]>
Commit | Line | Data |
---|---|---|
55a2c322 | 1 | /* Code for RTL transformations to satisfy insn constraints. |
d1e082c2 | 2 | Copyright (C) 2010-2013 Free Software Foundation, Inc. |
55a2c322 VM |
3 | Contributed by Vladimir Makarov <vmakarov@redhat.com>. |
4 | ||
5 | This file is part of GCC. | |
6 | ||
7 | GCC is free software; you can redistribute it and/or modify it under | |
8 | the terms of the GNU General Public License as published by the Free | |
9 | Software Foundation; either version 3, or (at your option) any later | |
10 | version. | |
11 | ||
12 | GCC is distributed in the hope that it will be useful, but WITHOUT ANY | |
13 | WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
14 | FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
15 | for more details. | |
16 | ||
17 | You should have received a copy of the GNU General Public License | |
18 | along with GCC; see the file COPYING3. If not see | |
19 | <http://www.gnu.org/licenses/>. */ | |
20 | ||
21 | ||
22 | /* This file contains code for 3 passes: constraint pass, | |
23 | inheritance/split pass, and pass for undoing failed inheritance and | |
24 | split. | |
25 | ||
26 | The major goal of constraint pass is to transform RTL to satisfy | |
27 | insn and address constraints by: | |
28 | o choosing insn alternatives; | |
29 | o generating *reload insns* (or reloads in brief) and *reload | |
30 | pseudos* which will get necessary hard registers later; | |
31 | o substituting pseudos with equivalent values and removing the | |
32 | instructions that initialized those pseudos. | |
33 | ||
34 | The constraint pass has biggest and most complicated code in LRA. | |
35 | There are a lot of important details like: | |
36 | o reuse of input reload pseudos to simplify reload pseudo | |
37 | allocations; | |
38 | o some heuristics to choose insn alternative to improve the | |
39 | inheritance; | |
40 | o early clobbers etc. | |
41 | ||
42 | The pass is mimicking former reload pass in alternative choosing | |
43 | because the reload pass is oriented to current machine description | |
44 | model. It might be changed if the machine description model is | |
45 | changed. | |
46 | ||
47 | There is special code for preventing all LRA and this pass cycling | |
48 | in case of bugs. | |
49 | ||
50 | On the first iteration of the pass we process every instruction and | |
51 | choose an alternative for each one. On subsequent iterations we try | |
52 | to avoid reprocessing instructions if we can be sure that the old | |
53 | choice is still valid. | |
54 | ||
55 | The inheritance/spilt pass is to transform code to achieve | |
56 | ineheritance and live range splitting. It is done on backward | |
57 | traversal of EBBs. | |
58 | ||
59 | The inheritance optimization goal is to reuse values in hard | |
60 | registers. There is analogous optimization in old reload pass. The | |
61 | inheritance is achieved by following transformation: | |
62 | ||
63 | reload_p1 <- p reload_p1 <- p | |
64 | ... new_p <- reload_p1 | |
65 | ... => ... | |
66 | reload_p2 <- p reload_p2 <- new_p | |
67 | ||
68 | where p is spilled and not changed between the insns. Reload_p1 is | |
69 | also called *original pseudo* and new_p is called *inheritance | |
70 | pseudo*. | |
71 | ||
72 | The subsequent assignment pass will try to assign the same (or | |
73 | another if it is not possible) hard register to new_p as to | |
74 | reload_p1 or reload_p2. | |
75 | ||
76 | If the assignment pass fails to assign a hard register to new_p, | |
77 | this file will undo the inheritance and restore the original code. | |
78 | This is because implementing the above sequence with a spilled | |
79 | new_p would make the code much worse. The inheritance is done in | |
80 | EBB scope. The above is just a simplified example to get an idea | |
81 | of the inheritance as the inheritance is also done for non-reload | |
82 | insns. | |
83 | ||
84 | Splitting (transformation) is also done in EBB scope on the same | |
85 | pass as the inheritance: | |
86 | ||
87 | r <- ... or ... <- r r <- ... or ... <- r | |
88 | ... s <- r (new insn -- save) | |
f4eafc30 | 89 | ... => |
55a2c322 VM |
90 | ... r <- s (new insn -- restore) |
91 | ... <- r ... <- r | |
92 | ||
93 | The *split pseudo* s is assigned to the hard register of the | |
94 | original pseudo or hard register r. | |
95 | ||
96 | Splitting is done: | |
97 | o In EBBs with high register pressure for global pseudos (living | |
98 | in at least 2 BBs) and assigned to hard registers when there | |
99 | are more one reloads needing the hard registers; | |
100 | o for pseudos needing save/restore code around calls. | |
101 | ||
102 | If the split pseudo still has the same hard register as the | |
103 | original pseudo after the subsequent assignment pass or the | |
104 | original pseudo was split, the opposite transformation is done on | |
105 | the same pass for undoing inheritance. */ | |
106 | ||
107 | #undef REG_OK_STRICT | |
108 | ||
109 | #include "config.h" | |
110 | #include "system.h" | |
111 | #include "coretypes.h" | |
112 | #include "tm.h" | |
113 | #include "hard-reg-set.h" | |
114 | #include "rtl.h" | |
115 | #include "tm_p.h" | |
116 | #include "regs.h" | |
117 | #include "insn-config.h" | |
118 | #include "insn-codes.h" | |
119 | #include "recog.h" | |
120 | #include "output.h" | |
121 | #include "addresses.h" | |
122 | #include "target.h" | |
123 | #include "function.h" | |
124 | #include "expr.h" | |
125 | #include "basic-block.h" | |
126 | #include "except.h" | |
127 | #include "optabs.h" | |
128 | #include "df.h" | |
129 | #include "ira.h" | |
130 | #include "rtl-error.h" | |
131 | #include "lra-int.h" | |
132 | ||
133 | /* Value of LRA_CURR_RELOAD_NUM at the beginning of BB of the current | |
134 | insn. Remember that LRA_CURR_RELOAD_NUM is the number of emitted | |
135 | reload insns. */ | |
136 | static int bb_reload_num; | |
137 | ||
138 | /* The current insn being processed and corresponding its data (basic | |
139 | block, the insn data, the insn static data, and the mode of each | |
140 | operand). */ | |
141 | static rtx curr_insn; | |
