]>
Commit | Line | Data |
---|---|---|
610d2478 | 1 | /* CPU mode switching |
23a5b65a | 2 | Copyright (C) 1998-2014 Free Software Foundation, Inc. |
610d2478 SB |
3 | |
4 | This file is part of GCC. | |
5 | ||
6 | GCC is free software; you can redistribute it and/or modify it under | |
7 | the terms of the GNU General Public License as published by the Free | |
9dcd6f09 | 8 | Software Foundation; either version 3, or (at your option) any later |
610d2478 SB |
9 | version. |
10 | ||
11 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
12 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
13 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
14 | for more details. | |
15 | ||
16 | You should have received a copy of the GNU General Public License | |
9dcd6f09 NC |
17 | along with GCC; see the file COPYING3. If not see |
18 | <http://www.gnu.org/licenses/>. */ | |
610d2478 SB |
19 | |
20 | #include "config.h" | |
21 | #include "system.h" | |
22 | #include "coretypes.h" | |
23 | #include "tm.h" | |
82f81f18 | 24 | #include "target.h" |
610d2478 SB |
25 | #include "rtl.h" |
26 | #include "regs.h" | |
27 | #include "hard-reg-set.h" | |
28 | #include "flags.h" | |
610d2478 SB |
29 | #include "insn-config.h" |
30 | #include "recog.h" | |
31 | #include "basic-block.h" | |
610d2478 SB |
32 | #include "tm_p.h" |
33 | #include "function.h" | |
ef330312 | 34 | #include "tree-pass.h" |
6fb5fa3c | 35 | #include "df.h" |
5936d944 | 36 | #include "emit-rtl.h" |
610d2478 SB |
37 | |
38 | /* We want target macros for the mode switching code to be able to refer | |
39 | to instruction attribute values. */ | |
40 | #include "insn-attr.h" | |
41 | ||
42 | #ifdef OPTIMIZE_MODE_SWITCHING | |
43 | ||
44 | /* The algorithm for setting the modes consists of scanning the insn list | |
45 | and finding all the insns which require a specific mode. Each insn gets | |
46 | a unique struct seginfo element. These structures are inserted into a list | |
47 | for each basic block. For each entity, there is an array of bb_info over | |
cee9defb | 48 | the flow graph basic blocks (local var 'bb_info'), which contains a list |
610d2478 SB |
49 | of all insns within that basic block, in the order they are encountered. |
50 | ||
51 | For each entity, any basic block WITHOUT any insns requiring a specific | |
cee9defb EB |
52 | mode are given a single entry without a mode (each basic block in the |
53 | flow graph must have at least one entry in the segment table). | |
610d2478 SB |
54 | |
55 | The LCM algorithm is then run over the flow graph to determine where to | |
cee9defb | 56 | place the sets to the highest-priority mode with respect to the first |
610d2478 SB |
57 | insn in any one block. Any adjustments required to the transparency |
58 | vectors are made, then the next iteration starts for the next-lower | |
59 | priority mode, till for each entity all modes are exhausted. | |
60 | ||
cee9defb | 61 | More details can be found in the code of optimize_mode_switching. */ |
610d2478 SB |
62 | \f |
63 | /* This structure contains the information for each insn which requires | |
64 | either single or double mode to be set. | |
65 | MODE is the mode this insn must be executed in. | |
66 | INSN_PTR is the insn to be executed (may be the note that marks the | |
67 | beginning of a basic block). | |
68 | BBNUM is the flow graph basic block this insn occurs in. | |
69 | NEXT is the next insn in the same basic block. */ | |
70 | struct seginfo | |
71 | { | |
72 | int mode; | |
73 | rtx insn_ptr; | |
74 | int bbnum; | |
75 | struct seginfo *next; | |
76 | HARD_REG_SET regs_live; | |
77 | }; | |
78 | ||
79 | struct bb_info | |
80 | { | |
81 | struct seginfo *seginfo; | |
82 | int computing; | |
83 | }; | |
84 | ||
85 | /* These bitmaps are used for the LCM algorithm. */ | |
86 | ||
87 | static sbitmap *antic; | |
88 | static sbitmap *transp; | |
89 | static sbitmap *comp; | |
90 | ||
91 | static struct seginfo * new_seginfo (int, rtx, int, HARD_REG_SET); | |
