]>
Commit | Line | Data |
---|---|---|
b4ead7d4 | 1 | /* Instruction scheduling pass. |
6fb5fa3c DB |
2 | Copyright (C) 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999, 2000, |
3 | 2001, 2002, 2003, 2004, 2005, 2006, 2007 | |
4 | Free Software Foundation, Inc. | |
b4ead7d4 BS |
5 | Contributed by Michael Tiemann (tiemann@cygnus.com) Enhanced by, |
6 | and currently maintained by, Jim Wilson (wilson@cygnus.com) | |
7 | ||
1322177d | 8 | This file is part of GCC. |
b4ead7d4 | 9 | |
1322177d LB |
10 | GCC is free software; you can redistribute it and/or modify it under |
11 | the terms of the GNU General Public License as published by the Free | |
9dcd6f09 | 12 | Software Foundation; either version 3, or (at your option) any later |
1322177d | 13 | version. |
b4ead7d4 | 14 | |
1322177d LB |
15 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY |
16 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
b4ead7d4 BS |
17 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
18 | for more details. | |
19 | ||
20 | You should have received a copy of the GNU General Public License | |
9dcd6f09 NC |
21 | along with GCC; see the file COPYING3. If not see |
22 | <http://www.gnu.org/licenses/>. */ | |
b4ead7d4 BS |
23 | |
24 | /* This pass implements list scheduling within basic blocks. It is | |
25 | run twice: (1) after flow analysis, but before register allocation, | |
26 | and (2) after register allocation. | |
27 | ||
28 | The first run performs interblock scheduling, moving insns between | |
29 | different blocks in the same "region", and the second runs only | |
30 | basic block scheduling. | |
31 | ||
32 | Interblock motions performed are useful motions and speculative | |
33 | motions, including speculative loads. Motions requiring code | |
34 | duplication are not supported. The identification of motion type | |
35 | and the check for validity of speculative motions requires | |
36 | construction and analysis of the function's control flow graph. | |
37 | ||
38 | The main entry point for this pass is schedule_insns(), called for | |
39 | each function. The work of the scheduler is organized in three | |
40 | levels: (1) function level: insns are subject to splitting, | |
41 | control-flow-graph is constructed, regions are computed (after | |
42 | reload, each region is of one block), (2) region level: control | |
43 | flow graph attributes required for interblock scheduling are | |
44 | computed (dominators, reachability, etc.), data dependences and | |
45 | priorities are computed, and (3) block level: insns in the block | |
46 | are actually scheduled. */ | |
47 | \f | |
48 | #include "config.h" | |
49 | #include "system.h" | |
4977bab6 ZW |
50 | #include "coretypes.h" |
51 | #include "tm.h" | |
b4ead7d4 BS |
52 | #include "toplev.h" |
53 | #include "rtl.h" | |
54 | #include "tm_p.h" | |
55 | #include "hard-reg-set.h" | |
b4ead7d4 BS |
56 | #include "regs.h" |
57 | #include "function.h" | |
58 | #include "flags.h" | |
59 | #include "insn-config.h" | |
60 | #include "insn-attr.h" | |
61 | #include "except.h" | |
62 | #include "toplev.h" | |
63 | #include "recog.h" | |
d73b1f07 | 64 | #include "cfglayout.h" |
f72c6b56 | 65 | #include "params.h" |
b4ead7d4 | 66 | #include "sched-int.h" |
fae15c93 | 67 | #include "target.h" |
ef330312 PB |
68 | #include "timevar.h" |
69 | #include "tree-pass.h" | |
6fb5fa3c | 70 | #include "dbgcnt.h" |
73991d6a | 71 | |
f56887a7 | 72 | #ifdef INSN_SCHEDULING |
b4ead7d4 BS |
73 | /* Some accessor macros for h_i_d members only used within this file. */ |
74 | #define INSN_REF_COUNT(INSN) (h_i_d[INSN_UID (INSN)].ref_count) | |
75 | #define FED_BY_SPEC_LOAD(insn) (h_i_d[INSN_UID (insn)].fed_by_spec_load) | |
76 | #define IS_LOAD_INSN(insn) (h_i_d[INSN_UID (insn)].is_load_insn) | |
77 | ||
b4ead7d4 BS |
78 | /* nr_inter/spec counts interblock/speculative motion for the function. */ |
79 | static int nr_inter, nr_spec; | |
80 | ||
46c5ad27 | 81 | static int is_cfg_nonregular (void); |
d72372e4 | 82 | static bool sched_is_disabled_for_current_region_p (void); |
b4ead7d4 BS |
83 | |
84 | /* A region is the main entity for interblock scheduling: insns | |
85 | are allowed to move between blocks in the same region, along | |
86 | control flow graph edges, in the 'up' direction. */ | |
87 | typedef struct | |
88 | { | |
496d7bb0 MK |
89 | /* Number of extended basic blocks in region. */ |
90 | int rgn_nr_blocks; | |
91 | /* cblocks in the region (actually index in rgn_bb_table). */ | |
92 | int rgn_blocks; | |
93 | /* Dependencies for this region are already computed. Basically, indicates, | |
94 | that this is a recovery block. */ | |
95 | unsigned int dont_calc_deps : 1; | |
96 | /* This region has at least one non-trivial ebb. */ | |
97 | unsigned int has_real_ebb : 1; | |
b4ead7d4 BS |
98 | } |
99 | region; | |
100 | ||
101 | /* Number of regions in the procedure. */ | |
102 | static int nr_regions; | |
103 | ||
104 | /* Table of region descriptions. */ | |
105 | static region *rgn_table; | |
106 | ||
107 | /* Array of lists of regions' blocks. */ | |
108 | static int *rgn_bb_table; | |
109 | ||
110 | /* Topological order of blocks in the region (if b2 is reachable from | |
111 | b1, block_to_bb[b2] > block_to_bb[b1]). Note: A basic block is | |
112 | always referred to by either block or b, while its topological | |
4d6922ee | 113 | order name (in the region) is referred to by bb. */ |
b4ead7d4 BS |
114 | static int *block_to_bb; |
115 | ||
116 | /* The number of the region containing a block. */ | |
117 | static int *containing_rgn; | |
118 | ||
36968131 PS |
119 | /* The minimum probability of reaching a source block so that it will be |
120 | considered for speculative scheduling. */ | |
121 | static int min_spec_prob; | |
122 | ||
b4ead7d4 BS |
123 | #define RGN_NR_BLOCKS(rgn) (rgn_table[rgn].rgn_nr_blocks) |
124 | #define RGN_BLOCKS(rgn) (rgn_table[rgn].rgn_blocks) | |
496d7bb0 MK |
125 | #define RGN_DONT_CALC_DEPS(rgn) (rgn_table[rgn].dont_calc_deps) |
126 | #define RGN_HAS_REAL_EBB(rgn) (rgn_table[rgn].has_real_ebb) | |
b4ead7d4 BS |
127 | #define BLOCK_TO_BB(block) (block_to_bb[block]) |
128 | #define CONTAINING_RGN(block) (containing_rgn[block]) | |
129 | ||
46c5ad27 AJ |
130 | void debug_regions (void); |
131 | static void find_single_block_region (void); | |
dcda8480 | 132 | static void find_rgns (void); |
d08eefb9 | 133 | static void extend_rgns (int *, int *, sbitmap, int *); |
f72c6b56 | 134 | static bool too_large (int, int *, int *); |
b4ead7d4 | 135 | |
46c5ad27 | 136 | extern void debug_live (int, int); |
b4ead7d4 BS |
137 | |
138 | /* Blocks of the current region being scheduled. */ | |
139 | static int current_nr_blocks; | |
140 | static int current_blocks; | |
141 | ||
496d7bb0 MK |
142 | static int rgn_n_insns; |
143 | ||
144 | /* The mapping from ebb to block. */ | |
145 | /* ebb_head [i] - is index in rgn_bb_table, while | |
146 | EBB_HEAD (i) - is basic block index. | |
147 | BASIC_BLOCK (EBB_HEAD (i)) - head of ebb. */ | |
148 | #define BB_TO_BLOCK(ebb) (rgn_bb_table[ebb_head[ebb]]) | |
149 | #define EBB_FIRST_BB(ebb) BASIC_BLOCK (BB_TO_BLOCK (ebb)) | |
150 | #define EBB_LAST_BB(ebb) BASIC_BLOCK (rgn_bb_table[ebb_head[ebb + 1] - 1]) | |
b4ead7d4 | 151 | |
b4ead7d4 BS |
152 | /* Target info declarations. |
153 | ||
154 | The block currently being scheduled is referred to as the "target" block, | |
155 | while other blocks in the region from which insns can be moved to the | |
156 | target are called "source" blocks. The candidate structure holds info | |
157 | about such sources: are they valid? Speculative? Etc. */ | |
dcda8480 UW |
158 | typedef struct |
159 | { | |
160 | basic_block *first_member; | |
161 | int nr_members; | |
162 | } | |
163 | bblst; | |
164 | ||
b4ead7d4 BS |
165 | typedef struct |
166 | { | |
167 | char is_valid; | |
168 | char is_speculative; | |
169 | int src_prob; | |
170 | bblst split_bbs; | |
171 | bblst update_bbs; | |
172 | } | |
173 | candidate; | |
174 | ||
175 | static candidate *candidate_table; | |
176 | ||
177 | /* A speculative motion requires checking live information on the path | |
178 | from 'source' to 'target'. The split blocks are those to be checked. | |
179 | After a speculative motion, live information should be modified in | |
180 | the 'update' blocks. | |
181 | ||
182 | Lists of split and update blocks for each candidate of the current | |
183 | target are in array bblst_table. */ | |
dcda8480 UW |
184 | static basic_block *bblst_table; |
185 | static int bblst_size, bblst_last; | |
b4ead7d4 BS |
186 | |
187 | #define IS_VALID(src) ( candidate_table[src].is_valid ) | |
188 | #define IS_SPECULATIVE(src) ( candidate_table[src].is_speculative ) | |
189 | #define SRC_PROB(src) ( candidate_table[src].src_prob ) | |
190 | ||
191 | /* The bb being currently scheduled. */ | |
192 | static int target_bb; | |
193 | ||
194 | /* List of edges. */ | |
dcda8480 UW |
195 | typedef struct |
196 | { | |
197 | edge *first_member; | |
198 | int nr_members; | |
199 | } | |
200 | edgelst; | |
201 | ||
202 | static edge *edgelst_table; | |
203 | static int edgelst_last; | |
204 | ||
205 | static void extract_edgelst (sbitmap, edgelst *); | |
206 | ||
b4ead7d4 BS |
207 | |
208 | /* Target info functions. */ | |
46c5ad27 AJ |
209 | static void split_edges (int, int, edgelst *); |
210 | static void compute_trg_info (int); | |
211 | void debug_candidate (int); | |
212 | void debug_candidates (int); | |
b4ead7d4 | 213 | |
bdfa170f | 214 | /* Dominators array: dom[i] contains the sbitmap of dominators of |
b4ead7d4 | 215 | bb i in the region. */ |
bdfa170f | 216 | static sbitmap *dom; |
b4ead7d4 BS |
217 | |
218 | /* bb 0 is the only region entry. */ | |
219 | #define IS_RGN_ENTRY(bb) (!bb) | |
220 | ||
221 | /* Is bb_src dominated by bb_trg. */ | |
222 | #define IS_DOMINATED(bb_src, bb_trg) \ | |
bdfa170f | 223 | ( TEST_BIT (dom[bb_src], bb_trg) ) |
b4ead7d4 | 224 | |
36968131 PS |
225 | /* Probability: Prob[i] is an int in [0, REG_BR_PROB_BASE] which is |
226 | the probability of bb i relative to the region entry. */ | |
227 | static int *prob; | |
b4ead7d4 BS |
228 | |
229 | /* Bit-set of edges, where bit i stands for edge i. */ | |
bdfa170f | 230 | typedef sbitmap edgeset; |
b4ead7d4 BS |
231 | |
232 | /* Number of edges in the region. */ | |
233 | static int rgn_nr_edges; | |
234 | ||
235 | /* Array of size rgn_nr_edges. */ | |
dcda8480 | 236 | static edge *rgn_edges; |
b4ead7d4 BS |
237 | |
238 | /* Mapping from each edge in the graph to its number in the rgn. */ | |
dcda8480 UW |
239 | #define EDGE_TO_BIT(edge) ((int)(size_t)(edge)->aux) |
240 | #define SET_EDGE_TO_BIT(edge,nr) ((edge)->aux = (void *)(size_t)(nr)) | |
b4ead7d4 BS |
241 | |
242 | /* The split edges of a source bb is different for each target | |
243 | bb. In order to compute this efficiently, the 'potential-split edges' | |
244 | are computed for each bb prior to scheduling a region. This is actually | |
245 | the split edges of each bb relative to the region entry. | |
246 | ||
247 | pot_split[bb] is the set of potential split edges of bb. */ | |
248 | static edgeset *pot_split; | |
249 | ||
250 | /* For every bb, a set of its ancestor edges. */ | |
251 | static edgeset *ancestor_edges; | |
252 | ||
496d7bb0 MK |
253 | /* Array of EBBs sizes. Currently we can get a ebb only through |
254 | splitting of currently scheduling block, therefore, we don't need | |
255 | ebb_head array for every region, its sufficient to hold it only | |
256 | for current one. */ | |
257 | static int *ebb_head; | |
258 | ||
46c5ad27 | 259 | static void compute_dom_prob_ps (int); |
b4ead7d4 | 260 | |
b4ead7d4 BS |
261 | #define INSN_PROBABILITY(INSN) (SRC_PROB (BLOCK_TO_BB (BLOCK_NUM (INSN)))) |
262 | #define IS_SPECULATIVE_INSN(INSN) (IS_SPECULATIVE (BLOCK_TO_BB (BLOCK_NUM (INSN)))) | |
263 | #define INSN_BB(INSN) (BLOCK_TO_BB (BLOCK_NUM (INSN))) | |
264 | ||
b4ead7d4 | 265 | /* Speculative scheduling functions. */ |
46c5ad27 AJ |
266 | static int check_live_1 (int, rtx); |
267 | static void update_live_1 (int, rtx); | |
268 | static int check_live (rtx, int); | |
269 | static void update_live (rtx, int); | |
270 | static void set_spec_fed (rtx); | |
271 | static int is_pfree (rtx, int, int); | |
272 | static int find_conditional_protection (rtx, int); | |
273 | static int is_conditionally_protected (rtx, int, int); | |
274 | static int is_prisky (rtx, int, int); | |
275 | static int is_exception_free (rtx, int, int); | |
276 | ||
277 | static bool sets_likely_spilled (rtx); | |
7bc980e1 | 278 | static void sets_likely_spilled_1 (rtx, const_rtx, void *); |
46c5ad27 | 279 | static void add_branch_dependences (rtx, rtx); |
e2f6ff94 | 280 | static void compute_block_dependences (int); |
46c5ad27 AJ |
281 | |
282 | static void init_regions (void); | |
283 | static void schedule_region (int); | |
284 | static rtx concat_INSN_LIST (rtx, rtx); | |
285 | static void concat_insn_mem_list (rtx, rtx, rtx *, rtx *); | |
286 | static void propagate_deps (int, struct deps *); | |
287 | static void free_pending_lists (void); | |
b4ead7d4 BS |
288 | |
289 | /* Functions for construction of the control flow graph. */ | |
290 | ||
291 | /* Return 1 if control flow graph should not be constructed, 0 otherwise. | |
292 | ||
293 | We decide not to build the control flow graph if there is possibly more | |
dcda8480 UW |
294 | than one entry to the function, if computed branches exist, if we |
295 | have nonlocal gotos, or if we have an unreachable loop. */ | |
b4ead7d4 BS |
296 | |
297 | static int | |
46c5ad27 | 298 | is_cfg_nonregular (void) |
b4ead7d4 | 299 | { |
e0082a72 | 300 | basic_block b; |
b4ead7d4 | 301 | rtx insn; |
b4ead7d4 BS |
302 | |
303 | /* If we have a label that could be the target of a nonlocal goto, then | |
304 | the cfg is not well structured. */ | |
305 | if (nonlocal_goto_handler_labels) | |
306 | return 1; | |
307 | ||
308 | /* If we have any forced labels, then the cfg is not well structured. */ | |
309 | if (forced_labels) | |
310 | return 1; | |
311 | ||
b4ead7d4 | 312 | /* If we have exception handlers, then we consider the cfg not well |
6fb5fa3c DB |
313 | structured. ?!? We should be able to handle this now that we |
314 | compute an accurate cfg for EH. */ | |
6a58eee9 | 315 | if (current_function_has_exception_handlers ()) |
b4ead7d4 BS |
316 | return 1; |
317 | ||
cf7c4aa6 HPN |
318 | /* If we have insns which refer to labels as non-jumped-to operands, |
319 | then we consider the cfg not well structured. */ | |
e0082a72 | 320 | FOR_EACH_BB (b) |
f7aa1423 | 321 | FOR_BB_INSNS (b, insn) |
b4ead7d4 | 322 | { |
cb2f563b | 323 | rtx note, next, set, dest; |
cf7c4aa6 | 324 | |
f7aa1423 SB |
325 | /* If this function has a computed jump, then we consider the cfg |
326 | not well structured. */ | |
cf7c4aa6 | 327 | if (JUMP_P (insn) && computed_jump_p (insn)) |
f7aa1423 | 328 | return 1; |
cb2f563b HPN |
329 | |
330 | if (!INSN_P (insn)) | |
331 | continue; | |
332 | ||
333 | note = find_reg_note (insn, REG_LABEL_OPERAND, NULL_RTX); | |
334 | if (note == NULL_RTX) | |
335 | continue; | |
336 | ||
337 | /* For that label not to be seen as a referred-to label, this | |
338 | must be a single-set which is feeding a jump *only*. This | |
339 | could be a conditional jump with the label split off for | |
340 | machine-specific reasons or a casesi/tablejump. */ | |
341 | next = next_nonnote_insn (insn); | |
342 | if (next == NULL_RTX | |
343 | || !JUMP_P (next) | |
344 | || (JUMP_LABEL (next) != XEXP (note, 0) | |
345 | && find_reg_note (next, REG_LABEL_TARGET, | |
346 | XEXP (note, 0)) == NULL_RTX) | |
347 | || BLOCK_FOR_INSN (insn) != BLOCK_FOR_INSN (next)) | |
