a31a7bd | 15 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
16 | for more details. | |
17 | ||
18 | You should have received a copy of the GNU General Public License | |
8c4c00c1 | 19 | along with GCC; see the file COPYING3. If not see |
20 | <http://www.gnu.org/licenses/>. */ | |
7a31a7bd | 21 | |
22 | /* This pass implements list scheduling within basic blocks. It is | |
23 | run twice: (1) after flow analysis, but before register allocation, | |
24 | and (2) after register allocation. | |
25 | ||
26 | The first run performs interblock scheduling, moving insns between | |
27 | different blocks in the same "region", and the second runs only | |
28 | basic block scheduling. | |
29 | ||
30 | Interblock motions performed are useful motions and speculative | |
31 | motions, including speculative loads. Motions requiring code | |
32 | duplication are not supported. The identification of motion type | |
33 | and the check for validity of speculative motions requires | |
34 | construction and analysis of the function's control flow graph. | |
35 | ||
36 | The main entry point for this pass is schedule_insns(), called for | |
37 | each function. The work of the scheduler is organized in three | |
38 | levels: (1) function level: insns are subject to splitting, | |
39 | control-flow-graph is constructed, regions are computed (after | |
40 | reload, each region is of one block), (2) region level: control | |
41 | flow graph attributes required for interblock scheduling are | |
42 | computed (dominators, reachability, etc.), data dependences and | |
43 | priorities are computed, and (3) block level: insns in the block | |
44 | are actually scheduled. */ | |
45 | \f | |
46 | #include "config.h" | |
47 | #include "system.h" | |
805e22b2 | 48 | #include "coretypes.h" |
49 | #include "tm.h" | |
0b205f4c | 50 | #include "diagnostic-core.h" |
7a31a7bd | 51 | #include "rtl.h" |
52 | #include "tm_p.h" | |
53 | #include "hard-reg-set.h" | |
7a31a7bd | 54 | #include "regs.h" |
55 | #include "function.h" | |
56 | #include "flags.h" | |
57 | #include "insn-config.h" | |
58 | #include "insn-attr.h" | |
59 | #include "except.h" | |
7a31a7bd | 60 | #include "recog.h" |
4c50e1f4 | 61 | #include "params.h" |
7a31a7bd | 62 | #include "sched-int.h" |
e1ab7874 | 63 | #include "sel-sched.h" |
bea4bad2 | 64 | #include "target.h" |
77fce4cd | 65 | #include "tree-pass.h" |
3072d30e | 66 | #include "dbgcnt.h" |
7fb47f9f | 67 | |
cda0a5f5 | 68 | #ifdef INSN_SCHEDULING |
e1ab7874 | 69 | |
7a31a7bd | 70 | /* Some accessor macros for h_i_d members only used within this file. */ |
e1ab7874 | 71 | #define FED_BY_SPEC_LOAD(INSN) (HID (INSN)->fed_by_spec_load) |
72 | #define IS_LOAD_INSN(INSN) (HID (insn)->is_load_insn) | |
7a31a7bd | 73 | |
7a31a7bd | 74 | /* nr_inter/spec counts interblock/speculative motion for the function. */ |
75 | static int nr_inter, nr_spec; | |
76 | ||
60b8c5b3 | 77 | static int is_cfg_nonregular (void); |
7a31a7bd | 78 | |
79 | /* Number of regions in the procedure. */ | |
e1ab7874 | 80 | int nr_regions = 0; |
7a31a7bd | 81 | |
82 | /* Table of region descriptions. */ | |
e1ab7874 | 83 | region *rgn_table = NULL; |
7a31a7bd | 84 | |
85 | /* Array of lists of regions' blocks. */ | |
e1ab7874 | 86 | int *rgn_bb_table = NULL; |
7a31a7bd | 87 | |
88 | /* Topological order of blocks in the region (if b2 is reachable from | |
89 | b1, block_to_bb[b2] > block_to_bb[b1]). Note: A basic block is | |
90 | always referred to by either block or b, while its topological | |
917bbcab | 91 | order name (in the region) is referred to by bb. */ |
e1ab7874 | 92 | int *block_to_bb = NULL; |
7a31a7bd | 93 | |
94 | /* The number of the region containing a block. */ | |
e1ab7874 | 95 | int *containing_rgn = NULL; |
96 | ||
97 | /* ebb_head [i] - is index in rgn_bb_table of the head basic block of i'th ebb. | |
98 | Currently we can get a ebb only through splitting of currently | |
99 | scheduling block, therefore, we don't need ebb_head array for every region, | |
100 | hence, its sufficient to hold it for current one only. */ | |
101 | int *ebb_head = NULL; | |
7a31a7bd | 102 | |
6c0b81cb | 103 | /* The minimum probability of reaching a source block so that it will be |
104 | considered for speculative scheduling. */ | |
105 | static int min_spec_prob; | |
106 | ||
e1ab7874 | 107 | static void find_single_block_region (bool); |
aae97b21 | 108 | static void find_rgns (void); |
4c50e1f4 | 109 | static bool too_large (int, int *, int *); |
7a31a7bd | 110 | |
7a31a7bd | 111 | /* Blocks of the current region being scheduled. */ |
e1ab7874 | 112 | int current_nr_blocks; |
113 | int current_blocks; | |
7a31a7bd | 114 | |
e1ab7874 | 115 | /* A speculative motion requires checking live information on the path |
116 | from 'source' to 'target'. The split blocks are those to be checked. | |
117 | After a speculative motion, live information should be modified in | |
118 | the 'update' blocks. | |
6a1cdb4d | 119 | |
e1ab7874 | 120 | Lists of split and update blocks for each candidate of the current |
121 | target are in array bblst_table. */ | |
122 | static basic_block *bblst_table; | |
123 | static int bblst_size, bblst_last; | |
7a31a7bd | 124 | |
bbd0cfb1 | 125 | /* Arrays that hold the DFA state at the end of a basic block, to re-use |
126 | as the initial state at the start of successor blocks. The BB_STATE | |
127 | array holds the actual DFA state, and BB_STATE_ARRAY[I] is a pointer | |
128 | into BB_STATE for basic block I. FIXME: This should be a vec. */ | |
129 | static char *bb_state_array = NULL; | |
130 | static state_t *bb_state = NULL; | |
0a15667c | 131 | |
7a31a7bd | 132 | /* Target info declarations. |
133 | ||
134 | The block currently being scheduled is referred to as the "target" block, | |
135 | while other blocks in the region from which insns can be moved to the | |
136 | target are called "source" blocks. The candidate structure holds info | |
137 | about such sources: are they valid? Speculative? Etc. */ | |
aae97b21 | 138 | typedef struct |
139 | { | |
140 | basic_block *first_member; | |
141 | int nr_members; | |
142 | } | |
143 | bblst; | |
144 | ||
7a31a7bd | 145 | typedef struct |
146 | { | |
147 | char is_valid; | |
148 | char is_speculative; | |
149 | int src_prob; | |
150 | bblst split_bbs; | |
151 | bblst update_bbs; | |
152 | } | |
153 | candidate; | |
154 | ||
155 | static candidate *candidate_table; | |
e1ab7874 | 156 | #define IS_VALID(src) (candidate_table[src].is_valid) |
157 | #define IS_SPECULATIVE(src) (candidate_table[src].is_speculative) | |
158 | #define IS_SPECULATIVE_INSN(INSN) \ | |
159 | (IS_SPECULATIVE (BLOCK_TO_BB (BLOCK_NUM (INSN)))) | |
7a31a7bd | 160 | #define SRC_PROB(src) ( candidate_table[src].src_prob ) |
161 | ||
162 | /* The bb being currently scheduled. */ | |
e1ab7874 | 163 | int target_bb; |
7a31a7bd | 164 | |
165 | /* List of edges. */ | |
aae97b21 | 166 | typedef struct |
167 | { | |
168 | edge *first_member; | |
169 | int nr_members; | |
170 | } | |
171 | edgelst; | |
172 | ||
173 | static edge *edgelst_table; | |
174 | static int edgelst_last; | |
175 | ||
176 | static void extract_edgelst (sbitmap, edgelst *); | |
177 | ||
7a31a7bd | 178 | /* Target info functions. */ |
60b8c5b3 | 179 | static void split_edges (int, int, edgelst *); |
180 | static void compute_trg_info (int); | |
181 | void debug_candidate (int); | |
182 | void debug_candidates (int); | |
7a31a7bd | 183 | |
79cafa9e | 184 | /* Dominators array: dom[i] contains the sbitmap of dominators of |
7a31a7bd | 185 | bb i in the region. */ |
79cafa9e | 186 | static sbitmap *dom; |
7a31a7bd | 187 | |
188 | /* bb 0 is the only region entry. */ | |
189 | #define IS_RGN_ENTRY(bb) (!bb) | |
190 | ||
191 | /* Is bb_src dominated by bb_trg. */ | |
192 | #define IS_DOMINATED(bb_src, bb_trg) \ | |
08b7917c | 193 | ( bitmap_bit_p (dom[bb_src], bb_trg) ) |
7a31a7bd | 194 | |
6c0b81cb | 195 | /* Probability: Prob[i] is an int in [0, REG_BR_PROB_BASE] which is |
196 | the probability of bb i relative to the region entry. */ | |
197 | static int *prob; | |
7a31a7bd | 198 | |
199 | /* Bit-set of edges, where bit i stands for edge i. */ | |
79cafa9e | 200 | typedef sbitmap edgeset; |
7a31a7bd | 201 | |
202 | /* Number of edges in the region. */ | |
203 | static int rgn_nr_edges; | |
204 | ||
205 | /* Array of size rgn_nr_edges. */ | |
aae97b21 | 206 | static edge *rgn_edges; |
7a31a7bd | 207 | |
208 | /* Mapping from each edge in the graph to its number in the rgn. */ | |
aae97b21 | 209 | #define EDGE_TO_BIT(edge) ((int)(size_t)(edge)->aux) |
210 | #define SET_EDGE_TO_BIT(edge,nr) ((edge)->aux = (void *)(size_t)(nr)) | |
7a31a7bd | 211 | |
212 | /* The split edges of a source bb is different for each target | |
213 | bb. In order to compute this efficiently, the 'potential-split edges' | |
214 | are computed for each bb prior to scheduling a region. This is actually | |
215 | the split edges of each bb relative to the region entry. | |
216 | ||
217 | pot_split[bb] is the set of potential split edges of bb. */ | |
218 | static edgeset *pot_split; | |
219 | ||
220 | /* For every bb, a set of its ancestor edges. */ | |
221 | static edgeset *ancestor_edges; | |
222 | ||
7a31a7bd | 223 | #define INSN_PROBABILITY(INSN) (SRC_PROB (BLOCK_TO_BB (BLOCK_NUM (INSN)))) |
7a31a7bd | 224 | |
7a31a7bd | 225 | /* Speculative scheduling functions. */ |
60b8c5b3 | 226 | static int check_live_1 (int, rtx); |
227 | static void update_live_1 (int, rtx); | |
60b8c5b3 | 228 | static int is_pfree (rtx, int, int); |
229 | static int find_conditional_protection (rtx, int); | |
230 | static int is_conditionally_protected (rtx, int, int); | |
231 | static int is_prisky (rtx, int, int); | |
232 | static int is_exception_free (rtx, int, int); | |
233 | ||
234 | static bool sets_likely_spilled (rtx); | |
81a410b1 | 235 | static void sets_likely_spilled_1 (rtx, const_rtx, void *); |
60b8c5b3 | 236 | static void add_branch_dependences (rtx, rtx); |
93f6b030 | 237 | static void compute_block_dependences (int); |
60b8c5b3 | 238 | |
60b8c5b3 | 239 | static void schedule_region (int); |
60b8c5b3 | 240 | static void concat_insn_mem_list (rtx, rtx, rtx *, rtx *); |
68e419a1 | 241 | static void propagate_deps (int, struct deps_desc *); |
60b8c5b3 | 242 | static void free_pending_lists (void); |
7a31a7bd | 243 | |
244 | /* Functions for construction of the control flow graph. */ | |
245 | ||
246 | /* Return 1 if control flow graph should not be constructed, 0 otherwise. | |
247 | ||
248 | We decide not to build the control flow graph if there is possibly more | |
aae97b21 | 249 | than one entry to the function, if computed branches exist, if we |
250 | have nonlocal gotos, or if we have an unreachable loop. */ | |
7a31a7bd | 251 | |
252 | static int | |
60b8c5b3 | 253 | is_cfg_nonregular (void) |
7a31a7bd | 254 | { |
4c26117a | 255 | basic_block b; |
7a31a7bd | 256 | rtx insn; |
7a31a7bd | 257 | |
258 | /* If we have a label that could be the target of a nonlocal goto, then | |
259 | the cfg is not well structured. */ | |
260 | if (nonlocal_goto_handler_labels) | |
261 | return 1; | |
262 | ||
263 | /* If we have any forced labels, then the cfg is not well structured. */ | |
264 | if (forced_labels) | |
265 | return 1; | |
266 | ||
7a31a7bd | 267 | /* If we have exception handlers, then we consider the cfg not well |
3072d30e | 268 | structured. ?!? We should be able to handle this now that we |
269 | compute an accurate cfg for EH. */ | |
8f8dcce4 | 270 | if (current_function_has_exception_handlers ()) |
7a31a7bd | 271 | return 1; |
272 | ||
19d2fe05 | 273 | /* If we have insns which refer to labels as non-jumped-to operands, |
274 | then we consider the cfg not well structured. */ | |
4c26117a | 275 | FOR_EACH_BB (b) |
cd6dccd3 | 276 | FOR_BB_INSNS (b, insn) |
7a31a7bd | 277 | { |
a8d1dae0 | 278 | rtx note, next, set, dest; |
19d2fe05 | 279 | |
cd6dccd3 | 280 | /* If this function has a computed jump, then we consider the cfg |
281 | not well structured. */ | |
19d2fe05 | 282 | if (JUMP_P (insn) && computed_jump_p (insn)) |
cd6dccd3 | 283 | return 1; |
a8d1dae0 | 284 | |
285 | if (!INSN_P (insn)) | |
286 | continue; | |
287 | ||
288 | note = find_reg_note (insn, REG_LABEL_OPERAND, NULL_RTX); | |
289 | if (note == NULL_RTX) | |
290 | continue; | |
291 | ||
292 | /* For that label not to be seen as a referred-to label, this | |
293 | must be a single-set which is feeding a jump *only*. This | |
294 | could be a conditional jump with the label split off for | |
295 | machine-specific reasons or a casesi/tablejump. */ | |
296 | next = next_nonnote_insn (insn); | |
297 | if (next == NULL_RTX | |
298 | || !JUMP_P (next) | |
299 | || (JUMP_LABEL (next) != XEXP (note, 0) | |
300 | && find_reg_note (next, REG_LABEL_TARGET, | |
301 | XEXP (note, 0)) == NULL_RTX) | |
302 | || BLOCK_FOR_INSN (insn) != BLOCK_FOR_INSN (next)) | |
303 | return 1; | |
304 | ||
305 | set = single_set (insn); | |
306 | if (set == NULL_RTX) | |
307 | return 1; | |
308 | ||
309 | dest = SET_DEST (set); | |
310 | if (!REG_P (dest) || !dead_or_set_p (next, dest)) | |
311 | return 1; | |
7a31a7bd | 312 | } |
313 | ||
7a31a7bd | 314 | /* Unreachable loops with more than one basic block are detected |
315 | during the DFS traversal in find_rgns. | |
316 | ||
317 | Unreachable loops with a single block are detected here. This | |
318 | test is redundant with the one in find_rgns, but it's much | |
aae97b21 | 319 | cheaper to go ahead and catch the trivial case here. */ |
4c26117a | 320 | FOR_EACH_BB (b) |
7a31a7bd | 321 | { |
cd665a06 | 322 | if (EDGE_COUNT (b->preds) == 0 |
ea091dfd | 323 | || (single_pred_p (b) |
324 | && single_pred (b) == b)) | |
aae97b21 | 325 | return 1; |
7a31a7bd | 326 | } |
327 | ||
aae97b21 | 328 | /* All the tests passed. Consider the cfg well structured. */ |
329 | return 0; | |
7a31a7bd | 330 | } |
331 | ||
aae97b21 | 332 | /* Extract list of edges from a bitmap containing EDGE_TO_BIT bits. */ |
7a31a7bd | 333 | |
334 | static void | |
aae97b21 | 335 | extract_edgelst (sbitmap set, edgelst *el) |
7a31a7bd | 336 | { |
86c1585a | 337 | unsigned int i = 0; |
3e790786 | 338 | sbitmap_iterator sbi; |
7a31a7bd | 339 | |
aae97b21 | 340 | /* edgelst table space is reused in each call to extract_edgelst. */ |
341 | edgelst_last = 0; | |
7a31a7bd | 342 | |
aae97b21 | 343 | el->first_member = &edgelst_table[edgelst_last]; |
344 | el->nr_members = 0; | |
7a31a7bd | 345 | |
346 | /* Iterate over each word in the bitset. */ | |
0d211963 | 347 | EXECUTE_IF_SET_IN_BITMAP (set, 0, i, sbi) |
3e790786 | 348 | { |
349 | edgelst_table[edgelst_last++] = rgn_edges[i]; | |
350 | el->nr_members++; | |
351 | } | |
7a31a7bd | 352 | } |
353 | ||
354 | /* Functions for the construction of regions. */ | |
355 | ||
356 | /* Print the regions, for debugging purposes. Callable from debugger. */ | |
357 | ||
4b987fac | 358 | DEBUG_FUNCTION void |
60b8c5b3 | 359 | debug_regions (void) |
7a31a7bd | 360 | { |
361 | int rgn, bb; | |
362 | ||
363 | fprintf (sched_dump, "\n;; ------------ REGIONS ----------\n\n"); | |
364 | for (rgn = 0; rgn < nr_regions; rgn++) | |
365 | { | |
366 | fprintf (sched_dump, ";;\trgn %d nr_blocks %d:\n", rgn, | |
367 | rgn_table[rgn].rgn_nr_blocks); | |
368 | fprintf (sched_dump, ";;\tbb/block: "); | |
369 | ||
6a1cdb4d | 370 | /* We don't have ebb_head initialized yet, so we can't use |
371 | BB_TO_BLOCK (). */ | |
372 | current_blocks = RGN_BLOCKS (rgn); | |
7a31a7bd | 373 | |
6a1cdb4d | 374 | for (bb = 0; bb < rgn_table[rgn].rgn_nr_blocks; bb++) |
375 | fprintf (sched_dump, " %d/%d ", bb, rgn_bb_table[current_blocks + bb]); | |
7a31a7bd | 376 | |
377 | fprintf (sched_dump, "\n\n"); | |
378 | } | |
379 | } | |
380 | ||
e1ab7874 | 381 | /* Print the region's basic blocks. */ |
382 | ||
4b987fac | 383 | DEBUG_FUNCTION void |
e1ab7874 | 384 | debug_region (int rgn) |
385 | { | |
386 | int bb; | |
387 | ||
388 | fprintf (stderr, "\n;; ------------ REGION %d ----------\n\n", rgn); | |
389 | fprintf (stderr, ";;\trgn %d nr_blocks %d:\n", rgn, | |
390 | rgn_table[rgn].rgn_nr_blocks); | |
391 | fprintf (stderr, ";;\tbb/block: "); | |
392 | ||
393 | /* We don't have ebb_head initialized yet, so we can't use | |
394 | BB_TO_BLOCK (). */ | |
395 | current_blocks = RGN_BLOCKS (rgn); | |
396 | ||
397 | for (bb = 0; bb < rgn_table[rgn].rgn_nr_blocks; bb++) | |
398 | fprintf (stderr, " %d/%d ", bb, rgn_bb_table[current_blocks + bb]); | |
399 | ||
400 | fprintf (stderr, "\n\n"); | |
401 | ||
402 | for (bb = 0; bb < rgn_table[rgn].rgn_nr_blocks; bb++) | |
403 | { | |
5147ec07 | 404 | dump_bb (stderr, BASIC_BLOCK (rgn_bb_table[current_blocks + bb]), |
bec2cf98 | 405 | 0, TDF_SLIM | TDF_BLOCKS); |
e1ab7874 | 406 | fprintf (stderr, "\n"); |
407 | } | |
408 | ||
409 | fprintf (stderr, "\n"); | |
410 | ||
411 | } | |
412 | ||
413 | /* True when a bb with index BB_INDEX contained in region RGN. */ | |
414 | static bool | |
415 | bb_in_region_p (int bb_index, int rgn) | |
416 | { | |
417 | int i; | |
418 | ||
419 | for (i = 0; i < rgn_table[rgn].rgn_nr_blocks; i++) | |
420 | if (rgn_bb_table[current_blocks + i] == bb_index) | |
421 | return true; | |
422 | ||
423 | return false; | |
424 | } | |
425 | ||
426 | /* Dump region RGN to file F using dot syntax. */ | |
427 | void | |
428 | dump_region_dot (FILE *f, int rgn) | |
429 | { | |
430 | int i; | |
431 | ||
432 | fprintf (f, "digraph Region_%d {\n", rgn); | |
433 | ||
434 | /* We don't have ebb_head initialized yet, so we can't use | |
435 | BB_TO_BLOCK (). */ | |
436 | current_blocks = RGN_BLOCKS (rgn); | |
437 | ||
438 | for (i = 0; i < rgn_table[rgn].rgn_nr_blocks; i++) | |
439 | { | |
440 | edge e; | |
441 | edge_iterator ei; | |
442 | int src_bb_num = rgn_bb_table[current_blocks + i]; | |
161dfa6e | 443 | basic_block bb = BASIC_BLOCK (src_bb_num); |
e1ab7874 | 444 | |
445 | FOR_EACH_EDGE (e, ei, bb->succs) | |
446 | if (bb_in_region_p (e->dest->index, rgn)) | |
447 | fprintf (f, "\t%d -> %d\n", src_bb_num, e->dest->index); | |