92 | static void add_seginfo (struct bb_info *, struct seginfo *); | |
408bed3c | 93 | static void reg_dies (rtx, HARD_REG_SET *); |
7bc980e1 | 94 | static void reg_becomes_live (rtx, const_rtx, void *); |
610d2478 SB |
95 | static void make_preds_opaque (basic_block, int); |
96 | \f | |
97 | ||
98 | /* This function will allocate a new BBINFO structure, initialized | |
99 | with the MODE, INSN, and basic block BB parameters. */ | |
100 | ||
101 | static struct seginfo * | |
102 | new_seginfo (int mode, rtx insn, int bb, HARD_REG_SET regs_live) | |
103 | { | |
104 | struct seginfo *ptr; | |
5ed6ace5 | 105 | ptr = XNEW (struct seginfo); |
610d2478 SB |
106 | ptr->mode = mode; |
107 | ptr->insn_ptr = insn; | |
108 | ptr->bbnum = bb; | |
109 | ptr->next = NULL; | |
110 | COPY_HARD_REG_SET (ptr->regs_live, regs_live); | |
111 | return ptr; | |
112 | } | |
113 | ||
114 | /* Add a seginfo element to the end of a list. | |
115 | HEAD is a pointer to the list beginning. | |
116 | INFO is the structure to be linked in. */ | |
117 | ||
118 | static void | |
119 | add_seginfo (struct bb_info *head, struct seginfo *info) | |
120 | { | |
121 | struct seginfo *ptr; | |
122 | ||
123 | if (head->seginfo == NULL) | |
124 | head->seginfo = info; | |
125 | else | |
126 | { | |
127 | ptr = head->seginfo; | |
128 | while (ptr->next != NULL) | |
129 | ptr = ptr->next; | |
130 | ptr->next = info; | |
131 | } | |
132 | } | |
133 | ||
134 | /* Make all predecessors of basic block B opaque, recursively, till we hit | |
135 | some that are already non-transparent, or an edge where aux is set; that | |
136 | denotes that a mode set is to be done on that edge. | |
137 | J is the bit number in the bitmaps that corresponds to the entity that | |
138 | we are currently handling mode-switching for. */ | |
139 | ||
140 | static void | |
141 | make_preds_opaque (basic_block b, int j) | |
142 | { | |
143 | edge e; | |
144 | edge_iterator ei; | |
145 | ||
146 | FOR_EACH_EDGE (e, ei, b->preds) | |
147 | { | |
148 | basic_block pb = e->src; | |
149 | ||
d7c028c0 | 150 | if (e->aux || ! bitmap_bit_p (transp[pb->index], j)) |
610d2478 SB |
151 | continue; |
152 | ||
d7c028c0 | 153 | bitmap_clear_bit (transp[pb->index], j); |
610d2478 SB |
154 | make_preds_opaque (pb, j); |
155 | } | |
156 | } | |
157 | ||
158 | /* Record in LIVE that register REG died. */ | |
159 | ||
160 | static void | |
408bed3c | 161 | reg_dies (rtx reg, HARD_REG_SET *live) |
610d2478 | 162 | { |
09e18274 | 163 | int regno; |
610d2478 SB |
164 | |
165 | if (!REG_P (reg)) | |
166 | return; | |
167 | ||
168 | regno = REGNO (reg); | |
169 | if (regno < FIRST_PSEUDO_REGISTER) | |
09e18274 | 170 | remove_from_hard_reg_set (live, GET_MODE (reg), regno); |
610d2478 SB |
171 | } |
172 | ||
173 | /* Record in LIVE that register REG became live. | |
174 | This is called via note_stores. */ | |
175 | ||
176 | static void | |
7bc980e1 | 177 | reg_becomes_live (rtx reg, const_rtx setter ATTRIBUTE_UNUSED, void *live) |
610d2478 | 178 | { |
09e18274 | 179 | int regno; |
610d2478 SB |
180 | |
181 | if (GET_CODE (reg) == SUBREG) | |
182 | reg = SUBREG_REG (reg); | |
183 | ||
184 | if (!REG_P (reg)) | |
185 | return; | |
186 | ||
187 | regno = REGNO (reg); | |
188 | if (regno < FIRST_PSEUDO_REGISTER) | |
09e18274 | 189 | add_to_hard_reg_set ((HARD_REG_SET *) live, GET_MODE (reg), regno); |
610d2478 SB |
190 | } |
191 | ||
192 | /* Make sure if MODE_ENTRY is defined the MODE_EXIT is defined | |
193 | and vice versa. */ | |
194 | #if defined (MODE_ENTRY) != defined (MODE_EXIT) | |
195 | #error "Both MODE_ENTRY and MODE_EXIT must be defined" | |
196 | #endif | |
197 | ||
198 | #if defined (MODE_ENTRY) && defined (MODE_EXIT) | |
199 | /* Split the fallthrough edge to the exit block, so that we can note | |
200 | that there NORMAL_MODE is required. Return the new block if it's | |
201 | inserted before the exit block. Otherwise return null. */ | |
202 | ||