348 | return 1; | |
349 | ||
350 | set = single_set (insn); | |
351 | if (set == NULL_RTX) | |
352 | return 1; | |
353 | ||
354 | dest = SET_DEST (set); | |
355 | if (!REG_P (dest) || !dead_or_set_p (next, dest)) | |
356 | return 1; | |
b4ead7d4 BS |
357 | } |
358 | ||
b4ead7d4 BS |
359 | /* Unreachable loops with more than one basic block are detected |
360 | during the DFS traversal in find_rgns. | |
361 | ||
362 | Unreachable loops with a single block are detected here. This | |
363 | test is redundant with the one in find_rgns, but it's much | |
dcda8480 | 364 | cheaper to go ahead and catch the trivial case here. */ |
e0082a72 | 365 | FOR_EACH_BB (b) |
b4ead7d4 | 366 | { |
628f6a4e | 367 | if (EDGE_COUNT (b->preds) == 0 |
c5cbcccf ZD |
368 | || (single_pred_p (b) |
369 | && single_pred (b) == b)) | |
dcda8480 | 370 | return 1; |
b4ead7d4 BS |
371 | } |
372 | ||
dcda8480 UW |
373 | /* All the tests passed. Consider the cfg well structured. */ |
374 | return 0; | |
b4ead7d4 BS |
375 | } |
376 | ||
dcda8480 | 377 | /* Extract list of edges from a bitmap containing EDGE_TO_BIT bits. */ |
b4ead7d4 BS |
378 | |
379 | static void | |
dcda8480 | 380 | extract_edgelst (sbitmap set, edgelst *el) |
b4ead7d4 | 381 | { |
dfea6c85 | 382 | unsigned int i = 0; |
b6e7e9af | 383 | sbitmap_iterator sbi; |
b4ead7d4 | 384 | |
dcda8480 UW |
385 | /* edgelst table space is reused in each call to extract_edgelst. */ |
386 | edgelst_last = 0; | |
b4ead7d4 | 387 | |
dcda8480 UW |
388 | el->first_member = &edgelst_table[edgelst_last]; |
389 | el->nr_members = 0; | |
b4ead7d4 BS |
390 | |
391 | /* Iterate over each word in the bitset. */ | |
b6e7e9af KH |
392 | EXECUTE_IF_SET_IN_SBITMAP (set, 0, i, sbi) |
393 | { | |
394 | edgelst_table[edgelst_last++] = rgn_edges[i]; | |
395 | el->nr_members++; | |
396 | } | |
b4ead7d4 BS |
397 | } |
398 | ||
399 | /* Functions for the construction of regions. */ | |
400 | ||
401 | /* Print the regions, for debugging purposes. Callable from debugger. */ | |
402 | ||
403 | void | |
46c5ad27 | 404 | debug_regions (void) |
b4ead7d4 BS |
405 | { |
406 | int rgn, bb; | |
407 | ||
408 | fprintf (sched_dump, "\n;; ------------ REGIONS ----------\n\n"); | |
409 | for (rgn = 0; rgn < nr_regions; rgn++) | |
410 | { | |
411 | fprintf (sched_dump, ";;\trgn %d nr_blocks %d:\n", rgn, | |
412 | rgn_table[rgn].rgn_nr_blocks); | |
413 | fprintf (sched_dump, ";;\tbb/block: "); | |
414 | ||
496d7bb0 MK |
415 | /* We don't have ebb_head initialized yet, so we can't use |
416 | BB_TO_BLOCK (). */ | |
417 | current_blocks = RGN_BLOCKS (rgn); | |
b4ead7d4 | 418 | |
496d7bb0 MK |
419 | for (bb = 0; bb < rgn_table[rgn].rgn_nr_blocks; bb++) |
420 | fprintf (sched_dump, " %d/%d ", bb, rgn_bb_table[current_blocks + bb]); | |
b4ead7d4 BS |
421 | |
422 | fprintf (sched_dump, "\n\n"); | |
423 | } | |
424 | } | |
425 | ||
426 | /* Build a single block region for each basic block in the function. | |
427 | This allows for using the same code for interblock and basic block | |
428 | scheduling. */ | |
429 | ||
430 | static void | |
46c5ad27 | 431 | find_single_block_region (void) |
b4ead7d4 | 432 | { |
e0082a72 | 433 | basic_block bb; |
355e4ec4 | 434 | |
e0082a72 ZD |
435 | nr_regions = 0; |
436 | ||
437 | FOR_EACH_BB (bb) | |
b4ead7d4 | 438 | { |
e0082a72 ZD |
439 | rgn_bb_table[nr_regions] = bb->index; |
440 | RGN_NR_BLOCKS (nr_regions) = 1; | |
441 | RGN_BLOCKS (nr_regions) = nr_regions; | |
496d7bb0 MK |
442 | RGN_DONT_CALC_DEPS (nr_regions) = 0; |
443 | RGN_HAS_REAL_EBB (nr_regions) = 0; | |
e0082a72 ZD |
444 | CONTAINING_RGN (bb->index) = nr_regions; |
445 | BLOCK_TO_BB (bb->index) = 0; | |
446 | nr_regions++; | |
b4ead7d4 | 447 | } |
b4ead7d4 BS |
448 | } |
449 | ||
450 | /* Update number of blocks and the estimate for number of insns | |
f72c6b56 DE |
451 | in the region. Return true if the region is "too large" for interblock |
452 | scheduling (compile time considerations). */ | |
b4ead7d4 | 453 | |
f72c6b56 | 454 | static bool |
46c5ad27 | 455 | too_large (int block, int *num_bbs, int *num_insns) |
b4ead7d4 BS |
456 | { |
457 | (*num_bbs)++; | |
f72c6b56 DE |
458 | (*num_insns) += (INSN_LUID (BB_END (BASIC_BLOCK (block))) |
459 | - INSN_LUID (BB_HEAD (BASIC_BLOCK (block)))); | |
460 | ||
461 | return ((*num_bbs > PARAM_VALUE (PARAM_MAX_SCHED_REGION_BLOCKS)) | |
462 | || (*num_insns > PARAM_VALUE (PARAM_MAX_SCHED_REGION_INSNS))); | |
b4ead7d4 BS |
463 | } |
464 | ||
465 | /* Update_loop_relations(blk, hdr): Check if the loop headed by max_hdr[blk] | |
466 | is still an inner loop. Put in max_hdr[blk] the header of the most inner | |
467 | loop containing blk. */ | |
786de7eb KH |
468 | #define UPDATE_LOOP_RELATIONS(blk, hdr) \ |
469 | { \ | |
470 | if (max_hdr[blk] == -1) \ | |
471 | max_hdr[blk] = hdr; \ | |
472 | else if (dfs_nr[max_hdr[blk]] > dfs_nr[hdr]) \ | |
473 | RESET_BIT (inner, hdr); \ | |
474 | else if (dfs_nr[max_hdr[blk]] < dfs_nr[hdr]) \ | |
475 | { \ | |
476 | RESET_BIT (inner,max_hdr[blk]); \ | |
477 | max_hdr[blk] = hdr; \ | |
478 | } \ | |
b4ead7d4 BS |
479 | } |
480 | ||
481 | /* Find regions for interblock scheduling. | |
482 | ||
483 | A region for scheduling can be: | |
484 | ||
485 | * A loop-free procedure, or | |
486 | ||
487 | * A reducible inner loop, or | |
488 | ||
489 | * A basic block not contained in any other region. | |
490 | ||
491 | ?!? In theory we could build other regions based on extended basic | |
492 | blocks or reverse extended basic blocks. Is it worth the trouble? | |
493 | ||
494 | Loop blocks that form a region are put into the region's block list | |
495 | in topological order. | |
496 | ||
497 | This procedure stores its results into the following global (ick) variables | |
498 | ||
499 | * rgn_nr | |
500 | * rgn_table | |
501 | * rgn_bb_table | |
502 | * block_to_bb | |
503 | * containing region | |
504 | ||
505 | We use dominator relationships to avoid making regions out of non-reducible | |
506 | loops. | |
507 | ||
508 | This procedure needs to be converted to work on pred/succ lists instead | |
509 | of edge tables. That would simplify it somewhat. */ | |
510 | ||
511 | static void | |
dcda8480 | 512 | find_rgns (void) |
b4ead7d4 | 513 | { |
dcda8480 | 514 | int *max_hdr, *dfs_nr, *degree; |
b4ead7d4 BS |
515 | char no_loops = 1; |
516 | int node, child, loop_head, i, head, tail; | |
8a6b9b7f | 517 | int count = 0, sp, idx = 0; |
dcda8480 UW |
518 | edge_iterator current_edge; |
519 | edge_iterator *stack; | |
b4ead7d4 BS |
520 | int num_bbs, num_insns, unreachable; |
521 | int too_large_failure; | |
e0082a72 | 522 | basic_block bb; |
b4ead7d4 | 523 | |
b4ead7d4 BS |
524 | /* Note if a block is a natural loop header. */ |
525 | sbitmap header; | |
526 | ||
09da1532 | 527 | /* Note if a block is a natural inner loop header. */ |
b4ead7d4 BS |
528 | sbitmap inner; |
529 | ||
530 | /* Note if a block is in the block queue. */ | |
531 | sbitmap in_queue; | |
532 | ||
533 | /* Note if a block is in the block queue. */ | |
534 | sbitmap in_stack; | |
535 | ||
b4ead7d4 BS |
536 | /* Perform a DFS traversal of the cfg. Identify loop headers, inner loops |
537 | and a mapping from block to its loop header (if the block is contained | |
538 | in a loop, else -1). | |
539 | ||
540 | Store results in HEADER, INNER, and MAX_HDR respectively, these will | |
541 | be used as inputs to the second traversal. | |
542 | ||
543 | STACK, SP and DFS_NR are only used during the first traversal. */ | |
544 | ||
545 | /* Allocate and initialize variables for the first traversal. */ | |
5ed6ace5 MD |
546 | max_hdr = XNEWVEC (int, last_basic_block); |
547 | dfs_nr = XCNEWVEC (int, last_basic_block); | |
548 | stack = XNEWVEC (edge_iterator, n_edges); | |
b4ead7d4 | 549 | |
d55bc081 | 550 | inner = sbitmap_alloc (last_basic_block); |
b4ead7d4 BS |
551 | sbitmap_ones (inner); |
552 | ||
d55bc081 | 553 | header = sbitmap_alloc (last_basic_block); |
b4ead7d4 BS |
554 | sbitmap_zero (header); |
555 | ||
d55bc081 | 556 | in_queue = sbitmap_alloc (last_basic_block); |
b4ead7d4 BS |
557 | sbitmap_zero (in_queue); |
558 | ||
d55bc081 | 559 | in_stack = sbitmap_alloc (last_basic_block); |
b4ead7d4 BS |
560 | sbitmap_zero (in_stack); |
561 | ||
bf77398c | 562 | for (i = 0; i < last_basic_block; i++) |
b4ead7d4 BS |
563 | max_hdr[i] = -1; |
564 | ||
dcda8480 UW |
565 | #define EDGE_PASSED(E) (ei_end_p ((E)) || ei_edge ((E))->aux) |
566 | #define SET_EDGE_PASSED(E) (ei_edge ((E))->aux = ei_edge ((E))) | |
567 | ||
b4ead7d4 BS |
568 | /* DFS traversal to find inner loops in the cfg. */ |
569 | ||
c5cbcccf | 570 | current_edge = ei_start (single_succ (ENTRY_BLOCK_PTR)->succs); |
b4ead7d4 | 571 | sp = -1; |
dcda8480 | 572 | |
b4ead7d4 BS |
573 | while (1) |
574 | { | |
dcda8480 | 575 | if (EDGE_PASSED (current_edge)) |
b4ead7d4 BS |
576 | { |
577 | /* We have reached a leaf node or a node that was already | |
578 | processed. Pop edges off the stack until we find | |
579 | an edge that has not yet been processed. */ | |
dcda8480 | 580 | while (sp >= 0 && EDGE_PASSED (current_edge)) |
b4ead7d4 BS |
581 | { |
582 | /* Pop entry off the stack. */ | |
583 | current_edge = stack[sp--]; | |
dcda8480 UW |
584 | node = ei_edge (current_edge)->src->index; |
585 | gcc_assert (node != ENTRY_BLOCK); | |
586 | child = ei_edge (current_edge)->dest->index; | |
587 | gcc_assert (child != EXIT_BLOCK); | |
b4ead7d4 BS |
588 | RESET_BIT (in_stack, child); |
589 | if (max_hdr[child] >= 0 && TEST_BIT (in_stack, max_hdr[child])) | |
590 | UPDATE_LOOP_RELATIONS (node, max_hdr[child]); | |
dcda8480 | 591 | ei_next (¤t_edge); |
b4ead7d4 BS |
592 | } |
593 | ||
594 | /* See if have finished the DFS tree traversal. */ | |
dcda8480 | 595 | if (sp < 0 && EDGE_PASSED (current_edge)) |
b4ead7d4 BS |
596 | break; |
597 | ||
598 | /* Nope, continue the traversal with the popped node. */ | |
599 | continue; | |
600 | } | |
601 | ||
602 | /* Process a node. */ | |
dcda8480 UW |
603 | node = ei_edge (current_edge)->src->index; |
604 | gcc_assert (node != ENTRY_BLOCK); | |
b4ead7d4 BS |
605 | SET_BIT (in_stack, node); |
606 | dfs_nr[node] = ++count; | |
607 | ||
dcda8480 UW |
608 | /* We don't traverse to the exit block. */ |
609 | child = ei_edge (current_edge)->dest->index; | |
610 | if (child == EXIT_BLOCK) | |
611 | { | |
612 | SET_EDGE_PASSED (current_edge); | |
613 | ei_next (¤t_edge); | |
614 | continue; | |
615 | } | |
616 | ||
b4ead7d4 BS |
617 | /* If the successor is in the stack, then we've found a loop. |
618 | Mark the loop, if it is not a natural loop, then it will | |
619 | be rejected during the second traversal. */ | |
620 | if (TEST_BIT (in_stack, child)) | |
621 | { | |
622 | no_loops = 0; | |
623 | SET_BIT (header, child); | |
624 | UPDATE_LOOP_RELATIONS (node, child); | |
dcda8480 UW |
625 | SET_EDGE_PASSED (current_edge); |
626 | ei_next (¤t_edge); | |
b4ead7d4 BS |
627 | continue; |
628 | } | |
629 | ||
630 | /* If the child was already visited, then there is no need to visit | |
631 | it again. Just update the loop relationships and restart | |
632 | with a new edge. */ | |
633 | if (dfs_nr[child]) | |
634 | { | |
635 | if (max_hdr[child] >= 0 && TEST_BIT (in_stack, max_hdr[child])) | |
636 | UPDATE_LOOP_RELATIONS (node, max_hdr[child]); | |
dcda8480 UW |
637 | SET_EDGE_PASSED (current_edge); |
638 | ei_next (¤t_edge); | |
b4ead7d4 BS |
639 | continue; |
640 | } | |
641 | ||
642 | /* Push an entry on the stack and continue DFS traversal. */ | |
643 | stack[++sp] = current_edge; | |
dcda8480 UW |
644 | SET_EDGE_PASSED (current_edge); |
645 | current_edge = ei_start (ei_edge (current_edge)->dest->succs); | |
646 | } | |
647 | ||
648 | /* Reset ->aux field used by EDGE_PASSED. */ | |
649 | FOR_ALL_BB (bb) | |
650 | { | |
651 | edge_iterator ei; | |
652 | edge e; | |
653 | FOR_EACH_EDGE (e, ei, bb->succs) | |
654 | e->aux = NULL; | |
b4ead7d4 BS |
655 | } |
656 | ||
dcda8480 | 657 | |
b4ead7d4 BS |
658 | /* Another check for unreachable blocks. The earlier test in |
659 | is_cfg_nonregular only finds unreachable blocks that do not | |
660 | form a loop. | |
661 | ||
662 | The DFS traversal will mark every block that is reachable from | |
663 | the entry node by placing a nonzero value in dfs_nr. Thus if | |
664 | dfs_nr is zero for any block, then it must be unreachable. */ | |
665 | unreachable = 0; | |
e0082a72 ZD |
666 | FOR_EACH_BB (bb) |
667 | if (dfs_nr[bb->index] == 0) | |
b4ead7d4 BS |
668 | { |
669 | unreachable = 1; | |
670 | break; | |
671 | } | |
672 | ||
673 | /* Gross. To avoid wasting memory, the second pass uses the dfs_nr array | |
674 | to hold degree counts. */ | |
675 | degree = dfs_nr; | |
676 | ||
e0082a72 | 677 | FOR_EACH_BB (bb) |
dcda8480 | 678 | degree[bb->index] = EDGE_COUNT (bb->preds); |
b4ead7d4 BS |
679 | |
680 | /* Do not perform region scheduling if there are any unreachable | |
681 | blocks. */ | |
682 | if (!unreachable) | |
683 | { | |
d08eefb9 MK |
684 | int *queue, *degree1 = NULL; |
685 | /* We use EXTENDED_RGN_HEADER as an addition to HEADER and put | |
686 | there basic blocks, which are forced to be region heads. | |
687 | This is done to try to assemble few smaller regions | |
688 | from a too_large region. */ | |
689 | sbitmap extended_rgn_header = NULL; | |
690 | bool extend_regions_p; | |
b4ead7d4 BS |
691 | |
692 | if (no_loops) | |
693 | SET_BIT (header, 0); | |
694 | ||
14b493d6 | 695 | /* Second traversal:find reducible inner loops and topologically sort |
b4ead7d4 BS |
696 | block of each region. */ |
697 | ||
5ed6ace5 | 698 | queue = XNEWVEC (int, n_basic_blocks); |
d08eefb9 MK |
699 | |
700 | extend_regions_p = PARAM_VALUE (PARAM_MAX_SCHED_EXTEND_REGIONS_ITERS) > 0; | |
701 | if (extend_regions_p) | |
702 | { | |
703 | degree1 = xmalloc (last_basic_block * sizeof (int)); | |
704 | extended_rgn_header = sbitmap_alloc (last_basic_block); | |
705 | sbitmap_zero (extended_rgn_header); | |
706 | } | |
b4ead7d4 BS |
707 | |
708 | /* Find blocks which are inner loop headers. We still have non-reducible | |
709 | loops to consider at this point. */ | |
e0082a72 | 710 | FOR_EACH_BB (bb) |
b4ead7d4 | 711 | { |
e0082a72 | 712 | if (TEST_BIT (header, bb->index) && TEST_BIT (inner, bb->index)) |
b4ead7d4 BS |
713 | { |
714 | edge e; | |
628f6a4e | 715 | edge_iterator ei; |
e0082a72 | 716 | basic_block jbb; |
b4ead7d4 BS |
717 | |
718 | /* Now check that the loop is reducible. We do this separate | |
719 | from finding inner loops so that we do not find a reducible | |
720 | loop which contains an inner non-reducible loop. | |
721 | ||
722 | A simple way to find reducible/natural loops is to verify | |
723 | that each block in the loop is dominated by the loop | |
724 | header. | |
725 | ||
726 | If there exists a block that is not dominated by the loop | |
727 | header, then the block is reachable from outside the loop | |
728 | and thus the loop is not a natural loop. */ | |
e0082a72 | 729 | FOR_EACH_BB (jbb) |
b4ead7d4 BS |
730 | { |
731 | /* First identify blocks in the loop, except for the loop | |
732 | entry block. */ | |
e0082a72 | 733 | if (bb->index == max_hdr[jbb->index] && bb != jbb) |
b4ead7d4 BS |
734 | { |
735 | /* Now verify that the block is dominated by the loop | |
736 | header. */ | |
d47cc544 | 737 | if (!dominated_by_p (CDI_DOMINATORS, jbb, bb)) |
b4ead7d4 BS |
738 | break; |
739 | } | |
740 | } | |
741 | ||
742 | /* If we exited the loop early, then I is the header of | |
743 | a non-reducible loop and we should quit processing it | |
744 | now. */ | |
e0082a72 | 745 | if (jbb != EXIT_BLOCK_PTR) |
b4ead7d4 BS |
746 | continue; |
747 | ||
748 | /* I is a header of an inner loop, or block 0 in a subroutine | |