448 | } | |
449 | fprintf (f, "}\n"); | |
450 | } | |
451 | ||
452 | /* The same, but first open a file specified by FNAME. */ | |
48e1416a | 453 | void |
e1ab7874 | 454 | dump_region_dot_file (const char *fname, int rgn) |
455 | { | |
456 | FILE *f = fopen (fname, "wt"); | |
457 | dump_region_dot (f, rgn); | |
458 | fclose (f); | |
459 | } | |
460 | ||
7a31a7bd | 461 | /* Build a single block region for each basic block in the function. |
462 | This allows for using the same code for interblock and basic block | |
463 | scheduling. */ | |
464 | ||
465 | static void | |
e1ab7874 | 466 | find_single_block_region (bool ebbs_p) |
7a31a7bd | 467 | { |
e1ab7874 | 468 | basic_block bb, ebb_start; |
469 | int i = 0; | |
4c5da238 | 470 | |
4c26117a | 471 | nr_regions = 0; |
472 | ||
e1ab7874 | 473 | if (ebbs_p) { |
474 | int probability_cutoff; | |
475 | if (profile_info && flag_branch_probabilities) | |
476 | probability_cutoff = PARAM_VALUE (TRACER_MIN_BRANCH_PROBABILITY_FEEDBACK); | |
477 | else | |
478 | probability_cutoff = PARAM_VALUE (TRACER_MIN_BRANCH_PROBABILITY); | |
479 | probability_cutoff = REG_BR_PROB_BASE / 100 * probability_cutoff; | |
480 | ||
481 | FOR_EACH_BB (ebb_start) | |
482 | { | |
483 | RGN_NR_BLOCKS (nr_regions) = 0; | |
484 | RGN_BLOCKS (nr_regions) = i; | |
485 | RGN_DONT_CALC_DEPS (nr_regions) = 0; | |
486 | RGN_HAS_REAL_EBB (nr_regions) = 0; | |
487 | ||
488 | for (bb = ebb_start; ; bb = bb->next_bb) | |
489 | { | |
490 | edge e; | |
e1ab7874 | 491 | |
492 | rgn_bb_table[i] = bb->index; | |
493 | RGN_NR_BLOCKS (nr_regions)++; | |
494 | CONTAINING_RGN (bb->index) = nr_regions; | |
495 | BLOCK_TO_BB (bb->index) = i - RGN_BLOCKS (nr_regions); | |
496 | i++; | |
497 | ||
498 | if (bb->next_bb == EXIT_BLOCK_PTR | |
499 | || LABEL_P (BB_HEAD (bb->next_bb))) | |
500 | break; | |
48e1416a | 501 | |
7f58c05e | 502 | e = find_fallthru_edge (bb->succs); |
e1ab7874 | 503 | if (! e) |
504 | break; | |
505 | if (e->probability <= probability_cutoff) | |
506 | break; | |
507 | } | |
508 | ||
509 | ebb_start = bb; | |
510 | nr_regions++; | |
511 | } | |
512 | } | |
513 | else | |
514 | FOR_EACH_BB (bb) | |
515 | { | |
516 | rgn_bb_table[nr_regions] = bb->index; | |
517 | RGN_NR_BLOCKS (nr_regions) = 1; | |
518 | RGN_BLOCKS (nr_regions) = nr_regions; | |
519 | RGN_DONT_CALC_DEPS (nr_regions) = 0; | |
520 | RGN_HAS_REAL_EBB (nr_regions) = 0; | |
521 | ||
522 | CONTAINING_RGN (bb->index) = nr_regions; | |
523 | BLOCK_TO_BB (bb->index) = 0; | |
524 | nr_regions++; | |
525 | } | |
526 | } | |
527 | ||
528 | /* Estimate number of the insns in the BB. */ | |
529 | static int | |
530 | rgn_estimate_number_of_insns (basic_block bb) | |
531 | { | |
9845d120 | 532 | int count; |
533 | ||
534 | count = INSN_LUID (BB_END (bb)) - INSN_LUID (BB_HEAD (bb)); | |
535 | ||
536 | if (MAY_HAVE_DEBUG_INSNS) | |
537 | { | |
538 | rtx insn; | |
539 | ||
540 | FOR_BB_INSNS (bb, insn) | |
541 | if (DEBUG_INSN_P (insn)) | |
542 | count--; | |
543 | } | |
544 | ||
545 | return count; | |
7a31a7bd | 546 | } |
547 | ||
548 | /* Update number of blocks and the estimate for number of insns | |
4c50e1f4 | 549 | in the region. Return true if the region is "too large" for interblock |
550 | scheduling (compile time considerations). */ | |
7a31a7bd | 551 | |
4c50e1f4 | 552 | static bool |
60b8c5b3 | 553 | too_large (int block, int *num_bbs, int *num_insns) |
7a31a7bd | 554 | { |
555 | (*num_bbs)++; | |
e1ab7874 | 556 | (*num_insns) += (common_sched_info->estimate_number_of_insns |
557 | (BASIC_BLOCK (block))); | |
4c50e1f4 | 558 | |
559 | return ((*num_bbs > PARAM_VALUE (PARAM_MAX_SCHED_REGION_BLOCKS)) | |
560 | || (*num_insns > PARAM_VALUE (PARAM_MAX_SCHED_REGION_INSNS))); | |
7a31a7bd | 561 | } |
562 | ||
563 | /* Update_loop_relations(blk, hdr): Check if the loop headed by max_hdr[blk] | |
564 | is still an inner loop. Put in max_hdr[blk] the header of the most inner | |
565 | loop containing blk. */ | |
40734805 | 566 | #define UPDATE_LOOP_RELATIONS(blk, hdr) \ |
567 | { \ | |
568 | if (max_hdr[blk] == -1) \ | |
569 | max_hdr[blk] = hdr; \ | |
570 | else if (dfs_nr[max_hdr[blk]] > dfs_nr[hdr]) \ | |
08b7917c | 571 | bitmap_clear_bit (inner, hdr); \ |
40734805 | 572 | else if (dfs_nr[max_hdr[blk]] < dfs_nr[hdr]) \ |
573 | { \ | |
08b7917c | 574 | bitmap_clear_bit (inner,max_hdr[blk]); \ |
40734805 | 575 | max_hdr[blk] = hdr; \ |
576 | } \ | |
7a31a7bd | 577 | } |
578 | ||
579 | /* Find regions for interblock scheduling. | |
580 | ||
581 | A region for scheduling can be: | |
582 | ||
583 | * A loop-free procedure, or | |
584 | ||
585 | * A reducible inner loop, or | |
586 | ||
587 | * A basic block not contained in any other region. | |
588 | ||
589 | ?!? In theory we could build other regions based on extended basic | |
590 | blocks or reverse extended basic blocks. Is it worth the trouble? | |
591 | ||
592 | Loop blocks that form a region are put into the region's block list | |
593 | in topological order. | |
594 | ||
595 | This procedure stores its results into the following global (ick) variables | |
596 | ||
597 | * rgn_nr | |
598 | * rgn_table | |
599 | * rgn_bb_table | |
600 | * block_to_bb | |
601 | * containing region | |
602 | ||
603 | We use dominator relationships to avoid making regions out of non-reducible | |
604 | loops. | |
605 | ||
606 | This procedure needs to be converted to work on pred/succ lists instead | |
607 | of edge tables. That would simplify it somewhat. */ | |
608 | ||
609 | static void | |
e1ab7874 | 610 | haifa_find_rgns (void) |
7a31a7bd | 611 | { |
aae97b21 | 612 | int *max_hdr, *dfs_nr, *degree; |
7a31a7bd | 613 | char no_loops = 1; |
614 | int node, child, loop_head, i, head, tail; | |
8deb7557 | 615 | int count = 0, sp, idx = 0; |
aae97b21 | 616 | edge_iterator current_edge; |
617 | edge_iterator *stack; | |
7a31a7bd | 618 | int num_bbs, num_insns, unreachable; |
619 | int too_large_failure; | |
4c26117a | 620 | basic_block bb; |
7a31a7bd | 621 | |
7a31a7bd | 622 | /* Note if a block is a natural loop header. */ |
623 | sbitmap header; | |
624 | ||
edc2a478 | 625 | /* Note if a block is a natural inner loop header. */ |
7a31a7bd | 626 | sbitmap inner; |
627 | ||
628 | /* Note if a block is in the block queue. */ | |
629 | sbitmap in_queue; | |
630 | ||
631 | /* Note if a block is in the block queue. */ | |
632 | sbitmap in_stack; | |
633 | ||
7a31a7bd | 634 | /* Perform a DFS traversal of the cfg. Identify loop headers, inner loops |
635 | and a mapping from block to its loop header (if the block is contained | |
636 | in a loop, else -1). | |
637 | ||
638 | Store results in HEADER, INNER, and MAX_HDR respectively, these will | |
639 | be used as inputs to the second traversal. | |
640 | ||
641 | STACK, SP and DFS_NR are only used during the first traversal. */ | |
642 | ||
643 | /* Allocate and initialize variables for the first traversal. */ | |
4c36ffe6 | 644 | max_hdr = XNEWVEC (int, last_basic_block); |
645 | dfs_nr = XCNEWVEC (int, last_basic_block); | |
646 | stack = XNEWVEC (edge_iterator, n_edges); | |
7a31a7bd | 647 | |
f20183e6 | 648 | inner = sbitmap_alloc (last_basic_block); |
53c5d9d4 | 649 | bitmap_ones (inner); |
7a31a7bd | 650 | |
f20183e6 | 651 | header = sbitmap_alloc (last_basic_block); |
53c5d9d4 | 652 | bitmap_clear (header); |
7a31a7bd | 653 | |
f20183e6 | 654 | in_queue = sbitmap_alloc (last_basic_block); |
53c5d9d4 | 655 | bitmap_clear (in_queue); |
7a31a7bd | 656 | |
f20183e6 | 657 | in_stack = sbitmap_alloc (last_basic_block); |
53c5d9d4 | 658 | bitmap_clear (in_stack); |
7a31a7bd | 659 | |
3c0a32c9 | 660 | for (i = 0; i < last_basic_block; i++) |
7a31a7bd | 661 | max_hdr[i] = -1; |
662 | ||
aae97b21 | 663 | #define EDGE_PASSED(E) (ei_end_p ((E)) || ei_edge ((E))->aux) |
664 | #define SET_EDGE_PASSED(E) (ei_edge ((E))->aux = ei_edge ((E))) | |
665 | ||
7a31a7bd | 666 | /* DFS traversal to find inner loops in the cfg. */ |
667 | ||
ea091dfd | 668 | current_edge = ei_start (single_succ (ENTRY_BLOCK_PTR)->succs); |
7a31a7bd | 669 | sp = -1; |
aae97b21 | 670 | |
7a31a7bd | 671 | while (1) |
672 | { | |
aae97b21 | 673 | if (EDGE_PASSED (current_edge)) |
7a31a7bd | 674 | { |
675 | /* We have reached a leaf node or a node that was already | |
676 | processed. Pop edges off the stack until we find | |
677 | an edge that has not yet been processed. */ | |
aae97b21 | 678 | while (sp >= 0 && EDGE_PASSED (current_edge)) |
7a31a7bd | 679 | { |
680 | /* Pop entry off the stack. */ | |
681 | current_edge = stack[sp--]; | |
aae97b21 | 682 | node = ei_edge (current_edge)->src->index; |
683 | gcc_assert (node != ENTRY_BLOCK); | |
684 | child = ei_edge (current_edge)->dest->index; | |
685 | gcc_assert (child != EXIT_BLOCK); | |
08b7917c | 686 | bitmap_clear_bit (in_stack, child); |
687 | if (max_hdr[child] >= 0 && bitmap_bit_p (in_stack, max_hdr[child])) | |
7a31a7bd | 688 | UPDATE_LOOP_RELATIONS (node, max_hdr[child]); |
aae97b21 | 689 | ei_next (¤t_edge); |
7a31a7bd | 690 | } |
691 | ||
692 | /* See if have finished the DFS tree traversal. */ | |
aae97b21 | 693 | if (sp < 0 && EDGE_PASSED (current_edge)) |
7a31a7bd | 694 | break; |
695 | ||
696 | /* Nope, continue the traversal with the popped node. */ | |
697 | continue; | |
698 | } | |
699 | ||
700 | /* Process a node. */ | |
aae97b21 | 701 | node = ei_edge (current_edge)->src->index; |
702 | gcc_assert (node != ENTRY_BLOCK); | |
08b7917c | 703 | bitmap_set_bit (in_stack, node); |
7a31a7bd | 704 | dfs_nr[node] = ++count; |
705 | ||
aae97b21 | 706 | /* We don't traverse to the exit block. */ |
707 | child = ei_edge (current_edge)->dest->index; | |
708 | if (child == EXIT_BLOCK) | |
709 | { | |
710 | SET_EDGE_PASSED (current_edge); | |
711 | ei_next (¤t_edge); | |
712 | continue; | |
713 | } | |
714 | ||
7a31a7bd | 715 | /* If the successor is in the stack, then we've found a loop. |
716 | Mark the loop, if it is not a natural loop, then it will | |
717 | be rejected during the second traversal. */ | |
08b7917c | 718 | if (bitmap_bit_p (in_stack, child)) |
7a31a7bd | 719 | { |
720 | no_loops = 0; | |
08b7917c | 721 | bitmap_set_bit (header, child); |
7a31a7bd | 722 | UPDATE_LOOP_RELATIONS (node, child); |
aae97b21 | 723 | SET_EDGE_PASSED (current_edge); |
724 | ei_next (¤t_edge); | |
7a31a7bd | 725 | continue; |
726 | } | |
727 | ||
728 | /* If the child was already visited, then there is no need to visit | |
729 | it again. Just update the loop relationships and restart | |
730 | with a new edge. */ | |
731 | if (dfs_nr[child]) | |
732 | { | |
08b7917c | 733 | if (max_hdr[child] >= 0 && bitmap_bit_p (in_stack, max_hdr[child])) |
7a31a7bd | 734 | UPDATE_LOOP_RELATIONS (node, max_hdr[child]); |
aae97b21 | 735 | SET_EDGE_PASSED (current_edge); |
736 | ei_next (¤t_edge); | |
7a31a7bd | 737 | continue; |
738 | } | |
739 | ||
740 | /* Push an entry on the stack and continue DFS traversal. */ | |
741 | stack[++sp] = current_edge; | |
aae97b21 | 742 | SET_EDGE_PASSED (current_edge); |
743 | current_edge = ei_start (ei_edge (current_edge)->dest->succs); | |
744 | } | |
745 | ||
746 | /* Reset ->aux field used by EDGE_PASSED. */ | |
747 | FOR_ALL_BB (bb) | |
748 | { | |
749 | edge_iterator ei; | |
750 | edge e; | |
751 | FOR_EACH_EDGE (e, ei, bb->succs) | |
752 | e->aux = NULL; | |
7a31a7bd | 753 | } |
754 | ||
aae97b21 | 755 | |
7a31a7bd | 756 | /* Another check for unreachable blocks. The earlier test in |
757 | is_cfg_nonregular only finds unreachable blocks that do not | |
758 | form a loop. | |
759 | ||
760 | The DFS traversal will mark every block that is reachable from | |
761 | the entry node by placing a nonzero value in dfs_nr. Thus if | |
762 | dfs_nr is zero for any block, then it must be unreachable. */ | |
763 | unreachable = 0; | |
4c26117a | 764 | FOR_EACH_BB (bb) |
765 | if (dfs_nr[bb->index] == 0) | |
7a31a7bd | 766 | { |
767 | unreachable = 1; | |
768 | break; | |
769 | } | |
770 | ||
771 | /* Gross. To avoid wasting memory, the second pass uses the dfs_nr array | |
772 | to hold degree counts. */ | |
773 | degree = dfs_nr; | |
774 | ||
4c26117a | 775 | FOR_EACH_BB (bb) |
aae97b21 | 776 | degree[bb->index] = EDGE_COUNT (bb->preds); |
7a31a7bd | 777 | |
778 | /* Do not perform region scheduling if there are any unreachable | |
779 | blocks. */ | |
780 | if (!unreachable) | |
781 | { | |
4bfe0e7b | 782 | int *queue, *degree1 = NULL; |
783 | /* We use EXTENDED_RGN_HEADER as an addition to HEADER and put | |
784 | there basic blocks, which are forced to be region heads. | |
48e1416a | 785 | This is done to try to assemble few smaller regions |
4bfe0e7b | 786 | from a too_large region. */ |
787 | sbitmap extended_rgn_header = NULL; | |
788 | bool extend_regions_p; | |
7a31a7bd | 789 | |
790 | if (no_loops) | |
08b7917c | 791 | bitmap_set_bit (header, 0); |
7a31a7bd | 792 | |
de132707 | 793 | /* Second traversal:find reducible inner loops and topologically sort |
7a31a7bd | 794 | block of each region. */ |
795 | ||
4c36ffe6 | 796 | queue = XNEWVEC (int, n_basic_blocks); |
48e1416a | 797 | |
4bfe0e7b | 798 | extend_regions_p = PARAM_VALUE (PARAM_MAX_SCHED_EXTEND_REGIONS_ITERS) > 0; |
799 | if (extend_regions_p) | |
800 | { | |
f7f3687c | 801 | degree1 = XNEWVEC (int, last_basic_block); |
4bfe0e7b | 802 | extended_rgn_header = sbitmap_alloc (last_basic_block); |
53c5d9d4 | 803 | bitmap_clear (extended_rgn_header); |
4bfe0e7b | 804 | } |
7a31a7bd | 805 | |
806 | /* Find blocks which are inner loop headers. We still have non-reducible | |
807 | loops to consider at this point. */ | |
4c26117a | 808 | FOR_EACH_BB (bb) |
7a31a7bd | 809 | { |
08b7917c | 810 | if (bitmap_bit_p (header, bb->index) && bitmap_bit_p (inner, bb->index)) |
7a31a7bd | 811 | { |
812 | edge e; | |
cd665a06 | 813 | edge_iterator ei; |
4c26117a | 814 | basic_block jbb; |
7a31a7bd | 815 | |
816 | /* Now check that the loop is reducible. We do this separate | |
817 | from finding inner loops so that we do not find a reducible | |
818 | loop which contains an inner non-reducible loop. | |
819 | ||
820 | A simple way to find reducible/natural loops is to verify | |
821 | that each block in the loop is dominated by the loop | |
822 | header. | |
823 | ||
824 | If there exists a block that is not dominated by the loop | |
825 | header, then the block is reachable from outside the loop | |
826 | and thus the loop is not a natural loop. */ | |
4c26117a | 827 | FOR_EACH_BB (jbb) |
7a31a7bd | 828 | { |
829 | /* First identify blocks in the loop, except for the loop | |
830 | entry block. */ | |
4c26117a | 831 | if (bb->index == max_hdr[jbb->index] && bb != jbb) |
7a31a7bd | 832 | { |
833 | /* Now verify that the block is dominated by the loop | |
834 | header. */ | |
0051c76a | 835 | if (!dominated_by_p (CDI_DOMINATORS, jbb, bb)) |
7a31a7bd | 836 | break; |
837 | } | |
838 | } | |
839 | ||
840 | /* If we exited the loop early, then I is the header of | |
841 | a non-reducible loop and we should quit processing it | |
842 | now. */ | |
4c26117a | 843 | if (jbb != EXIT_BLOCK_PTR) |
7a31a7bd | 844 | continue; |
845 | ||
846 | /* I is a header of an inner loop, or block 0 in a subroutine | |
847 | with no loops at all. */ | |
848 | head = tail = -1; | |
849 | too_large_failure = 0; | |
4c26117a | 850 | loop_head = max_hdr[bb->index]; |
7a31a7bd | 851 | |
4bfe0e7b | 852 | if (extend_regions_p) |
48e1416a | 853 | /* We save degree in case when we meet a too_large region |
854 | and cancel it. We need a correct degree later when | |
4bfe0e7b | 855 | calling extend_rgns. */ |
856 | memcpy (degree1, degree, last_basic_block * sizeof (int)); | |
48e1416a | 857 | |
7a31a7bd | 858 | /* Decrease degree of all I's successors for topological |
859 | ordering. */ | |
cd665a06 | 860 | FOR_EACH_EDGE (e, ei, bb->succs) |
7a31a7bd | 861 | if (e->dest != EXIT_BLOCK_PTR) |
b3d6de89 | 862 | --degree[e->dest->index]; |
7a31a7bd | 863 | |
864 | /* Estimate # insns, and count # blocks in the region. */ | |
865 | num_bbs = 1; | |
e1ab7874 | 866 | num_insns = common_sched_info->estimate_number_of_insns (bb); |
7a31a7bd | 867 | |
868 | /* Find all loop latches (blocks with back edges to the loop | |
869 | header) or all the leaf blocks in the cfg has no loops. | |
870 | ||
871 | Place those blocks into the queue. */ | |
872 | if (no_loops) | |
873 | { | |
4c26117a | 874 | FOR_EACH_BB (jbb) |
7a31a7bd | 875 | /* Leaf nodes have only a single successor which must |
876 | be EXIT_BLOCK. */ | |
ea091dfd | 877 | if (single_succ_p (jbb) |
878 | && single_succ (jbb) == EXIT_BLOCK_PTR) | |
7a31a7bd | 879 | { |
4c26117a | 880 | queue[++tail] = jbb->index; |
08b7917c | 881 | bitmap_set_bit (in_queue, jbb->index); |
7a31a7bd | 882 | |
4c26117a | 883 | if (too_large (jbb->index, &num_bbs, &num_insns)) |
7a31a7bd | 884 | { |
885 | too_large_failure = 1; | |
886 | break; | |
887 | } | |
888 | } | |
889 | } | |
890 | else | |
891 | { | |
892 | edge e; | |
893 | ||
cd665a06 | 894 | FOR_EACH_EDGE (e, ei, bb->preds) |
7a31a7bd | 895 | { |
896 | if (e->src == ENTRY_BLOCK_PTR) | |
897 | continue; | |
898 | ||
b3d6de89 | 899 | node = e->src->index; |
7a31a7bd | 900 | |
4c26117a | 901 | if (max_hdr[node] == loop_head && node != bb->index) |
7a31a7bd | 902 | { |
903 | /* This is a loop latch. */ | |
904 | queue[++tail] = node; | |
08b7917c | 905 | bitmap_set_bit (in_queue, node); |
7a31a7bd | 906 | |
907 | if (too_large (node, &num_bbs, &num_insns)) | |
908 | { | |
909 | too_large_failure = 1; | |
910 | break; | |
911 | } | |
912 | } | |
913 | } | |
914 | } | |
915 | ||
916 | /* Now add all the blocks in the loop to the queue. | |
917 | ||
918 | We know the loop is a natural loop; however the algorithm | |
919 | above will not always mark certain blocks as being in the | |
920 | loop. Consider: | |
921 | node children | |
922 | a b,c | |
923 | b c | |
924 | c a,d | |
925 | d b | |
926 | ||
927 | The algorithm in the DFS traversal may not mark B & D as part | |
0c6d8c36 | 928 | of the loop (i.e. they will not have max_hdr set to A). |