203 | static basic_block | |
204 | create_pre_exit (int n_entities, int *entity_map, const int *num_modes) | |
205 | { | |
206 | edge eg; | |
207 | edge_iterator ei; | |
208 | basic_block pre_exit; | |
209 | ||
210 | /* The only non-call predecessor at this stage is a block with a | |
211 | fallthrough edge; there can be at most one, but there could be | |
212 | none at all, e.g. when exit is called. */ | |
213 | pre_exit = 0; | |
fefa31b5 | 214 | FOR_EACH_EDGE (eg, ei, EXIT_BLOCK_PTR_FOR_FN (cfun)->preds) |
610d2478 SB |
215 | if (eg->flags & EDGE_FALLTHRU) |
216 | { | |
217 | basic_block src_bb = eg->src; | |
610d2478 SB |
218 | rtx last_insn, ret_reg; |
219 | ||
220 | gcc_assert (!pre_exit); | |
221 | /* If this function returns a value at the end, we have to | |
222 | insert the final mode switch before the return value copy | |
223 | to its hard register. */ | |
fefa31b5 | 224 | if (EDGE_COUNT (EXIT_BLOCK_PTR_FOR_FN (cfun)->preds) == 1 |
610d2478 SB |
225 | && NONJUMP_INSN_P ((last_insn = BB_END (src_bb))) |
226 | && GET_CODE (PATTERN (last_insn)) == USE | |
227 | && GET_CODE ((ret_reg = XEXP (PATTERN (last_insn), 0))) == REG) | |
228 | { | |
229 | int ret_start = REGNO (ret_reg); | |
230 | int nregs = hard_regno_nregs[ret_start][GET_MODE (ret_reg)]; | |
231 | int ret_end = ret_start + nregs; | |
c07757e5 UB |
232 | bool short_block = false; |
233 | bool multi_reg_return = false; | |
234 | bool forced_late_switch = false; | |
610d2478 SB |
235 | rtx before_return_copy; |
236 | ||
237 | do | |
238 | { | |
239 | rtx return_copy = PREV_INSN (last_insn); | |
240 | rtx return_copy_pat, copy_reg; | |
241 | int copy_start, copy_num; | |
242 | int j; | |
243 | ||
141a9e06 | 244 | if (NONDEBUG_INSN_P (return_copy)) |
610d2478 | 245 | { |
2bde7ae9 RS |
246 | /* When using SJLJ exceptions, the call to the |
247 | unregister function is inserted between the | |
248 | clobber of the return value and the copy. | |
249 | We do not want to split the block before this | |
250 | or any other call; if we have not found the | |
251 | copy yet, the copy must have been deleted. */ | |
252 | if (CALL_P (return_copy)) | |
253 | { | |
c07757e5 | 254 | short_block = true; |
2bde7ae9 RS |
255 | break; |
256 | } | |
89ab4659 KK |
257 | return_copy_pat = PATTERN (return_copy); |
258 | switch (GET_CODE (return_copy_pat)) | |
07288ab0 | 259 | { |
89ab4659 | 260 | case USE: |
c07757e5 UB |
261 | /* Skip USEs of multiple return registers. |
262 | __builtin_apply pattern is also handled here. */ | |
89ab4659 | 263 | if (GET_CODE (XEXP (return_copy_pat, 0)) == REG |
82f81f18 | 264 | && (targetm.calls.function_value_regno_p |
89ab4659 KK |
265 | (REGNO (XEXP (return_copy_pat, 0))))) |
266 | { | |
c07757e5 | 267 | multi_reg_return = true; |
89ab4659 KK |
268 | last_insn = return_copy; |
269 | continue; | |
270 | } | |
271 | break; | |
272 | ||
273 | case ASM_OPERANDS: | |
274 | /* Skip barrier insns. */ | |
275 | if (!MEM_VOLATILE_P (return_copy_pat)) | |
276 | break; | |
277 | ||
278 | /* Fall through. */ | |
279 | ||
280 | case ASM_INPUT: | |
281 | case UNSPEC_VOLATILE: | |
07288ab0 KK |
282 | last_insn = return_copy; |
283 | continue; | |
89ab4659 KK |
284 | |
285 | default: | |
286 | break; | |
07288ab0 | 287 | } |
89ab4659 | 288 | |
610d2478 SB |
289 | /* If the return register is not (in its entirety) |
290 | likely spilled, the return copy might be | |
291 | partially or completely optimized away. */ | |
292 | return_copy_pat = single_set (return_copy); | |
293 | if (!return_copy_pat) | |
294 | { | |
295 | return_copy_pat = PATTERN (return_copy); | |
296 | if (GET_CODE (return_copy_pat) != CLOBBER) | |
297 | break; | |
6fb5fa3c DB |
298 | else if (!optimize) |
299 | { | |
300 | /* This might be (clobber (reg [<result>])) | |
301 | when not optimizing. Then check if | |
302 | the previous insn is the clobber for | |
303 | the return register. */ | |
304 | copy_reg = SET_DEST (return_copy_pat); | |
305 | if (GET_CODE (copy_reg) == REG | |