749 | with no loops at all. */ | |
750 | head = tail = -1; | |
751 | too_large_failure = 0; | |
e0082a72 | 752 | loop_head = max_hdr[bb->index]; |
b4ead7d4 | 753 | |
d08eefb9 MK |
754 | if (extend_regions_p) |
755 | /* We save degree in case when we meet a too_large region | |
756 | and cancel it. We need a correct degree later when | |
757 | calling extend_rgns. */ | |
758 | memcpy (degree1, degree, last_basic_block * sizeof (int)); | |
759 | ||
b4ead7d4 BS |
760 | /* Decrease degree of all I's successors for topological |
761 | ordering. */ | |
628f6a4e | 762 | FOR_EACH_EDGE (e, ei, bb->succs) |
b4ead7d4 | 763 | if (e->dest != EXIT_BLOCK_PTR) |
0b17ab2f | 764 | --degree[e->dest->index]; |
b4ead7d4 BS |
765 | |
766 | /* Estimate # insns, and count # blocks in the region. */ | |
767 | num_bbs = 1; | |
a813c111 SB |
768 | num_insns = (INSN_LUID (BB_END (bb)) |
769 | - INSN_LUID (BB_HEAD (bb))); | |
b4ead7d4 BS |
770 | |
771 | /* Find all loop latches (blocks with back edges to the loop | |
772 | header) or all the leaf blocks in the cfg has no loops. | |
773 | ||
774 | Place those blocks into the queue. */ | |
775 | if (no_loops) | |
776 | { | |
e0082a72 | 777 | FOR_EACH_BB (jbb) |
b4ead7d4 BS |
778 | /* Leaf nodes have only a single successor which must |
779 | be EXIT_BLOCK. */ | |
c5cbcccf ZD |
780 | if (single_succ_p (jbb) |
781 | && single_succ (jbb) == EXIT_BLOCK_PTR) | |
b4ead7d4 | 782 | { |
e0082a72 ZD |
783 | queue[++tail] = jbb->index; |
784 | SET_BIT (in_queue, jbb->index); | |
b4ead7d4 | 785 | |
e0082a72 | 786 | if (too_large (jbb->index, &num_bbs, &num_insns)) |
b4ead7d4 BS |
787 | { |
788 | too_large_failure = 1; | |
789 | break; | |
790 | } | |
791 | } | |
792 | } | |
793 | else | |
794 | { | |
795 | edge e; | |
796 | ||
628f6a4e | 797 | FOR_EACH_EDGE (e, ei, bb->preds) |
b4ead7d4 BS |
798 | { |
799 | if (e->src == ENTRY_BLOCK_PTR) | |
800 | continue; | |
801 | ||
0b17ab2f | 802 | node = e->src->index; |
b4ead7d4 | 803 | |
e0082a72 | 804 | if (max_hdr[node] == loop_head && node != bb->index) |
b4ead7d4 BS |
805 | { |
806 | /* This is a loop latch. */ | |
807 | queue[++tail] = node; | |
808 | SET_BIT (in_queue, node); | |
809 | ||
810 | if (too_large (node, &num_bbs, &num_insns)) | |
811 | { | |
812 | too_large_failure = 1; | |
813 | break; | |
814 | } | |
815 | } | |
816 | } | |
817 | } | |
818 | ||
819 | /* Now add all the blocks in the loop to the queue. | |
820 | ||
821 | We know the loop is a natural loop; however the algorithm | |
822 | above will not always mark certain blocks as being in the | |
823 | loop. Consider: | |
824 | node children | |
825 | a b,c | |
826 | b c | |
827 | c a,d | |
828 | d b | |
829 | ||
830 | The algorithm in the DFS traversal may not mark B & D as part | |
454ff5cb | 831 | of the loop (i.e. they will not have max_hdr set to A). |
b4ead7d4 BS |
832 | |
833 | We know they can not be loop latches (else they would have | |
834 | had max_hdr set since they'd have a backedge to a dominator | |
835 | block). So we don't need them on the initial queue. | |
836 | ||
837 | We know they are part of the loop because they are dominated | |
838 | by the loop header and can be reached by a backwards walk of | |
839 | the edges starting with nodes on the initial queue. | |
840 | ||
841 | It is safe and desirable to include those nodes in the | |
842 | loop/scheduling region. To do so we would need to decrease | |
843 | the degree of a node if it is the target of a backedge | |
844 | within the loop itself as the node is placed in the queue. | |
845 | ||
846 | We do not do this because I'm not sure that the actual | |
847 | scheduling code will properly handle this case. ?!? */ | |
848 | ||
849 | while (head < tail && !too_large_failure) | |
850 | { | |
851 | edge e; | |
852 | child = queue[++head]; | |
853 | ||
628f6a4e | 854 | FOR_EACH_EDGE (e, ei, BASIC_BLOCK (child)->preds) |
b4ead7d4 | 855 | { |
0b17ab2f | 856 | node = e->src->index; |
b4ead7d4 BS |
857 | |
858 | /* See discussion above about nodes not marked as in | |
859 | this loop during the initial DFS traversal. */ | |
860 | if (e->src == ENTRY_BLOCK_PTR | |
861 | || max_hdr[node] != loop_head) | |
862 | { | |
863 | tail = -1; | |
864 | break; | |
865 | } | |
e0082a72 | 866 | else if (!TEST_BIT (in_queue, node) && node != bb->index) |
b4ead7d4 BS |
867 | { |
868 | queue[++tail] = node; | |
869 | SET_BIT (in_queue, node); | |
870 | ||
871 | if (too_large (node, &num_bbs, &num_insns)) | |
872 | { | |
873 | too_large_failure = 1; | |
874 | break; | |
875 | } | |
876 | } | |
877 | } | |
878 | } | |
879 | ||
880 | if (tail >= 0 && !too_large_failure) | |
881 | { | |
882 | /* Place the loop header into list of region blocks. */ | |
e0082a72 ZD |
883 | degree[bb->index] = -1; |
884 | rgn_bb_table[idx] = bb->index; | |
b4ead7d4 BS |
885 | RGN_NR_BLOCKS (nr_regions) = num_bbs; |
886 | RGN_BLOCKS (nr_regions) = idx++; | |
496d7bb0 MK |
887 | RGN_DONT_CALC_DEPS (nr_regions) = 0; |
888 | RGN_HAS_REAL_EBB (nr_regions) = 0; | |
e0082a72 ZD |
889 | CONTAINING_RGN (bb->index) = nr_regions; |
890 | BLOCK_TO_BB (bb->index) = count = 0; | |
b4ead7d4 BS |
891 | |
892 | /* Remove blocks from queue[] when their in degree | |
893 | becomes zero. Repeat until no blocks are left on the | |
894 | list. This produces a topological list of blocks in | |
895 | the region. */ | |
896 | while (tail >= 0) | |
897 | { | |
898 | if (head < 0) | |
899 | head = tail; | |
900 | child = queue[head]; | |
901 | if (degree[child] == 0) | |
902 | { | |
903 | edge e; | |
904 | ||
905 | degree[child] = -1; | |
906 | rgn_bb_table[idx++] = child; | |
907 | BLOCK_TO_BB (child) = ++count; | |
908 | CONTAINING_RGN (child) = nr_regions; | |
909 | queue[head] = queue[tail--]; | |
910 | ||
628f6a4e | 911 | FOR_EACH_EDGE (e, ei, BASIC_BLOCK (child)->succs) |
b4ead7d4 | 912 | if (e->dest != EXIT_BLOCK_PTR) |
0b17ab2f | 913 | --degree[e->dest->index]; |
b4ead7d4 BS |
914 | } |
915 | else | |
916 | --head; | |
917 | } | |
918 | ++nr_regions; | |
919 | } | |
d08eefb9 MK |
920 | else if (extend_regions_p) |
921 | { | |
922 | /* Restore DEGREE. */ | |
923 | int *t = degree; | |
924 | ||
925 | degree = degree1; | |
926 | degree1 = t; | |
927 | ||
928 | /* And force successors of BB to be region heads. | |
929 | This may provide several smaller regions instead | |
930 | of one too_large region. */ | |
931 | FOR_EACH_EDGE (e, ei, bb->succs) | |
932 | if (e->dest != EXIT_BLOCK_PTR) | |
933 | SET_BIT (extended_rgn_header, e->dest->index); | |
934 | } | |
b4ead7d4 BS |
935 | } |
936 | } | |
937 | free (queue); | |
d08eefb9 MK |
938 | |
939 | if (extend_regions_p) | |
940 | { | |
941 | free (degree1); | |
942 | ||
943 | sbitmap_a_or_b (header, header, extended_rgn_header); | |
944 | sbitmap_free (extended_rgn_header); | |
945 | ||
946 | extend_rgns (degree, &idx, header, max_hdr); | |
947 | } | |
b4ead7d4 BS |
948 | } |
949 | ||
950 | /* Any block that did not end up in a region is placed into a region | |
951 | by itself. */ | |
e0082a72 ZD |
952 | FOR_EACH_BB (bb) |
953 | if (degree[bb->index] >= 0) | |
b4ead7d4 | 954 | { |
e0082a72 | 955 | rgn_bb_table[idx] = bb->index; |
b4ead7d4 BS |
956 | RGN_NR_BLOCKS (nr_regions) = 1; |
957 | RGN_BLOCKS (nr_regions) = idx++; | |
496d7bb0 MK |
958 | RGN_DONT_CALC_DEPS (nr_regions) = 0; |
959 | RGN_HAS_REAL_EBB (nr_regions) = 0; | |
e0082a72 ZD |
960 | CONTAINING_RGN (bb->index) = nr_regions++; |
961 | BLOCK_TO_BB (bb->index) = 0; | |
b4ead7d4 BS |
962 | } |
963 | ||
964 | free (max_hdr); | |
d08eefb9 | 965 | free (degree); |
b4ead7d4 | 966 | free (stack); |
7b25b076 GS |
967 | sbitmap_free (header); |
968 | sbitmap_free (inner); | |
969 | sbitmap_free (in_queue); | |
970 | sbitmap_free (in_stack); | |
b4ead7d4 BS |
971 | } |
972 | ||
d08eefb9 MK |
973 | static int gather_region_statistics (int **); |
974 | static void print_region_statistics (int *, int, int *, int); | |
975 | ||
976 | /* Calculate the histogram that shows the number of regions having the | |
977 | given number of basic blocks, and store it in the RSP array. Return | |
978 | the size of this array. */ | |
979 | static int | |
980 | gather_region_statistics (int **rsp) | |
981 | { | |
982 | int i, *a = 0, a_sz = 0; | |
983 | ||
984 | /* a[i] is the number of regions that have (i + 1) basic blocks. */ | |
985 | for (i = 0; i < nr_regions; i++) | |
986 | { | |
987 | int nr_blocks = RGN_NR_BLOCKS (i); | |
988 | ||
989 | gcc_assert (nr_blocks >= 1); | |
990 | ||
991 | if (nr_blocks > a_sz) | |
992 | { | |
993 | a = xrealloc (a, nr_blocks * sizeof (*a)); | |
994 | do | |
995 | a[a_sz++] = 0; | |
996 | while (a_sz != nr_blocks); | |
997 | } | |
998 | ||
999 | a[nr_blocks - 1]++; | |
1000 | } | |
1001 | ||
1002 | *rsp = a; | |
1003 | return a_sz; | |
1004 | } | |
1005 | ||
1006 | /* Print regions statistics. S1 and S2 denote the data before and after | |
1007 | calling extend_rgns, respectively. */ | |
1008 | static void | |
1009 | print_region_statistics (int *s1, int s1_sz, int *s2, int s2_sz) | |
1010 | { | |
1011 | int i; | |
1012 | ||
1013 | /* We iterate until s2_sz because extend_rgns does not decrease | |
1014 | the maximal region size. */ | |
1015 | for (i = 1; i < s2_sz; i++) | |
1016 | { | |
1017 | int n1, n2; | |
1018 | ||
1019 | n2 = s2[i]; | |
1020 | ||
1021 | if (n2 == 0) | |
1022 | continue; | |
1023 | ||
1024 | if (i >= s1_sz) | |
1025 | n1 = 0; | |
1026 | else | |
1027 | n1 = s1[i]; | |
1028 | ||
1029 | fprintf (sched_dump, ";; Region extension statistics: size %d: " \ | |
1030 | "was %d + %d more\n", i + 1, n1, n2 - n1); | |
1031 | } | |
1032 | } | |
1033 | ||
1034 | /* Extend regions. | |
1035 | DEGREE - Array of incoming edge count, considering only | |
1036 | the edges, that don't have their sources in formed regions yet. | |
1037 | IDXP - pointer to the next available index in rgn_bb_table. | |
1038 | HEADER - set of all region heads. | |
1039 | LOOP_HDR - mapping from block to the containing loop | |
1040 | (two blocks can reside within one region if they have | |
1041 | the same loop header). */ | |
1042 | static void | |
1043 | extend_rgns (int *degree, int *idxp, sbitmap header, int *loop_hdr) | |
1044 | { | |
1045 | int *order, i, rescan = 0, idx = *idxp, iter = 0, max_iter, *max_hdr; | |
1046 | int nblocks = n_basic_blocks - NUM_FIXED_BLOCKS; | |
1047 | ||
1048 | max_iter = PARAM_VALUE (PARAM_MAX_SCHED_EXTEND_REGIONS_ITERS); | |
1049 | ||
1050 | max_hdr = xmalloc (last_basic_block * sizeof (*max_hdr)); | |
1051 | ||
1052 | order = xmalloc (last_basic_block * sizeof (*order)); | |
6fb5fa3c | 1053 | post_order_compute (order, false, false); |
d08eefb9 MK |
1054 | |
1055 | for (i = nblocks - 1; i >= 0; i--) | |
1056 | { | |
1057 | int bbn = order[i]; | |
1058 | if (degree[bbn] >= 0) | |
1059 | { | |
1060 | max_hdr[bbn] = bbn; | |
1061 | rescan = 1; | |
1062 | } | |
1063 | else | |
1064 | /* This block already was processed in find_rgns. */ | |
1065 | max_hdr[bbn] = -1; | |
1066 | } | |
1067 | ||
1068 | /* The idea is to topologically walk through CFG in top-down order. | |
1069 | During the traversal, if all the predecessors of a node are | |
1070 | marked to be in the same region (they all have the same max_hdr), | |
1071 | then current node is also marked to be a part of that region. | |
1072 | Otherwise the node starts its own region. | |
1073 | CFG should be traversed until no further changes are made. On each | |
1074 | iteration the set of the region heads is extended (the set of those | |
1075 | blocks that have max_hdr[bbi] == bbi). This set is upper bounded by the | |
1076 | set of all basic blocks, thus the algorithm is guaranteed to terminate. */ | |
1077 | ||
1078 | while (rescan && iter < max_iter) | |
1079 | { | |
1080 | rescan = 0; | |
1081 | ||
1082 | for (i = nblocks - 1; i >= 0; i--) | |
1083 | { | |
1084 | edge e; | |
1085 | edge_iterator ei; | |
1086 | int bbn = order[i]; | |
1087 | ||
1088 | if (max_hdr[bbn] != -1 && !TEST_BIT (header, bbn)) | |
1089 | { | |
1090 | int hdr = -1; | |
1091 | ||
1092 | FOR_EACH_EDGE (e, ei, BASIC_BLOCK (bbn)->preds) | |
1093 | { | |
1094 | int predn = e->src->index; | |
1095 | ||
1096 | if (predn != ENTRY_BLOCK | |
1097 | /* If pred wasn't processed in find_rgns. */ | |
1098 | && max_hdr[predn] != -1 | |
1099 | /* And pred and bb reside in the same loop. | |
1100 | (Or out of any loop). */ | |
1101 | && loop_hdr[bbn] == loop_hdr[predn]) | |
1102 | { | |
1103 | if (hdr == -1) | |
1104 | /* Then bb extends the containing region of pred. */ | |
1105 | hdr = max_hdr[predn]; | |
1106 | else if (hdr != max_hdr[predn]) | |
1107 | /* Too bad, there are at least two predecessors | |
1108 | that reside in different regions. Thus, BB should | |
1109 | begin its own region. */ | |
1110 | { | |
1111 | hdr = bbn; | |
1112 | break; | |
1113 | } | |
1114 | } | |
1115 | else | |
1116 | /* BB starts its own region. */ | |
1117 | { | |
1118 | hdr = bbn; | |
1119 | break; | |
1120 | } | |
1121 | } | |
1122 | ||
1123 | if (hdr == bbn) | |
1124 | { | |
1125 | /* If BB start its own region, | |
1126 | update set of headers with BB. */ | |
1127 | SET_BIT (header, bbn); | |
1128 | rescan = 1; | |
1129 | } | |
1130 | else | |
1131 | gcc_assert (hdr != -1); | |
1132 | ||
1133 | max_hdr[bbn] = hdr; | |
1134 | } | |
1135 | } | |
1136 | ||
1137 | iter++; | |
1138 | } | |
1139 | ||
1140 | /* Statistics were gathered on the SPEC2000 package of tests with | |
1141 | mainline weekly snapshot gcc-4.1-20051015 on ia64. | |
1142 | ||
1143 | Statistics for SPECint: | |
1144 | 1 iteration : 1751 cases (38.7%) | |
1145 | 2 iterations: 2770 cases (61.3%) | |
1146 | Blocks wrapped in regions by find_rgns without extension: 18295 blocks | |
1147 | Blocks wrapped in regions by 2 iterations in extend_rgns: 23821 blocks | |
1148 | (We don't count single block regions here). | |
1149 | ||
1150 | Statistics for SPECfp: | |
1151 | 1 iteration : 621 cases (35.9%) | |
1152 | 2 iterations: 1110 cases (64.1%) | |
1153 | Blocks wrapped in regions by find_rgns without extension: 6476 blocks | |
1154 | Blocks wrapped in regions by 2 iterations in extend_rgns: 11155 blocks | |
1155 | (We don't count single block regions here). | |
1156 | ||
1157 | By default we do at most 2 iterations. | |
917f1b7e | 1158 | This can be overridden with max-sched-extend-regions-iters parameter: |
d08eefb9 MK |
1159 | 0 - disable region extension, |
1160 | N > 0 - do at most N iterations. */ | |
1161 | ||
1162 | if (sched_verbose && iter != 0) | |
1163 | fprintf (sched_dump, ";; Region extension iterations: %d%s\n", iter, | |
1164 | rescan ? "... failed" : ""); | |
1165 | ||
1166 | if (!rescan && iter != 0) | |
1167 | { | |
1168 | int *s1 = NULL, s1_sz = 0; | |
1169 | ||
1170 | /* Save the old statistics for later printout. */ | |
1171 | if (sched_verbose >= 6) | |
1172 | s1_sz = gather_region_statistics (&s1); | |
1173 | ||
1174 | /* We have succeeded. Now assemble the regions. */ | |
1175 | for (i = nblocks - 1; i >= 0; i--) | |
1176 | { | |
1177 | int bbn = order[i]; | |
1178 | ||
1179 | if (max_hdr[bbn] == bbn) | |
1180 | /* BBN is a region head. */ | |
1181 | { | |
1182 | edge e; | |
1183 | edge_iterator ei; | |
1184 | int num_bbs = 0, j, num_insns = 0, large; | |
1185 | ||
1186 | large = too_large (bbn, &num_bbs, &num_insns); | |
1187 | ||
1188 | degree[bbn] = -1; | |
1189 | rgn_bb_table[idx] = bbn; | |
1190 | RGN_BLOCKS (nr_regions) = idx++; | |
496d7bb0 MK |
1191 | RGN_DONT_CALC_DEPS (nr_regions) = 0; |
1192 | RGN_HAS_REAL_EBB (nr_regions) = 0; | |
d08eefb9 MK |
1193 | CONTAINING_RGN (bbn) = nr_regions; |
1194 | BLOCK_TO_BB (bbn) = 0; | |
1195 | ||
1196 | FOR_EACH_EDGE (e, ei, BASIC_BLOCK (bbn)->succs) | |
1197 | if (e->dest != EXIT_BLOCK_PTR) | |