142 | static basic_block curr_bb; | |
143 | static lra_insn_recog_data_t curr_id; | |
144 | static struct lra_static_insn_data *curr_static_id; | |
145 | static enum machine_mode curr_operand_mode[MAX_RECOG_OPERANDS]; | |
146 | ||
147 | \f | |
148 | ||
149 | /* Start numbers for new registers and insns at the current constraints | |
150 | pass start. */ | |
151 | static int new_regno_start; | |
152 | static int new_insn_uid_start; | |
153 | ||
277f65de RS |
154 | /* If LOC is nonnull, strip any outer subreg from it. */ |
155 | static inline rtx * | |
156 | strip_subreg (rtx *loc) | |
157 | { | |
158 | return loc && GET_CODE (*loc) == SUBREG ? &SUBREG_REG (*loc) : loc; | |
159 | } | |
160 | ||
55a2c322 VM |
161 | /* Return hard regno of REGNO or if it is was not assigned to a hard |
162 | register, use a hard register from its allocno class. */ | |
163 | static int | |
164 | get_try_hard_regno (int regno) | |
165 | { | |
166 | int hard_regno; | |
167 | enum reg_class rclass; | |
168 | ||
169 | if ((hard_regno = regno) >= FIRST_PSEUDO_REGISTER) | |
170 | hard_regno = lra_get_regno_hard_regno (regno); | |
171 | if (hard_regno >= 0) | |
172 | return hard_regno; | |
173 | rclass = lra_get_allocno_class (regno); | |
174 | if (rclass == NO_REGS) | |
175 | return -1; | |
176 | return ira_class_hard_regs[rclass][0]; | |
177 | } | |
178 | ||
179 | /* Return final hard regno (plus offset) which will be after | |
180 | elimination. We do this for matching constraints because the final | |
181 | hard regno could have a different class. */ | |
182 | static int | |
183 | get_final_hard_regno (int hard_regno, int offset) | |
184 | { | |
185 | if (hard_regno < 0) | |
186 | return hard_regno; | |
187 | hard_regno = lra_get_elimination_hard_regno (hard_regno); | |
188 | return hard_regno + offset; | |
189 | } | |
190 | ||
191 | /* Return hard regno of X after removing subreg and making | |
192 | elimination. If X is not a register or subreg of register, return | |
193 | -1. For pseudo use its assignment. */ | |
194 | static int | |
195 | get_hard_regno (rtx x) | |
196 | { | |
197 | rtx reg; | |
198 | int offset, hard_regno; | |
199 | ||
200 | reg = x; | |
201 | if (GET_CODE (x) == SUBREG) | |
202 | reg = SUBREG_REG (x); | |
203 | if (! REG_P (reg)) | |
204 | return -1; | |
205 | if ((hard_regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER) | |
206 | hard_regno = lra_get_regno_hard_regno (hard_regno); | |
207 | if (hard_regno < 0) | |
208 | return -1; | |
209 | offset = 0; | |
210 | if (GET_CODE (x) == SUBREG) | |
211 | offset += subreg_regno_offset (hard_regno, GET_MODE (reg), | |
212 | SUBREG_BYTE (x), GET_MODE (x)); | |
213 | return get_final_hard_regno (hard_regno, offset); | |
214 | } | |
215 | ||
216 | /* If REGNO is a hard register or has been allocated a hard register, | |
217 | return the class of that register. If REGNO is a reload pseudo | |
218 | created by the current constraints pass, return its allocno class. | |
219 | Return NO_REGS otherwise. */ | |
220 | static enum reg_class | |
221 | get_reg_class (int regno) | |
222 | { | |
223 | int hard_regno; | |
224 | ||
225 | if ((hard_regno = regno) >= FIRST_PSEUDO_REGISTER) | |
226 | hard_regno = lra_get_regno_hard_regno (regno); | |
227 | if (hard_regno >= 0) | |
228 | { | |
229 | hard_regno = get_final_hard_regno (hard_regno, 0); | |
230 | return REGNO_REG_CLASS (hard_regno); | |
231 | } | |
232 | if (regno >= new_regno_start) | |
233 | return lra_get_allocno_class (regno); | |
234 | return NO_REGS; | |
235 | } | |
236 | ||
237 | /* Return true if REG satisfies (or will satisfy) reg class constraint | |
238 | CL. Use elimination first if REG is a hard register. If REG is a | |
239 | reload pseudo created by this constraints pass, assume that it will | |
240 | be allocated a hard register from its allocno class, but allow that | |
241 | class to be narrowed to CL if it is currently a superset of CL. | |
242 | ||
243 | If NEW_CLASS is nonnull, set *NEW_CLASS to the new allocno class of | |
244 | REGNO (reg), or NO_REGS if no change in its class was needed. */ | |
245 | static bool | |
246 | in_class_p (rtx reg, enum reg_class cl, enum reg_class *new_class) | |
247 | { | |
248 | enum reg_class rclass, common_class; | |
249 | enum machine_mode reg_mode; | |
250 | int class_size, hard_regno, nregs, i, j; | |
251 | int regno = REGNO (reg); | |
f4eafc30 | 252 | |
55a2c322 VM |
253 | if (new_class != NULL) |
254 | *new_class = NO_REGS; | |
255 | if (regno < FIRST_PSEUDO_REGISTER) | |
256 | { | |
257 | rtx final_reg = reg; | |
258 | rtx *final_loc = &final_reg; | |
f4eafc30 | 259 | |
55a2c322 VM |
260 | lra_eliminate_reg_if_possible (final_loc); |
261 | return TEST_HARD_REG_BIT (reg_class_contents[cl], REGNO (*final_loc)); | |
262 | } | |
263 | reg_mode = GET_MODE (reg); | |
264 | rclass = get_reg_class (regno); | |
265 | if (regno < new_regno_start | |
266 | /* Do not allow the constraints for reload instructions to | |
267 | influence the classes of new pseudos. These reloads are | |
268 | typically moves that have many alternatives, and restricting | |
269 | reload pseudos for one alternative may lead to situations | |
270 | where other reload pseudos are no longer allocatable. */ | |
271 | || INSN_UID (curr_insn) >= new_insn_uid_start) | |
272 | /* When we don't know what class will be used finally for reload | |
273 | pseudos, we use ALL_REGS. */ | |
274 | return ((regno >= new_regno_start && rclass == ALL_REGS) | |
275 | || (rclass != NO_REGS && ira_class_subset_p[rclass][cl] | |
276 | && ! hard_reg_set_subset_p (reg_class_contents[cl], | |
277 | lra_no_alloc_regs))); | |
278 | else | |
279 | { | |
280 | common_class = ira_reg_class_subset[rclass][cl]; | |
281 | if (new_class != NULL) | |
282 | *new_class = common_class; | |
283 | if (hard_reg_set_subset_p (reg_class_contents[common_class], | |
284 | lra_no_alloc_regs)) | |
285 | return false; | |
286 | /* Check that there are enough allocatable regs. */ | |
287 | class_size = ira_class_hard_regs_num[common_class]; | |
288 | for (i = 0; i < class_size; i++) | |
289 | { | |
290 | hard_regno = ira_class_hard_regs[common_class][i]; | |