306 | && !HARD_REGISTER_NUM_P (REGNO (copy_reg))) | |
307 | { | |
308 | if (INSN_P (PREV_INSN (return_copy))) | |
309 | { | |
310 | return_copy = PREV_INSN (return_copy); | |
311 | return_copy_pat = PATTERN (return_copy); | |
312 | if (GET_CODE (return_copy_pat) != CLOBBER) | |
313 | break; | |
314 | } | |
315 | } | |
316 | } | |
610d2478 SB |
317 | } |
318 | copy_reg = SET_DEST (return_copy_pat); | |
319 | if (GET_CODE (copy_reg) == REG) | |
320 | copy_start = REGNO (copy_reg); | |
321 | else if (GET_CODE (copy_reg) == SUBREG | |
322 | && GET_CODE (SUBREG_REG (copy_reg)) == REG) | |
323 | copy_start = REGNO (SUBREG_REG (copy_reg)); | |
324 | else | |
d78e64db KK |
325 | { |
326 | /* When control reaches end of non-void function, | |
327 | there are no return copy insns at all. This | |
328 | avoids an ice on that invalid function. */ | |
329 | if (ret_start + nregs == ret_end) | |
c07757e5 | 330 | short_block = true; |
d78e64db KK |
331 | break; |
332 | } | |
ffbbfaba | 333 | if (!targetm.calls.function_value_regno_p (copy_start)) |
ce4a9422 JR |
334 | copy_num = 0; |
335 | else | |
336 | copy_num | |
337 | = hard_regno_nregs[copy_start][GET_MODE (copy_reg)]; | |
610d2478 SB |
338 | |
339 | /* If the return register is not likely spilled, - as is | |
340 | the case for floating point on SH4 - then it might | |
341 | be set by an arithmetic operation that needs a | |
342 | different mode than the exit block. */ | |
343 | for (j = n_entities - 1; j >= 0; j--) | |
344 | { | |
345 | int e = entity_map[j]; | |
346 | int mode = MODE_NEEDED (e, return_copy); | |
347 | ||
348 | if (mode != num_modes[e] && mode != MODE_EXIT (e)) | |
349 | break; | |
350 | } | |
351 | if (j >= 0) | |
352 | { | |
b8435aa9 UB |
353 | /* __builtin_return emits a sequence of loads to all |
354 | return registers. One of them might require | |
355 | another mode than MODE_EXIT, even if it is | |
356 | unrelated to the return value, so we want to put | |
357 | the final mode switch after it. */ | |
c07757e5 | 358 | if (multi_reg_return |
b8435aa9 UB |
359 | && targetm.calls.function_value_regno_p |
360 | (copy_start)) | |
c07757e5 | 361 | forced_late_switch = true; |
b8435aa9 | 362 | |
610d2478 SB |
363 | /* For the SH4, floating point loads depend on fpscr, |
364 | thus we might need to put the final mode switch | |
365 | after the return value copy. That is still OK, | |
366 | because a floating point return value does not | |
367 | conflict with address reloads. */ | |
368 | if (copy_start >= ret_start | |
369 | && copy_start + copy_num <= ret_end | |
370 | && OBJECT_P (SET_SRC (return_copy_pat))) | |
c07757e5 | 371 | forced_late_switch = true; |
610d2478 SB |
372 | break; |
373 | } | |
ce4a9422 JR |
374 | if (copy_num == 0) |
375 | { | |
376 | last_insn = return_copy; | |
377 | continue; | |
378 | } | |
610d2478 SB |
379 | |
380 | if (copy_start >= ret_start | |
381 | && copy_start + copy_num <= ret_end) | |
382 | nregs -= copy_num; | |
c07757e5 | 383 | else if (!multi_reg_return |
82f81f18 AS |
384 | || !targetm.calls.function_value_regno_p |
385 | (copy_start)) | |
610d2478 SB |
386 | break; |
387 | last_insn = return_copy; | |
388 | } | |
389 | /* ??? Exception handling can lead to the return value | |
390 | copy being already separated from the return value use, | |
391 | as in unwind-dw2.c . | |
392 | Similarly, conditionally returning without a value, | |
393 | and conditionally using builtin_return can lead to an | |
394 | isolated use. */ | |
395 | if (return_copy == BB_HEAD (src_bb)) | |
396 | { | |
c07757e5 | 397 | short_block = true; |
610d2478 SB |
398 | break; |
399 | } | |
400 | last_insn = return_copy; | |
401 | } | |
402 | while (nregs); | |
b8698a0f | 403 | |
610d2478 SB |
404 | /* If we didn't see a full return value copy, verify that there |
405 | is a plausible reason for this. If some, but not all of the | |
406 | return register is likely spilled, we can expect that there | |