1198 | degree[e->dest->index]--; | |
1199 | ||
1200 | if (!large) | |
1201 | /* Here we check whether the region is too_large. */ | |
1202 | for (j = i - 1; j >= 0; j--) | |
1203 | { | |
1204 | int succn = order[j]; | |
1205 | if (max_hdr[succn] == bbn) | |
1206 | { | |
1207 | if ((large = too_large (succn, &num_bbs, &num_insns))) | |
1208 | break; | |
1209 | } | |
1210 | } | |
1211 | ||
1212 | if (large) | |
1213 | /* If the region is too_large, then wrap every block of | |
1214 | the region into single block region. | |
1215 | Here we wrap region head only. Other blocks are | |
1216 | processed in the below cycle. */ | |
1217 | { | |
1218 | RGN_NR_BLOCKS (nr_regions) = 1; | |
1219 | nr_regions++; | |
1220 | } | |
1221 | ||
1222 | num_bbs = 1; | |
1223 | ||
1224 | for (j = i - 1; j >= 0; j--) | |
1225 | { | |
1226 | int succn = order[j]; | |
1227 | ||
1228 | if (max_hdr[succn] == bbn) | |
1229 | /* This cycle iterates over all basic blocks, that | |
1230 | are supposed to be in the region with head BBN, | |
1231 | and wraps them into that region (or in single | |
1232 | block region). */ | |
1233 | { | |
1234 | gcc_assert (degree[succn] == 0); | |
1235 | ||
1236 | degree[succn] = -1; | |
1237 | rgn_bb_table[idx] = succn; | |
1238 | BLOCK_TO_BB (succn) = large ? 0 : num_bbs++; | |
1239 | CONTAINING_RGN (succn) = nr_regions; | |
1240 | ||
1241 | if (large) | |
1242 | /* Wrap SUCCN into single block region. */ | |
1243 | { | |
1244 | RGN_BLOCKS (nr_regions) = idx; | |
1245 | RGN_NR_BLOCKS (nr_regions) = 1; | |
496d7bb0 MK |
1246 | RGN_DONT_CALC_DEPS (nr_regions) = 0; |
1247 | RGN_HAS_REAL_EBB (nr_regions) = 0; | |
d08eefb9 MK |
1248 | nr_regions++; |
1249 | } | |
1250 | ||
1251 | idx++; | |
1252 | ||
1253 | FOR_EACH_EDGE (e, ei, BASIC_BLOCK (succn)->succs) | |
1254 | if (e->dest != EXIT_BLOCK_PTR) | |
1255 | degree[e->dest->index]--; | |
1256 | } | |
1257 | } | |
1258 | ||
1259 | if (!large) | |
1260 | { | |
1261 | RGN_NR_BLOCKS (nr_regions) = num_bbs; | |
1262 | nr_regions++; | |
1263 | } | |
1264 | } | |
1265 | } | |
1266 | ||
1267 | if (sched_verbose >= 6) | |
1268 | { | |
1269 | int *s2, s2_sz; | |
1270 | ||
1271 | /* Get the new statistics and print the comparison with the | |
1272 | one before calling this function. */ | |
1273 | s2_sz = gather_region_statistics (&s2); | |
1274 | print_region_statistics (s1, s1_sz, s2, s2_sz); | |
1275 | free (s1); | |
1276 | free (s2); | |
1277 | } | |
1278 | } | |
1279 | ||
1280 | free (order); | |
1281 | free (max_hdr); | |
1282 | ||
1283 | *idxp = idx; | |
1284 | } | |
1285 | ||
b4ead7d4 BS |
1286 | /* Functions for regions scheduling information. */ |
1287 | ||
1288 | /* Compute dominators, probability, and potential-split-edges of bb. | |
1289 | Assume that these values were already computed for bb's predecessors. */ | |
1290 | ||
1291 | static void | |
46c5ad27 | 1292 | compute_dom_prob_ps (int bb) |
b4ead7d4 | 1293 | { |
36968131 PS |
1294 | edge_iterator in_ei; |
1295 | edge in_edge; | |
b4ead7d4 | 1296 | |
496d7bb0 MK |
1297 | /* We shouldn't have any real ebbs yet. */ |
1298 | gcc_assert (ebb_head [bb] == bb + current_blocks); | |
1299 | ||
b4ead7d4 BS |
1300 | if (IS_RGN_ENTRY (bb)) |
1301 | { | |
bdfa170f | 1302 | SET_BIT (dom[bb], 0); |
36968131 | 1303 | prob[bb] = REG_BR_PROB_BASE; |
b4ead7d4 BS |
1304 | return; |
1305 | } | |
1306 | ||
36968131 PS |
1307 | prob[bb] = 0; |
1308 | ||
eaec9b3d | 1309 | /* Initialize dom[bb] to '111..1'. */ |
bdfa170f | 1310 | sbitmap_ones (dom[bb]); |
b4ead7d4 | 1311 | |
dcda8480 | 1312 | FOR_EACH_EDGE (in_edge, in_ei, BASIC_BLOCK (BB_TO_BLOCK (bb))->preds) |
b4ead7d4 | 1313 | { |
36968131 PS |
1314 | int pred_bb; |
1315 | edge out_edge; | |
1316 | edge_iterator out_ei; | |
1317 | ||
dcda8480 UW |
1318 | if (in_edge->src == ENTRY_BLOCK_PTR) |
1319 | continue; | |
b4ead7d4 | 1320 | |
dcda8480 UW |
1321 | pred_bb = BLOCK_TO_BB (in_edge->src->index); |
1322 | sbitmap_a_and_b (dom[bb], dom[bb], dom[pred_bb]); | |
1323 | sbitmap_a_or_b (ancestor_edges[bb], | |
1324 | ancestor_edges[bb], ancestor_edges[pred_bb]); | |
b4ead7d4 | 1325 | |
dcda8480 | 1326 | SET_BIT (ancestor_edges[bb], EDGE_TO_BIT (in_edge)); |
bdfa170f | 1327 | |
dcda8480 | 1328 | sbitmap_a_or_b (pot_split[bb], pot_split[bb], pot_split[pred_bb]); |
b4ead7d4 | 1329 | |
dcda8480 | 1330 | FOR_EACH_EDGE (out_edge, out_ei, in_edge->src->succs) |
36968131 | 1331 | SET_BIT (pot_split[bb], EDGE_TO_BIT (out_edge)); |
dcda8480 | 1332 | |
36968131 | 1333 | prob[bb] += ((prob[pred_bb] * in_edge->probability) / REG_BR_PROB_BASE); |
b4ead7d4 | 1334 | } |
b4ead7d4 | 1335 | |
bdfa170f DB |
1336 | SET_BIT (dom[bb], bb); |
1337 | sbitmap_difference (pot_split[bb], pot_split[bb], ancestor_edges[bb]); | |
b4ead7d4 BS |
1338 | |
1339 | if (sched_verbose >= 2) | |
1340 | fprintf (sched_dump, ";; bb_prob(%d, %d) = %3d\n", bb, BB_TO_BLOCK (bb), | |
36968131 | 1341 | (100 * prob[bb]) / REG_BR_PROB_BASE); |
b4ead7d4 BS |
1342 | } |
1343 | ||
1344 | /* Functions for target info. */ | |
1345 | ||
1346 | /* Compute in BL the list of split-edges of bb_src relatively to bb_trg. | |
1347 | Note that bb_trg dominates bb_src. */ | |
1348 | ||
1349 | static void | |
46c5ad27 | 1350 | split_edges (int bb_src, int bb_trg, edgelst *bl) |
b4ead7d4 | 1351 | { |
703ad42b | 1352 | sbitmap src = sbitmap_alloc (pot_split[bb_src]->n_bits); |
bdfa170f DB |
1353 | sbitmap_copy (src, pot_split[bb_src]); |
1354 | ||
1355 | sbitmap_difference (src, src, pot_split[bb_trg]); | |
dcda8480 | 1356 | extract_edgelst (src, bl); |
bdfa170f | 1357 | sbitmap_free (src); |
b4ead7d4 BS |
1358 | } |
1359 | ||
1360 | /* Find the valid candidate-source-blocks for the target block TRG, compute | |
1361 | their probability, and check if they are speculative or not. | |
1362 | For speculative sources, compute their update-blocks and split-blocks. */ | |
1363 | ||
1364 | static void | |
46c5ad27 | 1365 | compute_trg_info (int trg) |
b4ead7d4 | 1366 | { |
b3694847 | 1367 | candidate *sp; |
ba8a73e9 | 1368 | edgelst el = { NULL, 0 }; |
dcda8480 UW |
1369 | int i, j, k, update_idx; |
1370 | basic_block block; | |
740ce53d | 1371 | sbitmap visited; |
dcda8480 UW |
1372 | edge_iterator ei; |
1373 | edge e; | |
b4ead7d4 BS |
1374 | |
1375 | /* Define some of the fields for the target bb as well. */ | |
1376 | sp = candidate_table + trg; | |
1377 | sp->is_valid = 1; | |
1378 | sp->is_speculative = 0; | |
36968131 | 1379 | sp->src_prob = REG_BR_PROB_BASE; |
b4ead7d4 | 1380 | |
24bd1a0b | 1381 | visited = sbitmap_alloc (last_basic_block); |
740ce53d | 1382 | |
b4ead7d4 BS |
1383 | for (i = trg + 1; i < current_nr_blocks; i++) |
1384 | { | |
1385 | sp = candidate_table + i; | |
1386 | ||
1387 | sp->is_valid = IS_DOMINATED (i, trg); | |
1388 | if (sp->is_valid) | |
1389 | { | |
36968131 PS |
1390 | int tf = prob[trg], cf = prob[i]; |
1391 | ||
1392 | /* In CFGs with low probability edges TF can possibly be zero. */ | |
1393 | sp->src_prob = (tf ? ((cf * REG_BR_PROB_BASE) / tf) : 0); | |
1394 | sp->is_valid = (sp->src_prob >= min_spec_prob); | |
b4ead7d4 BS |
1395 | } |
1396 | ||
1397 | if (sp->is_valid) | |
1398 | { | |
1399 | split_edges (i, trg, &el); | |
1400 | sp->is_speculative = (el.nr_members) ? 1 : 0; | |
1401 | if (sp->is_speculative && !flag_schedule_speculative) | |
1402 | sp->is_valid = 0; | |
1403 | } | |
1404 | ||
1405 | if (sp->is_valid) | |
1406 | { | |
b4ead7d4 BS |
1407 | /* Compute split blocks and store them in bblst_table. |
1408 | The TO block of every split edge is a split block. */ | |
1409 | sp->split_bbs.first_member = &bblst_table[bblst_last]; | |
1410 | sp->split_bbs.nr_members = el.nr_members; | |
1411 | for (j = 0; j < el.nr_members; bblst_last++, j++) | |
dcda8480 | 1412 | bblst_table[bblst_last] = el.first_member[j]->dest; |
b4ead7d4 BS |
1413 | sp->update_bbs.first_member = &bblst_table[bblst_last]; |
1414 | ||
1415 | /* Compute update blocks and store them in bblst_table. | |
1416 | For every split edge, look at the FROM block, and check | |
1417 | all out edges. For each out edge that is not a split edge, | |
1418 | add the TO block to the update block list. This list can end | |
1419 | up with a lot of duplicates. We need to weed them out to avoid | |
1420 | overrunning the end of the bblst_table. */ | |
b4ead7d4 BS |
1421 | |
1422 | update_idx = 0; | |
740ce53d | 1423 | sbitmap_zero (visited); |
b4ead7d4 BS |
1424 | for (j = 0; j < el.nr_members; j++) |
1425 | { | |
dcda8480 UW |
1426 | block = el.first_member[j]->src; |
1427 | FOR_EACH_EDGE (e, ei, block->succs) | |
b4ead7d4 | 1428 | { |
24bd1a0b | 1429 | if (!TEST_BIT (visited, e->dest->index)) |
b4ead7d4 BS |
1430 | { |
1431 | for (k = 0; k < el.nr_members; k++) | |
dcda8480 | 1432 | if (e == el.first_member[k]) |
b4ead7d4 BS |
1433 | break; |
1434 | ||
1435 | if (k >= el.nr_members) | |
1436 | { | |
dcda8480 | 1437 | bblst_table[bblst_last++] = e->dest; |
24bd1a0b | 1438 | SET_BIT (visited, e->dest->index); |
b4ead7d4 BS |
1439 | update_idx++; |
1440 | } | |
1441 | } | |
b4ead7d4 | 1442 | } |
b4ead7d4 BS |
1443 | } |
1444 | sp->update_bbs.nr_members = update_idx; | |
1445 | ||
1446 | /* Make sure we didn't overrun the end of bblst_table. */ | |
41374e13 | 1447 | gcc_assert (bblst_last <= bblst_size); |
b4ead7d4 BS |
1448 | } |
1449 | else | |
1450 | { | |
1451 | sp->split_bbs.nr_members = sp->update_bbs.nr_members = 0; | |
1452 | ||
1453 | sp->is_speculative = 0; | |
1454 | sp->src_prob = 0; | |
1455 | } | |
1456 | } | |
740ce53d SB |
1457 | |
1458 | sbitmap_free (visited); | |
b4ead7d4 BS |
1459 | } |
1460 | ||
1461 | /* Print candidates info, for debugging purposes. Callable from debugger. */ | |
1462 | ||
1463 | void | |
46c5ad27 | 1464 | debug_candidate (int i) |
b4ead7d4 BS |
1465 | { |
1466 | if (!candidate_table[i].is_valid) | |
1467 | return; | |
1468 | ||
1469 | if (candidate_table[i].is_speculative) | |
1470 | { | |
1471 | int j; | |
1472 | fprintf (sched_dump, "src b %d bb %d speculative \n", BB_TO_BLOCK (i), i); | |
1473 | ||
1474 | fprintf (sched_dump, "split path: "); | |
1475 | for (j = 0; j < candidate_table[i].split_bbs.nr_members; j++) | |
1476 | { | |
dcda8480 | 1477 | int b = candidate_table[i].split_bbs.first_member[j]->index; |
b4ead7d4 BS |
1478 | |
1479 | fprintf (sched_dump, " %d ", b); | |
1480 | } | |
1481 | fprintf (sched_dump, "\n"); | |
1482 | ||
1483 | fprintf (sched_dump, "update path: "); | |
1484 | for (j = 0; j < candidate_table[i].update_bbs.nr_members; j++) | |
1485 | { | |
dcda8480 | 1486 | int b = candidate_table[i].update_bbs.first_member[j]->index; |
b4ead7d4 BS |
1487 | |
1488 | fprintf (sched_dump, " %d ", b); | |
1489 | } | |
1490 | fprintf (sched_dump, "\n"); | |
1491 | } | |
1492 | else | |
1493 | { | |
1494 | fprintf (sched_dump, " src %d equivalent\n", BB_TO_BLOCK (i)); | |
1495 | } | |
1496 | } | |
1497 | ||
1498 | /* Print candidates info, for debugging purposes. Callable from debugger. */ | |
1499 | ||
1500 | void | |
46c5ad27 | 1501 | debug_candidates (int trg) |
b4ead7d4 BS |
1502 | { |
1503 | int i; | |
1504 | ||
1505 | fprintf (sched_dump, "----------- candidate table: target: b=%d bb=%d ---\n", | |
1506 | BB_TO_BLOCK (trg), trg); | |
1507 | for (i = trg + 1; i < current_nr_blocks; i++) | |
1508 | debug_candidate (i); | |
1509 | } | |
1510 | ||
14b493d6 | 1511 | /* Functions for speculative scheduling. */ |
b4ead7d4 | 1512 | |
6fb5fa3c DB |
1513 | static bitmap_head not_in_df; |
1514 | ||
b4ead7d4 BS |
1515 | /* Return 0 if x is a set of a register alive in the beginning of one |
1516 | of the split-blocks of src, otherwise return 1. */ | |
1517 | ||
1518 | static int | |
46c5ad27 | 1519 | check_live_1 (int src, rtx x) |
b4ead7d4 | 1520 | { |
b3694847 SS |
1521 | int i; |
1522 | int regno; | |
1523 | rtx reg = SET_DEST (x); | |
b4ead7d4 BS |
1524 | |
1525 | if (reg == 0) | |
1526 | return 1; | |
1527 | ||
46d096a3 SB |
1528 | while (GET_CODE (reg) == SUBREG |
1529 | || GET_CODE (reg) == ZERO_EXTRACT | |
b4ead7d4 BS |
1530 | || GET_CODE (reg) == STRICT_LOW_PART) |
1531 | reg = XEXP (reg, 0); | |
1532 | ||
7193d1dc | 1533 | if (GET_CODE (reg) == PARALLEL) |
b4ead7d4 | 1534 | { |
b3694847 | 1535 | int i; |
90d036a0 | 1536 | |
b4ead7d4 | 1537 | for (i = XVECLEN (reg, 0) - 1; i >= 0; i--) |
7193d1dc RK |
1538 | if (XEXP (XVECEXP (reg, 0, i), 0) != 0) |
1539 | if (check_live_1 (src, XEXP (XVECEXP (reg, 0, i), 0))) | |
90d036a0 | 1540 | return 1; |
90d036a0 | 1541 | |
b4ead7d4 BS |
1542 | return 0; |
1543 | } | |
1544 | ||
f8cfc6aa | 1545 | if (!REG_P (reg)) |
b4ead7d4 BS |
1546 | return 1; |
1547 | ||
1548 | regno = REGNO (reg); | |
1549 | ||
1550 | if (regno < FIRST_PSEUDO_REGISTER && global_regs[regno]) | |
1551 | { | |
1552 | /* Global registers are assumed live. */ | |
1553 | return 0; | |
1554 | } | |
1555 | else | |
1556 | { | |
1557 | if (regno < FIRST_PSEUDO_REGISTER) | |
1558 | { | |
1559 | /* Check for hard registers. */ | |
66fd46b6 | 1560 | int j = hard_regno_nregs[regno][GET_MODE (reg)]; |
b4ead7d4 BS |
1561 | while (--j >= 0) |
1562 | { | |
1563 | for (i = 0; i < candidate_table[src].split_bbs.nr_members; i++) | |
1564 | { | |
dcda8480 | 1565 | basic_block b = candidate_table[src].split_bbs.first_member[i]; |
6fb5fa3c | 1566 | int t = bitmap_bit_p (¬_in_df, b->index); |
b4ead7d4 | 1567 | |
496d7bb0 MK |
1568 | /* We can have split blocks, that were recently generated. |
1569 | such blocks are always outside current region. */ | |
6fb5fa3c DB |
1570 | gcc_assert (!t || (CONTAINING_RGN (b->index) |
1571 | != CONTAINING_RGN (BB_TO_BLOCK (src)))); | |
1572 | ||
89a95777 | 1573 | if (t || REGNO_REG_SET_P (df_get_live_in (b), regno + j)) |
6fb5fa3c | 1574 | return 0; |
b4ead7d4 BS |
1575 | } |
1576 | } | |
1577 | } | |
1578 | else | |
1579 | { | |
2067c116 | 1580 | /* Check for pseudo registers. */ |
b4ead7d4 BS |
1581 | for (i = 0; i < candidate_table[src].split_bbs.nr_members; i++) |
1582 | { | |
dcda8480 | 1583 | basic_block b = candidate_table[src].split_bbs.first_member[i]; |
6fb5fa3c | 1584 | int t = bitmap_bit_p (¬_in_df, b->index); |
b4ead7d4 | 1585 | |
6fb5fa3c DB |
1586 | gcc_assert (!t || (CONTAINING_RGN (b->index) |
1587 | != CONTAINING_RGN (BB_TO_BLOCK (src)))); | |
1588 | ||
89a95777 | 1589 | if (t || REGNO_REG_SET_P (df_get_live_in (b), regno)) |
6fb5fa3c | 1590 | return 0; |
b4ead7d4 BS |
1591 | } |
1592 | } | |
1593 | } | |
1594 | ||
1595 | return 1; | |
1596 | } | |
1597 | ||
1598 | /* If x is a set of a register R, mark that R is alive in the beginning | |
1599 | of every update-block of src. */ | |
1600 | ||
1601 | static void | |
46c5ad27 | 1602 | update_live_1 (int src, rtx x) |
b4ead7d4 | 1603 | { |
b3694847 SS |
1604 | int i; |
1605 | int regno; | |
1606 | rtx reg = SET_DEST (x); | |
b4ead7d4 BS |
1607 | |
1608 | if (reg == 0) | |
1609 | return; | |
1610 | ||
46d096a3 SB |
1611 | while (GET_CODE (reg) == SUBREG |
1612 | || GET_CODE (reg) == ZERO_EXTRACT | |
b4ead7d4 BS |
1613 | || GET_CODE (reg) == STRICT_LOW_PART) |
1614 | reg = XEXP (reg, 0); | |
1615 | ||
7193d1dc | 1616 | if (GET_CODE (reg) == PARALLEL) |
b4ead7d4 | 1617 | { |
b3694847 | 1618 | int i; |
90d036a0 | 1619 | |
b4ead7d4 | 1620 | for (i = XVECLEN (reg, 0) - 1; i >= 0; i--) |
7193d1dc RK |
1621 | if (XEXP (XVECEXP (reg, 0, i), 0) != 0) |
1622 | update_live_1 (src, XEXP (XVECEXP (reg, 0, i), 0)); | |
90d036a0 | 1623 | |
b4ead7d4 BS |
1624 | return; |
1625 | } | |
1626 | ||
f8cfc6aa | 1627 | if (!REG_P (reg)) |
b4ead7d4 BS |
1628 | return; |
1629 | ||
1630 | /* Global registers are always live, so the code below does not apply | |
1631 | to them. */ | |
1632 | ||
1633 | regno = REGNO (reg); | |
1634 | ||
1635 | if (regno >= FIRST_PSEUDO_REGISTER || !global_regs[regno]) | |