291 | nregs = hard_regno_nregs[hard_regno][reg_mode]; | |
292 | if (nregs == 1) | |
293 | return true; | |
294 | for (j = 0; j < nregs; j++) | |
f421c426 VM |
295 | if (TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno + j) |
296 | || ! TEST_HARD_REG_BIT (reg_class_contents[common_class], | |
297 | hard_regno + j)) | |
55a2c322 VM |
298 | break; |
299 | if (j >= nregs) | |
300 | return true; | |
301 | } | |
302 | return false; | |
303 | } | |
304 | } | |
305 | ||
306 | /* Return true if REGNO satisfies a memory constraint. */ | |
307 | static bool | |
308 | in_mem_p (int regno) | |
309 | { | |
310 | return get_reg_class (regno) == NO_REGS; | |
311 | } | |
312 | ||
313 | /* If we have decided to substitute X with another value, return that | |
314 | value, otherwise return X. */ | |
315 | static rtx | |
316 | get_equiv_substitution (rtx x) | |
317 | { | |
318 | int regno; | |
319 | rtx res; | |
320 | ||
321 | if (! REG_P (x) || (regno = REGNO (x)) < FIRST_PSEUDO_REGISTER | |
322 | || ! ira_reg_equiv[regno].defined_p | |
323 | || ! ira_reg_equiv[regno].profitable_p | |
324 | || lra_get_regno_hard_regno (regno) >= 0) | |
325 | return x; | |
326 | if ((res = ira_reg_equiv[regno].memory) != NULL_RTX) | |
327 | return res; | |
328 | if ((res = ira_reg_equiv[regno].constant) != NULL_RTX) | |
329 | return res; | |
330 | if ((res = ira_reg_equiv[regno].invariant) != NULL_RTX) | |
331 | return res; | |
332 | gcc_unreachable (); | |
333 | } | |
334 | ||
335 | /* Set up curr_operand_mode. */ | |
336 | static void | |
337 | init_curr_operand_mode (void) | |
338 | { | |
339 | int nop = curr_static_id->n_operands; | |
340 | for (int i = 0; i < nop; i++) | |
341 | { | |
342 | enum machine_mode mode = GET_MODE (*curr_id->operand_loc[i]); | |
343 | if (mode == VOIDmode) | |
344 | { | |
345 | /* The .md mode for address operands is the mode of the | |
346 | addressed value rather than the mode of the address itself. */ | |
347 | if (curr_id->icode >= 0 && curr_static_id->operand[i].is_address) | |
348 | mode = Pmode; | |
349 | else | |
350 | mode = curr_static_id->operand[i].mode; | |
351 | } | |
352 | curr_operand_mode[i] = mode; | |
353 | } | |
354 | } | |
355 | ||
356 | \f | |
357 | ||
358 | /* The page contains code to reuse input reloads. */ | |
359 | ||
360 | /* Structure describes input reload of the current insns. */ | |
361 | struct input_reload | |
362 | { | |
363 | /* Reloaded value. */ | |
364 | rtx input; | |
365 | /* Reload pseudo used. */ | |
366 | rtx reg; | |
367 | }; | |
368 | ||
369 | /* The number of elements in the following array. */ | |
370 | static int curr_insn_input_reloads_num; | |
371 | /* Array containing info about input reloads. It is used to find the | |
372 | same input reload and reuse the reload pseudo in this case. */ | |
373 | static struct input_reload curr_insn_input_reloads[LRA_MAX_INSN_RELOADS]; | |
374 | ||
375 | /* Initiate data concerning reuse of input reloads for the current | |
376 | insn. */ | |
377 | static void | |
378 | init_curr_insn_input_reloads (void) | |
379 | { | |
380 | curr_insn_input_reloads_num = 0; | |
381 | } | |
382 | ||
383 | /* Change class of pseudo REGNO to NEW_CLASS. Print info about it | |
384 | using TITLE. Output a new line if NL_P. */ | |
385 | static void | |
386 | change_class (int regno, enum reg_class new_class, | |
387 | const char *title, bool nl_p) | |
388 | { | |
389 | lra_assert (regno >= FIRST_PSEUDO_REGISTER); | |
390 | if (lra_dump_file != NULL) | |
391 | fprintf (lra_dump_file, "%s to class %s for r%d", | |
392 | title, reg_class_names[new_class], regno); | |
393 | setup_reg_classes (regno, new_class, NO_REGS, new_class); | |
394 | if (lra_dump_file != NULL && nl_p) | |
395 | fprintf (lra_dump_file, "\n"); | |
396 | } | |
397 | ||
398 | /* Create a new pseudo using MODE, RCLASS, ORIGINAL or reuse already | |
399 | created input reload pseudo (only if TYPE is not OP_OUT). The | |
400 | result pseudo is returned through RESULT_REG. Return TRUE if we | |
401 | created a new pseudo, FALSE if we reused the already created input | |
402 | reload pseudo. Use TITLE to describe new registers for debug | |
403 | purposes. */ | |
404 | static bool | |
405 | get_reload_reg (enum op_type type, enum machine_mode mode, rtx original, | |
406 | enum reg_class rclass, const char *title, rtx *result_reg) | |
407 | { | |
408 | int i, regno; | |
409 | enum reg_class new_class; | |
410 | ||
411 | if (type == OP_OUT) | |
412 | { | |
413 | *result_reg | |
414 | = lra_create_new_reg_with_unique_value (mode, original, rclass, title); | |
415 | return true; | |
416 | } | |
73cca0cc VM |
417 | /* Prevent reuse value of expression with side effects, |
418 | e.g. volatile memory. */ | |
419 | if (! side_effects_p (original)) | |
420 | for (i = 0; i < curr_insn_input_reloads_num; i++) | |
421 | if (rtx_equal_p (curr_insn_input_reloads[i].input, original) | |
422 | && in_class_p (curr_insn_input_reloads[i].reg, rclass, &new_class)) | |
423 | { | |
424 | *result_reg = curr_insn_input_reloads[i].reg; | |
425 | regno = REGNO (*result_reg); | |
426 | if (lra_dump_file != NULL) | |
427 | { | |
428 | fprintf (lra_dump_file, " Reuse r%d for reload ", regno); | |
429 | dump_value_slim (lra_dump_file, original, 1); | |
430 | } | |
431 | if (new_class != lra_get_allocno_class (regno)) | |
432 | change_class (regno, new_class, ", change", false); | |
433 | if (lra_dump_file != NULL) | |
434 | fprintf (lra_dump_file, "\n"); | |
435 | return false; | |
436 | } | |
55a2c322 VM |
437 | *result_reg = lra_create_new_reg (mode, original, rclass, title); |
438 | lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS); | |
439 | curr_insn_input_reloads[curr_insn_input_reloads_num].input = original; | |
440 | curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = *result_reg; | |
441 | return true; | |
442 | } | |
443 | ||
444 | \f | |
445 | ||
446 | /* The page contains code to extract memory address parts. */ | |
447 | ||
55a2c322 VM |
448 | /* Wrapper around REGNO_OK_FOR_INDEX_P, to allow pseudos. */ |
449 | static inline bool | |
450 | ok_for_index_p_nonstrict (rtx reg) | |