407 | is a copy for the likely spilled part. */ | |
408 | gcc_assert (!nregs | |
409 | || forced_late_switch | |
410 | || short_block | |
07b8f0a8 | 411 | || !(targetm.class_likely_spilled_p |
610d2478 SB |
412 | (REGNO_REG_CLASS (ret_start))) |
413 | || (nregs | |
414 | != hard_regno_nregs[ret_start][GET_MODE (ret_reg)]) | |
415 | /* For multi-hard-register floating point | |
416 | values, sometimes the likely-spilled part | |
417 | is ordinarily copied first, then the other | |
418 | part is set with an arithmetic operation. | |
419 | This doesn't actually cause reload | |
420 | failures, so let it pass. */ | |
421 | || (GET_MODE_CLASS (GET_MODE (ret_reg)) != MODE_INT | |
422 | && nregs != 1)); | |
b8698a0f | 423 | |
bba33211 | 424 | if (!NOTE_INSN_BASIC_BLOCK_P (last_insn)) |
610d2478 SB |
425 | { |
426 | before_return_copy | |
427 | = emit_note_before (NOTE_INSN_DELETED, last_insn); | |
428 | /* Instructions preceding LAST_INSN in the same block might | |
429 | require a different mode than MODE_EXIT, so if we might | |
430 | have such instructions, keep them in a separate block | |
431 | from pre_exit. */ | |
bba33211 JR |
432 | src_bb = split_block (src_bb, |
433 | PREV_INSN (before_return_copy))->dest; | |
610d2478 SB |
434 | } |
435 | else | |
436 | before_return_copy = last_insn; | |
437 | pre_exit = split_block (src_bb, before_return_copy)->src; | |
438 | } | |
439 | else | |
440 | { | |
441 | pre_exit = split_edge (eg); | |
610d2478 SB |
442 | } |
443 | } | |
444 | ||
445 | return pre_exit; | |
446 | } | |
447 | #endif | |
448 | ||
449 | /* Find all insns that need a particular mode setting, and insert the | |
450 | necessary mode switches. Return true if we did work. */ | |
451 | ||
7399bcb0 | 452 | static int |
10d22567 | 453 | optimize_mode_switching (void) |
610d2478 SB |
454 | { |
455 | rtx insn; | |
456 | int e; | |
457 | basic_block bb; | |
458 | int need_commit = 0; | |
459 | sbitmap *kill; | |
460 | struct edge_list *edge_list; | |
461 | static const int num_modes[] = NUM_MODES_FOR_MODE_SWITCHING; | |
462 | #define N_ENTITIES ARRAY_SIZE (num_modes) | |
463 | int entity_map[N_ENTITIES]; | |
464 | struct bb_info *bb_info[N_ENTITIES]; | |
465 | int i, j; | |
466 | int n_entities; | |
467 | int max_num_modes = 0; | |
073a8998 | 468 | bool emitted ATTRIBUTE_UNUSED = false; |
610d2478 SB |
469 | basic_block post_entry ATTRIBUTE_UNUSED, pre_exit ATTRIBUTE_UNUSED; |
470 | ||
610d2478 SB |
471 | for (e = N_ENTITIES - 1, n_entities = 0; e >= 0; e--) |
472 | if (OPTIMIZE_MODE_SWITCHING (e)) | |
473 | { | |
474 | int entry_exit_extra = 0; | |
475 | ||
476 | /* Create the list of segments within each basic block. | |
477 | If NORMAL_MODE is defined, allow for two extra | |
478 | blocks split from the entry and exit block. */ | |
479 | #if defined (MODE_ENTRY) && defined (MODE_EXIT) | |
480 | entry_exit_extra = 3; | |
481 | #endif | |
482 | bb_info[n_entities] | |
8b1c6fd7 DM |
483 | = XCNEWVEC (struct bb_info, |
484 | last_basic_block_for_fn (cfun) + entry_exit_extra); | |
610d2478 SB |
485 | entity_map[n_entities++] = e; |
486 | if (num_modes[e] > max_num_modes) | |
487 | max_num_modes = num_modes[e]; | |
488 | } | |
489 | ||
490 | if (! n_entities) | |
491 | return 0; | |
492 | ||
493 | #if defined (MODE_ENTRY) && defined (MODE_EXIT) | |
494 | /* Split the edge from the entry block, so that we can note that | |
495 | there NORMAL_MODE is supplied. */ | |
fefa31b5 | 496 | post_entry = split_edge (single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun))); |
610d2478 SB |
497 | pre_exit = create_pre_exit (n_entities, entity_map, num_modes); |
498 | #endif | |
499 | ||
6fb5fa3c DB |
500 | df_analyze (); |
501 | ||
610d2478 SB |
502 | /* Create the bitmap vectors. */ |
503 | ||
8b1c6fd7 DM |
504 | antic = sbitmap_vector_alloc (last_basic_block_for_fn (cfun), n_entities); |
505 | transp = sbitmap_vector_alloc (last_basic_block_for_fn (cfun), n_entities); | |