1636 | { | |
1637 | if (regno < FIRST_PSEUDO_REGISTER) | |
1638 | { | |
66fd46b6 | 1639 | int j = hard_regno_nregs[regno][GET_MODE (reg)]; |
b4ead7d4 BS |
1640 | while (--j >= 0) |
1641 | { | |
1642 | for (i = 0; i < candidate_table[src].update_bbs.nr_members; i++) | |
1643 | { | |
dcda8480 | 1644 | basic_block b = candidate_table[src].update_bbs.first_member[i]; |
b4ead7d4 | 1645 | |
89a95777 | 1646 | SET_REGNO_REG_SET (df_get_live_in (b), regno + j); |
b4ead7d4 BS |
1647 | } |
1648 | } | |
1649 | } | |
1650 | else | |
1651 | { | |
1652 | for (i = 0; i < candidate_table[src].update_bbs.nr_members; i++) | |
1653 | { | |
dcda8480 | 1654 | basic_block b = candidate_table[src].update_bbs.first_member[i]; |
b4ead7d4 | 1655 | |
89a95777 | 1656 | SET_REGNO_REG_SET (df_get_live_in (b), regno); |
b4ead7d4 BS |
1657 | } |
1658 | } | |
1659 | } | |
1660 | } | |
1661 | ||
1662 | /* Return 1 if insn can be speculatively moved from block src to trg, | |
1663 | otherwise return 0. Called before first insertion of insn to | |
1664 | ready-list or before the scheduling. */ | |
1665 | ||
1666 | static int | |
46c5ad27 | 1667 | check_live (rtx insn, int src) |
b4ead7d4 BS |
1668 | { |
1669 | /* Find the registers set by instruction. */ | |
1670 | if (GET_CODE (PATTERN (insn)) == SET | |
1671 | || GET_CODE (PATTERN (insn)) == CLOBBER) | |
1672 | return check_live_1 (src, PATTERN (insn)); | |
1673 | else if (GET_CODE (PATTERN (insn)) == PARALLEL) | |
1674 | { | |
1675 | int j; | |
1676 | for (j = XVECLEN (PATTERN (insn), 0) - 1; j >= 0; j--) | |
1677 | if ((GET_CODE (XVECEXP (PATTERN (insn), 0, j)) == SET | |
1678 | || GET_CODE (XVECEXP (PATTERN (insn), 0, j)) == CLOBBER) | |
1679 | && !check_live_1 (src, XVECEXP (PATTERN (insn), 0, j))) | |
1680 | return 0; | |
1681 | ||
1682 | return 1; | |
1683 | } | |
1684 | ||
1685 | return 1; | |
1686 | } | |
1687 | ||
1688 | /* Update the live registers info after insn was moved speculatively from | |
1689 | block src to trg. */ | |
1690 | ||
1691 | static void | |
46c5ad27 | 1692 | update_live (rtx insn, int src) |
b4ead7d4 BS |
1693 | { |
1694 | /* Find the registers set by instruction. */ | |
1695 | if (GET_CODE (PATTERN (insn)) == SET | |
1696 | || GET_CODE (PATTERN (insn)) == CLOBBER) | |
1697 | update_live_1 (src, PATTERN (insn)); | |
1698 | else if (GET_CODE (PATTERN (insn)) == PARALLEL) | |
1699 | { | |
1700 | int j; | |
1701 | for (j = XVECLEN (PATTERN (insn), 0) - 1; j >= 0; j--) | |
1702 | if (GET_CODE (XVECEXP (PATTERN (insn), 0, j)) == SET | |
1703 | || GET_CODE (XVECEXP (PATTERN (insn), 0, j)) == CLOBBER) | |
1704 | update_live_1 (src, XVECEXP (PATTERN (insn), 0, j)); | |
1705 | } | |
1706 | } | |
1707 | ||
272d0bee | 1708 | /* Nonzero if block bb_to is equal to, or reachable from block bb_from. */ |
b4ead7d4 | 1709 | #define IS_REACHABLE(bb_from, bb_to) \ |
786de7eb | 1710 | (bb_from == bb_to \ |
b4ead7d4 | 1711 | || IS_RGN_ENTRY (bb_from) \ |
786de7eb | 1712 | || (TEST_BIT (ancestor_edges[bb_to], \ |
c5cbcccf | 1713 | EDGE_TO_BIT (single_pred_edge (BASIC_BLOCK (BB_TO_BLOCK (bb_from))))))) |
b4ead7d4 | 1714 | |
b4ead7d4 BS |
1715 | /* Turns on the fed_by_spec_load flag for insns fed by load_insn. */ |
1716 | ||
1717 | static void | |
46c5ad27 | 1718 | set_spec_fed (rtx load_insn) |
b4ead7d4 | 1719 | { |
e2f6ff94 MK |
1720 | sd_iterator_def sd_it; |
1721 | dep_t dep; | |
b4ead7d4 | 1722 | |
e2f6ff94 MK |
1723 | FOR_EACH_DEP (load_insn, SD_LIST_FORW, sd_it, dep) |
1724 | if (DEP_TYPE (dep) == REG_DEP_TRUE) | |
1725 | FED_BY_SPEC_LOAD (DEP_CON (dep)) = 1; | |
b198261f | 1726 | } |
b4ead7d4 BS |
1727 | |
1728 | /* On the path from the insn to load_insn_bb, find a conditional | |
1729 | branch depending on insn, that guards the speculative load. */ | |
1730 | ||
1731 | static int | |
46c5ad27 | 1732 | find_conditional_protection (rtx insn, int load_insn_bb) |
b4ead7d4 | 1733 | { |
e2f6ff94 MK |
1734 | sd_iterator_def sd_it; |
1735 | dep_t dep; | |
b4ead7d4 BS |
1736 | |
1737 | /* Iterate through DEF-USE forward dependences. */ | |
e2f6ff94 | 1738 | FOR_EACH_DEP (insn, SD_LIST_FORW, sd_it, dep) |
b4ead7d4 | 1739 | { |
e2f6ff94 | 1740 | rtx next = DEP_CON (dep); |
b198261f | 1741 | |
b4ead7d4 BS |
1742 | if ((CONTAINING_RGN (BLOCK_NUM (next)) == |
1743 | CONTAINING_RGN (BB_TO_BLOCK (load_insn_bb))) | |
1744 | && IS_REACHABLE (INSN_BB (next), load_insn_bb) | |
1745 | && load_insn_bb != INSN_BB (next) | |
e2f6ff94 | 1746 | && DEP_TYPE (dep) == REG_DEP_TRUE |
4b4bf941 | 1747 | && (JUMP_P (next) |
b4ead7d4 BS |
1748 | || find_conditional_protection (next, load_insn_bb))) |
1749 | return 1; | |
1750 | } | |
1751 | return 0; | |
1752 | } /* find_conditional_protection */ | |
1753 | ||
1754 | /* Returns 1 if the same insn1 that participates in the computation | |
1755 | of load_insn's address is feeding a conditional branch that is | |
1756 | guarding on load_insn. This is true if we find a the two DEF-USE | |
1757 | chains: | |
1758 | insn1 -> ... -> conditional-branch | |
1759 | insn1 -> ... -> load_insn, | |
1760 | and if a flow path exist: | |
1761 | insn1 -> ... -> conditional-branch -> ... -> load_insn, | |
1762 | and if insn1 is on the path | |
1763 | region-entry -> ... -> bb_trg -> ... load_insn. | |
1764 | ||
b198261f MK |
1765 | Locate insn1 by climbing on INSN_BACK_DEPS from load_insn. |
1766 | Locate the branch by following INSN_FORW_DEPS from insn1. */ | |
b4ead7d4 BS |
1767 | |
1768 | static int | |
46c5ad27 | 1769 | is_conditionally_protected (rtx load_insn, int bb_src, int bb_trg) |
b4ead7d4 | 1770 | { |
e2f6ff94 MK |
1771 | sd_iterator_def sd_it; |
1772 | dep_t dep; | |
b4ead7d4 | 1773 | |
e2f6ff94 | 1774 | FOR_EACH_DEP (load_insn, SD_LIST_BACK, sd_it, dep) |
b4ead7d4 | 1775 | { |
e2f6ff94 | 1776 | rtx insn1 = DEP_PRO (dep); |
b4ead7d4 BS |
1777 | |
1778 | /* Must be a DEF-USE dependence upon non-branch. */ | |
e2f6ff94 | 1779 | if (DEP_TYPE (dep) != REG_DEP_TRUE |
4b4bf941 | 1780 | || JUMP_P (insn1)) |
b4ead7d4 BS |
1781 | continue; |
1782 | ||
1783 | /* Must exist a path: region-entry -> ... -> bb_trg -> ... load_insn. */ | |
1784 | if (INSN_BB (insn1) == bb_src | |
1785 | || (CONTAINING_RGN (BLOCK_NUM (insn1)) | |
1786 | != CONTAINING_RGN (BB_TO_BLOCK (bb_src))) | |
1787 | || (!IS_REACHABLE (bb_trg, INSN_BB (insn1)) | |
1788 | && !IS_REACHABLE (INSN_BB (insn1), bb_trg))) | |
1789 | continue; | |
1790 | ||
1791 | /* Now search for the conditional-branch. */ | |
1792 | if (find_conditional_protection (insn1, bb_src)) | |
1793 | return 1; | |
1794 | ||
1795 | /* Recursive step: search another insn1, "above" current insn1. */ | |
1796 | return is_conditionally_protected (insn1, bb_src, bb_trg); | |
1797 | } | |
1798 | ||
1799 | /* The chain does not exist. */ | |
1800 | return 0; | |
1801 | } /* is_conditionally_protected */ | |
1802 | ||
1803 | /* Returns 1 if a clue for "similar load" 'insn2' is found, and hence | |
1804 | load_insn can move speculatively from bb_src to bb_trg. All the | |
1805 | following must hold: | |
1806 | ||
1807 | (1) both loads have 1 base register (PFREE_CANDIDATEs). | |
1808 | (2) load_insn and load1 have a def-use dependence upon | |
1809 | the same insn 'insn1'. | |
1810 | (3) either load2 is in bb_trg, or: | |
1811 | - there's only one split-block, and | |
1812 | - load1 is on the escape path, and | |
1813 | ||
1814 | From all these we can conclude that the two loads access memory | |
1815 | addresses that differ at most by a constant, and hence if moving | |
1816 | load_insn would cause an exception, it would have been caused by | |
1817 | load2 anyhow. */ | |
1818 | ||
1819 | static int | |
46c5ad27 | 1820 | is_pfree (rtx load_insn, int bb_src, int bb_trg) |
b4ead7d4 | 1821 | { |
e2f6ff94 MK |
1822 | sd_iterator_def back_sd_it; |
1823 | dep_t back_dep; | |
b3694847 | 1824 | candidate *candp = candidate_table + bb_src; |
b4ead7d4 BS |
1825 | |
1826 | if (candp->split_bbs.nr_members != 1) | |
1827 | /* Must have exactly one escape block. */ | |
1828 | return 0; | |
1829 | ||
e2f6ff94 | 1830 | FOR_EACH_DEP (load_insn, SD_LIST_BACK, back_sd_it, back_dep) |
b4ead7d4 | 1831 | { |
e2f6ff94 | 1832 | rtx insn1 = DEP_PRO (back_dep); |
b4ead7d4 | 1833 | |
e2f6ff94 MK |
1834 | if (DEP_TYPE (back_dep) == REG_DEP_TRUE) |
1835 | /* Found a DEF-USE dependence (insn1, load_insn). */ | |
b4ead7d4 | 1836 | { |
e2f6ff94 MK |
1837 | sd_iterator_def fore_sd_it; |
1838 | dep_t fore_dep; | |
b4ead7d4 | 1839 | |
e2f6ff94 | 1840 | FOR_EACH_DEP (insn1, SD_LIST_FORW, fore_sd_it, fore_dep) |
b4ead7d4 | 1841 | { |
e2f6ff94 | 1842 | rtx insn2 = DEP_CON (fore_dep); |
b198261f | 1843 | |
e2f6ff94 | 1844 | if (DEP_TYPE (fore_dep) == REG_DEP_TRUE) |
b4ead7d4 BS |
1845 | { |
1846 | /* Found a DEF-USE dependence (insn1, insn2). */ | |
1847 | if (haifa_classify_insn (insn2) != PFREE_CANDIDATE) | |
1848 | /* insn2 not guaranteed to be a 1 base reg load. */ | |
1849 | continue; | |
1850 | ||
1851 | if (INSN_BB (insn2) == bb_trg) | |
1852 | /* insn2 is the similar load, in the target block. */ | |
1853 | return 1; | |
1854 | ||
dcda8480 | 1855 | if (*(candp->split_bbs.first_member) == BLOCK_FOR_INSN (insn2)) |
b4ead7d4 BS |
1856 | /* insn2 is a similar load, in a split-block. */ |
1857 | return 1; | |
1858 | } | |
1859 | } | |
1860 | } | |
1861 | } | |
1862 | ||
1863 | /* Couldn't find a similar load. */ | |
1864 | return 0; | |
1865 | } /* is_pfree */ | |
1866 | ||
b4ead7d4 BS |
1867 | /* Return 1 if load_insn is prisky (i.e. if load_insn is fed by |
1868 | a load moved speculatively, or if load_insn is protected by | |
1869 | a compare on load_insn's address). */ | |
1870 | ||
1871 | static int | |
46c5ad27 | 1872 | is_prisky (rtx load_insn, int bb_src, int bb_trg) |
b4ead7d4 BS |
1873 | { |
1874 | if (FED_BY_SPEC_LOAD (load_insn)) | |
1875 | return 1; | |
1876 | ||
e2f6ff94 | 1877 | if (sd_lists_empty_p (load_insn, SD_LIST_BACK)) |
b4ead7d4 BS |
1878 | /* Dependence may 'hide' out of the region. */ |
1879 | return 1; | |
1880 | ||
1881 | if (is_conditionally_protected (load_insn, bb_src, bb_trg)) | |
1882 | return 1; | |
1883 | ||
1884 | return 0; | |
1885 | } | |
1886 | ||
1887 | /* Insn is a candidate to be moved speculatively from bb_src to bb_trg. | |
1888 | Return 1 if insn is exception-free (and the motion is valid) | |
1889 | and 0 otherwise. */ | |
1890 | ||
1891 | static int | |
46c5ad27 | 1892 | is_exception_free (rtx insn, int bb_src, int bb_trg) |
b4ead7d4 BS |
1893 | { |
1894 | int insn_class = haifa_classify_insn (insn); | |
1895 | ||
1896 | /* Handle non-load insns. */ | |
1897 | switch (insn_class) | |
1898 | { | |
1899 | case TRAP_FREE: | |
1900 | return 1; | |
1901 | case TRAP_RISKY: | |
1902 | return 0; | |
1903 | default:; | |
1904 | } | |
1905 | ||
1906 | /* Handle loads. */ | |
1907 | if (!flag_schedule_speculative_load) | |
1908 | return 0; | |
1909 | IS_LOAD_INSN (insn) = 1; | |
1910 | switch (insn_class) | |
1911 | { | |
1912 | case IFREE: | |
1913 | return (1); | |
1914 | case IRISKY: | |
1915 | return 0; | |
1916 | case PFREE_CANDIDATE: | |
1917 | if (is_pfree (insn, bb_src, bb_trg)) | |
1918 | return 1; | |
1919 | /* Don't 'break' here: PFREE-candidate is also PRISKY-candidate. */ | |
1920 | case PRISKY_CANDIDATE: | |
1921 | if (!flag_schedule_speculative_load_dangerous | |
1922 | || is_prisky (insn, bb_src, bb_trg)) | |
1923 | return 0; | |
1924 | break; | |
1925 | default:; | |
1926 | } | |
1927 | ||
1928 | return flag_schedule_speculative_load_dangerous; | |
1929 | } | |
1930 | \f | |
1931 | /* The number of insns from the current block scheduled so far. */ | |
1932 | static int sched_target_n_insns; | |
1933 | /* The number of insns from the current block to be scheduled in total. */ | |
1934 | static int target_n_insns; | |
1935 | /* The number of insns from the entire region scheduled so far. */ | |
1936 | static int sched_n_insns; | |
1937 | ||
1938 | /* Implementations of the sched_info functions for region scheduling. */ | |
63f54b1a | 1939 | static void init_ready_list (void); |
46c5ad27 | 1940 | static int can_schedule_ready_p (rtx); |
496d7bb0 MK |
1941 | static void begin_schedule_ready (rtx, rtx); |
1942 | static ds_t new_ready (rtx, ds_t); | |
46c5ad27 AJ |
1943 | static int schedule_more_p (void); |
1944 | static const char *rgn_print_insn (rtx, int); | |
1945 | static int rgn_rank (rtx, rtx); | |
1946 | static int contributes_to_priority (rtx, rtx); | |
5a257872 | 1947 | static void compute_jump_reg_dependencies (rtx, regset, regset, regset); |
b4ead7d4 | 1948 | |
496d7bb0 MK |
1949 | /* Functions for speculative scheduling. */ |
1950 | static void add_remove_insn (rtx, int); | |
1951 | static void extend_regions (void); | |
1952 | static void add_block1 (basic_block, basic_block); | |
1953 | static void fix_recovery_cfg (int, int, int); | |
1954 | static basic_block advance_target_bb (basic_block, rtx); | |
496d7bb0 | 1955 | |
b640bd8f MK |
1956 | static void debug_rgn_dependencies (int); |
1957 | ||
b4ead7d4 BS |
1958 | /* Return nonzero if there are more insns that should be scheduled. */ |
1959 | ||
1960 | static int | |
46c5ad27 | 1961 | schedule_more_p (void) |
b4ead7d4 | 1962 | { |
496d7bb0 | 1963 | return sched_target_n_insns < target_n_insns; |
b4ead7d4 BS |
1964 | } |
1965 | ||
1966 | /* Add all insns that are initially ready to the ready list READY. Called | |
1967 | once before scheduling a set of insns. */ | |
1968 | ||
1969 | static void | |
63f54b1a | 1970 | init_ready_list (void) |
b4ead7d4 BS |
1971 | { |
1972 | rtx prev_head = current_sched_info->prev_head; | |
1973 | rtx next_tail = current_sched_info->next_tail; | |
1974 | int bb_src; | |
1975 | rtx insn; | |
1976 | ||
1977 | target_n_insns = 0; | |
1978 | sched_target_n_insns = 0; | |
1979 | sched_n_insns = 0; | |
1980 | ||
1981 | /* Print debugging information. */ | |
1982 | if (sched_verbose >= 5) | |
b640bd8f | 1983 | debug_rgn_dependencies (target_bb); |
b4ead7d4 BS |
1984 | |
1985 | /* Prepare current target block info. */ | |
1986 | if (current_nr_blocks > 1) | |
1987 | { | |
5ed6ace5 | 1988 | candidate_table = XNEWVEC (candidate, current_nr_blocks); |
b4ead7d4 BS |
1989 | |
1990 | bblst_last = 0; | |
1991 | /* bblst_table holds split blocks and update blocks for each block after | |
1992 | the current one in the region. split blocks and update blocks are | |
1993 | the TO blocks of region edges, so there can be at most rgn_nr_edges | |
1994 | of them. */ | |
1995 | bblst_size = (current_nr_blocks - target_bb) * rgn_nr_edges; | |
5ed6ace5 | 1996 | bblst_table = XNEWVEC (basic_block, bblst_size); |
b4ead7d4 | 1997 | |
dcda8480 | 1998 | edgelst_last = 0; |
5ed6ace5 | 1999 | edgelst_table = XNEWVEC (edge, rgn_nr_edges); |
b4ead7d4 BS |
2000 | |
2001 | compute_trg_info (target_bb); | |
2002 | } | |
2003 | ||
2004 | /* Initialize ready list with all 'ready' insns in target block. | |
2005 | Count number of insns in the target block being scheduled. */ | |
2006 | for (insn = NEXT_INSN (prev_head); insn != next_tail; insn = NEXT_INSN (insn)) | |
63f54b1a MK |
2007 | { |
2008 | try_ready (insn); | |
58fb7809 | 2009 | target_n_insns++; |
496d7bb0 MK |
2010 | |
2011 | gcc_assert (!(TODO_SPEC (insn) & BEGIN_CONTROL)); | |
b4ead7d4 BS |
2012 | } |
2013 | ||
2014 | /* Add to ready list all 'ready' insns in valid source blocks. | |
2015 | For speculative insns, check-live, exception-free, and | |
2016 | issue-delay. */ | |
2017 | for (bb_src = target_bb + 1; bb_src < current_nr_blocks; bb_src++) | |
2018 | if (IS_VALID (bb_src)) | |
2019 | { | |
2020 | rtx src_head; | |
2021 | rtx src_next_tail; | |
2022 | rtx tail, head; | |
2023 | ||
496d7bb0 MK |
2024 | get_ebb_head_tail (EBB_FIRST_BB (bb_src), EBB_LAST_BB (bb_src), |