7a31a7bd | 929 | |
930 | We know they can not be loop latches (else they would have | |
931 | had max_hdr set since they'd have a backedge to a dominator | |
932 | block). So we don't need them on the initial queue. | |
933 | ||
934 | We know they are part of the loop because they are dominated | |
935 | by the loop header and can be reached by a backwards walk of | |
936 | the edges starting with nodes on the initial queue. | |
937 | ||
938 | It is safe and desirable to include those nodes in the | |
939 | loop/scheduling region. To do so we would need to decrease | |
940 | the degree of a node if it is the target of a backedge | |
941 | within the loop itself as the node is placed in the queue. | |
942 | ||
943 | We do not do this because I'm not sure that the actual | |
944 | scheduling code will properly handle this case. ?!? */ | |
945 | ||
946 | while (head < tail && !too_large_failure) | |
947 | { | |
948 | edge e; | |
949 | child = queue[++head]; | |
950 | ||
cd665a06 | 951 | FOR_EACH_EDGE (e, ei, BASIC_BLOCK (child)->preds) |
7a31a7bd | 952 | { |
b3d6de89 | 953 | node = e->src->index; |
7a31a7bd | 954 | |
955 | /* See discussion above about nodes not marked as in | |
956 | this loop during the initial DFS traversal. */ | |
957 | if (e->src == ENTRY_BLOCK_PTR | |
958 | || max_hdr[node] != loop_head) | |
959 | { | |
960 | tail = -1; | |
961 | break; | |
962 | } | |
08b7917c | 963 | else if (!bitmap_bit_p (in_queue, node) && node != bb->index) |
7a31a7bd | 964 | { |
965 | queue[++tail] = node; | |
08b7917c | 966 | bitmap_set_bit (in_queue, node); |
7a31a7bd | 967 | |
968 | if (too_large (node, &num_bbs, &num_insns)) | |
969 | { | |
970 | too_large_failure = 1; | |
971 | break; | |
972 | } | |
973 | } | |
974 | } | |
975 | } | |
976 | ||
977 | if (tail >= 0 && !too_large_failure) | |
978 | { | |
979 | /* Place the loop header into list of region blocks. */ | |
4c26117a | 980 | degree[bb->index] = -1; |
981 | rgn_bb_table[idx] = bb->index; | |
7a31a7bd | 982 | RGN_NR_BLOCKS (nr_regions) = num_bbs; |
983 | RGN_BLOCKS (nr_regions) = idx++; | |
6a1cdb4d | 984 | RGN_DONT_CALC_DEPS (nr_regions) = 0; |
985 | RGN_HAS_REAL_EBB (nr_regions) = 0; | |
4c26117a | 986 | CONTAINING_RGN (bb->index) = nr_regions; |
987 | BLOCK_TO_BB (bb->index) = count = 0; | |
7a31a7bd | 988 | |
989 | /* Remove blocks from queue[] when their in degree | |
990 | becomes zero. Repeat until no blocks are left on the | |
991 | list. This produces a topological list of blocks in | |
992 | the region. */ | |
993 | while (tail >= 0) | |
994 | { | |
995 | if (head < 0) | |
996 | head = tail; | |
997 | child = queue[head]; | |
998 | if (degree[child] == 0) | |
999 | { | |
1000 | edge e; | |
1001 | ||
1002 | degree[child] = -1; | |
1003 | rgn_bb_table[idx++] = child; | |
1004 | BLOCK_TO_BB (child) = ++count; | |
1005 | CONTAINING_RGN (child) = nr_regions; | |
1006 | queue[head] = queue[tail--]; | |
1007 | ||
cd665a06 | 1008 | FOR_EACH_EDGE (e, ei, BASIC_BLOCK (child)->succs) |
7a31a7bd | 1009 | if (e->dest != EXIT_BLOCK_PTR) |
b3d6de89 | 1010 | --degree[e->dest->index]; |
7a31a7bd | 1011 | } |
1012 | else | |
1013 | --head; | |
1014 | } | |
1015 | ++nr_regions; | |
1016 | } | |
4bfe0e7b | 1017 | else if (extend_regions_p) |
1018 | { | |
1019 | /* Restore DEGREE. */ | |
1020 | int *t = degree; | |
1021 | ||
1022 | degree = degree1; | |
1023 | degree1 = t; | |
48e1416a | 1024 | |
4bfe0e7b | 1025 | /* And force successors of BB to be region heads. |
1026 | This may provide several smaller regions instead | |
1027 | of one too_large region. */ | |
1028 | FOR_EACH_EDGE (e, ei, bb->succs) | |
1029 | if (e->dest != EXIT_BLOCK_PTR) | |
08b7917c | 1030 | bitmap_set_bit (extended_rgn_header, e->dest->index); |
4bfe0e7b | 1031 | } |
7a31a7bd | 1032 | } |
1033 | } | |
1034 | free (queue); | |
4bfe0e7b | 1035 | |
1036 | if (extend_regions_p) | |
1037 | { | |
1038 | free (degree1); | |
48e1416a | 1039 | |
53c5d9d4 | 1040 | bitmap_ior (header, header, extended_rgn_header); |
4bfe0e7b | 1041 | sbitmap_free (extended_rgn_header); |
48e1416a | 1042 | |
4bfe0e7b | 1043 | extend_rgns (degree, &idx, header, max_hdr); |
1044 | } | |
7a31a7bd | 1045 | } |
1046 | ||
1047 | /* Any block that did not end up in a region is placed into a region | |
1048 | by itself. */ | |
4c26117a | 1049 | FOR_EACH_BB (bb) |
1050 | if (degree[bb->index] >= 0) | |
7a31a7bd | 1051 | { |
4c26117a | 1052 | rgn_bb_table[idx] = bb->index; |
7a31a7bd | 1053 | RGN_NR_BLOCKS (nr_regions) = 1; |
1054 | RGN_BLOCKS (nr_regions) = idx++; | |
6a1cdb4d | 1055 | RGN_DONT_CALC_DEPS (nr_regions) = 0; |
1056 | RGN_HAS_REAL_EBB (nr_regions) = 0; | |
4c26117a | 1057 | CONTAINING_RGN (bb->index) = nr_regions++; |
1058 | BLOCK_TO_BB (bb->index) = 0; | |
7a31a7bd | 1059 | } |
1060 | ||
1061 | free (max_hdr); | |
4bfe0e7b | 1062 | free (degree); |
7a31a7bd | 1063 | free (stack); |
265a9759 | 1064 | sbitmap_free (header); |
1065 | sbitmap_free (inner); | |
1066 | sbitmap_free (in_queue); | |
1067 | sbitmap_free (in_stack); | |
7a31a7bd | 1068 | } |
1069 | ||
e1ab7874 | 1070 | |
1071 | /* Wrapper function. | |
1072 | If FLAG_SEL_SCHED_PIPELINING is set, then use custom function to form | |
1073 | regions. Otherwise just call find_rgns_haifa. */ | |
1074 | static void | |
1075 | find_rgns (void) | |
1076 | { | |
1077 | if (sel_sched_p () && flag_sel_sched_pipelining) | |
1078 | sel_find_rgns (); | |
1079 | else | |
1080 | haifa_find_rgns (); | |
1081 | } | |
1082 | ||
4bfe0e7b | 1083 | static int gather_region_statistics (int **); |
1084 | static void print_region_statistics (int *, int, int *, int); | |
1085 | ||
48e1416a | 1086 | /* Calculate the histogram that shows the number of regions having the |
1087 | given number of basic blocks, and store it in the RSP array. Return | |
4bfe0e7b | 1088 | the size of this array. */ |
1089 | static int | |
1090 | gather_region_statistics (int **rsp) | |
1091 | { | |
1092 | int i, *a = 0, a_sz = 0; | |
1093 | ||
1094 | /* a[i] is the number of regions that have (i + 1) basic blocks. */ | |
1095 | for (i = 0; i < nr_regions; i++) | |
1096 | { | |
1097 | int nr_blocks = RGN_NR_BLOCKS (i); | |
1098 | ||
1099 | gcc_assert (nr_blocks >= 1); | |
1100 | ||
1101 | if (nr_blocks > a_sz) | |
48e1416a | 1102 | { |
f7f3687c | 1103 | a = XRESIZEVEC (int, a, nr_blocks); |
4bfe0e7b | 1104 | do |
1105 | a[a_sz++] = 0; | |
1106 | while (a_sz != nr_blocks); | |
1107 | } | |
1108 | ||
1109 | a[nr_blocks - 1]++; | |
1110 | } | |
1111 | ||
1112 | *rsp = a; | |
1113 | return a_sz; | |
1114 | } | |
1115 | ||
48e1416a | 1116 | /* Print regions statistics. S1 and S2 denote the data before and after |
4bfe0e7b | 1117 | calling extend_rgns, respectively. */ |
1118 | static void | |
1119 | print_region_statistics (int *s1, int s1_sz, int *s2, int s2_sz) | |
1120 | { | |
1121 | int i; | |
48e1416a | 1122 | |
1123 | /* We iterate until s2_sz because extend_rgns does not decrease | |
4bfe0e7b | 1124 | the maximal region size. */ |
1125 | for (i = 1; i < s2_sz; i++) | |
1126 | { | |
1127 | int n1, n2; | |
1128 | ||
1129 | n2 = s2[i]; | |
1130 | ||
1131 | if (n2 == 0) | |
1132 | continue; | |
1133 | ||
1134 | if (i >= s1_sz) | |
1135 | n1 = 0; | |
1136 | else | |
1137 | n1 = s1[i]; | |
1138 | ||
1139 | fprintf (sched_dump, ";; Region extension statistics: size %d: " \ | |
1140 | "was %d + %d more\n", i + 1, n1, n2 - n1); | |
1141 | } | |
1142 | } | |
1143 | ||
1144 | /* Extend regions. | |
1145 | DEGREE - Array of incoming edge count, considering only | |
1146 | the edges, that don't have their sources in formed regions yet. | |
1147 | IDXP - pointer to the next available index in rgn_bb_table. | |
1148 | HEADER - set of all region heads. | |
1149 | LOOP_HDR - mapping from block to the containing loop | |
1150 | (two blocks can reside within one region if they have | |
1151 | the same loop header). */ | |
e1ab7874 | 1152 | void |
4bfe0e7b | 1153 | extend_rgns (int *degree, int *idxp, sbitmap header, int *loop_hdr) |
1154 | { | |
1155 | int *order, i, rescan = 0, idx = *idxp, iter = 0, max_iter, *max_hdr; | |
1156 | int nblocks = n_basic_blocks - NUM_FIXED_BLOCKS; | |
1157 | ||
1158 | max_iter = PARAM_VALUE (PARAM_MAX_SCHED_EXTEND_REGIONS_ITERS); | |
1159 | ||
f7f3687c | 1160 | max_hdr = XNEWVEC (int, last_basic_block); |
4bfe0e7b | 1161 | |
f7f3687c | 1162 | order = XNEWVEC (int, last_basic_block); |
3072d30e | 1163 | post_order_compute (order, false, false); |
4bfe0e7b | 1164 | |
1165 | for (i = nblocks - 1; i >= 0; i--) | |
1166 | { | |
1167 | int bbn = order[i]; | |
1168 | if (degree[bbn] >= 0) | |
1169 | { | |
1170 | max_hdr[bbn] = bbn; | |
1171 | rescan = 1; | |
1172 | } | |
1173 | else | |
1174 | /* This block already was processed in find_rgns. */ | |
1175 | max_hdr[bbn] = -1; | |
1176 | } | |
48e1416a | 1177 | |
4bfe0e7b | 1178 | /* The idea is to topologically walk through CFG in top-down order. |
1179 | During the traversal, if all the predecessors of a node are | |
1180 | marked to be in the same region (they all have the same max_hdr), | |
48e1416a | 1181 | then current node is also marked to be a part of that region. |
4bfe0e7b | 1182 | Otherwise the node starts its own region. |
48e1416a | 1183 | CFG should be traversed until no further changes are made. On each |
1184 | iteration the set of the region heads is extended (the set of those | |
1185 | blocks that have max_hdr[bbi] == bbi). This set is upper bounded by the | |
e1ab7874 | 1186 | set of all basic blocks, thus the algorithm is guaranteed to |
1187 | terminate. */ | |
4bfe0e7b | 1188 | |
1189 | while (rescan && iter < max_iter) | |
1190 | { | |
1191 | rescan = 0; | |
48e1416a | 1192 | |
4bfe0e7b | 1193 | for (i = nblocks - 1; i >= 0; i--) |
1194 | { | |
1195 | edge e; | |
1196 | edge_iterator ei; | |
1197 | int bbn = order[i]; | |
48e1416a | 1198 | |
08b7917c | 1199 | if (max_hdr[bbn] != -1 && !bitmap_bit_p (header, bbn)) |
4bfe0e7b | 1200 | { |
1201 | int hdr = -1; | |
1202 | ||
1203 | FOR_EACH_EDGE (e, ei, BASIC_BLOCK (bbn)->preds) | |
1204 | { | |
1205 | int predn = e->src->index; | |
1206 | ||
1207 | if (predn != ENTRY_BLOCK | |
1208 | /* If pred wasn't processed in find_rgns. */ | |
1209 | && max_hdr[predn] != -1 | |
1210 | /* And pred and bb reside in the same loop. | |
1211 | (Or out of any loop). */ | |
1212 | && loop_hdr[bbn] == loop_hdr[predn]) | |
1213 | { | |
1214 | if (hdr == -1) | |
1215 | /* Then bb extends the containing region of pred. */ | |
1216 | hdr = max_hdr[predn]; | |
1217 | else if (hdr != max_hdr[predn]) | |
1218 | /* Too bad, there are at least two predecessors | |
1219 | that reside in different regions. Thus, BB should | |
1220 | begin its own region. */ | |
1221 | { | |
1222 | hdr = bbn; | |
1223 | break; | |
48e1416a | 1224 | } |
4bfe0e7b | 1225 | } |
1226 | else | |
1227 | /* BB starts its own region. */ | |
1228 | { | |
1229 | hdr = bbn; | |
1230 | break; | |
48e1416a | 1231 | } |
4bfe0e7b | 1232 | } |
48e1416a | 1233 | |
4bfe0e7b | 1234 | if (hdr == bbn) |
1235 | { | |
1236 | /* If BB start its own region, | |
1237 | update set of headers with BB. */ | |
08b7917c | 1238 | bitmap_set_bit (header, bbn); |
4bfe0e7b | 1239 | rescan = 1; |
1240 | } | |
1241 | else | |
48e1416a | 1242 | gcc_assert (hdr != -1); |
4bfe0e7b | 1243 | |
1244 | max_hdr[bbn] = hdr; | |
1245 | } | |
1246 | } | |
1247 | ||
1248 | iter++; | |
1249 | } | |
48e1416a | 1250 | |
4bfe0e7b | 1251 | /* Statistics were gathered on the SPEC2000 package of tests with |
1252 | mainline weekly snapshot gcc-4.1-20051015 on ia64. | |
48e1416a | 1253 | |
4bfe0e7b | 1254 | Statistics for SPECint: |
1255 | 1 iteration : 1751 cases (38.7%) | |
1256 | 2 iterations: 2770 cases (61.3%) | |
1257 | Blocks wrapped in regions by find_rgns without extension: 18295 blocks | |
1258 | Blocks wrapped in regions by 2 iterations in extend_rgns: 23821 blocks | |
1259 | (We don't count single block regions here). | |
48e1416a | 1260 | |
4bfe0e7b | 1261 | Statistics for SPECfp: |
1262 | 1 iteration : 621 cases (35.9%) | |
1263 | 2 iterations: 1110 cases (64.1%) | |
1264 | Blocks wrapped in regions by find_rgns without extension: 6476 blocks | |
1265 | Blocks wrapped in regions by 2 iterations in extend_rgns: 11155 blocks | |
1266 | (We don't count single block regions here). | |
1267 | ||
1268 | By default we do at most 2 iterations. | |
9ca2c29a | 1269 | This can be overridden with max-sched-extend-regions-iters parameter: |
4bfe0e7b | 1270 | 0 - disable region extension, |
1271 | N > 0 - do at most N iterations. */ | |
48e1416a | 1272 | |
4bfe0e7b | 1273 | if (sched_verbose && iter != 0) |
1274 | fprintf (sched_dump, ";; Region extension iterations: %d%s\n", iter, | |
1275 | rescan ? "... failed" : ""); | |
48e1416a | 1276 | |
4bfe0e7b | 1277 | if (!rescan && iter != 0) |
1278 | { | |
1279 | int *s1 = NULL, s1_sz = 0; | |
1280 | ||
1281 | /* Save the old statistics for later printout. */ | |
1282 | if (sched_verbose >= 6) | |
1283 | s1_sz = gather_region_statistics (&s1); | |
1284 | ||
1285 | /* We have succeeded. Now assemble the regions. */ | |
1286 | for (i = nblocks - 1; i >= 0; i--) | |
1287 | { | |
1288 | int bbn = order[i]; | |
1289 | ||
1290 | if (max_hdr[bbn] == bbn) | |
1291 | /* BBN is a region head. */ | |
1292 | { | |
1293 | edge e; | |
1294 | edge_iterator ei; | |
1295 | int num_bbs = 0, j, num_insns = 0, large; | |
48e1416a | 1296 | |
4bfe0e7b | 1297 | large = too_large (bbn, &num_bbs, &num_insns); |
1298 | ||
1299 | degree[bbn] = -1; | |
1300 | rgn_bb_table[idx] = bbn; | |
1301 | RGN_BLOCKS (nr_regions) = idx++; | |
6a1cdb4d | 1302 | RGN_DONT_CALC_DEPS (nr_regions) = 0; |
1303 | RGN_HAS_REAL_EBB (nr_regions) = 0; | |
4bfe0e7b | 1304 | CONTAINING_RGN (bbn) = nr_regions; |
1305 | BLOCK_TO_BB (bbn) = 0; | |
1306 | ||
1307 | FOR_EACH_EDGE (e, ei, BASIC_BLOCK (bbn)->succs) | |
1308 | if (e->dest != EXIT_BLOCK_PTR) | |
1309 | degree[e->dest->index]--; | |
1310 | ||
1311 | if (!large) | |
1312 | /* Here we check whether the region is too_large. */ | |
1313 | for (j = i - 1; j >= 0; j--) | |
1314 | { | |
1315 | int succn = order[j]; | |
1316 | if (max_hdr[succn] == bbn) | |
1317 | { | |
1318 | if ((large = too_large (succn, &num_bbs, &num_insns))) | |
1319 | break; | |
1320 | } | |
1321 | } | |
1322 | ||
1323 | if (large) | |
1324 | /* If the region is too_large, then wrap every block of | |
1325 | the region into single block region. | |
1326 | Here we wrap region head only. Other blocks are | |
1327 | processed in the below cycle. */ | |
1328 | { | |
1329 | RGN_NR_BLOCKS (nr_regions) = 1; | |
1330 | nr_regions++; | |
48e1416a | 1331 | } |
4bfe0e7b | 1332 | |
1333 | num_bbs = 1; | |
1334 | ||
1335 | for (j = i - 1; j >= 0; j--) | |
1336 | { | |
1337 | int succn = order[j]; | |
1338 | ||
1339 | if (max_hdr[succn] == bbn) | |
48e1416a | 1340 | /* This cycle iterates over all basic blocks, that |
4bfe0e7b | 1341 | are supposed to be in the region with head BBN, |
1342 | and wraps them into that region (or in single | |
1343 | block region). */ | |
1344 | { | |
1345 | gcc_assert (degree[succn] == 0); | |
1346 | ||
1347 | degree[succn] = -1; | |
48e1416a | 1348 | rgn_bb_table[idx] = succn; |
4bfe0e7b | 1349 | BLOCK_TO_BB (succn) = large ? 0 : num_bbs++; |
1350 | CONTAINING_RGN (succn) = nr_regions; | |
1351 | ||
1352 | if (large) | |
1353 | /* Wrap SUCCN into single block region. */ | |
1354 | { | |
1355 | RGN_BLOCKS (nr_regions) = idx; | |
1356 | RGN_NR_BLOCKS (nr_regions) = 1; | |
6a1cdb4d | 1357 | RGN_DONT_CALC_DEPS (nr_regions) = 0; |
1358 | RGN_HAS_REAL_EBB (nr_regions) = 0; | |
4bfe0e7b | 1359 | nr_regions++; |
1360 | } | |
1361 | ||
1362 | idx++; | |
48e1416a | 1363 | |
4bfe0e7b | 1364 | FOR_EACH_EDGE (e, ei, BASIC_BLOCK (succn)->succs) |
1365 | if (e->dest != EXIT_BLOCK_PTR) | |
1366 | degree[e->dest->index]--; | |
1367 | } | |
1368 | } | |
1369 | ||
1370 | if (!large) | |
1371 | { | |
1372 | RGN_NR_BLOCKS (nr_regions) = num_bbs; | |
1373 | nr_regions++; | |
1374 | } | |
1375 | } | |
1376 | } | |
1377 | ||
1378 | if (sched_verbose >= 6) | |
1379 | { | |
1380 | int *s2, s2_sz; | |
1381 | ||
48e1416a | 1382 | /* Get the new statistics and print the comparison with the |
4bfe0e7b | 1383 | one before calling this function. */ |
1384 | s2_sz = gather_region_statistics (&s2); | |
1385 | print_region_statistics (s1, s1_sz, s2, s2_sz); | |
1386 | free (s1); | |
1387 | free (s2); | |
1388 | } | |
1389 | } | |
48e1416a | 1390 | |
4bfe0e7b | 1391 | free (order); |
1392 | free (max_hdr); | |
1393 | ||
48e1416a | 1394 | *idxp = idx; |
4bfe0e7b | 1395 | } |
1396 | ||
7a31a7bd | 1397 | /* Functions for regions scheduling information. */ |
1398 | ||
1399 | /* Compute dominators, probability, and potential-split-edges of bb. | |
1400 | Assume that these values were already computed for bb's predecessors. */ | |
1401 | ||
1402 | static void | |
60b8c5b3 | 1403 | compute_dom_prob_ps (int bb) |
7a31a7bd | 1404 | { |
6c0b81cb | 1405 | edge_iterator in_ei; |
1406 | edge in_edge; | |
7a31a7bd | 1407 | |
6a1cdb4d | 1408 | /* We shouldn't have any real ebbs yet. */ |
1409 | gcc_assert (ebb_head [bb] == bb + current_blocks); | |
48e1416a | 1410 | |
7a31a7bd | 1411 | if (IS_RGN_ENTRY (bb)) |
1412 | { | |
08b7917c | 1413 | bitmap_set_bit (dom[bb], 0); |
6c0b81cb | 1414 | prob[bb] = REG_BR_PROB_BASE; |
7a31a7bd | 1415 | return; |
1416 | } | |
1417 | ||
6c0b81cb | 1418 | prob[bb] = 0; |
1419 | ||
4a82352a | 1420 | /* Initialize dom[bb] to '111..1'. */ |
53c5d9d4 | 1421 | bitmap_ones (dom[bb]); |
7a31a7bd | 1422 | |
aae97b21 | 1423 | FOR_EACH_EDGE (in_edge, in_ei, BASIC_BLOCK (BB_TO_BLOCK (bb))->preds) |
7a31a7bd | 1424 | { |
6c0b81cb | 1425 | int pred_bb; |
1426 | edge out_edge; | |
1427 | edge_iterator out_ei; | |
1428 | ||
aae97b21 | 1429 | if (in_edge->src == ENTRY_BLOCK_PTR) |
1430 | continue; | |
7a31a7bd | 1431 | |
aae97b21 | 1432 | pred_bb = BLOCK_TO_BB (in_edge->src->index); |
53c5d9d4 | 1433 | bitmap_and (dom[bb], dom[bb], dom[pred_bb]); |
1434 | bitmap_ior (ancestor_edges[bb], | |