451 | { | |
452 | unsigned regno = REGNO (reg); | |
f4eafc30 | 453 | |
55a2c322 VM |
454 | return regno >= FIRST_PSEUDO_REGISTER || REGNO_OK_FOR_INDEX_P (regno); |
455 | } | |
456 | ||
457 | /* A version of regno_ok_for_base_p for use here, when all pseudos | |
458 | should count as OK. Arguments as for regno_ok_for_base_p. */ | |
459 | static inline bool | |
460 | ok_for_base_p_nonstrict (rtx reg, enum machine_mode mode, addr_space_t as, | |
461 | enum rtx_code outer_code, enum rtx_code index_code) | |
462 | { | |
463 | unsigned regno = REGNO (reg); | |
464 | ||
465 | if (regno >= FIRST_PSEUDO_REGISTER) | |
466 | return true; | |
467 | return ok_for_base_p_1 (regno, mode, as, outer_code, index_code); | |
468 | } | |
469 | ||
55a2c322 VM |
470 | \f |
471 | ||
472 | /* The page contains major code to choose the current insn alternative | |
473 | and generate reloads for it. */ | |
474 | ||
475 | /* Return the offset from REGNO of the least significant register | |
476 | in (reg:MODE REGNO). | |
477 | ||
478 | This function is used to tell whether two registers satisfy | |
479 | a matching constraint. (reg:MODE1 REGNO1) matches (reg:MODE2 REGNO2) if: | |
480 | ||
481 | REGNO1 + lra_constraint_offset (REGNO1, MODE1) | |
482 | == REGNO2 + lra_constraint_offset (REGNO2, MODE2) */ | |
483 | int | |
484 | lra_constraint_offset (int regno, enum machine_mode mode) | |
485 | { | |
486 | lra_assert (regno < FIRST_PSEUDO_REGISTER); | |
487 | if (WORDS_BIG_ENDIAN && GET_MODE_SIZE (mode) > UNITS_PER_WORD | |
488 | && SCALAR_INT_MODE_P (mode)) | |
489 | return hard_regno_nregs[regno][mode] - 1; | |
490 | return 0; | |
491 | } | |
492 | ||
493 | /* Like rtx_equal_p except that it allows a REG and a SUBREG to match | |
494 | if they are the same hard reg, and has special hacks for | |
495 | auto-increment and auto-decrement. This is specifically intended for | |
496 | process_alt_operands to use in determining whether two operands | |
497 | match. X is the operand whose number is the lower of the two. | |
498 | ||
499 | It is supposed that X is the output operand and Y is the input | |
500 | operand. Y_HARD_REGNO is the final hard regno of register Y or | |
501 | register in subreg Y as we know it now. Otherwise, it is a | |
502 | negative value. */ | |
503 | static bool | |
504 | operands_match_p (rtx x, rtx y, int y_hard_regno) | |
505 | { | |
506 | int i; | |
507 | RTX_CODE code = GET_CODE (x); | |
508 | const char *fmt; | |
509 | ||
510 | if (x == y) | |
511 | return true; | |
512 | if ((code == REG || (code == SUBREG && REG_P (SUBREG_REG (x)))) | |
513 | && (REG_P (y) || (GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y))))) | |
514 | { | |
515 | int j; | |
f4eafc30 | 516 | |
55a2c322 VM |
517 | i = get_hard_regno (x); |
518 | if (i < 0) | |
519 | goto slow; | |
520 | ||
521 | if ((j = y_hard_regno) < 0) | |
522 | goto slow; | |
523 | ||
524 | i += lra_constraint_offset (i, GET_MODE (x)); | |
525 | j += lra_constraint_offset (j, GET_MODE (y)); | |
526 | ||
527 | return i == j; | |
528 | } | |
529 | ||
530 | /* If two operands must match, because they are really a single | |
531 | operand of an assembler insn, then two post-increments are invalid | |
532 | because the assembler insn would increment only once. On the | |
533 | other hand, a post-increment matches ordinary indexing if the | |
534 | post-increment is the output operand. */ | |
535 | if (code == POST_DEC || code == POST_INC || code == POST_MODIFY) | |
536 | return operands_match_p (XEXP (x, 0), y, y_hard_regno); | |
537 | ||
538 | /* Two pre-increments are invalid because the assembler insn would | |
539 | increment only once. On the other hand, a pre-increment matches | |
540 | ordinary indexing if the pre-increment is the input operand. */ | |
541 | if (GET_CODE (y) == PRE_DEC || GET_CODE (y) == PRE_INC | |
542 | || GET_CODE (y) == PRE_MODIFY) | |
543 | return operands_match_p (x, XEXP (y, 0), -1); | |
f4eafc30 | 544 | |
55a2c322 VM |
545 | slow: |
546 | ||
547 | if (code == REG && GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y)) | |
548 | && x == SUBREG_REG (y)) | |
549 | return true; | |
550 | if (GET_CODE (y) == REG && code == SUBREG && REG_P (SUBREG_REG (x)) | |
551 | && SUBREG_REG (x) == y) | |
552 | return true; | |
553 | ||
554 | /* Now we have disposed of all the cases in which different rtx | |
555 | codes can match. */ | |
556 | if (code != GET_CODE (y)) | |
557 | return false; | |
558 | ||
559 | /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */ | |
560 | if (GET_MODE (x) != GET_MODE (y)) | |
561 | return false; | |
562 | ||
563 | switch (code) | |
564 | { | |
565 | CASE_CONST_UNIQUE: | |
566 | return false; | |
567 | ||
568 | case LABEL_REF: | |
569 | return XEXP (x, 0) == XEXP (y, 0); | |
570 | case SYMBOL_REF: | |
571 | return XSTR (x, 0) == XSTR (y, 0); | |
572 | ||
573 | default: | |
574 | break; | |
575 | } | |
576 | ||
577 | /* Compare the elements. If any pair of corresponding elements fail | |
578 | to match, return false for the whole things. */ | |
579 | ||
580 | fmt = GET_RTX_FORMAT (code); | |
581 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
582 | { | |
583 | int val, j; | |
584 | switch (fmt[i]) | |
585 | { | |
586 | case 'w': | |
587 | if (XWINT (x, i) != XWINT (y, i)) | |
588 | return false; | |
589 | break; | |
590 | ||
591 | case 'i': | |
592 | if (XINT (x, i) != XINT (y, i)) | |
593 | return false; | |
594 | break; | |
595 | ||
596 | case 'e': | |
597 | val = operands_match_p (XEXP (x, i), XEXP (y, i), -1); | |
598 | if (val == 0) | |
599 | return false; | |
600 | break; | |
601 | ||
602 | case '0': | |
603 | break; | |
604 | ||
605 | case 'E': | |
606 | if (XVECLEN (x, i) != XVECLEN (y, i)) | |
607 | return false; | |
608 | for (j = XVECLEN (x, i) - 1; j >= 0; --j) | |
609 | { | |
610 | val = operands_match_p (XVECEXP (x, i, j), XVECEXP (y, i, j), -1); | |
611 | if (val == 0) | |
612 | return false; | |
613 | } | |
614 | break; | |
615 | ||
616 | /* It is believed that rtx's at this level will never | |
617 | contain anything but integers and other rtx's, except for | |
618 | within LABEL_REFs and SYMBOL_REFs. */ | |