2025 | &head, &tail); | |
b4ead7d4 BS |
2026 | src_next_tail = NEXT_INSN (tail); |
2027 | src_head = head; | |
2028 | ||
2029 | for (insn = src_head; insn != src_next_tail; insn = NEXT_INSN (insn)) | |
63f54b1a MK |
2030 | if (INSN_P (insn)) |
2031 | try_ready (insn); | |
b4ead7d4 BS |
2032 | } |
2033 | } | |
2034 | ||
2035 | /* Called after taking INSN from the ready list. Returns nonzero if this | |
2036 | insn can be scheduled, nonzero if we should silently discard it. */ | |
2037 | ||
2038 | static int | |
46c5ad27 | 2039 | can_schedule_ready_p (rtx insn) |
b4ead7d4 | 2040 | { |
496d7bb0 MK |
2041 | /* An interblock motion? */ |
2042 | if (INSN_BB (insn) != target_bb | |
2043 | && IS_SPECULATIVE_INSN (insn) | |
2044 | && !check_live (insn, INSN_BB (insn))) | |
2045 | return 0; | |
2046 | else | |
2047 | return 1; | |
2048 | } | |
79c2ffde | 2049 | |
917f1b7e | 2050 | /* Updates counter and other information. Split from can_schedule_ready_p () |
496d7bb0 MK |
2051 | because when we schedule insn speculatively then insn passed to |
2052 | can_schedule_ready_p () differs from the one passed to | |
2053 | begin_schedule_ready (). */ | |
2054 | static void | |
2055 | begin_schedule_ready (rtx insn, rtx last ATTRIBUTE_UNUSED) | |
2056 | { | |
b4ead7d4 BS |
2057 | /* An interblock motion? */ |
2058 | if (INSN_BB (insn) != target_bb) | |
2059 | { | |
b4ead7d4 BS |
2060 | if (IS_SPECULATIVE_INSN (insn)) |
2061 | { | |
496d7bb0 MK |
2062 | gcc_assert (check_live (insn, INSN_BB (insn))); |
2063 | ||
b4ead7d4 BS |
2064 | update_live (insn, INSN_BB (insn)); |
2065 | ||
2066 | /* For speculative load, mark insns fed by it. */ | |
2067 | if (IS_LOAD_INSN (insn) || FED_BY_SPEC_LOAD (insn)) | |
2068 | set_spec_fed (insn); | |
2069 | ||
2070 | nr_spec++; | |
2071 | } | |
2072 | nr_inter++; | |
b4ead7d4 BS |
2073 | } |
2074 | else | |
2075 | { | |
2076 | /* In block motion. */ | |
2077 | sched_target_n_insns++; | |
2078 | } | |
2079 | sched_n_insns++; | |
b4ead7d4 BS |
2080 | } |
2081 | ||
496d7bb0 MK |
2082 | /* Called after INSN has all its hard dependencies resolved and the speculation |
2083 | of type TS is enough to overcome them all. | |
2084 | Return nonzero if it should be moved to the ready list or the queue, or zero | |
2085 | if we should silently discard it. */ | |
2086 | static ds_t | |
2087 | new_ready (rtx next, ds_t ts) | |
b4ead7d4 | 2088 | { |
496d7bb0 MK |
2089 | if (INSN_BB (next) != target_bb) |
2090 | { | |
2091 | int not_ex_free = 0; | |
2092 | ||
2093 | /* For speculative insns, before inserting to ready/queue, | |
2094 | check live, exception-free, and issue-delay. */ | |
2095 | if (!IS_VALID (INSN_BB (next)) | |
b4ead7d4 BS |
2096 | || CANT_MOVE (next) |
2097 | || (IS_SPECULATIVE_INSN (next) | |
fa0aee89 | 2098 | && ((recog_memoized (next) >= 0 |
496d7bb0 MK |
2099 | && min_insn_conflict_delay (curr_state, next, next) |
2100 | > PARAM_VALUE (PARAM_MAX_SCHED_INSN_CONFLICT_DELAY)) | |
d7bfd907 | 2101 | || IS_SPECULATION_CHECK_P (next) |
b4ead7d4 | 2102 | || !check_live (next, INSN_BB (next)) |
496d7bb0 MK |
2103 | || (not_ex_free = !is_exception_free (next, INSN_BB (next), |
2104 | target_bb))))) | |
2105 | { | |
2106 | if (not_ex_free | |
2107 | /* We are here because is_exception_free () == false. | |
2108 | But we possibly can handle that with control speculation. */ | |
e2f6ff94 MK |
2109 | && (current_sched_info->flags & DO_SPECULATION) |
2110 | && (spec_info->mask & BEGIN_CONTROL)) | |
496d7bb0 MK |
2111 | /* Here we got new control-speculative instruction. */ |
2112 | ts = set_dep_weak (ts, BEGIN_CONTROL, MAX_DEP_WEAK); | |
2113 | else | |
2114 | ts = (ts & ~SPECULATIVE) | HARD_DEP; | |
2115 | } | |
2116 | } | |
2117 | ||
2118 | return ts; | |
b4ead7d4 BS |
2119 | } |
2120 | ||
2121 | /* Return a string that contains the insn uid and optionally anything else | |
2122 | necessary to identify this insn in an output. It's valid to use a | |
2123 | static buffer for this. The ALIGNED parameter should cause the string | |
2124 | to be formatted so that multiple output lines will line up nicely. */ | |
2125 | ||
2126 | static const char * | |
46c5ad27 | 2127 | rgn_print_insn (rtx insn, int aligned) |
b4ead7d4 BS |
2128 | { |
2129 | static char tmp[80]; | |
2130 | ||
2131 | if (aligned) | |
2132 | sprintf (tmp, "b%3d: i%4d", INSN_BB (insn), INSN_UID (insn)); | |
2133 | else | |
2134 | { | |
b4ead7d4 | 2135 | if (current_nr_blocks > 1 && INSN_BB (insn) != target_bb) |
f56887a7 BS |
2136 | sprintf (tmp, "%d/b%d", INSN_UID (insn), INSN_BB (insn)); |
2137 | else | |
2138 | sprintf (tmp, "%d", INSN_UID (insn)); | |
b4ead7d4 BS |
2139 | } |
2140 | return tmp; | |
2141 | } | |
2142 | ||
2143 | /* Compare priority of two insns. Return a positive number if the second | |
2144 | insn is to be preferred for scheduling, and a negative one if the first | |
2145 | is to be preferred. Zero if they are equally good. */ | |
2146 | ||
2147 | static int | |
46c5ad27 | 2148 | rgn_rank (rtx insn1, rtx insn2) |
b4ead7d4 BS |
2149 | { |
2150 | /* Some comparison make sense in interblock scheduling only. */ | |
2151 | if (INSN_BB (insn1) != INSN_BB (insn2)) | |
2152 | { | |
2153 | int spec_val, prob_val; | |
2154 | ||
2155 | /* Prefer an inblock motion on an interblock motion. */ | |
2156 | if ((INSN_BB (insn2) == target_bb) && (INSN_BB (insn1) != target_bb)) | |
2157 | return 1; | |
2158 | if ((INSN_BB (insn1) == target_bb) && (INSN_BB (insn2) != target_bb)) | |
2159 | return -1; | |
2160 | ||
2161 | /* Prefer a useful motion on a speculative one. */ | |
2162 | spec_val = IS_SPECULATIVE_INSN (insn1) - IS_SPECULATIVE_INSN (insn2); | |
2163 | if (spec_val) | |
2164 | return spec_val; | |
2165 | ||
2166 | /* Prefer a more probable (speculative) insn. */ | |
2167 | prob_val = INSN_PROBABILITY (insn2) - INSN_PROBABILITY (insn1); | |
2168 | if (prob_val) | |
2169 | return prob_val; | |
2170 | } | |
2171 | return 0; | |
2172 | } | |
2173 | ||
18e720b3 BS |
2174 | /* NEXT is an instruction that depends on INSN (a backward dependence); |
2175 | return nonzero if we should include this dependence in priority | |
2176 | calculations. */ | |
2177 | ||
2178 | static int | |
46c5ad27 | 2179 | contributes_to_priority (rtx next, rtx insn) |
18e720b3 | 2180 | { |
496d7bb0 MK |
2181 | /* NEXT and INSN reside in one ebb. */ |
2182 | return BLOCK_TO_BB (BLOCK_NUM (next)) == BLOCK_TO_BB (BLOCK_NUM (insn)); | |
18e720b3 BS |
2183 | } |
2184 | ||
5a257872 EB |
2185 | /* INSN is a JUMP_INSN, COND_SET is the set of registers that are |
2186 | conditionally set before INSN. Store the set of registers that | |
2187 | must be considered as used by this jump in USED and that of | |
2188 | registers that must be considered as set in SET. */ | |
18e720b3 BS |
2189 | |
2190 | static void | |
46c5ad27 | 2191 | compute_jump_reg_dependencies (rtx insn ATTRIBUTE_UNUSED, |
5a257872 EB |
2192 | regset cond_exec ATTRIBUTE_UNUSED, |
2193 | regset used ATTRIBUTE_UNUSED, | |
46c5ad27 | 2194 | regset set ATTRIBUTE_UNUSED) |
18e720b3 BS |
2195 | { |
2196 | /* Nothing to do here, since we postprocess jumps in | |
2197 | add_branch_dependences. */ | |
2198 | } | |
2199 | ||
b4ead7d4 BS |
2200 | /* Used in schedule_insns to initialize current_sched_info for scheduling |
2201 | regions (or single basic blocks). */ | |
2202 | ||
2203 | static struct sched_info region_sched_info = | |
2204 | { | |
2205 | init_ready_list, | |
2206 | can_schedule_ready_p, | |
2207 | schedule_more_p, | |
2208 | new_ready, | |
2209 | rgn_rank, | |
2210 | rgn_print_insn, | |
18e720b3 BS |
2211 | contributes_to_priority, |
2212 | compute_jump_reg_dependencies, | |
b4ead7d4 BS |
2213 | |
2214 | NULL, NULL, | |
2215 | NULL, NULL, | |
ddbd5439 MK |
2216 | 0, 0, 0, |
2217 | ||
496d7bb0 MK |
2218 | add_remove_insn, |
2219 | begin_schedule_ready, | |
2220 | add_block1, | |
2221 | advance_target_bb, | |
2222 | fix_recovery_cfg, | |
6fb5fa3c | 2223 | SCHED_RGN |
b4ead7d4 BS |
2224 | }; |
2225 | ||
68c17f30 RH |
2226 | /* Determine if PAT sets a CLASS_LIKELY_SPILLED_P register. */ |
2227 | ||
2228 | static bool | |
46c5ad27 | 2229 | sets_likely_spilled (rtx pat) |
68c17f30 RH |
2230 | { |
2231 | bool ret = false; | |
2232 | note_stores (pat, sets_likely_spilled_1, &ret); | |
2233 | return ret; | |
2234 | } | |
2235 | ||
2236 | static void | |
7bc980e1 | 2237 | sets_likely_spilled_1 (rtx x, const_rtx pat, void *data) |
68c17f30 RH |
2238 | { |
2239 | bool *ret = (bool *) data; | |
2240 | ||
2241 | if (GET_CODE (pat) == SET | |
2242 | && REG_P (x) | |
2243 | && REGNO (x) < FIRST_PSEUDO_REGISTER | |
2244 | && CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (REGNO (x)))) | |
2245 | *ret = true; | |
2246 | } | |
2247 | ||
b4ead7d4 BS |
2248 | /* Add dependences so that branches are scheduled to run last in their |
2249 | block. */ | |
2250 | ||
2251 | static void | |
46c5ad27 | 2252 | add_branch_dependences (rtx head, rtx tail) |
b4ead7d4 BS |
2253 | { |
2254 | rtx insn, last; | |
2255 | ||
8d8a083e RH |
2256 | /* For all branches, calls, uses, clobbers, cc0 setters, and instructions |
2257 | that can throw exceptions, force them to remain in order at the end of | |
2258 | the block by adding dependencies and giving the last a high priority. | |
2259 | There may be notes present, and prev_head may also be a note. | |
b4ead7d4 BS |
2260 | |
2261 | Branches must obviously remain at the end. Calls should remain at the | |
2262 | end since moving them results in worse register allocation. Uses remain | |
68c17f30 RH |
2263 | at the end to ensure proper register allocation. |
2264 | ||
d91edf86 | 2265 | cc0 setters remain at the end because they can't be moved away from |
68c17f30 RH |
2266 | their cc0 user. |
2267 | ||
2bd1e239 SB |
2268 | COND_EXEC insns cannot be moved past a branch (see e.g. PR17808). |
2269 | ||
68c17f30 RH |
2270 | Insns setting CLASS_LIKELY_SPILLED_P registers (usually return values) |
2271 | are not moved before reload because we can wind up with register | |
2272 | allocation failures. */ | |
2273 | ||
b4ead7d4 BS |
2274 | insn = tail; |
2275 | last = 0; | |
4b4bf941 JQ |
2276 | while (CALL_P (insn) |
2277 | || JUMP_P (insn) | |
2278 | || (NONJUMP_INSN_P (insn) | |
b4ead7d4 BS |
2279 | && (GET_CODE (PATTERN (insn)) == USE |
2280 | || GET_CODE (PATTERN (insn)) == CLOBBER | |
8d8a083e | 2281 | || can_throw_internal (insn) |
b4ead7d4 BS |
2282 | #ifdef HAVE_cc0 |
2283 | || sets_cc0_p (PATTERN (insn)) | |
2284 | #endif | |
68c17f30 RH |
2285 | || (!reload_completed |
2286 | && sets_likely_spilled (PATTERN (insn))))) | |
4b4bf941 | 2287 | || NOTE_P (insn)) |
b4ead7d4 | 2288 | { |
4b4bf941 | 2289 | if (!NOTE_P (insn)) |
b4ead7d4 | 2290 | { |
b198261f | 2291 | if (last != 0 |
e2f6ff94 | 2292 | && sd_find_dep_between (insn, last, false) == NULL) |
b4ead7d4 | 2293 | { |
2bd1e239 SB |
2294 | if (! sched_insns_conditions_mutex_p (last, insn)) |
2295 | add_dependence (last, insn, REG_DEP_ANTI); | |
b4ead7d4 BS |
2296 | INSN_REF_COUNT (insn)++; |
2297 | } | |
2298 | ||
2299 | CANT_MOVE (insn) = 1; | |
2300 | ||
2301 | last = insn; | |
b4ead7d4 BS |
2302 | } |
2303 | ||
2304 | /* Don't overrun the bounds of the basic block. */ | |
2305 | if (insn == head) | |
2306 | break; | |
2307 | ||
2308 | insn = PREV_INSN (insn); | |
2309 | } | |
2310 | ||
2311 | /* Make sure these insns are scheduled last in their block. */ | |
2312 | insn = last; | |
2313 | if (insn != 0) | |
2314 | while (insn != head) | |
2315 | { | |
2316 | insn = prev_nonnote_insn (insn); | |
2317 | ||
2318 | if (INSN_REF_COUNT (insn) != 0) | |
2319 | continue; | |
2320 | ||
2bd1e239 SB |
2321 | if (! sched_insns_conditions_mutex_p (last, insn)) |
2322 | add_dependence (last, insn, REG_DEP_ANTI); | |
b4ead7d4 | 2323 | INSN_REF_COUNT (insn) = 1; |
b4ead7d4 | 2324 | } |
2bd1e239 SB |
2325 | |
2326 | #ifdef HAVE_conditional_execution | |
2327 | /* Finally, if the block ends in a jump, and we are doing intra-block | |
2328 | scheduling, make sure that the branch depends on any COND_EXEC insns | |
2329 | inside the block to avoid moving the COND_EXECs past the branch insn. | |
2330 | ||
2331 | We only have to do this after reload, because (1) before reload there | |
2332 | are no COND_EXEC insns, and (2) the region scheduler is an intra-block | |
2333 | scheduler after reload. | |
2334 | ||
2335 | FIXME: We could in some cases move COND_EXEC insns past the branch if | |
2336 | this scheduler would be a little smarter. Consider this code: | |
2337 | ||
2338 | T = [addr] | |
2339 | C ? addr += 4 | |
abf86bf2 | 2340 | !C ? X += 12 |
2bd1e239 | 2341 | C ? T += 1 |
abf86bf2 | 2342 | C ? jump foo |
2bd1e239 SB |
2343 | |
2344 | On a target with a one cycle stall on a memory access the optimal | |
2345 | sequence would be: | |
2346 | ||
2347 | T = [addr] | |
2348 | C ? addr += 4 | |
2349 | C ? T += 1 | |
2350 | C ? jump foo | |
2351 | !C ? X += 12 | |
2352 | ||
2353 | We don't want to put the 'X += 12' before the branch because it just | |
2354 | wastes a cycle of execution time when the branch is taken. | |
2355 | ||
2356 | Note that in the example "!C" will always be true. That is another | |
2357 | possible improvement for handling COND_EXECs in this scheduler: it | |
2358 | could remove always-true predicates. */ | |
2359 | ||
2360 | if (!reload_completed || ! JUMP_P (tail)) | |
2361 | return; | |
2362 | ||
abf86bf2 | 2363 | insn = tail; |
2bd1e239 SB |
2364 | while (insn != head) |
2365 | { | |
abf86bf2 RE |
2366 | insn = PREV_INSN (insn); |
2367 | ||
2bd1e239 SB |
2368 | /* Note that we want to add this dependency even when |
2369 | sched_insns_conditions_mutex_p returns true. The whole point | |
2370 | is that we _want_ this dependency, even if these insns really | |
2371 | are independent. */ | |
2372 | if (INSN_P (insn) && GET_CODE (PATTERN (insn)) == COND_EXEC) | |
2373 | add_dependence (tail, insn, REG_DEP_ANTI); | |
2bd1e239 SB |
2374 | } |
2375 | #endif | |
b4ead7d4 BS |
2376 | } |
2377 | ||
2378 | /* Data structures for the computation of data dependences in a regions. We | |
2379 | keep one `deps' structure for every basic block. Before analyzing the | |
2380 | data dependences for a bb, its variables are initialized as a function of | |
2381 | the variables of its predecessors. When the analysis for a bb completes, | |
2382 | we save the contents to the corresponding bb_deps[bb] variable. */ | |
2383 | ||
2384 | static struct deps *bb_deps; | |
2385 | ||
37a0f8a5 RH |
2386 | /* Duplicate the INSN_LIST elements of COPY and prepend them to OLD. */ |
2387 | ||
2388 | static rtx | |
46c5ad27 | 2389 | concat_INSN_LIST (rtx copy, rtx old) |
37a0f8a5 RH |
2390 | { |
2391 | rtx new = old; | |
2392 | for (; copy ; copy = XEXP (copy, 1)) | |
2393 | new = alloc_INSN_LIST (XEXP (copy, 0), new); | |
2394 | return new; | |
2395 | } | |
2396 | ||
2397 | static void | |
46c5ad27 AJ |
2398 | concat_insn_mem_list (rtx copy_insns, rtx copy_mems, rtx *old_insns_p, |
2399 | rtx *old_mems_p) | |
37a0f8a5 RH |
2400 | { |
2401 | rtx new_insns = *old_insns_p; | |
2402 | rtx new_mems = *old_mems_p; | |
2403 | ||
2404 | while (copy_insns) | |
2405 | { | |
2406 | new_insns = alloc_INSN_LIST (XEXP (copy_insns, 0), new_insns); | |
2407 | new_mems = alloc_EXPR_LIST (VOIDmode, XEXP (copy_mems, 0), new_mems); | |
2408 | copy_insns = XEXP (copy_insns, 1); | |
2409 | copy_mems = XEXP (copy_mems, 1); | |
2410 | } | |
2411 | ||
2412 | *old_insns_p = new_insns; | |
2413 | *old_mems_p = new_mems; | |
2414 | } | |
2415 | ||
b4ead7d4 | 2416 | /* After computing the dependencies for block BB, propagate the dependencies |
4ba478b8 | 2417 | found in TMP_DEPS to the successors of the block. */ |
b4ead7d4 | 2418 | static void |