aae97b21 | 1435 | ancestor_edges[bb], ancestor_edges[pred_bb]); |
7a31a7bd | 1436 | |
08b7917c | 1437 | bitmap_set_bit (ancestor_edges[bb], EDGE_TO_BIT (in_edge)); |
79cafa9e | 1438 | |
53c5d9d4 | 1439 | bitmap_ior (pot_split[bb], pot_split[bb], pot_split[pred_bb]); |
7a31a7bd | 1440 | |
aae97b21 | 1441 | FOR_EACH_EDGE (out_edge, out_ei, in_edge->src->succs) |
08b7917c | 1442 | bitmap_set_bit (pot_split[bb], EDGE_TO_BIT (out_edge)); |
aae97b21 | 1443 | |
70074000 | 1444 | prob[bb] += combine_probabilities (prob[pred_bb], in_edge->probability); |
df21e330 | 1445 | // The rounding divide in combine_probabilities can result in an extra |
1446 | // probability increment propagating along 50-50 edges. Eventually when | |
1447 | // the edges re-merge, the accumulated probability can go slightly above | |
1448 | // REG_BR_PROB_BASE. | |
1449 | if (prob[bb] > REG_BR_PROB_BASE) | |
1450 | prob[bb] = REG_BR_PROB_BASE; | |
7a31a7bd | 1451 | } |
7a31a7bd | 1452 | |
08b7917c | 1453 | bitmap_set_bit (dom[bb], bb); |
53c5d9d4 | 1454 | bitmap_and_compl (pot_split[bb], pot_split[bb], ancestor_edges[bb]); |
7a31a7bd | 1455 | |
1456 | if (sched_verbose >= 2) | |
1457 | fprintf (sched_dump, ";; bb_prob(%d, %d) = %3d\n", bb, BB_TO_BLOCK (bb), | |
6c0b81cb | 1458 | (100 * prob[bb]) / REG_BR_PROB_BASE); |
7a31a7bd | 1459 | } |
1460 | ||
1461 | /* Functions for target info. */ | |
1462 | ||
1463 | /* Compute in BL the list of split-edges of bb_src relatively to bb_trg. | |
1464 | Note that bb_trg dominates bb_src. */ | |
1465 | ||
1466 | static void | |
60b8c5b3 | 1467 | split_edges (int bb_src, int bb_trg, edgelst *bl) |
7a31a7bd | 1468 | { |
156093aa | 1469 | sbitmap src = sbitmap_alloc (SBITMAP_SIZE (pot_split[bb_src])); |
53c5d9d4 | 1470 | bitmap_copy (src, pot_split[bb_src]); |
79cafa9e | 1471 | |
53c5d9d4 | 1472 | bitmap_and_compl (src, src, pot_split[bb_trg]); |
aae97b21 | 1473 | extract_edgelst (src, bl); |
79cafa9e | 1474 | sbitmap_free (src); |
7a31a7bd | 1475 | } |
1476 | ||
1477 | /* Find the valid candidate-source-blocks for the target block TRG, compute | |
1478 | their probability, and check if they are speculative or not. | |
1479 | For speculative sources, compute their update-blocks and split-blocks. */ | |
1480 | ||
1481 | static void | |
60b8c5b3 | 1482 | compute_trg_info (int trg) |
7a31a7bd | 1483 | { |
19cb6b50 | 1484 | candidate *sp; |
91020a82 | 1485 | edgelst el = { NULL, 0 }; |
aae97b21 | 1486 | int i, j, k, update_idx; |
1487 | basic_block block; | |
d3129ae7 | 1488 | sbitmap visited; |
aae97b21 | 1489 | edge_iterator ei; |
1490 | edge e; | |
7a31a7bd | 1491 | |
e1ab7874 | 1492 | candidate_table = XNEWVEC (candidate, current_nr_blocks); |
1493 | ||
1494 | bblst_last = 0; | |
1495 | /* bblst_table holds split blocks and update blocks for each block after | |
1496 | the current one in the region. split blocks and update blocks are | |
1497 | the TO blocks of region edges, so there can be at most rgn_nr_edges | |
1498 | of them. */ | |
1499 | bblst_size = (current_nr_blocks - target_bb) * rgn_nr_edges; | |
1500 | bblst_table = XNEWVEC (basic_block, bblst_size); | |
1501 | ||
1502 | edgelst_last = 0; | |
1503 | edgelst_table = XNEWVEC (edge, rgn_nr_edges); | |
1504 | ||
7a31a7bd | 1505 | /* Define some of the fields for the target bb as well. */ |
1506 | sp = candidate_table + trg; | |
1507 | sp->is_valid = 1; | |
1508 | sp->is_speculative = 0; | |
6c0b81cb | 1509 | sp->src_prob = REG_BR_PROB_BASE; |
7a31a7bd | 1510 | |
4d2e5d52 | 1511 | visited = sbitmap_alloc (last_basic_block); |
d3129ae7 | 1512 | |
7a31a7bd | 1513 | for (i = trg + 1; i < current_nr_blocks; i++) |
1514 | { | |
1515 | sp = candidate_table + i; | |
1516 | ||
1517 | sp->is_valid = IS_DOMINATED (i, trg); | |
1518 | if (sp->is_valid) | |
1519 | { | |
6c0b81cb | 1520 | int tf = prob[trg], cf = prob[i]; |
1521 | ||
1522 | /* In CFGs with low probability edges TF can possibly be zero. */ | |
70074000 | 1523 | sp->src_prob = (tf ? GCOV_COMPUTE_SCALE (cf, tf) : 0); |
6c0b81cb | 1524 | sp->is_valid = (sp->src_prob >= min_spec_prob); |
7a31a7bd | 1525 | } |
1526 | ||
1527 | if (sp->is_valid) | |
1528 | { | |
1529 | split_edges (i, trg, &el); | |
1530 | sp->is_speculative = (el.nr_members) ? 1 : 0; | |
1531 | if (sp->is_speculative && !flag_schedule_speculative) | |
1532 | sp->is_valid = 0; | |
1533 | } | |
1534 | ||
1535 | if (sp->is_valid) | |
1536 | { | |
7a31a7bd | 1537 | /* Compute split blocks and store them in bblst_table. |
1538 | The TO block of every split edge is a split block. */ | |
1539 | sp->split_bbs.first_member = &bblst_table[bblst_last]; | |
1540 | sp->split_bbs.nr_members = el.nr_members; | |
1541 | for (j = 0; j < el.nr_members; bblst_last++, j++) | |
aae97b21 | 1542 | bblst_table[bblst_last] = el.first_member[j]->dest; |
7a31a7bd | 1543 | sp->update_bbs.first_member = &bblst_table[bblst_last]; |
1544 | ||
1545 | /* Compute update blocks and store them in bblst_table. | |
1546 | For every split edge, look at the FROM block, and check | |
1547 | all out edges. For each out edge that is not a split edge, | |
1548 | add the TO block to the update block list. This list can end | |
1549 | up with a lot of duplicates. We need to weed them out to avoid | |
1550 | overrunning the end of the bblst_table. */ | |
7a31a7bd | 1551 | |
1552 | update_idx = 0; | |
53c5d9d4 | 1553 | bitmap_clear (visited); |
7a31a7bd | 1554 | for (j = 0; j < el.nr_members; j++) |
1555 | { | |
aae97b21 | 1556 | block = el.first_member[j]->src; |
1557 | FOR_EACH_EDGE (e, ei, block->succs) | |
7a31a7bd | 1558 | { |
08b7917c | 1559 | if (!bitmap_bit_p (visited, e->dest->index)) |
7a31a7bd | 1560 | { |
1561 | for (k = 0; k < el.nr_members; k++) | |
aae97b21 | 1562 | if (e == el.first_member[k]) |
7a31a7bd | 1563 | break; |
1564 | ||
1565 | if (k >= el.nr_members) | |
1566 | { | |
aae97b21 | 1567 | bblst_table[bblst_last++] = e->dest; |
08b7917c | 1568 | bitmap_set_bit (visited, e->dest->index); |
7a31a7bd | 1569 | update_idx++; |
1570 | } | |
1571 | } | |
7a31a7bd | 1572 | } |
7a31a7bd | 1573 | } |
1574 | sp->update_bbs.nr_members = update_idx; | |
1575 | ||
1576 | /* Make sure we didn't overrun the end of bblst_table. */ | |
04e579b6 | 1577 | gcc_assert (bblst_last <= bblst_size); |
7a31a7bd | 1578 | } |
1579 | else | |
1580 | { | |
1581 | sp->split_bbs.nr_members = sp->update_bbs.nr_members = 0; | |
1582 | ||
1583 | sp->is_speculative = 0; | |
1584 | sp->src_prob = 0; | |
1585 | } | |
1586 | } | |
d3129ae7 | 1587 | |
1588 | sbitmap_free (visited); | |
7a31a7bd | 1589 | } |
1590 | ||
e1ab7874 | 1591 | /* Free the computed target info. */ |
1592 | static void | |
1593 | free_trg_info (void) | |
1594 | { | |
1595 | free (candidate_table); | |
1596 | free (bblst_table); | |
1597 | free (edgelst_table); | |
1598 | } | |
1599 | ||
7a31a7bd | 1600 | /* Print candidates info, for debugging purposes. Callable from debugger. */ |
1601 | ||
4b987fac | 1602 | DEBUG_FUNCTION void |
60b8c5b3 | 1603 | debug_candidate (int i) |
7a31a7bd | 1604 | { |
1605 | if (!candidate_table[i].is_valid) | |
1606 | return; | |
1607 | ||
1608 | if (candidate_table[i].is_speculative) | |
1609 | { | |
1610 | int j; | |
1611 | fprintf (sched_dump, "src b %d bb %d speculative \n", BB_TO_BLOCK (i), i); | |
1612 | ||
1613 | fprintf (sched_dump, "split path: "); | |
1614 | for (j = 0; j < candidate_table[i].split_bbs.nr_members; j++) | |
1615 | { | |
aae97b21 | 1616 | int b = candidate_table[i].split_bbs.first_member[j]->index; |
7a31a7bd | 1617 | |
1618 | fprintf (sched_dump, " %d ", b); | |
1619 | } | |
1620 | fprintf (sched_dump, "\n"); | |
1621 | ||
1622 | fprintf (sched_dump, "update path: "); | |
1623 | for (j = 0; j < candidate_table[i].update_bbs.nr_members; j++) | |
1624 | { | |
aae97b21 | 1625 | int b = candidate_table[i].update_bbs.first_member[j]->index; |
7a31a7bd | 1626 | |
1627 | fprintf (sched_dump, " %d ", b); | |
1628 | } | |
1629 | fprintf (sched_dump, "\n"); | |
1630 | } | |
1631 | else | |
1632 | { | |
1633 | fprintf (sched_dump, " src %d equivalent\n", BB_TO_BLOCK (i)); | |
1634 | } | |
1635 | } | |
1636 | ||
1637 | /* Print candidates info, for debugging purposes. Callable from debugger. */ | |
1638 | ||
4b987fac | 1639 | DEBUG_FUNCTION void |
60b8c5b3 | 1640 | debug_candidates (int trg) |
7a31a7bd | 1641 | { |
1642 | int i; | |
1643 | ||
1644 | fprintf (sched_dump, "----------- candidate table: target: b=%d bb=%d ---\n", | |
1645 | BB_TO_BLOCK (trg), trg); | |
1646 | for (i = trg + 1; i < current_nr_blocks; i++) | |
1647 | debug_candidate (i); | |
1648 | } | |
1649 | ||
de132707 | 1650 | /* Functions for speculative scheduling. */ |
7a31a7bd | 1651 | |
3072d30e | 1652 | static bitmap_head not_in_df; |
1653 | ||
7a31a7bd | 1654 | /* Return 0 if x is a set of a register alive in the beginning of one |
1655 | of the split-blocks of src, otherwise return 1. */ | |
1656 | ||
1657 | static int | |
60b8c5b3 | 1658 | check_live_1 (int src, rtx x) |
7a31a7bd | 1659 | { |
19cb6b50 | 1660 | int i; |
1661 | int regno; | |
1662 | rtx reg = SET_DEST (x); | |
7a31a7bd | 1663 | |
1664 | if (reg == 0) | |
1665 | return 1; | |
1666 | ||
476d094d | 1667 | while (GET_CODE (reg) == SUBREG |
1668 | || GET_CODE (reg) == ZERO_EXTRACT | |
7a31a7bd | 1669 | || GET_CODE (reg) == STRICT_LOW_PART) |
1670 | reg = XEXP (reg, 0); | |
1671 | ||
4b303227 | 1672 | if (GET_CODE (reg) == PARALLEL) |
7a31a7bd | 1673 | { |
19cb6b50 | 1674 | int i; |
216b2683 | 1675 | |
7a31a7bd | 1676 | for (i = XVECLEN (reg, 0) - 1; i >= 0; i--) |
4b303227 | 1677 | if (XEXP (XVECEXP (reg, 0, i), 0) != 0) |
1678 | if (check_live_1 (src, XEXP (XVECEXP (reg, 0, i), 0))) | |
216b2683 | 1679 | return 1; |
216b2683 | 1680 | |
7a31a7bd | 1681 | return 0; |
1682 | } | |
1683 | ||
8ad4c111 | 1684 | if (!REG_P (reg)) |
7a31a7bd | 1685 | return 1; |
1686 | ||
1687 | regno = REGNO (reg); | |
1688 | ||
1689 | if (regno < FIRST_PSEUDO_REGISTER && global_regs[regno]) | |
1690 | { | |
1691 | /* Global registers are assumed live. */ | |
1692 | return 0; | |
1693 | } | |
1694 | else | |
1695 | { | |
1696 | if (regno < FIRST_PSEUDO_REGISTER) | |
1697 | { | |
1698 | /* Check for hard registers. */ | |
67d6c12b | 1699 | int j = hard_regno_nregs[regno][GET_MODE (reg)]; |
7a31a7bd | 1700 | while (--j >= 0) |
1701 | { | |
1702 | for (i = 0; i < candidate_table[src].split_bbs.nr_members; i++) | |
1703 | { | |
aae97b21 | 1704 | basic_block b = candidate_table[src].split_bbs.first_member[i]; |
3072d30e | 1705 | int t = bitmap_bit_p (¬_in_df, b->index); |
7a31a7bd | 1706 | |
6a1cdb4d | 1707 | /* We can have split blocks, that were recently generated. |
f0b5f617 | 1708 | Such blocks are always outside current region. */ |
3072d30e | 1709 | gcc_assert (!t || (CONTAINING_RGN (b->index) |
1710 | != CONTAINING_RGN (BB_TO_BLOCK (src)))); | |
1711 | ||
deb2741b | 1712 | if (t || REGNO_REG_SET_P (df_get_live_in (b), regno + j)) |
3072d30e | 1713 | return 0; |
7a31a7bd | 1714 | } |
1715 | } | |
1716 | } | |
1717 | else | |
1718 | { | |
f024691d | 1719 | /* Check for pseudo registers. */ |
7a31a7bd | 1720 | for (i = 0; i < candidate_table[src].split_bbs.nr_members; i++) |
1721 | { | |
aae97b21 | 1722 | basic_block b = candidate_table[src].split_bbs.first_member[i]; |
3072d30e | 1723 | int t = bitmap_bit_p (¬_in_df, b->index); |
7a31a7bd | 1724 | |
3072d30e | 1725 | gcc_assert (!t || (CONTAINING_RGN (b->index) |
1726 | != CONTAINING_RGN (BB_TO_BLOCK (src)))); | |
1727 | ||
deb2741b | 1728 | if (t || REGNO_REG_SET_P (df_get_live_in (b), regno)) |
3072d30e | 1729 | return 0; |
7a31a7bd | 1730 | } |
1731 | } | |
1732 | } | |
1733 | ||
1734 | return 1; | |
1735 | } | |
1736 | ||
1737 | /* If x is a set of a register R, mark that R is alive in the beginning | |
1738 | of every update-block of src. */ | |
1739 | ||
1740 | static void | |
60b8c5b3 | 1741 | update_live_1 (int src, rtx x) |
7a31a7bd | 1742 | { |
19cb6b50 | 1743 | int i; |
1744 | int regno; | |
1745 | rtx reg = SET_DEST (x); | |
7a31a7bd | 1746 | |
1747 | if (reg == 0) | |
1748 | return; | |
1749 | ||
476d094d | 1750 | while (GET_CODE (reg) == SUBREG |
1751 | || GET_CODE (reg) == ZERO_EXTRACT | |
7a31a7bd | 1752 | || GET_CODE (reg) == STRICT_LOW_PART) |
1753 | reg = XEXP (reg, 0); | |
1754 | ||
4b303227 | 1755 | if (GET_CODE (reg) == PARALLEL) |
7a31a7bd | 1756 | { |
19cb6b50 | 1757 | int i; |
216b2683 | 1758 | |
7a31a7bd | 1759 | for (i = XVECLEN (reg, 0) - 1; i >= 0; i--) |
4b303227 | 1760 | if (XEXP (XVECEXP (reg, 0, i), 0) != 0) |
1761 | update_live_1 (src, XEXP (XVECEXP (reg, 0, i), 0)); | |
216b2683 | 1762 | |
7a31a7bd | 1763 | return; |
1764 | } | |
1765 | ||
8ad4c111 | 1766 | if (!REG_P (reg)) |
7a31a7bd | 1767 | return; |
1768 | ||
1769 | /* Global registers are always live, so the code below does not apply | |
1770 | to them. */ | |
1771 | ||
1772 | regno = REGNO (reg); | |
1773 | ||
771d4616 | 1774 | if (! HARD_REGISTER_NUM_P (regno) |
1775 | || !global_regs[regno]) | |
7a31a7bd | 1776 | { |
771d4616 | 1777 | for (i = 0; i < candidate_table[src].update_bbs.nr_members; i++) |
7a31a7bd | 1778 | { |
771d4616 | 1779 | basic_block b = candidate_table[src].update_bbs.first_member[i]; |
7a31a7bd | 1780 | |
771d4616 | 1781 | if (HARD_REGISTER_NUM_P (regno)) |
1782 | bitmap_set_range (df_get_live_in (b), regno, | |
1783 | hard_regno_nregs[regno][GET_MODE (reg)]); | |
1784 | else | |
1785 | bitmap_set_bit (df_get_live_in (b), regno); | |
7a31a7bd | 1786 | } |
1787 | } | |
1788 | } | |
1789 | ||
1790 | /* Return 1 if insn can be speculatively moved from block src to trg, | |
1791 | otherwise return 0. Called before first insertion of insn to | |
1792 | ready-list or before the scheduling. */ | |
1793 | ||
1794 | static int | |
60b8c5b3 | 1795 | check_live (rtx insn, int src) |
7a31a7bd | 1796 | { |
1797 | /* Find the registers set by instruction. */ | |
1798 | if (GET_CODE (PATTERN (insn)) == SET | |
1799 | || GET_CODE (PATTERN (insn)) == CLOBBER) | |
1800 | return check_live_1 (src, PATTERN (insn)); | |
1801 | else if (GET_CODE (PATTERN (insn)) == PARALLEL) | |
1802 | { | |
1803 | int j; | |
1804 | for (j = XVECLEN (PATTERN (insn), 0) - 1; j >= 0; j--) | |
1805 | if ((GET_CODE (XVECEXP (PATTERN (insn), 0, j)) == SET | |
1806 | || GET_CODE (XVECEXP (PATTERN (insn), 0, j)) == CLOBBER) | |
1807 | && !check_live_1 (src, XVECEXP (PATTERN (insn), 0, j))) | |
1808 | return 0; | |
1809 | ||
1810 | return 1; | |
1811 | } | |
1812 | ||
1813 | return 1; | |
1814 | } | |
1815 | ||
1816 | /* Update the live registers info after insn was moved speculatively from | |
1817 | block src to trg. */ | |
1818 | ||
1819 | static void | |
60b8c5b3 | 1820 | update_live (rtx insn, int src) |
7a31a7bd | 1821 | { |
1822 | /* Find the registers set by instruction. */ | |
1823 | if (GET_CODE (PATTERN (insn)) == SET | |
1824 | || GET_CODE (PATTERN (insn)) == CLOBBER) | |
1825 | update_live_1 (src, PATTERN (insn)); | |
1826 | else if (GET_CODE (PATTERN (insn)) == PARALLEL) | |
1827 | { | |
1828 | int j; | |
1829 | for (j = XVECLEN (PATTERN (insn), 0) - 1; j >= 0; j--) | |
1830 | if (GET_CODE (XVECEXP (PATTERN (insn), 0, j)) == SET | |
1831 | || GET_CODE (XVECEXP (PATTERN (insn), 0, j)) == CLOBBER) | |
1832 | update_live_1 (src, XVECEXP (PATTERN (insn), 0, j)); | |
1833 | } | |
1834 | } | |
1835 | ||
a8b24921 | 1836 | /* Nonzero if block bb_to is equal to, or reachable from block bb_from. */ |
7a31a7bd | 1837 | #define IS_REACHABLE(bb_from, bb_to) \ |
40734805 | 1838 | (bb_from == bb_to \ |
7a31a7bd | 1839 | || IS_RGN_ENTRY (bb_from) \ |
08b7917c | 1840 | || (bitmap_bit_p (ancestor_edges[bb_to], \ |
ea091dfd | 1841 | EDGE_TO_BIT (single_pred_edge (BASIC_BLOCK (BB_TO_BLOCK (bb_from))))))) |
7a31a7bd | 1842 | |
7a31a7bd | 1843 | /* Turns on the fed_by_spec_load flag for insns fed by load_insn. */ |
1844 | ||
1845 | static void | |
60b8c5b3 | 1846 | set_spec_fed (rtx load_insn) |
7a31a7bd | 1847 | { |
93f6b030 | 1848 | sd_iterator_def sd_it; |
1849 | dep_t dep; | |
7a31a7bd | 1850 | |
93f6b030 | 1851 | FOR_EACH_DEP (load_insn, SD_LIST_FORW, sd_it, dep) |
1852 | if (DEP_TYPE (dep) == REG_DEP_TRUE) | |
1853 | FED_BY_SPEC_LOAD (DEP_CON (dep)) = 1; | |
9997bd27 | 1854 | } |
7a31a7bd | 1855 | |
1856 | /* On the path from the insn to load_insn_bb, find a conditional | |
1857 | branch depending on insn, that guards the speculative load. */ | |
1858 | ||
1859 | static int | |
60b8c5b3 | 1860 | find_conditional_protection (rtx insn, int load_insn_bb) |
7a31a7bd | 1861 | { |
93f6b030 | 1862 | sd_iterator_def sd_it; |
1863 | dep_t dep; | |
7a31a7bd | 1864 | |
1865 | /* Iterate through DEF-USE forward dependences. */ | |
93f6b030 | 1866 | FOR_EACH_DEP (insn, SD_LIST_FORW, sd_it, dep) |
7a31a7bd | 1867 | { |
93f6b030 | 1868 | rtx next = DEP_CON (dep); |
9997bd27 | 1869 | |
7a31a7bd | 1870 | if ((CONTAINING_RGN (BLOCK_NUM (next)) == |
1871 | CONTAINING_RGN (BB_TO_BLOCK (load_insn_bb))) | |
1872 | && IS_REACHABLE (INSN_BB (next), load_insn_bb) | |
1873 | && load_insn_bb != INSN_BB (next) | |
93f6b030 | 1874 | && DEP_TYPE (dep) == REG_DEP_TRUE |
6d7dc5b9 | 1875 | && (JUMP_P (next) |
7a31a7bd | 1876 | || find_conditional_protection (next, load_insn_bb))) |
1877 | return 1; | |
1878 | } | |
1879 | return 0; | |
1880 | } /* find_conditional_protection */ | |
1881 | ||
1882 | /* Returns 1 if the same insn1 that participates in the computation | |
1883 | of load_insn's address is feeding a conditional branch that is | |
f0b5f617 | 1884 | guarding on load_insn. This is true if we find two DEF-USE |
7a31a7bd | 1885 | chains: |
1886 | insn1 -> ... -> conditional-branch | |
1887 | insn1 -> ... -> load_insn, | |
f0b5f617 | 1888 | and if a flow path exists: |
7a31a7bd | 1889 | insn1 -> ... -> conditional-branch -> ... -> load_insn, |
1890 | and if insn1 is on the path | |
1891 | region-entry -> ... -> bb_trg -> ... load_insn. | |
1892 | ||
9997bd27 | 1893 | Locate insn1 by climbing on INSN_BACK_DEPS from load_insn. |
1894 | Locate the branch by following INSN_FORW_DEPS from insn1. */ | |
7a31a7bd | 1895 | |
1896 | static int | |