619 | default: | |
620 | gcc_unreachable (); | |
621 | } | |
622 | } | |
623 | return true; | |
624 | } | |
625 | ||
626 | /* True if X is a constant that can be forced into the constant pool. | |
627 | MODE is the mode of the operand, or VOIDmode if not known. */ | |
628 | #define CONST_POOL_OK_P(MODE, X) \ | |
629 | ((MODE) != VOIDmode \ | |
630 | && CONSTANT_P (X) \ | |
631 | && GET_CODE (X) != HIGH \ | |
632 | && !targetm.cannot_force_const_mem (MODE, X)) | |
633 | ||
634 | /* True if C is a non-empty register class that has too few registers | |
635 | to be safely used as a reload target class. */ | |
636 | #define SMALL_REGISTER_CLASS_P(C) \ | |
637 | (reg_class_size [(C)] == 1 \ | |
638 | || (reg_class_size [(C)] >= 1 && targetm.class_likely_spilled_p (C))) | |
639 | ||
640 | /* If REG is a reload pseudo, try to make its class satisfying CL. */ | |
641 | static void | |
642 | narrow_reload_pseudo_class (rtx reg, enum reg_class cl) | |
643 | { | |
644 | enum reg_class rclass; | |
645 | ||
646 | /* Do not make more accurate class from reloads generated. They are | |
647 | mostly moves with a lot of constraints. Making more accurate | |
648 | class may results in very narrow class and impossibility of find | |
649 | registers for several reloads of one insn. */ | |
650 | if (INSN_UID (curr_insn) >= new_insn_uid_start) | |
651 | return; | |
652 | if (GET_CODE (reg) == SUBREG) | |
653 | reg = SUBREG_REG (reg); | |
654 | if (! REG_P (reg) || (int) REGNO (reg) < new_regno_start) | |
655 | return; | |
656 | if (in_class_p (reg, cl, &rclass) && rclass != cl) | |
657 | change_class (REGNO (reg), rclass, " Change", true); | |
658 | } | |
659 | ||
660 | /* Generate reloads for matching OUT and INS (array of input operand | |
661 | numbers with end marker -1) with reg class GOAL_CLASS. Add input | |
511dcace VM |
662 | and output reloads correspondingly to the lists *BEFORE and *AFTER. |
663 | OUT might be negative. In this case we generate input reloads for | |
664 | matched input operands INS. */ | |
55a2c322 VM |
665 | static void |
666 | match_reload (signed char out, signed char *ins, enum reg_class goal_class, | |
667 | rtx *before, rtx *after) | |
668 | { | |
669 | int i, in; | |
c5cd5a7e | 670 | rtx new_in_reg, new_out_reg, reg, clobber; |
55a2c322 VM |
671 | enum machine_mode inmode, outmode; |
672 | rtx in_rtx = *curr_id->operand_loc[ins[0]]; | |
511dcace | 673 | rtx out_rtx = out < 0 ? in_rtx : *curr_id->operand_loc[out]; |
55a2c322 | 674 | |
55a2c322 | 675 | inmode = curr_operand_mode[ins[0]]; |
511dcace | 676 | outmode = out < 0 ? inmode : curr_operand_mode[out]; |
55a2c322 VM |
677 | push_to_sequence (*before); |
678 | if (inmode != outmode) | |
679 | { | |
680 | if (GET_MODE_SIZE (inmode) > GET_MODE_SIZE (outmode)) | |
681 | { | |
682 | reg = new_in_reg | |
683 | = lra_create_new_reg_with_unique_value (inmode, in_rtx, | |
684 | goal_class, ""); | |
685 | if (SCALAR_INT_MODE_P (inmode)) | |
686 | new_out_reg = gen_lowpart_SUBREG (outmode, reg); | |
687 | else | |
688 | new_out_reg = gen_rtx_SUBREG (outmode, reg, 0); | |
350c0fe7 | 689 | /* If the input reg is dying here, we can use the same hard |
f681cf95 VM |
690 | register for REG and IN_RTX. We do it only for original |
691 | pseudos as reload pseudos can die although original | |
692 | pseudos still live where reload pseudos dies. */ | |
693 | if (REG_P (in_rtx) && (int) REGNO (in_rtx) < lra_new_regno_start | |
350c0fe7 VM |
694 | && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))) |
695 | lra_reg_info[REGNO (reg)].val = lra_reg_info[REGNO (in_rtx)].val; | |
55a2c322 VM |
696 | } |
697 | else | |
698 | { | |
699 | reg = new_out_reg | |
700 | = lra_create_new_reg_with_unique_value (outmode, out_rtx, | |
701 | goal_class, ""); | |
702 | if (SCALAR_INT_MODE_P (outmode)) | |
703 | new_in_reg = gen_lowpart_SUBREG (inmode, reg); | |
704 | else | |
705 | new_in_reg = gen_rtx_SUBREG (inmode, reg, 0); | |
706 | /* NEW_IN_REG is non-paradoxical subreg. We don't want | |
707 | NEW_OUT_REG living above. We add clobber clause for | |
c5cd5a7e VM |
708 | this. This is just a temporary clobber. We can remove |
709 | it at the end of LRA work. */ | |
710 | clobber = emit_clobber (new_out_reg); | |
711 | LRA_TEMP_CLOBBER_P (PATTERN (clobber)) = 1; | |
350c0fe7 VM |
712 | if (GET_CODE (in_rtx) == SUBREG) |
713 | { | |
714 | rtx subreg_reg = SUBREG_REG (in_rtx); | |
715 | ||
716 | /* If SUBREG_REG is dying here and sub-registers IN_RTX | |
717 | and NEW_IN_REG are similar, we can use the same hard | |
718 | register for REG and SUBREG_REG. */ | |
f681cf95 VM |
719 | if (REG_P (subreg_reg) |
720 | && (int) REGNO (subreg_reg) < lra_new_regno_start | |
721 | && GET_MODE (subreg_reg) == outmode | |
350c0fe7 VM |
722 | && SUBREG_BYTE (in_rtx) == SUBREG_BYTE (new_in_reg) |
723 | && find_regno_note (curr_insn, REG_DEAD, REGNO (subreg_reg))) | |
724 | lra_reg_info[REGNO (reg)].val | |
725 | = lra_reg_info[REGNO (subreg_reg)].val; | |
726 | } | |
55a2c322 VM |
727 | } |
728 | } | |
729 | else | |
730 | { | |
731 | /* Pseudos have values -- see comments for lra_reg_info. | |
732 | Different pseudos with the same value do not conflict even if | |
733 | they live in the same place. When we create a pseudo we | |
734 | assign value of original pseudo (if any) from which we | |
735 | created the new pseudo. If we create the pseudo from the | |
736 | input pseudo, the new pseudo will no conflict with the input | |
737 | pseudo which is wrong when the input pseudo lives after the | |
738 | insn and as the new pseudo value is changed by the insn | |
739 | output. Therefore we create the new pseudo from the output. | |
f4eafc30 | 740 | |
55a2c322 VM |
741 | We cannot reuse the current output register because we might |
742 | have a situation like "a <- a op b", where the constraints | |
743 | force the second input operand ("b") to match the output | |
744 | operand ("a"). "b" must then be copied into a new register | |
745 | so that it doesn't clobber the current value of "a". */ | |
f4eafc30 | 746 | |
55a2c322 VM |
747 | new_in_reg = new_out_reg |