46c5ad27 | 2419 | propagate_deps (int bb, struct deps *pred_deps) |
b4ead7d4 | 2420 | { |
dcda8480 UW |
2421 | basic_block block = BASIC_BLOCK (BB_TO_BLOCK (bb)); |
2422 | edge_iterator ei; | |
2423 | edge e; | |
b4ead7d4 BS |
2424 | |
2425 | /* bb's structures are inherited by its successors. */ | |
dcda8480 UW |
2426 | FOR_EACH_EDGE (e, ei, block->succs) |
2427 | { | |
2428 | struct deps *succ_deps; | |
3cd8c58a | 2429 | unsigned reg; |
a2041967 | 2430 | reg_set_iterator rsi; |
b4ead7d4 | 2431 | |
dcda8480 UW |
2432 | /* Only bbs "below" bb, in the same region, are interesting. */ |
2433 | if (e->dest == EXIT_BLOCK_PTR | |
2434 | || CONTAINING_RGN (block->index) != CONTAINING_RGN (e->dest->index) | |
2435 | || BLOCK_TO_BB (e->dest->index) <= bb) | |
2436 | continue; | |
37a0f8a5 | 2437 | |
dcda8480 | 2438 | succ_deps = bb_deps + BLOCK_TO_BB (e->dest->index); |
37a0f8a5 | 2439 | |
dcda8480 | 2440 | /* The reg_last lists are inherited by successor. */ |
a2041967 | 2441 | EXECUTE_IF_SET_IN_REG_SET (&pred_deps->reg_last_in_use, 0, reg, rsi) |
dcda8480 UW |
2442 | { |
2443 | struct deps_reg *pred_rl = &pred_deps->reg_last[reg]; | |
2444 | struct deps_reg *succ_rl = &succ_deps->reg_last[reg]; | |
2445 | ||
2446 | succ_rl->uses = concat_INSN_LIST (pred_rl->uses, succ_rl->uses); | |
2447 | succ_rl->sets = concat_INSN_LIST (pred_rl->sets, succ_rl->sets); | |
2448 | succ_rl->clobbers = concat_INSN_LIST (pred_rl->clobbers, | |
2449 | succ_rl->clobbers); | |
2450 | succ_rl->uses_length += pred_rl->uses_length; | |
2451 | succ_rl->clobbers_length += pred_rl->clobbers_length; | |
a2041967 | 2452 | } |
dcda8480 UW |
2453 | IOR_REG_SET (&succ_deps->reg_last_in_use, &pred_deps->reg_last_in_use); |
2454 | ||
2455 | /* Mem read/write lists are inherited by successor. */ | |
2456 | concat_insn_mem_list (pred_deps->pending_read_insns, | |
2457 | pred_deps->pending_read_mems, | |
2458 | &succ_deps->pending_read_insns, | |
2459 | &succ_deps->pending_read_mems); | |
2460 | concat_insn_mem_list (pred_deps->pending_write_insns, | |
2461 | pred_deps->pending_write_mems, | |
2462 | &succ_deps->pending_write_insns, | |
2463 | &succ_deps->pending_write_mems); | |
2464 | ||
2465 | succ_deps->last_pending_memory_flush | |
2466 | = concat_INSN_LIST (pred_deps->last_pending_memory_flush, | |
2467 | succ_deps->last_pending_memory_flush); | |
2468 | ||
bdbf40a5 MK |
2469 | succ_deps->pending_read_list_length |
2470 | += pred_deps->pending_read_list_length; | |
2471 | succ_deps->pending_write_list_length | |
2472 | += pred_deps->pending_write_list_length; | |
dcda8480 UW |
2473 | succ_deps->pending_flush_length += pred_deps->pending_flush_length; |
2474 | ||
2475 | /* last_function_call is inherited by successor. */ | |
2476 | succ_deps->last_function_call | |
2477 | = concat_INSN_LIST (pred_deps->last_function_call, | |
2478 | succ_deps->last_function_call); | |
2479 | ||
2480 | /* sched_before_next_call is inherited by successor. */ | |
2481 | succ_deps->sched_before_next_call | |
2482 | = concat_INSN_LIST (pred_deps->sched_before_next_call, | |
2483 | succ_deps->sched_before_next_call); | |
2484 | } | |
b4ead7d4 | 2485 | |
37a0f8a5 RH |
2486 | /* These lists should point to the right place, for correct |
2487 | freeing later. */ | |
2488 | bb_deps[bb].pending_read_insns = pred_deps->pending_read_insns; | |
2489 | bb_deps[bb].pending_read_mems = pred_deps->pending_read_mems; | |
2490 | bb_deps[bb].pending_write_insns = pred_deps->pending_write_insns; | |
2491 | bb_deps[bb].pending_write_mems = pred_deps->pending_write_mems; | |
2492 | ||
2493 | /* Can't allow these to be freed twice. */ | |
2494 | pred_deps->pending_read_insns = 0; | |
2495 | pred_deps->pending_read_mems = 0; | |
2496 | pred_deps->pending_write_insns = 0; | |
2497 | pred_deps->pending_write_mems = 0; | |
b4ead7d4 BS |
2498 | } |
2499 | ||
e2f6ff94 | 2500 | /* Compute dependences inside bb. In a multiple blocks region: |
b4ead7d4 BS |
2501 | (1) a bb is analyzed after its predecessors, and (2) the lists in |
2502 | effect at the end of bb (after analyzing for bb) are inherited by | |
14b493d6 | 2503 | bb's successors. |
b4ead7d4 BS |
2504 | |
2505 | Specifically for reg-reg data dependences, the block insns are | |
2506 | scanned by sched_analyze () top-to-bottom. Two lists are | |
4ba478b8 RH |
2507 | maintained by sched_analyze (): reg_last[].sets for register DEFs, |
2508 | and reg_last[].uses for register USEs. | |
b4ead7d4 BS |
2509 | |
2510 | When analysis is completed for bb, we update for its successors: | |
2511 | ; - DEFS[succ] = Union (DEFS [succ], DEFS [bb]) | |
2512 | ; - USES[succ] = Union (USES [succ], DEFS [bb]) | |
2513 | ||
2514 | The mechanism for computing mem-mem data dependence is very | |
2515 | similar, and the result is interblock dependences in the region. */ | |
2516 | ||
2517 | static void | |
e2f6ff94 | 2518 | compute_block_dependences (int bb) |
b4ead7d4 BS |
2519 | { |
2520 | rtx head, tail; | |
b4ead7d4 BS |
2521 | struct deps tmp_deps; |
2522 | ||
2523 | tmp_deps = bb_deps[bb]; | |
2524 | ||
2525 | /* Do the analysis for this block. */ | |
496d7bb0 MK |
2526 | gcc_assert (EBB_FIRST_BB (bb) == EBB_LAST_BB (bb)); |
2527 | get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail); | |
e2f6ff94 | 2528 | |
b4ead7d4 BS |
2529 | sched_analyze (&tmp_deps, head, tail); |
2530 | add_branch_dependences (head, tail); | |
2531 | ||
2532 | if (current_nr_blocks > 1) | |
4ba478b8 | 2533 | propagate_deps (bb, &tmp_deps); |
b4ead7d4 BS |
2534 | |
2535 | /* Free up the INSN_LISTs. */ | |
2536 | free_deps (&tmp_deps); | |
e2f6ff94 MK |
2537 | |
2538 | if (targetm.sched.dependencies_evaluation_hook) | |
2539 | targetm.sched.dependencies_evaluation_hook (head, tail); | |
2540 | } | |
2541 | ||
2542 | /* Free dependencies of instructions inside BB. */ | |
2543 | static void | |
2544 | free_block_dependencies (int bb) | |
2545 | { | |
2546 | rtx head; | |
2547 | rtx tail; | |
2548 | ||
2549 | get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail); | |
2550 | ||
2551 | sched_free_deps (head, tail, true); | |
b4ead7d4 | 2552 | } |
4ba478b8 | 2553 | |
b4ead7d4 BS |
2554 | /* Remove all INSN_LISTs and EXPR_LISTs from the pending lists and add |
2555 | them to the unused_*_list variables, so that they can be reused. */ | |
2556 | ||
2557 | static void | |
46c5ad27 | 2558 | free_pending_lists (void) |
b4ead7d4 BS |
2559 | { |
2560 | int bb; | |
2561 | ||
2562 | for (bb = 0; bb < current_nr_blocks; bb++) | |
2563 | { | |
2564 | free_INSN_LIST_list (&bb_deps[bb].pending_read_insns); | |
2565 | free_INSN_LIST_list (&bb_deps[bb].pending_write_insns); | |
2566 | free_EXPR_LIST_list (&bb_deps[bb].pending_read_mems); | |
2567 | free_EXPR_LIST_list (&bb_deps[bb].pending_write_mems); | |
2568 | } | |
2569 | } | |
2570 | \f | |
e2f6ff94 MK |
2571 | /* Print dependences for debugging starting from FROM_BB. |
2572 | Callable from debugger. */ | |
b640bd8f MK |
2573 | /* Print dependences for debugging starting from FROM_BB. |
2574 | Callable from debugger. */ | |
b4ead7d4 | 2575 | void |
b640bd8f | 2576 | debug_rgn_dependencies (int from_bb) |
b4ead7d4 BS |
2577 | { |
2578 | int bb; | |
2579 | ||
b640bd8f MK |
2580 | fprintf (sched_dump, |
2581 | ";; --------------- forward dependences: ------------ \n"); | |
2582 | ||
2583 | for (bb = from_bb; bb < current_nr_blocks; bb++) | |
b4ead7d4 | 2584 | { |
fa0aee89 | 2585 | rtx head, tail; |
fa0aee89 | 2586 | |
496d7bb0 MK |
2587 | gcc_assert (EBB_FIRST_BB (bb) == EBB_LAST_BB (bb)); |
2588 | get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail); | |
fa0aee89 PB |
2589 | fprintf (sched_dump, "\n;; --- Region Dependences --- b %d bb %d \n", |
2590 | BB_TO_BLOCK (bb), bb); | |
2591 | ||
b640bd8f MK |
2592 | debug_dependencies (head, tail); |
2593 | } | |
2594 | } | |
fa0aee89 | 2595 | |
b640bd8f MK |
2596 | /* Print dependencies information for instructions between HEAD and TAIL. |
2597 | ??? This function would probably fit best in haifa-sched.c. */ | |
2598 | void debug_dependencies (rtx head, rtx tail) | |
2599 | { | |
2600 | rtx insn; | |
2601 | rtx next_tail = NEXT_INSN (tail); | |
2602 | ||
2603 | fprintf (sched_dump, ";; %7s%6s%6s%6s%6s%6s%14s\n", | |
2604 | "insn", "code", "bb", "dep", "prio", "cost", | |
2605 | "reservation"); | |
2606 | fprintf (sched_dump, ";; %7s%6s%6s%6s%6s%6s%14s\n", | |
2607 | "----", "----", "--", "---", "----", "----", | |
2608 | "-----------"); | |
b4ead7d4 | 2609 | |
b640bd8f MK |
2610 | for (insn = head; insn != next_tail; insn = NEXT_INSN (insn)) |
2611 | { | |
b640bd8f MK |
2612 | if (! INSN_P (insn)) |
2613 | { | |
2614 | int n; | |
2615 | fprintf (sched_dump, ";; %6d ", INSN_UID (insn)); | |
2616 | if (NOTE_P (insn)) | |
b4ead7d4 | 2617 | { |
a38e7aa5 JH |
2618 | n = NOTE_KIND (insn); |
2619 | fprintf (sched_dump, "%s\n", GET_NOTE_INSN_NAME (n)); | |
b4ead7d4 | 2620 | } |
fa0aee89 | 2621 | else |
b640bd8f MK |
2622 | fprintf (sched_dump, " {%s}\n", GET_RTX_NAME (GET_CODE (insn))); |
2623 | continue; | |
b4ead7d4 | 2624 | } |
b640bd8f MK |
2625 | |
2626 | fprintf (sched_dump, | |
2627 | ";; %s%5d%6d%6d%6d%6d%6d ", | |
2628 | (SCHED_GROUP_P (insn) ? "+" : " "), | |
2629 | INSN_UID (insn), | |
2630 | INSN_CODE (insn), | |
2631 | BLOCK_NUM (insn), | |
e2f6ff94 | 2632 | sd_lists_size (insn, SD_LIST_BACK), |
b640bd8f MK |
2633 | INSN_PRIORITY (insn), |
2634 | insn_cost (insn)); | |
2635 | ||
2636 | if (recog_memoized (insn) < 0) | |
2637 | fprintf (sched_dump, "nothing"); | |
2638 | else | |
2639 | print_reservation (sched_dump, insn); | |
2640 | ||
2641 | fprintf (sched_dump, "\t: "); | |
e2f6ff94 MK |
2642 | { |
2643 | sd_iterator_def sd_it; | |
2644 | dep_t dep; | |
2645 | ||
2646 | FOR_EACH_DEP (insn, SD_LIST_FORW, sd_it, dep) | |
2647 | fprintf (sched_dump, "%d ", INSN_UID (DEP_CON (dep))); | |
2648 | } | |
b640bd8f | 2649 | fprintf (sched_dump, "\n"); |
b4ead7d4 | 2650 | } |
b640bd8f | 2651 | |
b4ead7d4 BS |
2652 | fprintf (sched_dump, "\n"); |
2653 | } | |
2654 | \f | |
d72372e4 MH |
2655 | /* Returns true if all the basic blocks of the current region have |
2656 | NOTE_DISABLE_SCHED_OF_BLOCK which means not to schedule that region. */ | |
2657 | static bool | |
2658 | sched_is_disabled_for_current_region_p (void) | |
2659 | { | |
d72372e4 MH |
2660 | int bb; |
2661 | ||
2662 | for (bb = 0; bb < current_nr_blocks; bb++) | |
076c7ab8 ZW |
2663 | if (!(BASIC_BLOCK (BB_TO_BLOCK (bb))->flags & BB_DISABLE_SCHEDULE)) |
2664 | return false; | |
d72372e4 MH |
2665 | |
2666 | return true; | |
2667 | } | |
2668 | ||
b4ead7d4 BS |
2669 | /* Schedule a region. A region is either an inner loop, a loop-free |
2670 | subroutine, or a single basic block. Each bb in the region is | |
2671 | scheduled after its flow predecessors. */ | |
2672 | ||
2673 | static void | |
46c5ad27 | 2674 | schedule_region (int rgn) |
b4ead7d4 | 2675 | { |
dcda8480 UW |
2676 | basic_block block; |
2677 | edge_iterator ei; | |
2678 | edge e; | |
b4ead7d4 | 2679 | int bb; |
b4ead7d4 BS |
2680 | int sched_rgn_n_insns = 0; |
2681 | ||
496d7bb0 | 2682 | rgn_n_insns = 0; |
b4ead7d4 BS |
2683 | /* Set variables for the current region. */ |
2684 | current_nr_blocks = RGN_NR_BLOCKS (rgn); | |
2685 | current_blocks = RGN_BLOCKS (rgn); | |
496d7bb0 MK |
2686 | |
2687 | /* See comments in add_block1, for what reasons we allocate +1 element. */ | |
2688 | ebb_head = xrealloc (ebb_head, (current_nr_blocks + 1) * sizeof (*ebb_head)); | |
2689 | for (bb = 0; bb <= current_nr_blocks; bb++) | |
2690 | ebb_head[bb] = current_blocks + bb; | |
b4ead7d4 | 2691 | |
d72372e4 MH |
2692 | /* Don't schedule region that is marked by |
2693 | NOTE_DISABLE_SCHED_OF_BLOCK. */ | |
2694 | if (sched_is_disabled_for_current_region_p ()) | |
2695 | return; | |
2696 | ||
496d7bb0 MK |
2697 | if (!RGN_DONT_CALC_DEPS (rgn)) |
2698 | { | |
2699 | init_deps_global (); | |
b4ead7d4 | 2700 | |
496d7bb0 MK |
2701 | /* Initializations for region data dependence analysis. */ |
2702 | bb_deps = XNEWVEC (struct deps, current_nr_blocks); | |
2703 | for (bb = 0; bb < current_nr_blocks; bb++) | |
2704 | init_deps (bb_deps + bb); | |
b4ead7d4 | 2705 | |
e2f6ff94 | 2706 | /* Compute dependencies. */ |
496d7bb0 | 2707 | for (bb = 0; bb < current_nr_blocks; bb++) |
e2f6ff94 | 2708 | compute_block_dependences (bb); |
b4ead7d4 | 2709 | |
496d7bb0 MK |
2710 | free_pending_lists (); |
2711 | ||
2712 | finish_deps_global (); | |
2713 | ||
2714 | free (bb_deps); | |
2715 | } | |
2716 | else | |
2717 | /* This is a recovery block. It is always a single block region. */ | |
2718 | gcc_assert (current_nr_blocks == 1); | |
2719 | ||
b4ead7d4 | 2720 | /* Set priorities. */ |
63f54b1a | 2721 | current_sched_info->sched_max_insns_priority = 0; |
b4ead7d4 | 2722 | for (bb = 0; bb < current_nr_blocks; bb++) |
79c2ffde BS |
2723 | { |
2724 | rtx head, tail; | |
496d7bb0 MK |
2725 | |
2726 | gcc_assert (EBB_FIRST_BB (bb) == EBB_LAST_BB (bb)); | |
2727 | get_ebb_head_tail (EBB_FIRST_BB (bb), EBB_LAST_BB (bb), &head, &tail); | |
79c2ffde BS |
2728 | |
2729 | rgn_n_insns += set_priorities (head, tail); | |
2730 | } | |
63f54b1a | 2731 | current_sched_info->sched_max_insns_priority++; |
b4ead7d4 BS |
2732 | |
2733 | /* Compute interblock info: probabilities, split-edges, dominators, etc. */ | |
2734 | if (current_nr_blocks > 1) | |
2735 | { | |
36968131 | 2736 | prob = XNEWVEC (int, current_nr_blocks); |
b4ead7d4 | 2737 | |
bdfa170f DB |
2738 | dom = sbitmap_vector_alloc (current_nr_blocks, current_nr_blocks); |
2739 | sbitmap_vector_zero (dom, current_nr_blocks); | |
dcda8480 UW |
2740 | |
2741 | /* Use ->aux to implement EDGE_TO_BIT mapping. */ | |
b4ead7d4 | 2742 | rgn_nr_edges = 0; |
dcda8480 UW |
2743 | FOR_EACH_BB (block) |
2744 | { | |
2745 | if (CONTAINING_RGN (block->index) != rgn) | |
2746 | continue; | |
2747 | FOR_EACH_EDGE (e, ei, block->succs) | |
2748 | SET_EDGE_TO_BIT (e, rgn_nr_edges++); | |
2749 | } | |
b4ead7d4 | 2750 | |
5ed6ace5 | 2751 | rgn_edges = XNEWVEC (edge, rgn_nr_edges); |
b4ead7d4 | 2752 | rgn_nr_edges = 0; |
dcda8480 UW |
2753 | FOR_EACH_BB (block) |
2754 | { | |
2755 | if (CONTAINING_RGN (block->index) != rgn) | |
2756 | continue; | |
2757 | FOR_EACH_EDGE (e, ei, block->succs) | |
2758 | rgn_edges[rgn_nr_edges++] = e; | |
2759 | } | |
b4ead7d4 BS |
2760 | |
2761 | /* Split edges. */ | |
bdfa170f DB |
2762 | pot_split = sbitmap_vector_alloc (current_nr_blocks, rgn_nr_edges); |
2763 | sbitmap_vector_zero (pot_split, current_nr_blocks); | |
2764 | ancestor_edges = sbitmap_vector_alloc (current_nr_blocks, rgn_nr_edges); | |
2765 | sbitmap_vector_zero (ancestor_edges, current_nr_blocks); | |
b4ead7d4 BS |
2766 | |
2767 | /* Compute probabilities, dominators, split_edges. */ | |
2768 | for (bb = 0; bb < current_nr_blocks; bb++) | |
2769 | compute_dom_prob_ps (bb); | |
496d7bb0 MK |
2770 | |
2771 | /* Cleanup ->aux used for EDGE_TO_BIT mapping. */ | |
917f1b7e | 2772 | /* We don't need them anymore. But we want to avoid duplication of |
496d7bb0 MK |
2773 | aux fields in the newly created edges. */ |
2774 | FOR_EACH_BB (block) | |
2775 | { | |
2776 | if (CONTAINING_RGN (block->index) != rgn) | |
2777 | continue; | |
2778 | FOR_EACH_EDGE (e, ei, block->succs) | |
2779 | e->aux = NULL; | |
2780 | } | |
b4ead7d4 BS |
2781 | } |
2782 | ||
2783 | /* Now we can schedule all blocks. */ | |
2784 | for (bb = 0; bb < current_nr_blocks; bb++) | |
2785 | { | |
496d7bb0 | 2786 | basic_block first_bb, last_bb, curr_bb; |
b4ead7d4 | 2787 | rtx head, tail; |
b4ead7d4 | 2788 | |
496d7bb0 MK |
2789 | first_bb = EBB_FIRST_BB (bb); |
2790 | last_bb = EBB_LAST_BB (bb); | |
2791 | ||
2792 | get_ebb_head_tail (first_bb, last_bb, &head, &tail); | |
b4ead7d4 BS |
2793 | |