506 | comp = sbitmap_vector_alloc (last_basic_block_for_fn (cfun), n_entities); | |
610d2478 | 507 | |
8b1c6fd7 | 508 | bitmap_vector_ones (transp, last_basic_block_for_fn (cfun)); |
610d2478 SB |
509 | |
510 | for (j = n_entities - 1; j >= 0; j--) | |
511 | { | |
512 | int e = entity_map[j]; | |
513 | int no_mode = num_modes[e]; | |
514 | struct bb_info *info = bb_info[j]; | |
515 | ||
516 | /* Determine what the first use (if any) need for a mode of entity E is. | |
517 | This will be the mode that is anticipatable for this block. | |
518 | Also compute the initial transparency settings. */ | |
11cd3bed | 519 | FOR_EACH_BB_FN (bb, cfun) |
610d2478 SB |
520 | { |
521 | struct seginfo *ptr; | |
522 | int last_mode = no_mode; | |
a44250f4 | 523 | bool any_set_required = false; |
610d2478 SB |
524 | HARD_REG_SET live_now; |
525 | ||
6fb5fa3c | 526 | REG_SET_TO_HARD_REG_SET (live_now, df_get_live_in (bb)); |
24c2fde2 RH |
527 | |
528 | /* Pretend the mode is clobbered across abnormal edges. */ | |
529 | { | |
530 | edge_iterator ei; | |
531 | edge e; | |
532 | FOR_EACH_EDGE (e, ei, bb->preds) | |
533 | if (e->flags & EDGE_COMPLEX) | |
534 | break; | |
535 | if (e) | |
650a59ef R |
536 | { |
537 | ptr = new_seginfo (no_mode, BB_HEAD (bb), bb->index, live_now); | |
538 | add_seginfo (info + bb->index, ptr); | |
d7c028c0 | 539 | bitmap_clear_bit (transp[bb->index], j); |
650a59ef | 540 | } |
24c2fde2 RH |
541 | } |
542 | ||
0f346928 | 543 | FOR_BB_INSNS (bb, insn) |
610d2478 SB |
544 | { |
545 | if (INSN_P (insn)) | |
546 | { | |
547 | int mode = MODE_NEEDED (e, insn); | |
548 | rtx link; | |
549 | ||
550 | if (mode != no_mode && mode != last_mode) | |
551 | { | |
a44250f4 | 552 | any_set_required = true; |
610d2478 SB |
553 | last_mode = mode; |
554 | ptr = new_seginfo (mode, insn, bb->index, live_now); | |
555 | add_seginfo (info + bb->index, ptr); | |
d7c028c0 | 556 | bitmap_clear_bit (transp[bb->index], j); |
610d2478 SB |
557 | } |
558 | #ifdef MODE_AFTER | |
9786913b | 559 | last_mode = MODE_AFTER (e, last_mode, insn); |
610d2478 SB |
560 | #endif |
561 | /* Update LIVE_NOW. */ | |
562 | for (link = REG_NOTES (insn); link; link = XEXP (link, 1)) | |
563 | if (REG_NOTE_KIND (link) == REG_DEAD) | |
408bed3c | 564 | reg_dies (XEXP (link, 0), &live_now); |
610d2478 SB |
565 | |
566 | note_stores (PATTERN (insn), reg_becomes_live, &live_now); | |
567 | for (link = REG_NOTES (insn); link; link = XEXP (link, 1)) | |
568 | if (REG_NOTE_KIND (link) == REG_UNUSED) | |
408bed3c | 569 | reg_dies (XEXP (link, 0), &live_now); |
610d2478 SB |
570 | } |
571 | } | |
572 | ||
573 | info[bb->index].computing = last_mode; | |
a44250f4 | 574 | /* Check for blocks without ANY mode requirements. |
611a4849 UB |
575 | N.B. because of MODE_AFTER, last_mode might still |
576 | be different from no_mode, in which case we need to | |
577 | mark the block as nontransparent. */ | |
a44250f4 | 578 | if (!any_set_required) |
610d2478 SB |
579 | { |
580 | ptr = new_seginfo (no_mode, BB_END (bb), bb->index, live_now); | |
581 | add_seginfo (info + bb->index, ptr); | |
611a4849 UB |
582 | if (last_mode != no_mode) |
583 | bitmap_clear_bit (transp[bb->index], j); | |
610d2478 SB |
584 | } |
585 | } | |
586 | #if defined (MODE_ENTRY) && defined (MODE_EXIT) | |
587 | { | |
588 | int mode = MODE_ENTRY (e); | |
589 | ||
590 | if (mode != no_mode) | |
591 | { | |
592 | bb = post_entry; | |
593 | ||
594 | /* By always making this nontransparent, we save | |
595 | an extra check in make_preds_opaque. We also | |
596 | need this to avoid confusing pre_edge_lcm when | |
597 | antic is cleared but transp and comp are set. */ | |
d7c028c0 | 598 | bitmap_clear_bit (transp[bb->index], j); |
610d2478 SB |
599 | |
600 | /* Insert a fake computing definition of MODE into entry | |
601 | blocks which compute no mode. This represents the mode on | |
602 | entry. */ | |
603 | info[bb->index].computing = mode; | |
604 | ||
605 | if (pre_exit) | |