748 | = lra_create_new_reg_with_unique_value (outmode, out_rtx, | |
749 | goal_class, ""); | |
750 | } | |
511dcace VM |
751 | /* In operand can be got from transformations before processing insn |
752 | constraints. One example of such transformations is subreg | |
753 | reloading (see function simplify_operand_subreg). The new | |
754 | pseudos created by the transformations might have inaccurate | |
55a2c322 VM |
755 | class (ALL_REGS) and we should make their classes more |
756 | accurate. */ | |
757 | narrow_reload_pseudo_class (in_rtx, goal_class); | |
55a2c322 VM |
758 | lra_emit_move (copy_rtx (new_in_reg), in_rtx); |
759 | *before = get_insns (); | |
760 | end_sequence (); | |
761 | for (i = 0; (in = ins[i]) >= 0; i++) | |
762 | { | |
763 | lra_assert | |
764 | (GET_MODE (*curr_id->operand_loc[in]) == VOIDmode | |
765 | || GET_MODE (new_in_reg) == GET_MODE (*curr_id->operand_loc[in])); | |
766 | *curr_id->operand_loc[in] = new_in_reg; | |
767 | } | |
768 | lra_update_dups (curr_id, ins); | |
511dcace VM |
769 | if (out < 0) |
770 | return; | |
771 | /* See a comment for the input operand above. */ | |
772 | narrow_reload_pseudo_class (out_rtx, goal_class); | |
55a2c322 VM |
773 | if (find_reg_note (curr_insn, REG_UNUSED, out_rtx) == NULL_RTX) |
774 | { | |
775 | start_sequence (); | |
776 | lra_emit_move (out_rtx, copy_rtx (new_out_reg)); | |
777 | emit_insn (*after); | |
778 | *after = get_insns (); | |
779 | end_sequence (); | |
780 | } | |
781 | *curr_id->operand_loc[out] = new_out_reg; | |
782 | lra_update_dup (curr_id, out); | |
783 | } | |
784 | ||
785 | /* Return register class which is union of all reg classes in insn | |
786 | constraint alternative string starting with P. */ | |
787 | static enum reg_class | |
788 | reg_class_from_constraints (const char *p) | |
789 | { | |
790 | int c, len; | |
791 | enum reg_class op_class = NO_REGS; | |
792 | ||
793 | do | |
794 | switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c) | |
795 | { | |
796 | case '#': | |
797 | case ',': | |
798 | return op_class; | |
799 | ||
800 | case 'p': | |
801 | op_class = (reg_class_subunion | |
802 | [op_class][base_reg_class (VOIDmode, ADDR_SPACE_GENERIC, | |
803 | ADDRESS, SCRATCH)]); | |
804 | break; | |
f4eafc30 | 805 | |
55a2c322 VM |
806 | case 'g': |
807 | case 'r': | |
808 | op_class = reg_class_subunion[op_class][GENERAL_REGS]; | |
809 | break; | |
f4eafc30 | 810 | |
55a2c322 VM |
811 | default: |
812 | if (REG_CLASS_FROM_CONSTRAINT (c, p) == NO_REGS) | |
813 | { | |
814 | #ifdef EXTRA_CONSTRAINT_STR | |
815 | if (EXTRA_ADDRESS_CONSTRAINT (c, p)) | |
816 | op_class | |
817 | = (reg_class_subunion | |
818 | [op_class][base_reg_class (VOIDmode, ADDR_SPACE_GENERIC, | |
819 | ADDRESS, SCRATCH)]); | |
820 | #endif | |
821 | break; | |
822 | } | |
f4eafc30 | 823 | |
55a2c322 VM |
824 | op_class |
825 | = reg_class_subunion[op_class][REG_CLASS_FROM_CONSTRAINT (c, p)]; | |
826 | break; | |
827 | } | |
828 | while ((p += len), c); | |
829 | return op_class; | |
830 | } | |
831 | ||
832 | /* If OP is a register, return the class of the register as per | |
833 | get_reg_class, otherwise return NO_REGS. */ | |
834 | static inline enum reg_class | |
835 | get_op_class (rtx op) | |
836 | { | |
837 | return REG_P (op) ? get_reg_class (REGNO (op)) : NO_REGS; | |
838 | } | |
839 | ||
840 | /* Return generated insn mem_pseudo:=val if TO_P or val:=mem_pseudo | |
841 | otherwise. If modes of MEM_PSEUDO and VAL are different, use | |
842 | SUBREG for VAL to make them equal. */ | |
843 | static rtx | |
844 | emit_spill_move (bool to_p, rtx mem_pseudo, rtx val) | |
845 | { | |
846 | if (GET_MODE (mem_pseudo) != GET_MODE (val)) | |
847 | val = gen_rtx_SUBREG (GET_MODE (mem_pseudo), | |
848 | GET_CODE (val) == SUBREG ? SUBREG_REG (val) : val, | |
849 | 0); | |
850 | return (to_p | |
851 | ? gen_move_insn (mem_pseudo, val) | |
852 | : gen_move_insn (val, mem_pseudo)); | |
853 | } | |
854 | ||
855 | /* Process a special case insn (register move), return true if we | |
856 | don't need to process it anymore. Return that RTL was changed | |
857 | through CHANGE_P and macro SECONDARY_MEMORY_NEEDED says to use | |
858 | secondary memory through SEC_MEM_P. */ | |
859 | static bool | |
860 | check_and_process_move (bool *change_p, bool *sec_mem_p) | |
861 | { | |
862 | int sregno, dregno; | |
863 | rtx set, dest, src, dreg, sreg, old_sreg, new_reg, before, scratch_reg; | |
864 | enum reg_class dclass, sclass, secondary_class; | |
865 | enum machine_mode sreg_mode; | |
866 | secondary_reload_info sri; | |
867 | ||
868 | *sec_mem_p = *change_p = false; | |
869 | if ((set = single_set (curr_insn)) == NULL) | |
870 | return false; | |
871 | dreg = dest = SET_DEST (set); | |
872 | sreg = src = SET_SRC (set); | |
873 | /* Quick check on the right move insn which does not need | |
874 | reloads. */ | |
875 | if ((dclass = get_op_class (dest)) != NO_REGS | |
876 | && (sclass = get_op_class (src)) != NO_REGS | |
877 | /* The backend guarantees that register moves of cost 2 never | |
878 | need reloads. */ | |
879 | && targetm.register_move_cost (GET_MODE (src), dclass, sclass) == 2) | |
880 | return true; | |
881 | if (GET_CODE (dest) == SUBREG) | |
882 | dreg = SUBREG_REG (dest); | |
883 | if (GET_CODE (src) == SUBREG) | |
884 | sreg = SUBREG_REG (src); | |
885 | if (! REG_P (dreg) || ! REG_P (sreg)) | |
886 | return false; | |
887 | sclass = dclass = NO_REGS; | |
888 | dreg = get_equiv_substitution (dreg); | |
889 | if (REG_P (dreg)) | |
890 | dclass = get_reg_class (REGNO (dreg)); | |
891 | if (dclass == ALL_REGS) | |
892 | /* ALL_REGS is used for new pseudos created by transformations | |
893 | like reload of SUBREG_REG (see function | |
894 | simplify_operand_subreg). We don't know their class yet. We | |
895 | should figure out the class from processing the insn | |
896 | constraints not in this fast path function. Even if ALL_REGS | |
897 | were a right class for the pseudo, secondary_... hooks usually | |
898 | are not define for ALL_REGS. */ | |
899 | return false; | |
900 | sreg_mode = GET_MODE (sreg); | |
901 | old_sreg = sreg; | |
902 | sreg = get_equiv_substitution (sreg); | |