2794 | if (no_real_insns_p (head, tail)) | |
496d7bb0 MK |
2795 | { |
2796 | gcc_assert (first_bb == last_bb); | |
2797 | continue; | |
2798 | } | |
b4ead7d4 BS |
2799 | |
2800 | current_sched_info->prev_head = PREV_INSN (head); | |
2801 | current_sched_info->next_tail = NEXT_INSN (tail); | |
2802 | ||
b4ead7d4 BS |
2803 | |
2804 | /* rm_other_notes only removes notes which are _inside_ the | |
2805 | block---that is, it won't remove notes before the first real insn | |
46c5ad27 | 2806 | or after the last real insn of the block. So if the first insn |
b4ead7d4 BS |
2807 | has a REG_SAVE_NOTE which would otherwise be emitted before the |
2808 | insn, it is redundant with the note before the start of the | |
570a98eb | 2809 | block, and so we have to take it out. */ |
b4ead7d4 BS |
2810 | if (INSN_P (head)) |
2811 | { | |
2812 | rtx note; | |
2813 | ||
2814 | for (note = REG_NOTES (head); note; note = XEXP (note, 1)) | |
2815 | if (REG_NOTE_KIND (note) == REG_SAVE_NOTE) | |
668707f7 | 2816 | remove_note (head, note); |
b4ead7d4 | 2817 | } |
496d7bb0 MK |
2818 | else |
2819 | /* This means that first block in ebb is empty. | |
2820 | It looks to me as an impossible thing. There at least should be | |
2821 | a recovery check, that caused the splitting. */ | |
2822 | gcc_unreachable (); | |
b4ead7d4 BS |
2823 | |
2824 | /* Remove remaining note insns from the block, save them in | |
2825 | note_list. These notes are restored at the end of | |
2826 | schedule_block (). */ | |
2827 | rm_other_notes (head, tail); | |
2828 | ||
496d7bb0 MK |
2829 | unlink_bb_notes (first_bb, last_bb); |
2830 | ||
b4ead7d4 BS |
2831 | target_bb = bb; |
2832 | ||
63f54b1a MK |
2833 | gcc_assert (flag_schedule_interblock || current_nr_blocks == 1); |
2834 | current_sched_info->queue_must_finish_empty = current_nr_blocks == 1; | |
b4ead7d4 | 2835 | |
496d7bb0 | 2836 | curr_bb = first_bb; |
6fb5fa3c DB |
2837 | if (dbg_cnt (sched_block)) |
2838 | { | |
2839 | schedule_block (&curr_bb, rgn_n_insns); | |
2840 | gcc_assert (EBB_FIRST_BB (bb) == first_bb); | |
2841 | sched_rgn_n_insns += sched_n_insns; | |
2842 | } | |
2843 | else | |
2844 | { | |
2845 | sched_rgn_n_insns += rgn_n_insns; | |
2846 | } | |
b4ead7d4 | 2847 | |
b4ead7d4 BS |
2848 | /* Clean up. */ |
2849 | if (current_nr_blocks > 1) | |
2850 | { | |
2851 | free (candidate_table); | |
2852 | free (bblst_table); | |
dcda8480 | 2853 | free (edgelst_table); |
b4ead7d4 BS |
2854 | } |
2855 | } | |
2856 | ||
2857 | /* Sanity check: verify that all region insns were scheduled. */ | |
41374e13 | 2858 | gcc_assert (sched_rgn_n_insns == rgn_n_insns); |
b4ead7d4 | 2859 | |
b4ead7d4 | 2860 | /* Done with this region. */ |
b4ead7d4 BS |
2861 | |
2862 | if (current_nr_blocks > 1) | |
2863 | { | |
b4ead7d4 | 2864 | free (prob); |
bdfa170f DB |
2865 | sbitmap_vector_free (dom); |
2866 | sbitmap_vector_free (pot_split); | |
2867 | sbitmap_vector_free (ancestor_edges); | |
b4ead7d4 | 2868 | free (rgn_edges); |
b4ead7d4 | 2869 | } |
e2f6ff94 MK |
2870 | |
2871 | /* Free dependencies. */ | |
2872 | for (bb = 0; bb < current_nr_blocks; ++bb) | |
2873 | free_block_dependencies (bb); | |
2874 | ||
2875 | gcc_assert (haifa_recovery_bb_ever_added_p | |
2876 | || deps_pools_are_empty_p ()); | |
b4ead7d4 BS |
2877 | } |
2878 | ||
b4ead7d4 BS |
2879 | /* Initialize data structures for region scheduling. */ |
2880 | ||
2881 | static void | |
46c5ad27 | 2882 | init_regions (void) |
b4ead7d4 | 2883 | { |
b4ead7d4 | 2884 | nr_regions = 0; |
496d7bb0 MK |
2885 | rgn_table = 0; |
2886 | rgn_bb_table = 0; | |
2887 | block_to_bb = 0; | |
2888 | containing_rgn = 0; | |
2889 | extend_regions (); | |
b4ead7d4 | 2890 | |
b4ead7d4 BS |
2891 | /* Compute regions for scheduling. */ |
2892 | if (reload_completed | |
24bd1a0b | 2893 | || n_basic_blocks == NUM_FIXED_BLOCKS + 1 |
dcda8480 UW |
2894 | || !flag_schedule_interblock |
2895 | || is_cfg_nonregular ()) | |
b4ead7d4 BS |
2896 | { |
2897 | find_single_block_region (); | |
2898 | } | |
2899 | else | |
2900 | { | |
dcda8480 UW |
2901 | /* Compute the dominators and post dominators. */ |
2902 | calculate_dominance_info (CDI_DOMINATORS); | |
b4ead7d4 | 2903 | |
dcda8480 UW |
2904 | /* Find regions. */ |
2905 | find_rgns (); | |
b4ead7d4 | 2906 | |
dcda8480 UW |
2907 | if (sched_verbose >= 3) |
2908 | debug_regions (); | |
b4ead7d4 | 2909 | |
dcda8480 | 2910 | /* For now. This will move as more and more of haifa is converted |
6fb5fa3c | 2911 | to using the cfg code. */ |
dcda8480 | 2912 | free_dominance_info (CDI_DOMINATORS); |
b4ead7d4 | 2913 | } |
496d7bb0 MK |
2914 | RGN_BLOCKS (nr_regions) = RGN_BLOCKS (nr_regions - 1) + |
2915 | RGN_NR_BLOCKS (nr_regions - 1); | |
b4ead7d4 BS |
2916 | } |
2917 | ||
10d22567 | 2918 | /* The one entry point in this file. */ |
b4ead7d4 BS |
2919 | |
2920 | void | |
10d22567 | 2921 | schedule_insns (void) |
b4ead7d4 | 2922 | { |
b4ead7d4 | 2923 | int rgn; |
b4ead7d4 BS |
2924 | |
2925 | /* Taking care of this degenerate case makes the rest of | |
2926 | this code simpler. */ | |
24bd1a0b | 2927 | if (n_basic_blocks == NUM_FIXED_BLOCKS) |
b4ead7d4 BS |
2928 | return; |
2929 | ||
2930 | nr_inter = 0; | |
2931 | nr_spec = 0; | |
ddbd5439 MK |
2932 | |
2933 | /* We need current_sched_info in init_dependency_caches, which is | |
2934 | invoked via sched_init. */ | |
2935 | current_sched_info = ®ion_sched_info; | |
2936 | ||
6fb5fa3c DB |
2937 | df_set_flags (DF_LR_RUN_DCE); |
2938 | df_note_add_problem (); | |
2939 | df_analyze (); | |
2940 | regstat_compute_calls_crossed (); | |
2941 | ||
10d22567 | 2942 | sched_init (); |
b4ead7d4 | 2943 | |
6fb5fa3c DB |
2944 | bitmap_initialize (¬_in_df, 0); |
2945 | bitmap_clear (¬_in_df); | |
2946 | ||
36968131 PS |
2947 | min_spec_prob = ((PARAM_VALUE (PARAM_MIN_SPEC_PROB) * REG_BR_PROB_BASE) |
2948 | / 100); | |
2949 | ||
b4ead7d4 BS |
2950 | init_regions (); |
2951 | ||
917f1b7e | 2952 | /* EBB_HEAD is a region-scope structure. But we realloc it for |
496d7bb0 MK |
2953 | each region to save time/memory/something else. */ |
2954 | ebb_head = 0; | |
2955 | ||
b4ead7d4 BS |
2956 | /* Schedule every region in the subroutine. */ |
2957 | for (rgn = 0; rgn < nr_regions; rgn++) | |
6fb5fa3c DB |
2958 | if (dbg_cnt (sched_region)) |
2959 | schedule_region (rgn); | |
496d7bb0 MK |
2960 | |
2961 | free(ebb_head); | |
b4ead7d4 BS |
2962 | /* Reposition the prologue and epilogue notes in case we moved the |
2963 | prologue/epilogue insns. */ | |
2964 | if (reload_completed) | |
6fb5fa3c | 2965 | reposition_prologue_and_epilogue_notes (); |
b4ead7d4 | 2966 | |
b4ead7d4 BS |
2967 | if (sched_verbose) |
2968 | { | |
2969 | if (reload_completed == 0 && flag_schedule_interblock) | |
2970 | { | |
2971 | fprintf (sched_dump, | |
2972 | "\n;; Procedure interblock/speculative motions == %d/%d \n", | |
2973 | nr_inter, nr_spec); | |
2974 | } | |
2975 | else | |
41374e13 | 2976 | gcc_assert (nr_inter <= 0); |
b4ead7d4 BS |
2977 | fprintf (sched_dump, "\n\n"); |
2978 | } | |
2979 | ||
2980 | /* Clean up. */ | |
2981 | free (rgn_table); | |
2982 | free (rgn_bb_table); | |
2983 | free (block_to_bb); | |
2984 | free (containing_rgn); | |
2985 | ||
6fb5fa3c DB |
2986 | regstat_free_calls_crossed (); |
2987 | ||
2988 | bitmap_clear (¬_in_df); | |
b4ead7d4 | 2989 | |
6fb5fa3c | 2990 | sched_finish (); |
b4ead7d4 | 2991 | } |
496d7bb0 MK |
2992 | |
2993 | /* INSN has been added to/removed from current region. */ | |
2994 | static void | |
2995 | add_remove_insn (rtx insn, int remove_p) | |
2996 | { | |
2997 | if (!remove_p) | |
2998 | rgn_n_insns++; | |
2999 | else | |
3000 | rgn_n_insns--; | |
3001 | ||
3002 | if (INSN_BB (insn) == target_bb) | |
3003 | { | |
3004 | if (!remove_p) | |
3005 | target_n_insns++; | |
3006 | else | |
3007 | target_n_insns--; | |
3008 | } | |
3009 | } | |
3010 | ||
3011 | /* Extend internal data structures. */ | |
3012 | static void | |
3013 | extend_regions (void) | |
3014 | { | |
3015 | rgn_table = XRESIZEVEC (region, rgn_table, n_basic_blocks); | |
3016 | rgn_bb_table = XRESIZEVEC (int, rgn_bb_table, n_basic_blocks); | |
3017 | block_to_bb = XRESIZEVEC (int, block_to_bb, last_basic_block); | |
3018 | containing_rgn = XRESIZEVEC (int, containing_rgn, last_basic_block); | |
3019 | } | |
3020 | ||
3021 | /* BB was added to ebb after AFTER. */ | |
3022 | static void | |
3023 | add_block1 (basic_block bb, basic_block after) | |
3024 | { | |
3025 | extend_regions (); | |
3026 | ||
6fb5fa3c DB |
3027 | bitmap_set_bit (¬_in_df, bb->index); |
3028 | ||
496d7bb0 MK |
3029 | if (after == 0 || after == EXIT_BLOCK_PTR) |
3030 | { | |
3031 | int i; | |
3032 | ||
3033 | i = RGN_BLOCKS (nr_regions); | |
3034 | /* I - first free position in rgn_bb_table. */ | |
3035 | ||
3036 | rgn_bb_table[i] = bb->index; | |
3037 | RGN_NR_BLOCKS (nr_regions) = 1; | |
3038 | RGN_DONT_CALC_DEPS (nr_regions) = after == EXIT_BLOCK_PTR; | |
3039 | RGN_HAS_REAL_EBB (nr_regions) = 0; | |
3040 | CONTAINING_RGN (bb->index) = nr_regions; | |
3041 | BLOCK_TO_BB (bb->index) = 0; | |
3042 | ||
3043 | nr_regions++; | |
3044 | ||
3045 | RGN_BLOCKS (nr_regions) = i + 1; | |
496d7bb0 MK |
3046 | } |
3047 | else | |
3048 | { | |
3049 | int i, pos; | |
3050 | ||
3051 | /* We need to fix rgn_table, block_to_bb, containing_rgn | |
3052 | and ebb_head. */ | |
3053 | ||
3054 | BLOCK_TO_BB (bb->index) = BLOCK_TO_BB (after->index); | |
3055 | ||
3056 | /* We extend ebb_head to one more position to | |
3057 | easily find the last position of the last ebb in | |
3058 | the current region. Thus, ebb_head[BLOCK_TO_BB (after) + 1] | |
3059 | is _always_ valid for access. */ | |
3060 | ||
3061 | i = BLOCK_TO_BB (after->index) + 1; | |
1d49ee6a MK |
3062 | pos = ebb_head[i] - 1; |
3063 | /* Now POS is the index of the last block in the region. */ | |
3064 | ||
3065 | /* Find index of basic block AFTER. */ | |
3066 | for (; rgn_bb_table[pos] != after->index; pos--); | |
3067 | ||
496d7bb0 MK |
3068 | pos++; |
3069 | gcc_assert (pos > ebb_head[i - 1]); | |
1d49ee6a | 3070 | |
496d7bb0 MK |
3071 | /* i - ebb right after "AFTER". */ |
3072 | /* ebb_head[i] - VALID. */ | |
3073 | ||
3074 | /* Source position: ebb_head[i] | |
917f1b7e | 3075 | Destination position: ebb_head[i] + 1 |
496d7bb0 MK |
3076 | Last position: |
3077 | RGN_BLOCKS (nr_regions) - 1 | |
3078 | Number of elements to copy: (last_position) - (source_position) + 1 | |
3079 | */ | |
3080 | ||
3081 | memmove (rgn_bb_table + pos + 1, | |
3082 | rgn_bb_table + pos, | |
3083 | ((RGN_BLOCKS (nr_regions) - 1) - (pos) + 1) | |
3084 | * sizeof (*rgn_bb_table)); | |
3085 | ||
3086 | rgn_bb_table[pos] = bb->index; | |
3087 | ||
3088 | for (; i <= current_nr_blocks; i++) | |
3089 | ebb_head [i]++; | |
3090 | ||
3091 | i = CONTAINING_RGN (after->index); | |
3092 | CONTAINING_RGN (bb->index) = i; | |
3093 | ||
3094 | RGN_HAS_REAL_EBB (i) = 1; | |
3095 | ||
3096 | for (++i; i <= nr_regions; i++) | |
3097 | RGN_BLOCKS (i)++; | |
496d7bb0 MK |
3098 | } |
3099 | } | |
3100 | ||
3101 | /* Fix internal data after interblock movement of jump instruction. | |
3102 | For parameter meaning please refer to | |
3103 | sched-int.h: struct sched_info: fix_recovery_cfg. */ | |
3104 | static void | |
3105 | fix_recovery_cfg (int bbi, int check_bbi, int check_bb_nexti) | |
3106 | { | |
3107 | int old_pos, new_pos, i; | |
3108 | ||
3109 | BLOCK_TO_BB (check_bb_nexti) = BLOCK_TO_BB (bbi); | |
3110 | ||
3111 | for (old_pos = ebb_head[BLOCK_TO_BB (check_bbi) + 1] - 1; | |
3112 | rgn_bb_table[old_pos] != check_bb_nexti; | |
3113 | old_pos--); | |
3114 | gcc_assert (old_pos > ebb_head[BLOCK_TO_BB (check_bbi)]); | |
3115 | ||
3116 | for (new_pos = ebb_head[BLOCK_TO_BB (bbi) + 1] - 1; | |
3117 | rgn_bb_table[new_pos] != bbi; | |
3118 | new_pos--); | |
3119 | new_pos++; | |
3120 | gcc_assert (new_pos > ebb_head[BLOCK_TO_BB (bbi)]); | |
3121 | ||
3122 | gcc_assert (new_pos < old_pos); | |
3123 | ||
3124 | memmove (rgn_bb_table + new_pos + 1, | |
3125 | rgn_bb_table + new_pos, | |
3126 | (old_pos - new_pos) * sizeof (*rgn_bb_table)); | |
3127 | ||
3128 | rgn_bb_table[new_pos] = check_bb_nexti; | |
3129 | ||
3130 | for (i = BLOCK_TO_BB (bbi) + 1; i <= BLOCK_TO_BB (check_bbi); i++) | |
3131 | ebb_head[i]++; | |
3132 | } | |
3133 | ||
3134 | /* Return next block in ebb chain. For parameter meaning please refer to | |
3135 | sched-int.h: struct sched_info: advance_target_bb. */ | |
3136 | static basic_block | |
3137 | advance_target_bb (basic_block bb, rtx insn) | |
3138 | { | |
3139 | if (insn) | |
3140 | return 0; | |
3141 | ||
3142 | gcc_assert (BLOCK_TO_BB (bb->index) == target_bb | |
3143 | && BLOCK_TO_BB (bb->next_bb->index) == target_bb); | |
3144 | return bb->next_bb; | |
3145 | } | |
3146 | ||
f56887a7 | 3147 | #endif |
ef330312 PB |
3148 | \f |
3149 | static bool | |
3150 | gate_handle_sched (void) | |
3151 | { | |
3152 | #ifdef INSN_SCHEDULING | |
6fb5fa3c | 3153 | return flag_schedule_insns && dbg_cnt (sched_func); |
ef330312 PB |
3154 | #else |
3155 | return 0; | |
3156 | #endif | |
3157 | } | |
3158 | ||
3159 | /* Run instruction scheduler. */ | |
c2924966 | 3160 | static unsigned int |
ef330312 PB |
3161 | rest_of_handle_sched (void) |
3162 | { | |
3163 | #ifdef INSN_SCHEDULING | |
10d22567 | 3164 | schedule_insns (); |
ef330312 | 3165 | #endif |
c2924966 | 3166 | return 0; |
ef330312 PB |
3167 | } |
3168 | ||
3169 | static bool | |
3170 | gate_handle_sched2 (void) | |
3171 | { | |
3172 | #ifdef INSN_SCHEDULING | |
6fb5fa3c DB |
3173 | return optimize > 0 && flag_schedule_insns_after_reload |
3174 | && dbg_cnt (sched2_func); | |
ef330312 PB |
3175 | #else |
3176 | return 0; | |
3177 | #endif | |
3178 | } | |
3179 | ||
3180 | /* Run second scheduling pass after reload. */ | |
c2924966 | 3181 | static unsigned int |
ef330312 PB |
3182 | rest_of_handle_sched2 (void) |
3183 | { | |
3184 | #ifdef INSN_SCHEDULING | |
3185 | /* Do control and data sched analysis again, | |
3186 | and write some more of the results to dump file. */ | |
ef330312 | 3187 | if (flag_sched2_use_superblocks || flag_sched2_use_traces) |
6fb5fa3c | 3188 | schedule_ebbs (); |
ef330312 | 3189 | else |
10d22567 | 3190 | schedule_insns (); |
ef330312 | 3191 | #endif |
c2924966 | 3192 | return 0; |
ef330312 PB |
3193 | } |
3194 | ||
3195 | struct tree_opt_pass pass_sched = | |
3196 | { | |
3197 | "sched1", /* name */ | |
3198 | gate_handle_sched, /* gate */ | |
3199 | rest_of_handle_sched, /* execute */ | |
3200 | NULL, /* sub */ | |
3201 | NULL, /* next */ | |
3202 | 0, /* static_pass_number */ | |
3203 | TV_SCHED, /* tv_id */ | |
3204 | 0, /* properties_required */ | |
3205 | 0, /* properties_provided */ | |
3206 | 0, /* properties_destroyed */ | |
3207 | 0, /* todo_flags_start */ | |
a36b8a1e | 3208 | TODO_df_finish | TODO_verify_rtl_sharing | |
ef330312 | 3209 | TODO_dump_func | |
6fb5fa3c | 3210 | TODO_verify_flow | |
ef330312 PB |
3211 | TODO_ggc_collect, /* todo_flags_finish */ |
3212 | 'S' /* letter */ | |
3213 | }; | |
3214 | ||
3215 | struct tree_opt_pass pass_sched2 = | |
3216 | { | |
3217 | "sched2", /* name */ | |
3218 | gate_handle_sched2, /* gate */ | |
3219 | rest_of_handle_sched2, /* execute */ | |
3220 | NULL, /* sub */ | |
3221 | NULL, /* next */ | |
3222 | 0, /* static_pass_number */ | |
3223 | TV_SCHED2, /* tv_id */ | |
3224 | 0, /* properties_required */ | |
3225 | 0, /* properties_provided */ | |
3226 | 0, /* properties_destroyed */ | |
3227 | 0, /* todo_flags_start */ | |
a36b8a1e | 3228 | TODO_df_finish | TODO_verify_rtl_sharing | |
ef330312 | 3229 | TODO_dump_func | |
6fb5fa3c | 3230 | TODO_verify_flow | |
ef330312 PB |
3231 | TODO_ggc_collect, /* todo_flags_finish */ |
3232 | 'R' /* letter */ | |
3233 | }; | |
3234 |