60b8c5b3 | 1897 | is_conditionally_protected (rtx load_insn, int bb_src, int bb_trg) |
7a31a7bd | 1898 | { |
93f6b030 | 1899 | sd_iterator_def sd_it; |
1900 | dep_t dep; | |
7a31a7bd | 1901 | |
93f6b030 | 1902 | FOR_EACH_DEP (load_insn, SD_LIST_BACK, sd_it, dep) |
7a31a7bd | 1903 | { |
93f6b030 | 1904 | rtx insn1 = DEP_PRO (dep); |
7a31a7bd | 1905 | |
1906 | /* Must be a DEF-USE dependence upon non-branch. */ | |
93f6b030 | 1907 | if (DEP_TYPE (dep) != REG_DEP_TRUE |
6d7dc5b9 | 1908 | || JUMP_P (insn1)) |
7a31a7bd | 1909 | continue; |
1910 | ||
1911 | /* Must exist a path: region-entry -> ... -> bb_trg -> ... load_insn. */ | |
1912 | if (INSN_BB (insn1) == bb_src | |
1913 | || (CONTAINING_RGN (BLOCK_NUM (insn1)) | |
1914 | != CONTAINING_RGN (BB_TO_BLOCK (bb_src))) | |
1915 | || (!IS_REACHABLE (bb_trg, INSN_BB (insn1)) | |
1916 | && !IS_REACHABLE (INSN_BB (insn1), bb_trg))) | |
1917 | continue; | |
1918 | ||
1919 | /* Now search for the conditional-branch. */ | |
1920 | if (find_conditional_protection (insn1, bb_src)) | |
1921 | return 1; | |
1922 | ||
1923 | /* Recursive step: search another insn1, "above" current insn1. */ | |
1924 | return is_conditionally_protected (insn1, bb_src, bb_trg); | |
1925 | } | |
1926 | ||
1927 | /* The chain does not exist. */ | |
1928 | return 0; | |
1929 | } /* is_conditionally_protected */ | |
1930 | ||
1931 | /* Returns 1 if a clue for "similar load" 'insn2' is found, and hence | |
1932 | load_insn can move speculatively from bb_src to bb_trg. All the | |
1933 | following must hold: | |
1934 | ||
1935 | (1) both loads have 1 base register (PFREE_CANDIDATEs). | |
1936 | (2) load_insn and load1 have a def-use dependence upon | |
1937 | the same insn 'insn1'. | |
1938 | (3) either load2 is in bb_trg, or: | |
1939 | - there's only one split-block, and | |
1940 | - load1 is on the escape path, and | |
1941 | ||
1942 | From all these we can conclude that the two loads access memory | |
1943 | addresses that differ at most by a constant, and hence if moving | |
1944 | load_insn would cause an exception, it would have been caused by | |
1945 | load2 anyhow. */ | |
1946 | ||
1947 | static int | |
60b8c5b3 | 1948 | is_pfree (rtx load_insn, int bb_src, int bb_trg) |
7a31a7bd | 1949 | { |
93f6b030 | 1950 | sd_iterator_def back_sd_it; |
1951 | dep_t back_dep; | |
19cb6b50 | 1952 | candidate *candp = candidate_table + bb_src; |
7a31a7bd | 1953 | |
1954 | if (candp->split_bbs.nr_members != 1) | |
1955 | /* Must have exactly one escape block. */ | |
1956 | return 0; | |
1957 | ||
93f6b030 | 1958 | FOR_EACH_DEP (load_insn, SD_LIST_BACK, back_sd_it, back_dep) |
7a31a7bd | 1959 | { |
93f6b030 | 1960 | rtx insn1 = DEP_PRO (back_dep); |
7a31a7bd | 1961 | |
93f6b030 | 1962 | if (DEP_TYPE (back_dep) == REG_DEP_TRUE) |
1963 | /* Found a DEF-USE dependence (insn1, load_insn). */ | |
7a31a7bd | 1964 | { |
93f6b030 | 1965 | sd_iterator_def fore_sd_it; |
1966 | dep_t fore_dep; | |
7a31a7bd | 1967 | |
93f6b030 | 1968 | FOR_EACH_DEP (insn1, SD_LIST_FORW, fore_sd_it, fore_dep) |
7a31a7bd | 1969 | { |
93f6b030 | 1970 | rtx insn2 = DEP_CON (fore_dep); |
9997bd27 | 1971 | |
93f6b030 | 1972 | if (DEP_TYPE (fore_dep) == REG_DEP_TRUE) |
7a31a7bd | 1973 | { |
1974 | /* Found a DEF-USE dependence (insn1, insn2). */ | |
1975 | if (haifa_classify_insn (insn2) != PFREE_CANDIDATE) | |
1976 | /* insn2 not guaranteed to be a 1 base reg load. */ | |
1977 | continue; | |
1978 | ||
1979 | if (INSN_BB (insn2) == bb_trg) | |
1980 | /* insn2 is the similar load, in the target block. */ | |
1981 | return 1; | |
1982 | ||
aae97b21 | 1983 | if (*(candp->split_bbs.first_member) == BLOCK_FOR_INSN (insn2)) |
7a31a7bd | 1984 | /* insn2 is a similar load, in a split-block. */ |
1985 | return 1; | |
1986 | } | |
1987 | } | |
1988 | } | |
1989 | } | |
1990 | ||
1991 | /* Couldn't find a similar load. */ | |
1992 | return 0; | |
1993 | } /* is_pfree */ | |
1994 | ||
7a31a7bd | 1995 | /* Return 1 if load_insn is prisky (i.e. if load_insn is fed by |
1996 | a load moved speculatively, or if load_insn is protected by | |
1997 | a compare on load_insn's address). */ | |
1998 | ||
1999 | static int | |
60b8c5b3 | 2000 | is_prisky (rtx load_insn, int bb_src, int bb_trg) |
7a31a7bd | 2001 | { |
2002 | if (FED_BY_SPEC_LOAD (load_insn)) | |
2003 | return 1; | |
2004 | ||
93f6b030 | 2005 | if (sd_lists_empty_p (load_insn, SD_LIST_BACK)) |
7a31a7bd | 2006 | /* Dependence may 'hide' out of the region. */ |
2007 | return 1; | |
2008 | ||
2009 | if (is_conditionally_protected (load_insn, bb_src, bb_trg)) | |
2010 | return 1; | |
2011 | ||
2012 | return 0; | |
2013 | } | |
2014 | ||
2015 | /* Insn is a candidate to be moved speculatively from bb_src to bb_trg. | |
2016 | Return 1 if insn is exception-free (and the motion is valid) | |
2017 | and 0 otherwise. */ | |
2018 | ||
2019 | static int | |
60b8c5b3 | 2020 | is_exception_free (rtx insn, int bb_src, int bb_trg) |
7a31a7bd | 2021 | { |
2022 | int insn_class = haifa_classify_insn (insn); | |
2023 | ||
2024 | /* Handle non-load insns. */ | |
2025 | switch (insn_class) | |
2026 | { | |
2027 | case TRAP_FREE: | |
2028 | return 1; | |
2029 | case TRAP_RISKY: | |
2030 | return 0; | |
2031 | default:; | |
2032 | } | |
2033 | ||
2034 | /* Handle loads. */ | |
2035 | if (!flag_schedule_speculative_load) | |
2036 | return 0; | |
2037 | IS_LOAD_INSN (insn) = 1; | |
2038 | switch (insn_class) | |
2039 | { | |
2040 | case IFREE: | |
2041 | return (1); | |
2042 | case IRISKY: | |
2043 | return 0; | |
2044 | case PFREE_CANDIDATE: | |
2045 | if (is_pfree (insn, bb_src, bb_trg)) | |
2046 | return 1; | |
2047 | /* Don't 'break' here: PFREE-candidate is also PRISKY-candidate. */ | |
2048 | case PRISKY_CANDIDATE: | |
2049 | if (!flag_schedule_speculative_load_dangerous | |
2050 | || is_prisky (insn, bb_src, bb_trg)) | |
2051 | return 0; | |
2052 | break; | |
2053 | default:; | |
2054 | } | |
2055 | ||
2056 | return flag_schedule_speculative_load_dangerous; | |
2057 | } | |
2058 | \f | |
2059 | /* The number of insns from the current block scheduled so far. */ | |
2060 | static int sched_target_n_insns; | |
2061 | /* The number of insns from the current block to be scheduled in total. */ | |
2062 | static int target_n_insns; | |
2063 | /* The number of insns from the entire region scheduled so far. */ | |
2064 | static int sched_n_insns; | |
2065 | ||
2066 | /* Implementations of the sched_info functions for region scheduling. */ | |
e4897000 | 2067 | static void init_ready_list (void); |
60b8c5b3 | 2068 | static int can_schedule_ready_p (rtx); |
d2412f57 | 2069 | static void begin_schedule_ready (rtx); |
6a1cdb4d | 2070 | static ds_t new_ready (rtx, ds_t); |
60b8c5b3 | 2071 | static int schedule_more_p (void); |
e1ab7874 | 2072 | static const char *rgn_print_insn (const_rtx, int); |
60b8c5b3 | 2073 | static int rgn_rank (rtx, rtx); |
6aed13f1 | 2074 | static void compute_jump_reg_dependencies (rtx, regset); |
7a31a7bd | 2075 | |
6a1cdb4d | 2076 | /* Functions for speculative scheduling. */ |
e1ab7874 | 2077 | static void rgn_add_remove_insn (rtx, int); |
2078 | static void rgn_add_block (basic_block, basic_block); | |
2079 | static void rgn_fix_recovery_cfg (int, int, int); | |
6a1cdb4d | 2080 | static basic_block advance_target_bb (basic_block, rtx); |
6a1cdb4d | 2081 | |
7a31a7bd | 2082 | /* Return nonzero if there are more insns that should be scheduled. */ |
2083 | ||
2084 | static int | |
60b8c5b3 | 2085 | schedule_more_p (void) |
7a31a7bd | 2086 | { |
6a1cdb4d | 2087 | return sched_target_n_insns < target_n_insns; |
7a31a7bd | 2088 | } |
2089 | ||
2090 | /* Add all insns that are initially ready to the ready list READY. Called | |
2091 | once before scheduling a set of insns. */ | |
2092 | ||
2093 | static void | |
e4897000 | 2094 | init_ready_list (void) |
7a31a7bd | 2095 | { |
2096 | rtx prev_head = current_sched_info->prev_head; | |
2097 | rtx next_tail = current_sched_info->next_tail; | |
2098 | int bb_src; | |
2099 | rtx insn; | |
2100 | ||
2101 | target_n_insns = 0; | |
2102 | sched_target_n_insns = 0; | |
2103 | sched_n_insns = 0; | |
2104 | ||
2105 | /* Print debugging information. */ | |
2106 | if (sched_verbose >= 5) | |
a2819fc2 | 2107 | debug_rgn_dependencies (target_bb); |
7a31a7bd | 2108 | |
2109 | /* Prepare current target block info. */ | |
2110 | if (current_nr_blocks > 1) | |
e1ab7874 | 2111 | compute_trg_info (target_bb); |
7a31a7bd | 2112 | |
2113 | /* Initialize ready list with all 'ready' insns in target block. | |
2114 | Count number of insns in the target block being scheduled. */ | |
2115 | for (insn = NEXT_INSN (prev_head); insn != next_tail; insn = NEXT_INSN (insn)) | |
48e1416a | 2116 | { |
d452a169 | 2117 | gcc_assert (TODO_SPEC (insn) == HARD_DEP || TODO_SPEC (insn) == DEP_POSTPONED); |
2118 | TODO_SPEC (insn) = HARD_DEP; | |
e4897000 | 2119 | try_ready (insn); |
e26579fc | 2120 | target_n_insns++; |
6a1cdb4d | 2121 | |
2122 | gcc_assert (!(TODO_SPEC (insn) & BEGIN_CONTROL)); | |
7a31a7bd | 2123 | } |
2124 | ||
2125 | /* Add to ready list all 'ready' insns in valid source blocks. | |
2126 | For speculative insns, check-live, exception-free, and | |
2127 | issue-delay. */ | |
2128 | for (bb_src = target_bb + 1; bb_src < current_nr_blocks; bb_src++) | |
2129 | if (IS_VALID (bb_src)) | |
2130 | { | |
2131 | rtx src_head; | |
2132 | rtx src_next_tail; | |
2133 | rtx tail, head; | |
2134 | ||
6a1cdb4d | 2135 | get_ebb_head_tail (EBB_FIRST_BB (bb_src), EBB_LAST_BB (bb_src), |
2136 | &head, &tail); | |
7a31a7bd | 2137 | src_next_tail = NEXT_INSN (tail); |
2138 | src_head = head; | |
2139 | ||
2140 | for (insn = src_head; insn != src_next_tail; insn = NEXT_INSN (insn)) | |
9221ec44 | 2141 | if (INSN_P (insn)) |
d452a169 | 2142 | { |
2143 | gcc_assert (TODO_SPEC (insn) == HARD_DEP || TODO_SPEC (insn) == DEP_POSTPONED); | |
2144 | TODO_SPEC (insn) = HARD_DEP; | |
2145 | try_ready (insn); | |
2146 | } | |
7a31a7bd | 2147 | } |
2148 | } | |
2149 | ||
2150 | /* Called after taking INSN from the ready list. Returns nonzero if this | |
2151 | insn can be scheduled, nonzero if we should silently discard it. */ | |
2152 | ||
2153 | static int | |
60b8c5b3 | 2154 | can_schedule_ready_p (rtx insn) |
7a31a7bd | 2155 | { |
6a1cdb4d | 2156 | /* An interblock motion? */ |
2157 | if (INSN_BB (insn) != target_bb | |
2158 | && IS_SPECULATIVE_INSN (insn) | |
2159 | && !check_live (insn, INSN_BB (insn))) | |
48e1416a | 2160 | return 0; |
6a1cdb4d | 2161 | else |
2162 | return 1; | |
2163 | } | |
2295df67 | 2164 | |
9ca2c29a | 2165 | /* Updates counter and other information. Split from can_schedule_ready_p () |
6a1cdb4d | 2166 | because when we schedule insn speculatively then insn passed to |
2167 | can_schedule_ready_p () differs from the one passed to | |
2168 | begin_schedule_ready (). */ | |
2169 | static void | |
d2412f57 | 2170 | begin_schedule_ready (rtx insn) |
6a1cdb4d | 2171 | { |
7a31a7bd | 2172 | /* An interblock motion? */ |
2173 | if (INSN_BB (insn) != target_bb) | |
2174 | { | |
7a31a7bd | 2175 | if (IS_SPECULATIVE_INSN (insn)) |
2176 | { | |
6a1cdb4d | 2177 | gcc_assert (check_live (insn, INSN_BB (insn))); |
2178 | ||
7a31a7bd | 2179 | update_live (insn, INSN_BB (insn)); |
2180 | ||
2181 | /* For speculative load, mark insns fed by it. */ | |
2182 | if (IS_LOAD_INSN (insn) || FED_BY_SPEC_LOAD (insn)) | |
2183 | set_spec_fed (insn); | |
2184 | ||
2185 | nr_spec++; | |
2186 | } | |
2187 | nr_inter++; | |
7a31a7bd | 2188 | } |
2189 | else | |
2190 | { | |
2191 | /* In block motion. */ | |
2192 | sched_target_n_insns++; | |
2193 | } | |
2194 | sched_n_insns++; | |
7a31a7bd | 2195 | } |
2196 | ||
6a1cdb4d | 2197 | /* Called after INSN has all its hard dependencies resolved and the speculation |
2198 | of type TS is enough to overcome them all. | |
2199 | Return nonzero if it should be moved to the ready list or the queue, or zero | |
2200 | if we should silently discard it. */ | |
2201 | static ds_t | |
2202 | new_ready (rtx next, ds_t ts) | |
7a31a7bd | 2203 | { |
6a1cdb4d | 2204 | if (INSN_BB (next) != target_bb) |
2205 | { | |
2206 | int not_ex_free = 0; | |
2207 | ||
2208 | /* For speculative insns, before inserting to ready/queue, | |
48e1416a | 2209 | check live, exception-free, and issue-delay. */ |
6a1cdb4d | 2210 | if (!IS_VALID (INSN_BB (next)) |
7a31a7bd | 2211 | || CANT_MOVE (next) |
2212 | || (IS_SPECULATIVE_INSN (next) | |
67900a4f | 2213 | && ((recog_memoized (next) >= 0 |
48e1416a | 2214 | && min_insn_conflict_delay (curr_state, next, next) |
6a1cdb4d | 2215 | > PARAM_VALUE (PARAM_MAX_SCHED_INSN_CONFLICT_DELAY)) |
fd27912f | 2216 | || IS_SPECULATION_CHECK_P (next) |
7a31a7bd | 2217 | || !check_live (next, INSN_BB (next)) |
6a1cdb4d | 2218 | || (not_ex_free = !is_exception_free (next, INSN_BB (next), |
2219 | target_bb))))) | |
2220 | { | |
2221 | if (not_ex_free | |
2222 | /* We are here because is_exception_free () == false. | |
2223 | But we possibly can handle that with control speculation. */ | |
e1ab7874 | 2224 | && sched_deps_info->generate_spec_deps |
2225 | && spec_info->mask & BEGIN_CONTROL) | |
42de5f34 | 2226 | { |
2227 | ds_t new_ds; | |
2228 | ||
2229 | /* Add control speculation to NEXT's dependency type. */ | |
2230 | new_ds = set_dep_weak (ts, BEGIN_CONTROL, MAX_DEP_WEAK); | |
2231 | ||
2232 | /* Check if NEXT can be speculated with new dependency type. */ | |
2233 | if (sched_insn_is_legitimate_for_speculation_p (next, new_ds)) | |
2234 | /* Here we got new control-speculative instruction. */ | |
2235 | ts = new_ds; | |
2236 | else | |
2237 | /* NEXT isn't ready yet. */ | |
d452a169 | 2238 | ts = DEP_POSTPONED; |
42de5f34 | 2239 | } |
6a1cdb4d | 2240 | else |
42de5f34 | 2241 | /* NEXT isn't ready yet. */ |
d452a169 | 2242 | ts = DEP_POSTPONED; |
6a1cdb4d | 2243 | } |
2244 | } | |
48e1416a | 2245 | |
6a1cdb4d | 2246 | return ts; |
7a31a7bd | 2247 | } |
2248 | ||
2249 | /* Return a string that contains the insn uid and optionally anything else | |
2250 | necessary to identify this insn in an output. It's valid to use a | |
2251 | static buffer for this. The ALIGNED parameter should cause the string | |
2252 | to be formatted so that multiple output lines will line up nicely. */ | |
2253 | ||
2254 | static const char * | |
e1ab7874 | 2255 | rgn_print_insn (const_rtx insn, int aligned) |
7a31a7bd | 2256 | { |
2257 | static char tmp[80]; | |
2258 | ||
2259 | if (aligned) | |
2260 | sprintf (tmp, "b%3d: i%4d", INSN_BB (insn), INSN_UID (insn)); | |
2261 | else | |
2262 | { | |
7a31a7bd | 2263 | if (current_nr_blocks > 1 && INSN_BB (insn) != target_bb) |
cda0a5f5 | 2264 | sprintf (tmp, "%d/b%d", INSN_UID (insn), INSN_BB (insn)); |
2265 | else | |
2266 | sprintf (tmp, "%d", INSN_UID (insn)); | |
7a31a7bd | 2267 | } |
2268 | return tmp; | |
2269 | } | |
2270 | ||
2271 | /* Compare priority of two insns. Return a positive number if the second | |
2272 | insn is to be preferred for scheduling, and a negative one if the first | |
2273 | is to be preferred. Zero if they are equally good. */ | |
2274 | ||
2275 | static int | |
60b8c5b3 | 2276 | rgn_rank (rtx insn1, rtx insn2) |
7a31a7bd | 2277 | { |
2278 | /* Some comparison make sense in interblock scheduling only. */ | |
2279 | if (INSN_BB (insn1) != INSN_BB (insn2)) | |
2280 | { | |
2281 | int spec_val, prob_val; | |
2282 | ||
2283 | /* Prefer an inblock motion on an interblock motion. */ | |
2284 | if ((INSN_BB (insn2) == target_bb) && (INSN_BB (insn1) != target_bb)) | |
2285 | return 1; | |
2286 | if ((INSN_BB (insn1) == target_bb) && (INSN_BB (insn2) != target_bb)) | |
2287 | return -1; | |
2288 | ||
2289 | /* Prefer a useful motion on a speculative one. */ | |
2290 | spec_val = IS_SPECULATIVE_INSN (insn1) - IS_SPECULATIVE_INSN (insn2); | |
2291 | if (spec_val) | |
2292 | return spec_val; | |
2293 | ||
2294 | /* Prefer a more probable (speculative) insn. */ | |
2295 | prob_val = INSN_PROBABILITY (insn2) - INSN_PROBABILITY (insn1); | |
2296 | if (prob_val) | |
2297 | return prob_val; | |
2298 | } | |
2299 | return 0; | |
2300 | } | |
2301 | ||
d6141c0c | 2302 | /* NEXT is an instruction that depends on INSN (a backward dependence); |
2303 | return nonzero if we should include this dependence in priority | |
2304 | calculations. */ | |
2305 | ||
e1ab7874 | 2306 | int |
60b8c5b3 | 2307 | contributes_to_priority (rtx next, rtx insn) |
d6141c0c | 2308 | { |
6a1cdb4d | 2309 | /* NEXT and INSN reside in one ebb. */ |
2310 | return BLOCK_TO_BB (BLOCK_NUM (next)) == BLOCK_TO_BB (BLOCK_NUM (insn)); | |
d6141c0c | 2311 | } |
2312 | ||
6aed13f1 | 2313 | /* INSN is a JUMP_INSN. Store the set of registers that must be |
2314 | considered as used by this jump in USED. */ | |
d6141c0c | 2315 | |
2316 | static void | |
60b8c5b3 | 2317 | compute_jump_reg_dependencies (rtx insn ATTRIBUTE_UNUSED, |
6aed13f1 | 2318 | regset used ATTRIBUTE_UNUSED) |
d6141c0c | 2319 | { |
2320 | /* Nothing to do here, since we postprocess jumps in | |
2321 | add_branch_dependences. */ | |
2322 | } | |
2323 | ||
48e1416a | 2324 | /* This variable holds common_sched_info hooks and data relevant to |
e1ab7874 | 2325 | the interblock scheduler. */ |
2326 | static struct common_sched_info_def rgn_common_sched_info; | |
2327 | ||
2328 | ||
2329 | /* This holds data for the dependence analysis relevant to | |
2330 | the interblock scheduler. */ | |
2331 | static struct sched_deps_info_def rgn_sched_deps_info; | |
2332 | ||
2333 | /* This holds constant data used for initializing the above structure | |
2334 | for the Haifa scheduler. */ | |
2335 | static const struct sched_deps_info_def rgn_const_sched_deps_info = | |
2336 | { | |
2337 | compute_jump_reg_dependencies, | |
2338 | NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, | |
2339 | 0, 0, 0 | |
2340 | }; | |
2341 | ||
2342 | /* Same as above, but for the selective scheduler. */ | |
2343 | static const struct sched_deps_info_def rgn_const_sel_sched_deps_info = | |
2344 | { | |
2345 | compute_jump_reg_dependencies, | |