903 | if (REG_P (sreg)) | |
904 | sclass = get_reg_class (REGNO (sreg)); | |
905 | if (sclass == ALL_REGS) | |
906 | /* See comments above. */ | |
907 | return false; | |
908 | #ifdef SECONDARY_MEMORY_NEEDED | |
909 | if (dclass != NO_REGS && sclass != NO_REGS | |
910 | && SECONDARY_MEMORY_NEEDED (sclass, dclass, GET_MODE (src))) | |
911 | { | |
912 | *sec_mem_p = true; | |
913 | return false; | |
914 | } | |
915 | #endif | |
916 | sri.prev_sri = NULL; | |
917 | sri.icode = CODE_FOR_nothing; | |
918 | sri.extra_cost = 0; | |
919 | secondary_class = NO_REGS; | |
920 | /* Set up hard register for a reload pseudo for hook | |
921 | secondary_reload because some targets just ignore unassigned | |
922 | pseudos in the hook. */ | |
923 | if (dclass != NO_REGS && lra_get_regno_hard_regno (REGNO (dreg)) < 0) | |
924 | { | |
925 | dregno = REGNO (dreg); | |
926 | reg_renumber[dregno] = ira_class_hard_regs[dclass][0]; | |
927 | } | |
928 | else | |
929 | dregno = -1; | |
930 | if (sclass != NO_REGS && lra_get_regno_hard_regno (REGNO (sreg)) < 0) | |
931 | { | |
932 | sregno = REGNO (sreg); | |
933 | reg_renumber[sregno] = ira_class_hard_regs[sclass][0]; | |
934 | } | |
935 | else | |
936 | sregno = -1; | |
937 | if (sclass != NO_REGS) | |
938 | secondary_class | |
939 | = (enum reg_class) targetm.secondary_reload (false, dest, | |
940 | (reg_class_t) sclass, | |
941 | GET_MODE (src), &sri); | |
942 | if (sclass == NO_REGS | |
943 | || ((secondary_class != NO_REGS || sri.icode != CODE_FOR_nothing) | |
944 | && dclass != NO_REGS)) | |
945 | { | |
55a2c322 VM |
946 | enum reg_class old_sclass = secondary_class; |
947 | secondary_reload_info old_sri = sri; | |
55a2c322 VM |
948 | |
949 | sri.prev_sri = NULL; | |
950 | sri.icode = CODE_FOR_nothing; | |
951 | sri.extra_cost = 0; | |
952 | secondary_class | |
953 | = (enum reg_class) targetm.secondary_reload (true, sreg, | |
954 | (reg_class_t) dclass, | |
955 | sreg_mode, &sri); | |
956 | /* Check the target hook consistency. */ | |
957 | lra_assert | |
958 | ((secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing) | |
959 | || (old_sclass == NO_REGS && old_sri.icode == CODE_FOR_nothing) | |
960 | || (secondary_class == old_sclass && sri.icode == old_sri.icode)); | |
961 | } | |
962 | if (sregno >= 0) | |
963 | reg_renumber [sregno] = -1; | |
964 | if (dregno >= 0) | |
965 | reg_renumber [dregno] = -1; | |
966 | if (secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing) | |
967 | return false; | |
968 | *change_p = true; | |
969 | new_reg = NULL_RTX; | |
970 | if (secondary_class != NO_REGS) | |
971 | new_reg = lra_create_new_reg_with_unique_value (sreg_mode, NULL_RTX, | |
972 | secondary_class, | |
973 | "secondary"); | |
974 | start_sequence (); | |
975 | if (old_sreg != sreg) | |
976 | sreg = copy_rtx (sreg); | |
977 | if (sri.icode == CODE_FOR_nothing) | |
978 | lra_emit_move (new_reg, sreg); | |
979 | else | |
980 | { | |
981 | enum reg_class scratch_class; | |
982 | ||
983 | scratch_class = (reg_class_from_constraints | |
984 | (insn_data[sri.icode].operand[2].constraint)); | |
985 | scratch_reg = (lra_create_new_reg_with_unique_value | |
986 | (insn_data[sri.icode].operand[2].mode, NULL_RTX, | |
987 | scratch_class, "scratch")); | |
988 | emit_insn (GEN_FCN (sri.icode) (new_reg != NULL_RTX ? new_reg : dest, | |
989 | sreg, scratch_reg)); | |
990 | } | |
991 | before = get_insns (); | |
992 | end_sequence (); | |
993 | lra_process_new_insns (curr_insn, before, NULL_RTX, "Inserting the move"); | |
994 | if (new_reg != NULL_RTX) | |
995 | { | |
996 | if (GET_CODE (src) == SUBREG) | |
997 | SUBREG_REG (src) = new_reg; | |
998 | else | |
999 | SET_SRC (set) = new_reg; | |
1000 | } | |
1001 | else | |
1002 | { | |
1003 | if (lra_dump_file != NULL) | |
1004 | { | |
1005 | fprintf (lra_dump_file, "Deleting move %u\n", INSN_UID (curr_insn)); | |
cfbeaedf | 1006 | dump_insn_slim (lra_dump_file, curr_insn); |
55a2c322 VM |
1007 | } |
1008 | lra_set_insn_deleted (curr_insn); | |
1009 | return true; | |
1010 | } | |
1011 | return false; | |
1012 | } | |
1013 | ||
1014 | /* The following data describe the result of process_alt_operands. | |
1015 | The data are used in curr_insn_transform to generate reloads. */ | |
1016 | ||
1017 | /* The chosen reg classes which should be used for the corresponding | |
1018 | operands. */ | |
1019 | static enum reg_class goal_alt[MAX_RECOG_OPERANDS]; | |
1020 | /* True if the operand should be the same as another operand and that | |
1021 | other operand does not need a reload. */ | |
1022 | static bool goal_alt_match_win[MAX_RECOG_OPERANDS]; | |
1023 | /* True if the operand does not need a reload. */ | |
1024 | static bool goal_alt_win[MAX_RECOG_OPERANDS]; | |
1025 | /* True if the operand can be offsetable memory. */ | |
1026 | static bool goal_alt_offmemok[MAX_RECOG_OPERANDS]; | |
1027 | /* The number of an operand to which given operand can be matched to. */ | |
1028 | static int goal_alt_matches[MAX_RECOG_OPERANDS]; | |
1029 | /* The number of elements in the following array. */ | |
1030 | static int goal_alt_dont_inherit_ops_num; | |
1031 | /* Numbers of operands whose reload pseudos should not be inherited. */ | |
1032 | static int goal_alt_dont_inherit_ops[MAX_RECOG_OPERANDS]; | |
1033 | /* True if the insn commutative operands should be swapped. */ | |
1034 | static bool goal_alt_swapped; | |
1035 | /* The chosen insn alternative. */ | |
1036 | static int goal_alt_number; | |
1037 | ||
1038 | /* The following five variables are used to choose the best insn | |
1039 | alternative. They reflect final characteristics of the best | |
1040 | alternative. */ | |
1041 | ||
1042 | /* Number of necessary reloads and overall cost reflecting the | |
1043 | previous value and other unpleasantness of the best alternative. */ | |
1044 | static int best_losers, best_overall; | |
1045 | /* Number of small register classes used for operands of the best | |
1046 | alternative. */ | |
1047 | static int best_small_class_operands_num; | |
1048 | /* Overall number hard registers used for reloads. For example, on | |