606 | info[pre_exit->index].seginfo->mode = MODE_EXIT (e); | |
607 | } | |
608 | } | |
609 | #endif /* NORMAL_MODE */ | |
610 | } | |
611 | ||
8b1c6fd7 | 612 | kill = sbitmap_vector_alloc (last_basic_block_for_fn (cfun), n_entities); |
610d2478 SB |
613 | for (i = 0; i < max_num_modes; i++) |
614 | { | |
615 | int current_mode[N_ENTITIES]; | |
60564289 | 616 | sbitmap *del; |
610d2478 SB |
617 | sbitmap *insert; |
618 | ||
619 | /* Set the anticipatable and computing arrays. */ | |
8b1c6fd7 DM |
620 | bitmap_vector_clear (antic, last_basic_block_for_fn (cfun)); |
621 | bitmap_vector_clear (comp, last_basic_block_for_fn (cfun)); | |
610d2478 SB |
622 | for (j = n_entities - 1; j >= 0; j--) |
623 | { | |
624 | int m = current_mode[j] = MODE_PRIORITY_TO_MODE (entity_map[j], i); | |
625 | struct bb_info *info = bb_info[j]; | |
626 | ||
11cd3bed | 627 | FOR_EACH_BB_FN (bb, cfun) |
610d2478 SB |
628 | { |
629 | if (info[bb->index].seginfo->mode == m) | |
d7c028c0 | 630 | bitmap_set_bit (antic[bb->index], j); |
610d2478 SB |
631 | |
632 | if (info[bb->index].computing == m) | |
d7c028c0 | 633 | bitmap_set_bit (comp[bb->index], j); |
610d2478 SB |
634 | } |
635 | } | |
636 | ||
637 | /* Calculate the optimal locations for the | |
638 | placement mode switches to modes with priority I. */ | |
639 | ||
11cd3bed | 640 | FOR_EACH_BB_FN (bb, cfun) |
f61e445a | 641 | bitmap_not (kill[bb->index], transp[bb->index]); |
10d22567 | 642 | edge_list = pre_edge_lcm (n_entities, transp, comp, antic, |
60564289 | 643 | kill, &insert, &del); |
610d2478 SB |
644 | |
645 | for (j = n_entities - 1; j >= 0; j--) | |
646 | { | |
647 | /* Insert all mode sets that have been inserted by lcm. */ | |
648 | int no_mode = num_modes[entity_map[j]]; | |
649 | ||
650 | /* Wherever we have moved a mode setting upwards in the flow graph, | |
651 | the blocks between the new setting site and the now redundant | |
652 | computation ceases to be transparent for any lower-priority | |
653 | mode of the same entity. First set the aux field of each | |
654 | insertion site edge non-transparent, then propagate the new | |
655 | non-transparency from the redundant computation upwards till | |
656 | we hit an insertion site or an already non-transparent block. */ | |
657 | for (e = NUM_EDGES (edge_list) - 1; e >= 0; e--) | |
658 | { | |
659 | edge eg = INDEX_EDGE (edge_list, e); | |
660 | int mode; | |
661 | basic_block src_bb; | |
662 | HARD_REG_SET live_at_edge; | |
663 | rtx mode_set; | |
664 | ||
665 | eg->aux = 0; | |
666 | ||
d7c028c0 | 667 | if (! bitmap_bit_p (insert[e], j)) |
610d2478 SB |
668 | continue; |
669 | ||
670 | eg->aux = (void *)1; | |
671 | ||
672 | mode = current_mode[j]; | |
673 | src_bb = eg->src; | |
674 | ||
6fb5fa3c | 675 | REG_SET_TO_HARD_REG_SET (live_at_edge, df_get_live_out (src_bb)); |
610d2478 | 676 | |
5f28524a | 677 | rtl_profile_for_edge (eg); |
610d2478 SB |
678 | start_sequence (); |
679 | EMIT_MODE_SET (entity_map[j], mode, live_at_edge); | |
680 | mode_set = get_insns (); | |
681 | end_sequence (); | |
5f28524a | 682 | default_rtl_profile (); |
610d2478 SB |
683 | |
684 | /* Do not bother to insert empty sequence. */ | |
685 | if (mode_set == NULL_RTX) | |
686 | continue; | |
687 | ||
650a59ef R |
688 | /* We should not get an abnormal edge here. */ |
689 | gcc_assert (! (eg->flags & EDGE_ABNORMAL)); | |
690 | ||
691 | need_commit = 1; | |
692 | insert_insn_on_edge (mode_set, eg); | |
610d2478 SB |
693 | } |
694 | ||
4f42035e | 695 | FOR_EACH_BB_REVERSE_FN (bb, cfun) |
d7c028c0 | 696 | if (bitmap_bit_p (del[bb->index], j)) |
610d2478 SB |
697 | { |
698 | make_preds_opaque (bb, j); | |
699 | /* Cancel the 'deleted' mode set. */ | |
700 | bb_info[j][bb->index].seginfo->mode = no_mode; | |
701 | } | |
702 | } | |
703 | ||
60564289 | 704 | sbitmap_vector_free (del); |
610d2478 SB |
705 | sbitmap_vector_free (insert); |
706 | clear_aux_for_edges (); | |
707 | free_edge_list (edge_list); | |
708 | } | |
709 | ||