2346 | NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, NULL, | |
2347 | 0, 0, 0 | |
2348 | }; | |
2349 | ||
4db82bc9 | 2350 | /* Return true if scheduling INSN will trigger finish of scheduling |
2351 | current block. */ | |
2352 | static bool | |
2353 | rgn_insn_finishes_block_p (rtx insn) | |
2354 | { | |
2355 | if (INSN_BB (insn) == target_bb | |
2356 | && sched_target_n_insns + 1 == target_n_insns) | |
2357 | /* INSN is the last not-scheduled instruction in the current block. */ | |
2358 | return true; | |
2359 | ||
2360 | return false; | |
2361 | } | |
2362 | ||
7a31a7bd | 2363 | /* Used in schedule_insns to initialize current_sched_info for scheduling |
2364 | regions (or single basic blocks). */ | |
2365 | ||
e1ab7874 | 2366 | static const struct haifa_sched_info rgn_const_sched_info = |
7a31a7bd | 2367 | { |
2368 | init_ready_list, | |
2369 | can_schedule_ready_p, | |
2370 | schedule_more_p, | |
2371 | new_ready, | |
2372 | rgn_rank, | |
2373 | rgn_print_insn, | |
d6141c0c | 2374 | contributes_to_priority, |
4db82bc9 | 2375 | rgn_insn_finishes_block_p, |
7a31a7bd | 2376 | |
2377 | NULL, NULL, | |
2378 | NULL, NULL, | |
e1ab7874 | 2379 | 0, 0, |
4d64d9a4 | 2380 | |
e1ab7874 | 2381 | rgn_add_remove_insn, |
6a1cdb4d | 2382 | begin_schedule_ready, |
d2412f57 | 2383 | NULL, |
6a1cdb4d | 2384 | advance_target_bb, |
e2f4a6ff | 2385 | NULL, NULL, |
3072d30e | 2386 | SCHED_RGN |
7a31a7bd | 2387 | }; |
2388 | ||
e1ab7874 | 2389 | /* This variable holds the data and hooks needed to the Haifa scheduler backend |
2390 | for the interblock scheduler frontend. */ | |
2391 | static struct haifa_sched_info rgn_sched_info; | |
2392 | ||
2393 | /* Returns maximum priority that an insn was assigned to. */ | |
2394 | ||
2395 | int | |
2396 | get_rgn_sched_max_insns_priority (void) | |
2397 | { | |
2398 | return rgn_sched_info.sched_max_insns_priority; | |
2399 | } | |
2400 | ||
24dd0668 | 2401 | /* Determine if PAT sets a TARGET_CLASS_LIKELY_SPILLED_P register. */ |
cbf780cc | 2402 | |
2403 | static bool | |
60b8c5b3 | 2404 | sets_likely_spilled (rtx pat) |
cbf780cc | 2405 | { |
2406 | bool ret = false; | |
2407 | note_stores (pat, sets_likely_spilled_1, &ret); | |
2408 | return ret; | |
2409 | } | |
2410 | ||
2411 | static void | |
81a410b1 | 2412 | sets_likely_spilled_1 (rtx x, const_rtx pat, void *data) |
cbf780cc | 2413 | { |
2414 | bool *ret = (bool *) data; | |
2415 | ||
2416 | if (GET_CODE (pat) == SET | |
2417 | && REG_P (x) | |
24dd0668 | 2418 | && HARD_REGISTER_P (x) |
2419 | && targetm.class_likely_spilled_p (REGNO_REG_CLASS (REGNO (x)))) | |
cbf780cc | 2420 | *ret = true; |
2421 | } | |
2422 | ||
566e7db2 | 2423 | /* A bitmap to note insns that participate in any dependency. Used in |
2424 | add_branch_dependences. */ | |
2425 | static sbitmap insn_referenced; | |
e1ab7874 | 2426 | |
7a31a7bd | 2427 | /* Add dependences so that branches are scheduled to run last in their |
2428 | block. */ | |
7a31a7bd | 2429 | static void |
60b8c5b3 | 2430 | add_branch_dependences (rtx head, rtx tail) |
7a31a7bd | 2431 | { |
2432 | rtx insn, last; | |
2433 | ||
cbaab9a3 | 2434 | /* For all branches, calls, uses, clobbers, cc0 setters, and instructions |
2435 | that can throw exceptions, force them to remain in order at the end of | |
2436 | the block by adding dependencies and giving the last a high priority. | |
2437 | There may be notes present, and prev_head may also be a note. | |
7a31a7bd | 2438 | |
2439 | Branches must obviously remain at the end. Calls should remain at the | |
2440 | end since moving them results in worse register allocation. Uses remain | |
cbf780cc | 2441 | at the end to ensure proper register allocation. |
2442 | ||
40e55fbb | 2443 | cc0 setters remain at the end because they can't be moved away from |
cbf780cc | 2444 | their cc0 user. |
2445 | ||
681b9609 | 2446 | Predecessors of SCHED_GROUP_P instructions at the end remain at the end. |
2447 | ||
e6a25dc9 | 2448 | COND_EXEC insns cannot be moved past a branch (see e.g. PR17808). |
2449 | ||
24dd0668 | 2450 | Insns setting TARGET_CLASS_LIKELY_SPILLED_P registers (usually return |
2451 | values) are not moved before reload because we can wind up with register | |
cbf780cc | 2452 | allocation failures. */ |
2453 | ||
9845d120 | 2454 | while (tail != head && DEBUG_INSN_P (tail)) |
2455 | tail = PREV_INSN (tail); | |
2456 | ||
7a31a7bd | 2457 | insn = tail; |
2458 | last = 0; | |
6d7dc5b9 | 2459 | while (CALL_P (insn) |
91f71fa3 | 2460 | || JUMP_P (insn) || JUMP_TABLE_DATA_P (insn) |
6d7dc5b9 | 2461 | || (NONJUMP_INSN_P (insn) |
7a31a7bd | 2462 | && (GET_CODE (PATTERN (insn)) == USE |
2463 | || GET_CODE (PATTERN (insn)) == CLOBBER | |
cbaab9a3 | 2464 | || can_throw_internal (insn) |
7a31a7bd | 2465 | #ifdef HAVE_cc0 |
2466 | || sets_cc0_p (PATTERN (insn)) | |
2467 | #endif | |
cbf780cc | 2468 | || (!reload_completed |
2469 | && sets_likely_spilled (PATTERN (insn))))) | |
681b9609 | 2470 | || NOTE_P (insn) |
2471 | || (last != 0 && SCHED_GROUP_P (last))) | |
7a31a7bd | 2472 | { |
6d7dc5b9 | 2473 | if (!NOTE_P (insn)) |
7a31a7bd | 2474 | { |
9997bd27 | 2475 | if (last != 0 |
93f6b030 | 2476 | && sd_find_dep_between (insn, last, false) == NULL) |
7a31a7bd | 2477 | { |
e6a25dc9 | 2478 | if (! sched_insns_conditions_mutex_p (last, insn)) |
2479 | add_dependence (last, insn, REG_DEP_ANTI); | |
08b7917c | 2480 | bitmap_set_bit (insn_referenced, INSN_LUID (insn)); |
7a31a7bd | 2481 | } |
2482 | ||
2483 | CANT_MOVE (insn) = 1; | |
2484 | ||
2485 | last = insn; | |
7a31a7bd | 2486 | } |
2487 | ||
2488 | /* Don't overrun the bounds of the basic block. */ | |
2489 | if (insn == head) | |
2490 | break; | |
2491 | ||
9845d120 | 2492 | do |
2493 | insn = PREV_INSN (insn); | |
2494 | while (insn != head && DEBUG_INSN_P (insn)); | |
7a31a7bd | 2495 | } |
2496 | ||
2497 | /* Make sure these insns are scheduled last in their block. */ | |
2498 | insn = last; | |
2499 | if (insn != 0) | |
2500 | while (insn != head) | |
2501 | { | |
2502 | insn = prev_nonnote_insn (insn); | |
2503 | ||
08b7917c | 2504 | if (bitmap_bit_p (insn_referenced, INSN_LUID (insn)) |
9845d120 | 2505 | || DEBUG_INSN_P (insn)) |
7a31a7bd | 2506 | continue; |
2507 | ||
e6a25dc9 | 2508 | if (! sched_insns_conditions_mutex_p (last, insn)) |
2509 | add_dependence (last, insn, REG_DEP_ANTI); | |
7a31a7bd | 2510 | } |
e6a25dc9 | 2511 | |
751d3ba7 | 2512 | if (!targetm.have_conditional_execution ()) |
2513 | return; | |
2514 | ||
e6a25dc9 | 2515 | /* Finally, if the block ends in a jump, and we are doing intra-block |
2516 | scheduling, make sure that the branch depends on any COND_EXEC insns | |
2517 | inside the block to avoid moving the COND_EXECs past the branch insn. | |
2518 | ||
2519 | We only have to do this after reload, because (1) before reload there | |
2520 | are no COND_EXEC insns, and (2) the region scheduler is an intra-block | |
2521 | scheduler after reload. | |
2522 | ||
2523 | FIXME: We could in some cases move COND_EXEC insns past the branch if | |
2524 | this scheduler would be a little smarter. Consider this code: | |
2525 | ||
2526 | T = [addr] | |
2527 | C ? addr += 4 | |
f19b9016 | 2528 | !C ? X += 12 |
e6a25dc9 | 2529 | C ? T += 1 |
f19b9016 | 2530 | C ? jump foo |
e6a25dc9 | 2531 | |
2532 | On a target with a one cycle stall on a memory access the optimal | |
2533 | sequence would be: | |
2534 | ||
2535 | T = [addr] | |
2536 | C ? addr += 4 | |
2537 | C ? T += 1 | |
2538 | C ? jump foo | |
2539 | !C ? X += 12 | |
2540 | ||
2541 | We don't want to put the 'X += 12' before the branch because it just | |
2542 | wastes a cycle of execution time when the branch is taken. | |
2543 | ||
2544 | Note that in the example "!C" will always be true. That is another | |
2545 | possible improvement for handling COND_EXECs in this scheduler: it | |
2546 | could remove always-true predicates. */ | |
2547 | ||
91f71fa3 | 2548 | if (!reload_completed || ! (JUMP_P (tail) || JUMP_TABLE_DATA_P (tail))) |
e6a25dc9 | 2549 | return; |
2550 | ||
f19b9016 | 2551 | insn = tail; |
e6a25dc9 | 2552 | while (insn != head) |
2553 | { | |
f19b9016 | 2554 | insn = PREV_INSN (insn); |
2555 | ||
e6a25dc9 | 2556 | /* Note that we want to add this dependency even when |
2557 | sched_insns_conditions_mutex_p returns true. The whole point | |
2558 | is that we _want_ this dependency, even if these insns really | |
2559 | are independent. */ | |
2560 | if (INSN_P (insn) && GET_CODE (PATTERN (insn)) == COND_EXEC) | |
2561 | add_dependence (tail, insn, REG_DEP_ANTI); | |
e6a25dc9 | 2562 | } |
7a31a7bd | 2563 | } |
2564 | ||
2565 | /* Data structures for the computation of data dependences in a regions. We | |
2566 | keep one `deps' structure for every basic block. Before analyzing the | |
2567 | data dependences for a bb, its variables are initialized as a function of | |
2568 | the variables of its predecessors. When the analysis for a bb completes, | |
2569 | we save the contents to the corresponding bb_deps[bb] variable. */ | |
2570 | ||
68e419a1 | 2571 | static struct deps_desc *bb_deps; |
7a31a7bd | 2572 | |
5deaeb50 | 2573 | static void |
60b8c5b3 | 2574 | concat_insn_mem_list (rtx copy_insns, rtx copy_mems, rtx *old_insns_p, |
2575 | rtx *old_mems_p) | |
5deaeb50 | 2576 | { |
2577 | rtx new_insns = *old_insns_p; | |
2578 | rtx new_mems = *old_mems_p; | |
2579 | ||
2580 | while (copy_insns) | |
2581 | { | |
2582 | new_insns = alloc_INSN_LIST (XEXP (copy_insns, 0), new_insns); | |
2583 | new_mems = alloc_EXPR_LIST (VOIDmode, XEXP (copy_mems, 0), new_mems); | |
2584 | copy_insns = XEXP (copy_insns, 1); | |
2585 | copy_mems = XEXP (copy_mems, 1); | |
2586 | } | |
2587 | ||
2588 | *old_insns_p = new_insns; | |
2589 | *old_mems_p = new_mems; | |
2590 | } | |
2591 | ||
e1ab7874 | 2592 | /* Join PRED_DEPS to the SUCC_DEPS. */ |
2593 | void | |
68e419a1 | 2594 | deps_join (struct deps_desc *succ_deps, struct deps_desc *pred_deps) |
e1ab7874 | 2595 | { |
2596 | unsigned reg; | |
2597 | reg_set_iterator rsi; | |
2598 | ||
2599 | /* The reg_last lists are inherited by successor. */ | |
2600 | EXECUTE_IF_SET_IN_REG_SET (&pred_deps->reg_last_in_use, 0, reg, rsi) | |
2601 | { | |
2602 | struct deps_reg *pred_rl = &pred_deps->reg_last[reg]; | |
2603 | struct deps_reg *succ_rl = &succ_deps->reg_last[reg]; | |
2604 | ||
2605 | succ_rl->uses = concat_INSN_LIST (pred_rl->uses, succ_rl->uses); | |
2606 | succ_rl->sets = concat_INSN_LIST (pred_rl->sets, succ_rl->sets); | |
a7dcf969 | 2607 | succ_rl->implicit_sets |
2608 | = concat_INSN_LIST (pred_rl->implicit_sets, succ_rl->implicit_sets); | |
e1ab7874 | 2609 | succ_rl->clobbers = concat_INSN_LIST (pred_rl->clobbers, |
2610 | succ_rl->clobbers); | |
2611 | succ_rl->uses_length += pred_rl->uses_length; | |
2612 | succ_rl->clobbers_length += pred_rl->clobbers_length; | |
2613 | } | |
2614 | IOR_REG_SET (&succ_deps->reg_last_in_use, &pred_deps->reg_last_in_use); | |
2615 | ||
2616 | /* Mem read/write lists are inherited by successor. */ | |
2617 | concat_insn_mem_list (pred_deps->pending_read_insns, | |
2618 | pred_deps->pending_read_mems, | |
2619 | &succ_deps->pending_read_insns, | |
2620 | &succ_deps->pending_read_mems); | |
2621 | concat_insn_mem_list (pred_deps->pending_write_insns, | |
2622 | pred_deps->pending_write_mems, | |
2623 | &succ_deps->pending_write_insns, | |
2624 | &succ_deps->pending_write_mems); | |
2625 | ||
effd1640 | 2626 | succ_deps->pending_jump_insns |
2627 | = concat_INSN_LIST (pred_deps->pending_jump_insns, | |
2628 | succ_deps->pending_jump_insns); | |
e1ab7874 | 2629 | succ_deps->last_pending_memory_flush |
2630 | = concat_INSN_LIST (pred_deps->last_pending_memory_flush, | |
2631 | succ_deps->last_pending_memory_flush); | |
2632 | ||
2633 | succ_deps->pending_read_list_length += pred_deps->pending_read_list_length; | |
2634 | succ_deps->pending_write_list_length += pred_deps->pending_write_list_length; | |
2635 | succ_deps->pending_flush_length += pred_deps->pending_flush_length; | |
2636 | ||
2637 | /* last_function_call is inherited by successor. */ | |
2638 | succ_deps->last_function_call | |
2639 | = concat_INSN_LIST (pred_deps->last_function_call, | |
2640 | succ_deps->last_function_call); | |
2641 | ||
326d0c19 | 2642 | /* last_function_call_may_noreturn is inherited by successor. */ |
2643 | succ_deps->last_function_call_may_noreturn | |
2644 | = concat_INSN_LIST (pred_deps->last_function_call_may_noreturn, | |
2645 | succ_deps->last_function_call_may_noreturn); | |
2646 | ||
e1ab7874 | 2647 | /* sched_before_next_call is inherited by successor. */ |
2648 | succ_deps->sched_before_next_call | |
2649 | = concat_INSN_LIST (pred_deps->sched_before_next_call, | |
2650 | succ_deps->sched_before_next_call); | |
2651 | } | |
2652 | ||
7a31a7bd | 2653 | /* After computing the dependencies for block BB, propagate the dependencies |
749c6f58 | 2654 | found in TMP_DEPS to the successors of the block. */ |
7a31a7bd | 2655 | static void |
68e419a1 | 2656 | propagate_deps (int bb, struct deps_desc *pred_deps) |
7a31a7bd | 2657 | { |
aae97b21 | 2658 | basic_block block = BASIC_BLOCK (BB_TO_BLOCK (bb)); |
2659 | edge_iterator ei; | |
2660 | edge e; | |
7a31a7bd | 2661 | |
2662 | /* bb's structures are inherited by its successors. */ | |
aae97b21 | 2663 | FOR_EACH_EDGE (e, ei, block->succs) |
2664 | { | |
aae97b21 | 2665 | /* Only bbs "below" bb, in the same region, are interesting. */ |
2666 | if (e->dest == EXIT_BLOCK_PTR | |
2667 | || CONTAINING_RGN (block->index) != CONTAINING_RGN (e->dest->index) | |
2668 | || BLOCK_TO_BB (e->dest->index) <= bb) | |
2669 | continue; | |
5deaeb50 | 2670 | |
e1ab7874 | 2671 | deps_join (bb_deps + BLOCK_TO_BB (e->dest->index), pred_deps); |
aae97b21 | 2672 | } |
7a31a7bd | 2673 | |
5deaeb50 | 2674 | /* These lists should point to the right place, for correct |
2675 | freeing later. */ | |
2676 | bb_deps[bb].pending_read_insns = pred_deps->pending_read_insns; | |
2677 | bb_deps[bb].pending_read_mems = pred_deps->pending_read_mems; | |
2678 | bb_deps[bb].pending_write_insns = pred_deps->pending_write_insns; | |
2679 | bb_deps[bb].pending_write_mems = pred_deps->pending_write_mems; | |
effd1640 | 2680 | bb_deps[bb].pending_jump_insns = pred_deps->pending_jump_insns; |
5deaeb50 | 2681 | |
2682 | /* Can't allow these to be freed twice. */ | |
2683 | pred_deps->pending_read_insns = 0; | |
2684 | pred_deps->pending_read_mems = 0; | |
2685 | pred_deps->pending_write_insns = 0; | |
2686 | pred_deps->pending_write_mems = 0; | |
effd1640 | 2687 | pred_deps->pending_jump_insns = 0; |
7a31a7bd | 2688 | } |
2689 | ||
93f6b030 | 2690 | /* Compute dependences inside bb. In a multiple blocks region: |
7a31a7bd | 2691 | (1) a bb is analyzed after its predecessors, and (2) the lists in |
2692 | effect at the end of bb (after analyzing for bb) are inherited by | |
de132707 | 2693 | bb's successors. |
7a31a7bd | 2694 | |
2695 | Specifically for reg-reg data dependences, the block insns are | |
a7dcf969 | 2696 | scanned by sched_analyze () top-to-bottom. Three lists are |
749c6f58 | 2697 | maintained by sched_analyze (): reg_last[].sets for register DEFs, |
a7dcf969 | 2698 | reg_last[].implicit_sets for implicit hard register DEFs, and |
2699 | reg_last[].uses for register USEs. | |
7a31a7bd | 2700 | |
2701 | When analysis is completed for bb, we update for its successors: | |
2702 | ; - DEFS[succ] = Union (DEFS [succ], DEFS [bb]) | |
a7dcf969 | 2703 | ; - IMPLICIT_DEFS[succ] = Union (IMPLICIT_DEFS [succ], IMPLICIT_DEFS [bb]) |
7a31a7bd | 2704 | ; - USES[succ] = Union (USES [succ], DEFS [bb]) |
2705 | ||
2706 | The mechanism for computing mem-mem data dependence is very | |
2707 | similar, and the result is interblock dependences in the region. */ | |
2708 | ||
2709 | static void | |
93f6b030 | 2710 | compute_block_dependences (int bb) |
7a31a7bd | 2711 | { |
2712 | rtx head, tail; | |
68e419a1 | 2713 | struct deps_desc tmp_deps; |
7a31a7bd | 2714 | |
2715 | tmp_deps = bb_deps[bb]; | |
2716 | ||
2717 | /* Do the analysis for this block. */ | |
6a1cdb4d | 2718 | gcc_assert (EBB_FIRST_BB (bb) == EBB_LAST_BB (bb)); |
2719 | get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail); | |
93f6b030 | 2720 | |
7a31a7bd | 2721 | sched_analyze (&tmp_deps, head, tail); |
e1ab7874 | 2722 | |
2723 | /* Selective scheduling handles control dependencies by itself. */ | |
2724 | if (!sel_sched_p ()) | |
2725 | add_branch_dependences (head, tail); | |
7a31a7bd | 2726 | |
2727 | if (current_nr_blocks > 1) | |
749c6f58 | 2728 | propagate_deps (bb, &tmp_deps); |
7a31a7bd | 2729 | |
2730 | /* Free up the INSN_LISTs. */ | |
2731 | free_deps (&tmp_deps); | |
93f6b030 | 2732 | |
2733 | if (targetm.sched.dependencies_evaluation_hook) | |
2734 | targetm.sched.dependencies_evaluation_hook (head, tail); | |
2735 | } | |
2736 | ||
2737 | /* Free dependencies of instructions inside BB. */ | |
2738 | static void | |
2739 | free_block_dependencies (int bb) | |
2740 | { | |
2741 | rtx head; | |
2742 | rtx tail; | |
2743 | ||
2744 | get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail); | |
2745 | ||
9845d120 | 2746 | if (no_real_insns_p (head, tail)) |
2747 | return; | |
2748 | ||
93f6b030 | 2749 | sched_free_deps (head, tail, true); |
7a31a7bd | 2750 | } |
749c6f58 | 2751 | |
7a31a7bd | 2752 | /* Remove all INSN_LISTs and EXPR_LISTs from the pending lists and add |
2753 | them to the unused_*_list variables, so that they can be reused. */ | |
2754 | ||
2755 | static void | |
60b8c5b3 | 2756 | free_pending_lists (void) |
7a31a7bd | 2757 | { |
2758 | int bb; | |
2759 | ||
2760 | for (bb = 0; bb < current_nr_blocks; bb++) | |
2761 | { | |
2762 | free_INSN_LIST_list (&bb_deps[bb].pending_read_insns); | |
2763 | free_INSN_LIST_list (&bb_deps[bb].pending_write_insns); | |
2764 | free_EXPR_LIST_list (&bb_deps[bb].pending_read_mems); | |
2765 | free_EXPR_LIST_list (&bb_deps[bb].pending_write_mems); | |
effd1640 | 2766 | free_INSN_LIST_list (&bb_deps[bb].pending_jump_insns); |
7a31a7bd | 2767 | } |
2768 | } | |
2769 | \f | |
93f6b030 | 2770 | /* Print dependences for debugging starting from FROM_BB. |
2771 | Callable from debugger. */ | |
a2819fc2 | 2772 | /* Print dependences for debugging starting from FROM_BB. |
2773 | Callable from debugger. */ | |
4b987fac | 2774 | DEBUG_FUNCTION void |
a2819fc2 | 2775 | debug_rgn_dependencies (int from_bb) |
7a31a7bd | 2776 | { |
2777 | int bb; | |
2778 | ||
a2819fc2 | 2779 | fprintf (sched_dump, |
2780 | ";; --------------- forward dependences: ------------ \n"); | |
2781 | ||
2782 | for (bb = from_bb; bb < current_nr_blocks; bb++) | |
7a31a7bd | 2783 | { |
67900a4f | 2784 | rtx head, tail; |
67900a4f | 2785 | |
6a1cdb4d | 2786 | get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail); |
67900a4f | 2787 | fprintf (sched_dump, "\n;; --- Region Dependences --- b %d bb %d \n", |
2788 | BB_TO_BLOCK (bb), bb); | |
2789 | ||
a2819fc2 | 2790 | debug_dependencies (head, tail); |
2791 | } | |
2792 | } | |
67900a4f | 2793 | |
a2819fc2 | 2794 | /* Print dependencies information for instructions between HEAD and TAIL. |
2795 | ??? This function would probably fit best in haifa-sched.c. */ | |
2796 | void debug_dependencies (rtx head, rtx tail) | |
2797 | { | |
2798 | rtx insn; | |
2799 | rtx next_tail = NEXT_INSN (tail); | |
2800 | ||
2801 | fprintf (sched_dump, ";; %7s%6s%6s%6s%6s%6s%14s\n", | |
2802 | "insn", "code", "bb", "dep", "prio", "cost", | |
2803 | "reservation"); | |
2804 | fprintf (sched_dump, ";; %7s%6s%6s%6s%6s%6s%14s\n", | |
2805 | "----", "----", "--", "---", "----", "----", | |
2806 | "-----------"); | |
7a31a7bd | 2807 | |
a2819fc2 | 2808 | for (insn = head; insn != next_tail; insn = NEXT_INSN (insn)) |
2809 | { | |
a2819fc2 | 2810 | if (! INSN_P (insn)) |
2811 | { | |
2812 | int n; | |
2813 | fprintf (sched_dump, ";; %6d ", INSN_UID (insn)); | |
2814 | if (NOTE_P (insn)) | |
7a31a7bd | 2815 | { |
ad4583d9 | 2816 | n = NOTE_KIND (insn); |
2817 | fprintf (sched_dump, "%s\n", GET_NOTE_INSN_NAME (n)); | |
7a31a7bd | 2818 | } |
67900a4f | 2819 | else |
a2819fc2 | 2820 | fprintf (sched_dump, " {%s}\n", GET_RTX_NAME (GET_CODE (insn))); |
2821 | continue; | |
7a31a7bd | 2822 | } |
a2819fc2 | 2823 | |
2824 | fprintf (sched_dump, | |
2825 | ";; %s%5d%6d%6d%6d%6d%6d ", | |
2826 | (SCHED_GROUP_P (insn) ? "+" : " "), | |
2827 | INSN_UID (insn), | |
2828 | INSN_CODE (insn), | |
2829 | BLOCK_NUM (insn), | |
e1ab7874 | 2830 | sched_emulate_haifa_p ? -1 : sd_lists_size (insn, SD_LIST_BACK), |
2831 | (sel_sched_p () ? (sched_emulate_haifa_p ? -1 | |
2832 | : INSN_PRIORITY (insn)) | |
2833 | : INSN_PRIORITY (insn)), | |
2834 | (sel_sched_p () ? (sched_emulate_haifa_p ? -1 | |
2835 | : insn_cost (insn)) | |
2836 | : insn_cost (insn))); | |
a2819fc2 | 2837 | |
2838 | if (recog_memoized (insn) < 0) | |
2839 | fprintf (sched_dump, "nothing"); | |
2840 | else | |
2841 | print_reservation (sched_dump, insn); | |
2842 | ||
2843 | fprintf (sched_dump, "\t: "); | |
93f6b030 | 2844 | { |
2845 | sd_iterator_def sd_it; | |
2846 | dep_t dep; | |
2847 | ||
2848 | FOR_EACH_DEP (insn, SD_LIST_FORW, sd_it, dep) | |
d452a169 | 2849 | fprintf (sched_dump, "%d%s%s ", INSN_UID (DEP_CON (dep)), |
2850 | DEP_NONREG (dep) ? "n" : "", | |
2851 | DEP_MULTIPLE (dep) ? "m" : ""); | |
93f6b030 | 2852 | } |
a2819fc2 | 2853 | fprintf (sched_dump, "\n"); |
7a31a7bd | 2854 | } |
a2819fc2 | 2855 | |
7a31a7bd | 2856 | fprintf (sched_dump, "\n"); |
2857 | } | |
2858 | \f | |
f045d41d | 2859 | /* Returns true if all the basic blocks of the current region have |
2860 | NOTE_DISABLE_SCHED_OF_BLOCK which means not to schedule that region. */ | |
e1ab7874 | 2861 | bool |
f045d41d | 2862 | sched_is_disabled_for_current_region_p (void) |
2863 | { | |
f045d41d | 2864 | int bb; |
2865 | ||
2866 | for (bb = 0; bb < current_nr_blocks; bb++) | |
7562ed74 | 2867 | if (!(BASIC_BLOCK (BB_TO_BLOCK (bb))->flags & BB_DISABLE_SCHEDULE)) |
2868 | return false; | |
f045d41d | 2869 | |
2870 | return true; | |
2871 | } | |
2872 | ||
48e1416a | 2873 | /* Free all region dependencies saved in INSN_BACK_DEPS and |
e1ab7874 | 2874 | INSN_RESOLVED_BACK_DEPS. The Haifa scheduler does this on the fly |
48e1416a | 2875 | when scheduling, so this function is supposed to be called from |
e1ab7874 | 2876 | the selective scheduling only. */ |
2877 | void | |
2878 | free_rgn_deps (void) | |
7a31a7bd | 2879 | { |
2880 | int bb; | |
f045d41d | 2881 | |
e1ab7874 | 2882 | for (bb = 0; bb < current_nr_blocks; bb++) |
6a1cdb4d | 2883 | { |
e1ab7874 | 2884 | rtx head, tail; |
48e1416a | 2885 | |
e1ab7874 | 2886 | gcc_assert (EBB_FIRST_BB (bb) == EBB_LAST_BB (bb)); |
2887 | get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail); | |
7a31a7bd | 2888 | |
e1ab7874 | 2889 | sched_free_deps (head, tail, false); |
2890 | } | |
2891 | } | |
7a31a7bd | 2892 | |
e1ab7874 | 2893 | static int rgn_n_insns; |
6a1cdb4d | 2894 | |
e1ab7874 | 2895 | /* Compute insn priority for a current region. */ |
2896 | void | |
48e1416a | 2897 | compute_priorities (void) |
e1ab7874 | 2898 | { |
2899 | int bb; | |
6a1cdb4d | 2900 | |
e4897000 | 2901 | current_sched_info->sched_max_insns_priority = 0; |
7a31a7bd | 2902 | for (bb = 0; bb < current_nr_blocks; bb++) |
2295df67 | 2903 | { |
2904 | rtx head, tail; | |
48e1416a | 2905 | |
6a1cdb4d | 2906 | gcc_assert (EBB_FIRST_BB (bb) == EBB_LAST_BB (bb)); |
2907 | get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail); | |
2295df67 | 2908 | |
9845d120 | 2909 | if (no_real_insns_p (head, tail)) |
2910 | continue; | |
2911 | ||
2295df67 | 2912 | rgn_n_insns += set_priorities (head, tail); |
2913 | } | |
e4897000 | 2914 | current_sched_info->sched_max_insns_priority++; |
e1ab7874 | 2915 | } |
7a31a7bd | 2916 | |
bbd0cfb1 | 2917 | /* (Re-)initialize the arrays of DFA states at the end of each basic block. |
2918 | ||
2919 | SAVED_LAST_BASIC_BLOCK is the previous length of the arrays. It must be | |
2920 | zero for the first call to this function, to allocate the arrays for the | |
2921 | first time. | |
2922 | ||
2923 | This function is called once during initialization of the scheduler, and | |
2924 | called again to resize the arrays if new basic blocks have been created, | |
2925 | for example for speculation recovery code. */ | |
2926 | ||
2927 | static void | |
2928 | realloc_bb_state_array (int saved_last_basic_block) | |
2929 | { | |
2930 | char *old_bb_state_array = bb_state_array; | |
2931 | size_t lbb = (size_t) last_basic_block; | |
2932 | size_t slbb = (size_t) saved_last_basic_block; | |
2933 | ||
2934 | /* Nothing to do if nothing changed since the last time this was called. */ | |
2935 | if (saved_last_basic_block == last_basic_block) | |
2936 | return; | |
2937 | ||
2938 | /* The selective scheduler doesn't use the state arrays. */ | |
2939 | if (sel_sched_p ()) | |
2940 | { | |
2941 | gcc_assert (bb_state_array == NULL && bb_state == NULL); | |
2942 | return; | |
2943 | } | |
2944 | ||
2945 | gcc_checking_assert (saved_last_basic_block == 0 | |
2946 | || (bb_state_array != NULL && bb_state != NULL)); | |
2947 | ||
2948 | bb_state_array = XRESIZEVEC (char, bb_state_array, lbb * dfa_state_size); | |
2949 | bb_state = XRESIZEVEC (state_t, bb_state, lbb); | |
2950 | ||
2951 | /* If BB_STATE_ARRAY has moved, fixup all the state pointers array. | |
2952 | Otherwise only fixup the newly allocated ones. For the state | |
2953 | array itself, only initialize the new entries. */ | |
2954 | bool bb_state_array_moved = (bb_state_array != old_bb_state_array); | |
2955 | for (size_t i = bb_state_array_moved ? 0 : slbb; i < lbb; i++) | |
2956 | bb_state[i] = (state_t) (bb_state_array + i * dfa_state_size); | |
2957 | for (size_t i = slbb; i < lbb; i++) | |
2958 | state_reset (bb_state[i]); | |
2959 | } | |
2960 | ||
2961 | /* Free the arrays of DFA states at the end of each basic block. */ | |
2962 | ||
2963 | static void | |
2964 | free_bb_state_array (void) | |
2965 | { | |
2966 | free (bb_state_array); | |
2967 | free (bb_state); | |
2968 | bb_state_array = NULL; | |
2969 | bb_state = NULL; | |
2970 | } | |
2971 | ||
e1ab7874 | 2972 | /* Schedule a region. A region is either an inner loop, a loop-free |
2973 | subroutine, or a single basic block. Each bb in the region is | |
2974 | scheduled after its flow predecessors. */ | |
7a31a7bd | 2975 | |
e1ab7874 | 2976 | static void |
2977 | schedule_region (int rgn) | |
2978 | { | |
2979 | int bb; | |
2980 | int sched_rgn_n_insns = 0; | |
aae97b21 | 2981 | |
e1ab7874 | 2982 | rgn_n_insns = 0; |
7a31a7bd | 2983 | |
e1ab7874 | 2984 | rgn_setup_region (rgn); |
7a31a7bd | 2985 | |
e1ab7874 | 2986 | /* Don't schedule region that is marked by |
2987 | NOTE_DISABLE_SCHED_OF_BLOCK. */ | |
2988 | if (sched_is_disabled_for_current_region_p ()) | |
2989 | return; | |
7a31a7bd | 2990 | |
e1ab7874 | 2991 | sched_rgn_compute_dependencies (rgn); |
6a1cdb4d | 2992 | |
e1ab7874 | 2993 | sched_rgn_local_init (rgn); |
2994 | ||
2995 | /* Set priorities. */ | |
2996 | compute_priorities (); | |
2997 | ||
2998 | sched_extend_ready_list (rgn_n_insns); | |
7a31a7bd | 2999 | |
11189c7a | 3000 | if (sched_pressure == SCHED_PRESSURE_WEIGHTED) |
a7dcf969 | 3001 | { |
3002 | sched_init_region_reg_pressure_info (); | |
3003 | for (bb = 0; bb < current_nr_blocks; bb++) | |
3004 | { | |
3005 | basic_block first_bb, last_bb; | |
3006 | rtx head, tail; | |
48e1416a | 3007 | |
a7dcf969 | 3008 | first_bb = EBB_FIRST_BB (bb); |
3009 | last_bb = EBB_LAST_BB (bb); | |
48e1416a | 3010 | |
a7dcf969 | 3011 | get_ebb_head_tail (first_bb, last_bb, &head, &tail); |
48e1416a | 3012 | |
a7dcf969 | 3013 | if (no_real_insns_p (head, tail)) |
3014 | { | |
3015 | gcc_assert (first_bb == last_bb); | |
3016 | continue; | |
3017 | } | |
3018 | sched_setup_bb_reg_pressure_info (first_bb, PREV_INSN (head)); | |
3019 | } | |
3020 | } | |
3021 | ||
7a31a7bd | 3022 | /* Now we can schedule all blocks. */ |
3023 | for (bb = 0; bb < current_nr_blocks; bb++) | |
3024 | { | |
6a1cdb4d | 3025 | basic_block first_bb, last_bb, curr_bb; |
7a31a7bd | 3026 | rtx head, tail; |
7a31a7bd | 3027 | |
6a1cdb4d | 3028 | first_bb = EBB_FIRST_BB (bb); |
3029 | last_bb = EBB_LAST_BB (bb); | |
3030 | ||
3031 | get_ebb_head_tail (first_bb, last_bb, &head, &tail); | |
7a31a7bd | 3032 | |
3033 | if (no_real_insns_p (head, tail)) | |
6a1cdb4d | 3034 | { |
3035 | gcc_assert (first_bb == last_bb); | |
3036 | continue; | |
3037 | } | |
7a31a7bd | 3038 | |
3039 | current_sched_info->prev_head = PREV_INSN (head); | |
3040 | current_sched_info->next_tail = NEXT_INSN (tail); | |
3041 | ||
e1ab7874 | 3042 | remove_notes (head, tail); |
7a31a7bd | 3043 | |
6a1cdb4d | 3044 | unlink_bb_notes (first_bb, last_bb); |
3045 | ||
7a31a7bd | 3046 | target_bb = bb; |
3047 | ||
e4897000 | 3048 | gcc_assert (flag_schedule_interblock || current_nr_blocks == 1); |
3049 | current_sched_info->queue_must_finish_empty = current_nr_blocks == 1; | |
7a31a7bd | 3050 | |
6a1cdb4d | 3051 | curr_bb = first_bb; |
3072d30e | 3052 | if (dbg_cnt (sched_block)) |
3053 | { | |
0a15667c | 3054 | edge f; |
bbd0cfb1 | 3055 | int saved_last_basic_block = last_basic_block; |
0a15667c | 3056 | |
bbd0cfb1 | 3057 | schedule_block (&curr_bb, bb_state[first_bb->index]); |
3058 | gcc_assert (EBB_FIRST_BB (bb) == first_bb); | |
3059 | sched_rgn_n_insns += sched_n_insns; | |
3060 | realloc_bb_state_array (saved_last_basic_block); | |
0a15667c | 3061 | f = find_fallthru_edge (last_bb->succs); |
3062 | if (f && f->probability * 100 / REG_BR_PROB_BASE >= | |
3063 | PARAM_VALUE (PARAM_SCHED_STATE_EDGE_PROB_CUTOFF)) | |
3064 | { | |
3065 | memcpy (bb_state[f->dest->index], curr_state, | |
3066 | dfa_state_size); | |
3067 | if (sched_verbose >= 5) | |
3068 | fprintf (sched_dump, "saving state for edge %d->%d\n", | |
3069 | f->src->index, f->dest->index); | |
3070 | } | |
3072d30e | 3071 | } |
3072 | else | |
3073 | { | |
3074 | sched_rgn_n_insns += rgn_n_insns; | |
3075 | } | |
7a31a7bd | 3076 | |
7a31a7bd | 3077 | /* Clean up. */ |
3078 | if (current_nr_blocks > 1) | |
e1ab7874 | 3079 | free_trg_info (); |
7a31a7bd | 3080 | } |
3081 | ||
3082 | /* Sanity check: verify that all region insns were scheduled. */ | |
04e579b6 | 3083 | gcc_assert (sched_rgn_n_insns == rgn_n_insns); |
7a31a7bd | 3084 | |
e1ab7874 | 3085 | sched_finish_ready_list (); |
7a31a7bd | 3086 | |
e1ab7874 | 3087 | /* Done with this region. */ |
3088 | sched_rgn_local_finish (); | |
93f6b030 | 3089 | |
3090 | /* Free dependencies. */ | |
3091 | for (bb = 0; bb < current_nr_blocks; ++bb) | |
3092 | free_block_dependencies (bb); | |
3093 | ||
3094 | gcc_assert (haifa_recovery_bb_ever_added_p | |
3095 | || deps_pools_are_empty_p ()); | |
7a31a7bd | 3096 | } |
3097 | ||
7a31a7bd | 3098 | /* Initialize data structures for region scheduling. */ |
3099 | ||
e1ab7874 | 3100 | void |
3101 | sched_rgn_init (bool single_blocks_p) | |
7a31a7bd | 3102 | { |
e1ab7874 | 3103 | min_spec_prob = ((PARAM_VALUE (PARAM_MIN_SPEC_PROB) * REG_BR_PROB_BASE) |
3104 | / 100); | |
3105 | ||
3106 | nr_inter = 0; | |
3107 | nr_spec = 0; | |
3108 | ||
6a1cdb4d | 3109 | extend_regions (); |
7a31a7bd | 3110 | |
e1ab7874 | 3111 | CONTAINING_RGN (ENTRY_BLOCK) = -1; |
3112 | CONTAINING_RGN (EXIT_BLOCK) = -1; | |
3113 | ||
bbd0cfb1 | 3114 | realloc_bb_state_array (0); |
0a15667c | 3115 | |
7a31a7bd | 3116 | /* Compute regions for scheduling. */ |
e1ab7874 | 3117 | if (single_blocks_p |
4d2e5d52 | 3118 | || n_basic_blocks == NUM_FIXED_BLOCKS + 1 |
aae97b21 | 3119 | || !flag_schedule_interblock |
3120 | || is_cfg_nonregular ()) | |
7a31a7bd | 3121 | { |
e1ab7874 | 3122 | find_single_block_region (sel_sched_p ()); |
7a31a7bd | 3123 | } |
3124 | else | |
3125 | { | |
aae97b21 | 3126 | /* Compute the dominators and post dominators. */ |
e1ab7874 | 3127 | if (!sel_sched_p ()) |
3128 | calculate_dominance_info (CDI_DOMINATORS); | |
7a31a7bd | 3129 | |
aae97b21 | 3130 | /* Find regions. */ |
3131 | find_rgns (); | |
7a31a7bd | 3132 | |
aae97b21 | 3133 | if (sched_verbose >= 3) |
3134 | debug_regions (); | |
7a31a7bd | 3135 | |
aae97b21 | 3136 | /* For now. This will move as more and more of haifa is converted |
3072d30e | 3137 | to using the cfg code. */ |
e1ab7874 | 3138 | if (!sel_sched_p ()) |
3139 | free_dominance_info (CDI_DOMINATORS); | |
7a31a7bd | 3140 | } |
7a31a7bd | 3141 | |
e1ab7874 | 3142 | gcc_assert (0 < nr_regions && nr_regions <= n_basic_blocks); |
7a31a7bd | 3143 | |
e1ab7874 | 3144 | RGN_BLOCKS (nr_regions) = (RGN_BLOCKS (nr_regions - 1) + |
3145 | RGN_NR_BLOCKS (nr_regions - 1)); | |
3146 | } | |
3147 | ||
3148 | /* Free data structures for region scheduling. */ | |
7a31a7bd | 3149 | void |
e1ab7874 | 3150 | sched_rgn_finish (void) |
7a31a7bd | 3151 | { |
bbd0cfb1 | 3152 | free_bb_state_array (); |
0a15667c | 3153 | |
7a31a7bd | 3154 | /* Reposition the prologue and epilogue notes in case we moved the |
3155 | prologue/epilogue insns. */ | |
3156 | if (reload_completed) | |
3072d30e | 3157 | reposition_prologue_and_epilogue_notes (); |
7a31a7bd | 3158 | |
7a31a7bd | 3159 | if (sched_verbose) |
3160 | { | |
e1ab7874 | 3161 | if (reload_completed == 0 |
3162 | && flag_schedule_interblock) | |
7a31a7bd | 3163 | { |
3164 | fprintf (sched_dump, | |
3165 | "\n;; Procedure interblock/speculative motions == %d/%d \n", | |
3166 | nr_inter, nr_spec); | |
3167 | } | |
3168 | else | |
04e579b6 | 3169 | gcc_assert (nr_inter <= 0); |
7a31a7bd | 3170 | fprintf (sched_dump, "\n\n"); |
3171 | } | |
3172 | ||
e1ab7874 | 3173 | nr_regions = 0; |
3174 | ||
7a31a7bd | 3175 | free (rgn_table); |
e1ab7874 | 3176 | rgn_table = NULL; |
3177 | ||
7a31a7bd | 3178 | free (rgn_bb_table); |
e1ab7874 | 3179 | rgn_bb_table = NULL; |
3180 | ||
7a31a7bd | 3181 | free (block_to_bb); |
e1ab7874 | 3182 | block_to_bb = NULL; |
3183 | ||
7a31a7bd | 3184 | free (containing_rgn); |
e1ab7874 | 3185 | containing_rgn = NULL; |
3186 | ||
3187 | free (ebb_head); | |
3188 | ebb_head = NULL; | |
3189 | } | |
3190 | ||
3191 | /* Setup global variables like CURRENT_BLOCKS and CURRENT_NR_BLOCK to | |
3192 | point to the region RGN. */ | |
3193 | void | |
3194 | rgn_setup_region (int rgn) | |
3195 | { | |
3196 | int bb; | |
3197 | ||
3198 | /* Set variables for the current region. */ | |
3199 | current_nr_blocks = RGN_NR_BLOCKS (rgn); | |
3200 | current_blocks = RGN_BLOCKS (rgn); | |
48e1416a | 3201 | |
e1ab7874 | 3202 | /* EBB_HEAD is a region-scope structure. But we realloc it for |
3203 | each region to save time/memory/something else. | |
3204 | See comments in add_block1, for what reasons we allocate +1 element. */ | |
3205 | ebb_head = XRESIZEVEC (int, ebb_head, current_nr_blocks + 1); | |
3206 | for (bb = 0; bb <= current_nr_blocks; bb++) | |
3207 | ebb_head[bb] = current_blocks + bb; | |
3208 | } | |
3209 | ||
3210 | /* Compute instruction dependencies in region RGN. */ | |
3211 | void | |
3212 | sched_rgn_compute_dependencies (int rgn) | |
3213 | { | |
3214 | if (!RGN_DONT_CALC_DEPS (rgn)) | |
3215 | { | |
3216 | int bb; | |
3217 | ||
3218 | if (sel_sched_p ()) | |
3219 | sched_emulate_haifa_p = 1; | |
3220 | ||
3221 | init_deps_global (); | |
3222 | ||
3223 | /* Initializations for region data dependence analysis. */ | |
68e419a1 | 3224 | bb_deps = XNEWVEC (struct deps_desc, current_nr_blocks); |
e1ab7874 | 3225 | for (bb = 0; bb < current_nr_blocks; bb++) |
d9ab2038 | 3226 | init_deps (bb_deps + bb, false); |
e1ab7874 | 3227 | |
566e7db2 | 3228 | /* Initialize bitmap used in add_branch_dependences. */ |
3229 | insn_referenced = sbitmap_alloc (sched_max_luid); | |
53c5d9d4 | 3230 | bitmap_clear (insn_referenced); |
48e1416a | 3231 | |
e1ab7874 | 3232 | /* Compute backward dependencies. */ |
3233 | for (bb = 0; bb < current_nr_blocks; bb++) | |
3234 | compute_block_dependences (bb); | |
48e1416a | 3235 | |
566e7db2 | 3236 | sbitmap_free (insn_referenced); |
e1ab7874 | 3237 | free_pending_lists (); |
3238 | finish_deps_global (); | |
3239 | free (bb_deps); | |
7a31a7bd | 3240 | |
e1ab7874 | 3241 | /* We don't want to recalculate this twice. */ |
3242 | RGN_DONT_CALC_DEPS (rgn) = 1; | |