710 | /* Now output the remaining mode sets in all the segments. */ | |
711 | for (j = n_entities - 1; j >= 0; j--) | |
712 | { | |
713 | int no_mode = num_modes[entity_map[j]]; | |
714 | ||
4f42035e | 715 | FOR_EACH_BB_REVERSE_FN (bb, cfun) |
610d2478 SB |
716 | { |
717 | struct seginfo *ptr, *next; | |
718 | for (ptr = bb_info[j][bb->index].seginfo; ptr; ptr = next) | |
719 | { | |
720 | next = ptr->next; | |
721 | if (ptr->mode != no_mode) | |
722 | { | |
723 | rtx mode_set; | |
724 | ||
5f28524a | 725 | rtl_profile_for_bb (bb); |
610d2478 SB |
726 | start_sequence (); |
727 | EMIT_MODE_SET (entity_map[j], ptr->mode, ptr->regs_live); | |
728 | mode_set = get_insns (); | |
729 | end_sequence (); | |
730 | ||
731 | /* Insert MODE_SET only if it is nonempty. */ | |
732 | if (mode_set != NULL_RTX) | |
733 | { | |
073a8998 | 734 | emitted = true; |
a38e7aa5 | 735 | if (NOTE_INSN_BASIC_BLOCK_P (ptr->insn_ptr)) |
610d2478 SB |
736 | emit_insn_after (mode_set, ptr->insn_ptr); |
737 | else | |
738 | emit_insn_before (mode_set, ptr->insn_ptr); | |
739 | } | |
5f28524a JH |
740 | |
741 | default_rtl_profile (); | |
610d2478 SB |
742 | } |
743 | ||
744 | free (ptr); | |
745 | } | |
746 | } | |
747 | ||
748 | free (bb_info[j]); | |
749 | } | |
750 | ||
751 | /* Finished. Free up all the things we've allocated. */ | |
610d2478 SB |
752 | sbitmap_vector_free (kill); |
753 | sbitmap_vector_free (antic); | |
754 | sbitmap_vector_free (transp); | |
755 | sbitmap_vector_free (comp); | |
756 | ||
757 | if (need_commit) | |
758 | commit_edge_insertions (); | |
759 | ||
760 | #if defined (MODE_ENTRY) && defined (MODE_EXIT) | |
761 | cleanup_cfg (CLEANUP_NO_INSN_DEL); | |
762 | #else | |
073a8998 | 763 | if (!need_commit && !emitted) |
610d2478 SB |
764 | return 0; |
765 | #endif | |
766 | ||
610d2478 SB |
767 | return 1; |
768 | } | |
ef330312 | 769 | |
610d2478 | 770 | #endif /* OPTIMIZE_MODE_SWITCHING */ |
ef330312 PB |
771 | \f |
772 | static bool | |
773 | gate_mode_switching (void) | |
774 | { | |
775 | #ifdef OPTIMIZE_MODE_SWITCHING | |
776 | return true; | |
777 | #else | |
778 | return false; | |
779 | #endif | |
780 | } | |
781 | ||
c2924966 | 782 | static unsigned int |
ef330312 PB |
783 | rest_of_handle_mode_switching (void) |
784 | { | |
785 | #ifdef OPTIMIZE_MODE_SWITCHING | |
10d22567 | 786 | optimize_mode_switching (); |
ef330312 | 787 | #endif /* OPTIMIZE_MODE_SWITCHING */ |
c2924966 | 788 | return 0; |
ef330312 PB |
789 | } |
790 | ||
791 | ||
27a4cd48 DM |
792 | namespace { |
793 | ||
794 | const pass_data pass_data_mode_switching = | |
ef330312 | 795 | { |
27a4cd48 DM |
796 | RTL_PASS, /* type */ |
797 | "mode_sw", /* name */ | |
798 | OPTGROUP_NONE, /* optinfo_flags */ | |
799 | true, /* has_gate */ | |
800 | true, /* has_execute */ | |
801 | TV_MODE_SWITCH, /* tv_id */ | |
802 | 0, /* properties_required */ | |
803 | 0, /* properties_provided */ | |
804 | 0, /* properties_destroyed */ | |
805 | 0, /* todo_flags_start */ | |
806 | ( TODO_df_finish | TODO_verify_rtl_sharing | 0 ), /* todo_flags_finish */ | |
ef330312 | 807 | }; |
27a4cd48 DM |
808 | |
809 | class pass_mode_switching : public rtl_opt_pass | |
810 | { | |
811 | public: | |
c3284718 RS |
812 | pass_mode_switching (gcc::context *ctxt) |
813 | : rtl_opt_pass (pass_data_mode_switching, ctxt) | |
27a4cd48 DM |
814 | {} |
815 | ||
816 | /* opt_pass methods: */ | |
05555c4a DM |
817 | /* The epiphany backend creates a second instance of this pass, so we need |
818 | a clone method. */ | |
65d3284b | 819 | opt_pass * clone () { return new pass_mode_switching (m_ctxt); } |
27a4cd48 DM |
820 | bool gate () { return gate_mode_switching (); } |
821 | unsigned int execute () { return rest_of_handle_mode_switching (); } | |
822 | ||
823 | }; // class pass_mode_switching | |
824 | ||
825 | } // anon namespace | |
826 | ||
827 | rtl_opt_pass * | |
828 | make_pass_mode_switching (gcc::context *ctxt) | |
829 | { | |
830 | return new pass_mode_switching (ctxt); | |
831 | } |