3072d30e | 3243 | |
e1ab7874 | 3244 | if (sel_sched_p ()) |
3245 | sched_emulate_haifa_p = 0; | |
3246 | } | |
3247 | else | |
3248 | /* (This is a recovery block. It is always a single block region.) | |
3249 | OR (We use selective scheduling.) */ | |
3250 | gcc_assert (current_nr_blocks == 1 || sel_sched_p ()); | |
3251 | } | |
3252 | ||
3253 | /* Init region data structures. Returns true if this region should | |
3254 | not be scheduled. */ | |
3255 | void | |
3256 | sched_rgn_local_init (int rgn) | |
3257 | { | |
3258 | int bb; | |
48e1416a | 3259 | |
e1ab7874 | 3260 | /* Compute interblock info: probabilities, split-edges, dominators, etc. */ |
3261 | if (current_nr_blocks > 1) | |
3262 | { | |
3263 | basic_block block; | |
3264 | edge e; | |
3265 | edge_iterator ei; | |
3266 | ||
3267 | prob = XNEWVEC (int, current_nr_blocks); | |
3268 | ||
3269 | dom = sbitmap_vector_alloc (current_nr_blocks, current_nr_blocks); | |
53c5d9d4 | 3270 | bitmap_vector_clear (dom, current_nr_blocks); |
e1ab7874 | 3271 | |
3272 | /* Use ->aux to implement EDGE_TO_BIT mapping. */ | |
3273 | rgn_nr_edges = 0; | |
3274 | FOR_EACH_BB (block) | |
3275 | { | |
3276 | if (CONTAINING_RGN (block->index) != rgn) | |
3277 | continue; | |
3278 | FOR_EACH_EDGE (e, ei, block->succs) | |
3279 | SET_EDGE_TO_BIT (e, rgn_nr_edges++); | |
3280 | } | |
3281 | ||
3282 | rgn_edges = XNEWVEC (edge, rgn_nr_edges); | |
3283 | rgn_nr_edges = 0; | |
3284 | FOR_EACH_BB (block) | |
3285 | { | |
3286 | if (CONTAINING_RGN (block->index) != rgn) | |
3287 | continue; | |
3288 | FOR_EACH_EDGE (e, ei, block->succs) | |
3289 | rgn_edges[rgn_nr_edges++] = e; | |
3290 | } | |
3291 | ||
3292 | /* Split edges. */ | |
3293 | pot_split = sbitmap_vector_alloc (current_nr_blocks, rgn_nr_edges); | |
53c5d9d4 | 3294 | bitmap_vector_clear (pot_split, current_nr_blocks); |
e1ab7874 | 3295 | ancestor_edges = sbitmap_vector_alloc (current_nr_blocks, rgn_nr_edges); |
53c5d9d4 | 3296 | bitmap_vector_clear (ancestor_edges, current_nr_blocks); |
e1ab7874 | 3297 | |
3298 | /* Compute probabilities, dominators, split_edges. */ | |
3299 | for (bb = 0; bb < current_nr_blocks; bb++) | |
3300 | compute_dom_prob_ps (bb); | |
3301 | ||
3302 | /* Cleanup ->aux used for EDGE_TO_BIT mapping. */ | |
3303 | /* We don't need them anymore. But we want to avoid duplication of | |
3304 | aux fields in the newly created edges. */ | |
3305 | FOR_EACH_BB (block) | |
3306 | { | |
3307 | if (CONTAINING_RGN (block->index) != rgn) | |
3308 | continue; | |
3309 | FOR_EACH_EDGE (e, ei, block->succs) | |
3310 | e->aux = NULL; | |
3311 | } | |
3312 | } | |
3313 | } | |
3314 | ||
3315 | /* Free data computed for the finished region. */ | |
48e1416a | 3316 | void |
e1ab7874 | 3317 | sched_rgn_local_free (void) |
3318 | { | |
3319 | free (prob); | |
3320 | sbitmap_vector_free (dom); | |
3321 | sbitmap_vector_free (pot_split); | |
3322 | sbitmap_vector_free (ancestor_edges); | |
3323 | free (rgn_edges); | |
3324 | } | |
3325 | ||
3326 | /* Free data computed for the finished region. */ | |
3327 | void | |
3328 | sched_rgn_local_finish (void) | |
3329 | { | |
3330 | if (current_nr_blocks > 1 && !sel_sched_p ()) | |
3331 | { | |
3332 | sched_rgn_local_free (); | |
3333 | } | |
3334 | } | |
3335 | ||
3336 | /* Setup scheduler infos. */ | |
3337 | void | |
3338 | rgn_setup_common_sched_info (void) | |
3339 | { | |
3340 | memcpy (&rgn_common_sched_info, &haifa_common_sched_info, | |
3341 | sizeof (rgn_common_sched_info)); | |
3342 | ||
3343 | rgn_common_sched_info.fix_recovery_cfg = rgn_fix_recovery_cfg; | |
3344 | rgn_common_sched_info.add_block = rgn_add_block; | |
3345 | rgn_common_sched_info.estimate_number_of_insns | |
3346 | = rgn_estimate_number_of_insns; | |
3347 | rgn_common_sched_info.sched_pass_id = SCHED_RGN_PASS; | |
3348 | ||
3349 | common_sched_info = &rgn_common_sched_info; | |
3350 | } | |
3351 | ||
3352 | /* Setup all *_sched_info structures (for the Haifa frontend | |
3353 | and for the dependence analysis) in the interblock scheduler. */ | |
3354 | void | |
3355 | rgn_setup_sched_infos (void) | |
3356 | { | |
3357 | if (!sel_sched_p ()) | |
3358 | memcpy (&rgn_sched_deps_info, &rgn_const_sched_deps_info, | |
3359 | sizeof (rgn_sched_deps_info)); | |
3360 | else | |
3361 | memcpy (&rgn_sched_deps_info, &rgn_const_sel_sched_deps_info, | |
3362 | sizeof (rgn_sched_deps_info)); | |
3363 | ||
3364 | sched_deps_info = &rgn_sched_deps_info; | |
3365 | ||
3366 | memcpy (&rgn_sched_info, &rgn_const_sched_info, sizeof (rgn_sched_info)); | |
3367 | current_sched_info = &rgn_sched_info; | |
3368 | } | |
3369 | ||
3370 | /* The one entry point in this file. */ | |
3371 | void | |
3372 | schedule_insns (void) | |
3373 | { | |
3374 | int rgn; | |
3375 | ||
3376 | /* Taking care of this degenerate case makes the rest of | |
3377 | this code simpler. */ | |
3378 | if (n_basic_blocks == NUM_FIXED_BLOCKS) | |
3379 | return; | |
3380 | ||
3381 | rgn_setup_common_sched_info (); | |
3382 | rgn_setup_sched_infos (); | |
3383 | ||
3384 | haifa_sched_init (); | |
3385 | sched_rgn_init (reload_completed); | |
3386 | ||
3387 | bitmap_initialize (¬_in_df, 0); | |
3072d30e | 3388 | bitmap_clear (¬_in_df); |
7a31a7bd | 3389 | |
e1ab7874 | 3390 | /* Schedule every region in the subroutine. */ |
3391 | for (rgn = 0; rgn < nr_regions; rgn++) | |
3392 | if (dbg_cnt (sched_region)) | |
3393 | schedule_region (rgn); | |
3394 | ||
3395 | /* Clean up. */ | |
3396 | sched_rgn_finish (); | |
3397 | bitmap_clear (¬_in_df); | |
3398 | ||
3399 | haifa_sched_finish (); | |
7a31a7bd | 3400 | } |
6a1cdb4d | 3401 | |
3402 | /* INSN has been added to/removed from current region. */ | |
3403 | static void | |
e1ab7874 | 3404 | rgn_add_remove_insn (rtx insn, int remove_p) |
6a1cdb4d | 3405 | { |
3406 | if (!remove_p) | |
3407 | rgn_n_insns++; | |
3408 | else | |
3409 | rgn_n_insns--; | |
3410 | ||
3411 | if (INSN_BB (insn) == target_bb) | |
3412 | { | |
3413 | if (!remove_p) | |
3414 | target_n_insns++; | |
3415 | else | |
3416 | target_n_insns--; | |
3417 | } | |
3418 | } | |
3419 | ||
3420 | /* Extend internal data structures. */ | |
e1ab7874 | 3421 | void |
6a1cdb4d | 3422 | extend_regions (void) |
3423 | { | |
3424 | rgn_table = XRESIZEVEC (region, rgn_table, n_basic_blocks); | |
3425 | rgn_bb_table = XRESIZEVEC (int, rgn_bb_table, n_basic_blocks); | |
3426 | block_to_bb = XRESIZEVEC (int, block_to_bb, last_basic_block); | |
3427 | containing_rgn = XRESIZEVEC (int, containing_rgn, last_basic_block); | |
3428 | } | |
3429 | ||
e1ab7874 | 3430 | void |
3431 | rgn_make_new_region_out_of_new_block (basic_block bb) | |
3432 | { | |
3433 | int i; | |
3434 | ||
3435 | i = RGN_BLOCKS (nr_regions); | |
3436 | /* I - first free position in rgn_bb_table. */ | |
3437 | ||
3438 | rgn_bb_table[i] = bb->index; | |
3439 | RGN_NR_BLOCKS (nr_regions) = 1; | |
3440 | RGN_HAS_REAL_EBB (nr_regions) = 0; | |
3441 | RGN_DONT_CALC_DEPS (nr_regions) = 0; | |
3442 | CONTAINING_RGN (bb->index) = nr_regions; | |
3443 | BLOCK_TO_BB (bb->index) = 0; | |
3444 | ||
3445 | nr_regions++; | |
48e1416a | 3446 | |
e1ab7874 | 3447 | RGN_BLOCKS (nr_regions) = i + 1; |
3448 | } | |
3449 | ||
6a1cdb4d | 3450 | /* BB was added to ebb after AFTER. */ |
3451 | static void | |
e1ab7874 | 3452 | rgn_add_block (basic_block bb, basic_block after) |
6a1cdb4d | 3453 | { |
3454 | extend_regions (); | |
3072d30e | 3455 | bitmap_set_bit (¬_in_df, bb->index); |
3456 | ||
6a1cdb4d | 3457 | if (after == 0 || after == EXIT_BLOCK_PTR) |
3458 | { | |
e1ab7874 | 3459 | rgn_make_new_region_out_of_new_block (bb); |
3460 | RGN_DONT_CALC_DEPS (nr_regions - 1) = (after == EXIT_BLOCK_PTR); | |
6a1cdb4d | 3461 | } |
3462 | else | |
48e1416a | 3463 | { |
6a1cdb4d | 3464 | int i, pos; |
3465 | ||
3466 | /* We need to fix rgn_table, block_to_bb, containing_rgn | |
3467 | and ebb_head. */ | |
3468 | ||
3469 | BLOCK_TO_BB (bb->index) = BLOCK_TO_BB (after->index); | |
3470 | ||
3471 | /* We extend ebb_head to one more position to | |
48e1416a | 3472 | easily find the last position of the last ebb in |
6a1cdb4d | 3473 | the current region. Thus, ebb_head[BLOCK_TO_BB (after) + 1] |
3474 | is _always_ valid for access. */ | |
3475 | ||
3476 | i = BLOCK_TO_BB (after->index) + 1; | |
90bf0a00 | 3477 | pos = ebb_head[i] - 1; |
3478 | /* Now POS is the index of the last block in the region. */ | |
3479 | ||
3480 | /* Find index of basic block AFTER. */ | |
3c802a1e | 3481 | for (; rgn_bb_table[pos] != after->index; pos--) |
3482 | ; | |
90bf0a00 | 3483 | |
6a1cdb4d | 3484 | pos++; |
3485 | gcc_assert (pos > ebb_head[i - 1]); | |
90bf0a00 | 3486 | |
6a1cdb4d | 3487 | /* i - ebb right after "AFTER". */ |
3488 | /* ebb_head[i] - VALID. */ | |
3489 | ||
3490 | /* Source position: ebb_head[i] | |
9ca2c29a | 3491 | Destination position: ebb_head[i] + 1 |
48e1416a | 3492 | Last position: |
6a1cdb4d | 3493 | RGN_BLOCKS (nr_regions) - 1 |
3494 | Number of elements to copy: (last_position) - (source_position) + 1 | |
3495 | */ | |
48e1416a | 3496 | |
6a1cdb4d | 3497 | memmove (rgn_bb_table + pos + 1, |
3498 | rgn_bb_table + pos, | |
3499 | ((RGN_BLOCKS (nr_regions) - 1) - (pos) + 1) | |
3500 | * sizeof (*rgn_bb_table)); | |
3501 | ||
3502 | rgn_bb_table[pos] = bb->index; | |
48e1416a | 3503 | |
6a1cdb4d | 3504 | for (; i <= current_nr_blocks; i++) |
3505 | ebb_head [i]++; | |
3506 | ||
3507 | i = CONTAINING_RGN (after->index); | |
3508 | CONTAINING_RGN (bb->index) = i; | |
48e1416a | 3509 | |
6a1cdb4d | 3510 | RGN_HAS_REAL_EBB (i) = 1; |
3511 | ||
3512 | for (++i; i <= nr_regions; i++) | |
3513 | RGN_BLOCKS (i)++; | |
6a1cdb4d | 3514 | } |
3515 | } | |
3516 | ||
3517 | /* Fix internal data after interblock movement of jump instruction. | |
3518 | For parameter meaning please refer to | |
3519 | sched-int.h: struct sched_info: fix_recovery_cfg. */ | |
3520 | static void | |
e1ab7874 | 3521 | rgn_fix_recovery_cfg (int bbi, int check_bbi, int check_bb_nexti) |
6a1cdb4d | 3522 | { |
3523 | int old_pos, new_pos, i; | |
3524 | ||
3525 | BLOCK_TO_BB (check_bb_nexti) = BLOCK_TO_BB (bbi); | |
48e1416a | 3526 | |
6a1cdb4d | 3527 | for (old_pos = ebb_head[BLOCK_TO_BB (check_bbi) + 1] - 1; |
3528 | rgn_bb_table[old_pos] != check_bb_nexti; | |
3c802a1e | 3529 | old_pos--) |
3530 | ; | |
6a1cdb4d | 3531 | gcc_assert (old_pos > ebb_head[BLOCK_TO_BB (check_bbi)]); |
3532 | ||
3533 | for (new_pos = ebb_head[BLOCK_TO_BB (bbi) + 1] - 1; | |
3534 | rgn_bb_table[new_pos] != bbi; | |
3c802a1e | 3535 | new_pos--) |
3536 | ; | |
6a1cdb4d | 3537 | new_pos++; |
3538 | gcc_assert (new_pos > ebb_head[BLOCK_TO_BB (bbi)]); | |
48e1416a | 3539 | |
6a1cdb4d | 3540 | gcc_assert (new_pos < old_pos); |
3541 | ||
3542 | memmove (rgn_bb_table + new_pos + 1, | |
3543 | rgn_bb_table + new_pos, | |
3544 | (old_pos - new_pos) * sizeof (*rgn_bb_table)); | |
3545 | ||
3546 | rgn_bb_table[new_pos] = check_bb_nexti; | |
3547 | ||
3548 | for (i = BLOCK_TO_BB (bbi) + 1; i <= BLOCK_TO_BB (check_bbi); i++) | |
3549 | ebb_head[i]++; | |
3550 | } | |
3551 | ||
3552 | /* Return next block in ebb chain. For parameter meaning please refer to | |
3553 | sched-int.h: struct sched_info: advance_target_bb. */ | |
3554 | static basic_block | |
3555 | advance_target_bb (basic_block bb, rtx insn) | |
3556 | { | |
3557 | if (insn) | |
3558 | return 0; | |
3559 | ||
3560 | gcc_assert (BLOCK_TO_BB (bb->index) == target_bb | |
3561 | && BLOCK_TO_BB (bb->next_bb->index) == target_bb); | |
3562 | return bb->next_bb; | |
3563 | } | |
3564 | ||
cda0a5f5 | 3565 | #endif |
77fce4cd | 3566 | \f |
57a8bf1b | 3567 | static bool |
3568 | gate_handle_live_range_shrinkage (void) | |
3569 | { | |
3570 | #ifdef INSN_SCHEDULING | |
3571 | return flag_live_range_shrinkage; | |
3572 | #else | |
3573 | return 0; | |
3574 | #endif | |
3575 | } | |
3576 | ||
3577 | /* Run instruction scheduler. */ | |
3578 | static unsigned int | |
3579 | rest_of_handle_live_range_shrinkage (void) | |
3580 | { | |
3581 | #ifdef INSN_SCHEDULING | |
3582 | int saved; | |
3583 | ||
3584 | initialize_live_range_shrinkage (); | |
3585 | saved = flag_schedule_interblock; | |
3586 | flag_schedule_interblock = false; | |
3587 | schedule_insns (); | |
3588 | flag_schedule_interblock = saved; | |
3589 | finish_live_range_shrinkage (); | |
3590 | #endif | |
3591 | return 0; | |
3592 | } | |
3593 | ||
77fce4cd | 3594 | static bool |
3595 | gate_handle_sched (void) | |
3596 | { | |
3597 | #ifdef INSN_SCHEDULING | |
1c3660b1 | 3598 | return optimize > 0 && flag_schedule_insns && dbg_cnt (sched_func); |
77fce4cd | 3599 | #else |
3600 | return 0; | |
3601 | #endif | |
3602 | } | |
3603 | ||
3604 | /* Run instruction scheduler. */ | |
2a1990e9 | 3605 | static unsigned int |
77fce4cd | 3606 | rest_of_handle_sched (void) |
3607 | { | |
3608 | #ifdef INSN_SCHEDULING | |
e1ab7874 | 3609 | if (flag_selective_scheduling |
3610 | && ! maybe_skip_selective_scheduling ()) | |
3611 | run_selective_scheduling (); | |
3612 | else | |
3613 | schedule_insns (); | |
77fce4cd | 3614 | #endif |
2a1990e9 | 3615 | return 0; |
77fce4cd | 3616 | } |
3617 | ||
3618 | static bool | |
3619 | gate_handle_sched2 (void) | |
3620 | { | |
3621 | #ifdef INSN_SCHEDULING | |
48e1416a | 3622 | return optimize > 0 && flag_schedule_insns_after_reload |
8a42230a | 3623 | && !targetm.delay_sched2 && dbg_cnt (sched2_func); |
77fce4cd | 3624 | #else |
3625 | return 0; | |
3626 | #endif | |
3627 | } | |
3628 | ||
3629 | /* Run second scheduling pass after reload. */ | |
2a1990e9 | 3630 | static unsigned int |
77fce4cd | 3631 | rest_of_handle_sched2 (void) |
3632 | { | |
3633 | #ifdef INSN_SCHEDULING | |
e1ab7874 | 3634 | if (flag_selective_scheduling2 |
3635 | && ! maybe_skip_selective_scheduling ()) | |
3636 | run_selective_scheduling (); | |
77fce4cd | 3637 | else |
e1ab7874 | 3638 | { |
3639 | /* Do control and data sched analysis again, | |
3640 | and write some more of the results to dump file. */ | |
fda153ea | 3641 | if (flag_sched2_use_superblocks) |
e1ab7874 | 3642 | schedule_ebbs (); |
3643 | else | |
3644 | schedule_insns (); | |
3645 | } | |
77fce4cd | 3646 | #endif |
2a1990e9 | 3647 | return 0; |
77fce4cd | 3648 | } |
3649 | ||
cbe8bda8 | 3650 | namespace { |
3651 | ||
57a8bf1b | 3652 | const pass_data pass_data_live_range_shrinkage = |
3653 | { | |
3654 | RTL_PASS, /* type */ | |
3655 | "lr_shrinkage", /* name */ | |
3656 | OPTGROUP_NONE, /* optinfo_flags */ | |
3657 | true, /* has_gate */ | |
3658 | true, /* has_execute */ | |
3659 | TV_LIVE_RANGE_SHRINKAGE, /* tv_id */ | |
3660 | 0, /* properties_required */ | |
3661 | 0, /* properties_provided */ | |
3662 | 0, /* properties_destroyed */ | |
3663 | 0, /* todo_flags_start */ | |
3664 | ( TODO_df_finish | TODO_verify_rtl_sharing | |
3665 | | TODO_verify_flow ), /* todo_flags_finish */ | |
3666 | }; | |
3667 | ||
3668 | class pass_live_range_shrinkage : public rtl_opt_pass | |
3669 | { | |
3670 | public: | |
3671 | pass_live_range_shrinkage(gcc::context *ctxt) | |
3672 | : rtl_opt_pass(pass_data_live_range_shrinkage, ctxt) | |
3673 | {} | |
3674 | ||
3675 | /* opt_pass methods: */ | |
3676 | bool gate () { return gate_handle_live_range_shrinkage (); } | |
3677 | unsigned int execute () { return rest_of_handle_live_range_shrinkage (); } | |
3678 | ||
3679 | }; // class pass_live_range_shrinkage | |
3680 | ||
3681 | } // anon namespace | |
3682 | ||
3683 | rtl_opt_pass * | |
3684 | make_pass_live_range_shrinkage (gcc::context *ctxt) | |
3685 | { | |
3686 | return new pass_live_range_shrinkage (ctxt); | |
3687 | } | |
3688 | ||
3689 | namespace { | |
3690 | ||
cbe8bda8 | 3691 | const pass_data pass_data_sched = |
3692 | { | |
3693 | RTL_PASS, /* type */ | |
3694 | "sched1", /* name */ | |
3695 | OPTGROUP_NONE, /* optinfo_flags */ | |
3696 | true, /* has_gate */ | |
3697 | true, /* has_execute */ | |
3698 | TV_SCHED, /* tv_id */ | |
3699 | 0, /* properties_required */ | |
3700 | 0, /* properties_provided */ | |
3701 | 0, /* properties_destroyed */ | |
3702 | 0, /* todo_flags_start */ | |
3703 | ( TODO_df_finish | TODO_verify_rtl_sharing | |
3704 | | TODO_verify_flow ), /* todo_flags_finish */ | |
77fce4cd | 3705 | }; |
3706 | ||
cbe8bda8 | 3707 | class pass_sched : public rtl_opt_pass |
3708 | { | |
3709 | public: | |
9af5ce0c | 3710 | pass_sched (gcc::context *ctxt) |
3711 | : rtl_opt_pass (pass_data_sched, ctxt) | |
cbe8bda8 | 3712 | {} |
3713 | ||
3714 | /* opt_pass methods: */ | |
3715 | bool gate () { return gate_handle_sched (); } | |
3716 | unsigned int execute () { return rest_of_handle_sched (); } | |
3717 | ||
3718 | }; // class pass_sched | |
3719 | ||
3720 | } // anon namespace | |
3721 | ||
3722 | rtl_opt_pass * | |
3723 | make_pass_sched (gcc::context *ctxt) | |
3724 | { | |
3725 | return new pass_sched (ctxt); | |
3726 | } | |
3727 | ||
3728 | namespace { | |
3729 | ||
3730 | const pass_data pass_data_sched2 = | |
3731 | { | |
3732 | RTL_PASS, /* type */ | |
3733 | "sched2", /* name */ | |
3734 | OPTGROUP_NONE, /* optinfo_flags */ | |
3735 | true, /* has_gate */ | |
3736 | true, /* has_execute */ | |
3737 | TV_SCHED2, /* tv_id */ | |
3738 | 0, /* properties_required */ | |
3739 | 0, /* properties_provided */ | |
3740 | 0, /* properties_destroyed */ | |
3741 | 0, /* todo_flags_start */ | |
3742 | ( TODO_df_finish | TODO_verify_rtl_sharing | |
3743 | | TODO_verify_flow ), /* todo_flags_finish */ | |
77fce4cd | 3744 | }; |
cbe8bda8 | 3745 | |
3746 | class pass_sched2 : public rtl_opt_pass | |
3747 | { | |
3748 | public: | |
9af5ce0c | 3749 | pass_sched2 (gcc::context *ctxt) |
3750 | : rtl_opt_pass (pass_data_sched2, ctxt) | |
cbe8bda8 | 3751 | {} |
3752 | ||
3753 | /* opt_pass methods: */ | |
3754 | bool gate () { return gate_handle_sched2 (); } | |
3755 | unsigned int execute () { return rest_of_handle_sched2 (); } | |
3756 | ||
3757 | }; // class pass_sched2 | |
3758 | ||
3759 | } // anon namespace | |
3760 | ||
3761 | rtl_opt_pass * | |
3762 | make_pass_sched2 (gcc::context *ctxt) | |
3763 | { | |
3764 | return new pass_sched2 (ctxt); | |
3765 | } |