1049 | some targets we need 2 general registers to reload DFmode and only | |
1050 | one floating point register. */ | |
1051 | static int best_reload_nregs; | |
1052 | /* Overall number reflecting distances of previous reloading the same | |
1053 | value. The distances are counted from the current BB start. It is | |
1054 | used to improve inheritance chances. */ | |
1055 | static int best_reload_sum; | |
1056 | ||
1057 | /* True if the current insn should have no correspondingly input or | |
1058 | output reloads. */ | |
1059 | static bool no_input_reloads_p, no_output_reloads_p; | |
1060 | ||
1061 | /* True if we swapped the commutative operands in the current | |
1062 | insn. */ | |
1063 | static int curr_swapped; | |
1064 | ||
1065 | /* Arrange for address element *LOC to be a register of class CL. | |
1066 | Add any input reloads to list BEFORE. AFTER is nonnull if *LOC is an | |
1067 | automodified value; handle that case by adding the required output | |
1068 | reloads to list AFTER. Return true if the RTL was changed. */ | |
1069 | static bool | |
1070 | process_addr_reg (rtx *loc, rtx *before, rtx *after, enum reg_class cl) | |
1071 | { | |
1072 | int regno; | |
1073 | enum reg_class rclass, new_class; | |
277f65de | 1074 | rtx reg; |
55a2c322 VM |
1075 | rtx new_reg; |
1076 | enum machine_mode mode; | |
1077 | bool before_p = false; | |
1078 | ||
277f65de RS |
1079 | loc = strip_subreg (loc); |
1080 | reg = *loc; | |
55a2c322 VM |
1081 | mode = GET_MODE (reg); |
1082 | if (! REG_P (reg)) | |
1083 | { | |
1084 | /* Always reload memory in an address even if the target supports | |
1085 | such addresses. */ | |
1086 | new_reg = lra_create_new_reg_with_unique_value (mode, reg, cl, "address"); | |
1087 | before_p = true; | |
1088 | } | |
1089 | else | |
1090 | { | |
1091 | regno = REGNO (reg); | |
1092 | rclass = get_reg_class (regno); | |
1093 | if ((*loc = get_equiv_substitution (reg)) != reg) | |
1094 | { | |
1095 | if (lra_dump_file != NULL) | |
1096 | { | |
1097 | fprintf (lra_dump_file, | |
1098 | "Changing pseudo %d in address of insn %u on equiv ", | |
1099 | REGNO (reg), INSN_UID (curr_insn)); | |
cfbeaedf | 1100 | dump_value_slim (lra_dump_file, *loc, 1); |
55a2c322 VM |
1101 | fprintf (lra_dump_file, "\n"); |
1102 | } | |
1103 | *loc = copy_rtx (*loc); | |
1104 | } | |
1105 | if (*loc != reg || ! in_class_p (reg, cl, &new_class)) | |
1106 | { | |
1107 | reg = *loc; | |
1108 | if (get_reload_reg (after == NULL ? OP_IN : OP_INOUT, | |
1109 | mode, reg, cl, "address", &new_reg)) | |
1110 | before_p = true; | |
1111 | } | |
1112 | else if (new_class != NO_REGS && rclass != new_class) | |
1113 | { | |
1114 | change_class (regno, new_class, " Change", true); | |
1115 | return false; | |
1116 | } | |
1117 | else | |
1118 | return false; | |
1119 | } | |
1120 | if (before_p) | |
1121 | { | |
1122 | push_to_sequence (*before); | |
1123 | lra_emit_move (new_reg, reg); | |
1124 | *before = get_insns (); | |
1125 | end_sequence (); | |
1126 | } | |
1127 | *loc = new_reg; | |
1128 | if (after != NULL) | |
1129 | { | |
1130 | start_sequence (); | |
1131 | lra_emit_move (reg, new_reg); | |
1132 | emit_insn (*after); | |
1133 | *after = get_insns (); | |
1134 | end_sequence (); | |
1135 | } | |
1136 | return true; | |
1137 | } | |
1138 | ||
55a2c322 VM |
1139 | /* Make reloads for subreg in operand NOP with internal subreg mode |
1140 | REG_MODE, add new reloads for further processing. Return true if | |
1141 | any reload was generated. */ | |
1142 | static bool | |
1143 | simplify_operand_subreg (int nop, enum machine_mode reg_mode) | |
1144 | { | |
1145 | int hard_regno; | |
1146 | rtx before, after; | |
1147 | enum machine_mode mode; | |
1148 | rtx reg, new_reg; | |
1149 | rtx operand = *curr_id->operand_loc[nop]; | |
1150 | ||
1151 | before = after = NULL_RTX; | |
1152 | ||
1153 | if (GET_CODE (operand) != SUBREG) | |
1154 | return false; | |
f4eafc30 | 1155 | |
55a2c322 VM |
1156 | mode = GET_MODE (operand); |
1157 | reg = SUBREG_REG (operand); | |
1158 | /* If we change address for paradoxical subreg of memory, the | |
1159 | address might violate the necessary alignment or the access might | |
b28ece32 VM |
1160 | be slow. So take this into consideration. We should not worry |
1161 | about access beyond allocated memory for paradoxical memory | |
1162 | subregs as we don't substitute such equiv memory (see processing | |
1163 | equivalences in function lra_constraints) and because for spilled | |
1164 | pseudos we allocate stack memory enough for the biggest | |
1165 | corresponding paradoxical subreg. */ | |
55a2c322 | 1166 | if ((MEM_P (reg) |
08e931f3 | 1167 | && (! SLOW_UNALIGNED_ACCESS (mode, MEM_ALIGN (reg)) |
55a2c322 VM |
1168 | || MEM_ALIGN (reg) >= GET_MODE_ALIGNMENT (mode))) |
1169 | || (REG_P (reg) && REGNO (reg) < FIRST_PSEUDO_REGISTER)) | |
1170 | { | |
1171 | alter_subreg (curr_id->operand_loc[nop], false); | |
1172 | return true; | |
1173 | } | |
1174 | /* Put constant into memory when we have mixed modes. It generates | |
1175 | a better code in most cases as it does not need a secondary | |
1176 | reload memory. It also prevents LRA looping when LRA is using | |
1177 | secondary reload memory again and again. */ | |
1178 | if (CONSTANT_P (reg) && CONST_POOL_OK_P (reg_mode, reg) | |
1179 | && SCALAR_INT_MODE_P (reg_mode) != SCALAR_INT_MODE_P (mode)) | |
1180 | { | |
1181 | SUBREG_REG (operand) = force_const_mem (reg_mode, reg); | |
1182 | alter_subreg (curr_id->operand_loc[nop], false); | |
1183 | return true; | |
1184 | } | |
1185 | /* Force a reload of the SUBREG_REG if this is a constant or PLUS or | |
1186 | if there may be a problem accessing OPERAND in the outer | |
1187 | mode. */ | |
1188 | if ((REG_P (reg) | |
1189 | && REGNO (reg) >= FIRST_PSEUDO_REGISTER | |
1190 | && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0 | |
1191 | /* Don't reload paradoxical subregs because we could be looping | |
1192 | having repeatedly final regno out of hard regs range. */ | |
1193 | && (hard_regno_nregs[hard_regno][GET_MODE (reg)] | |
1194 | >= hard_regno_nregs[hard_regno][mode]) | |
1195 | && simplify_subreg_regno (hard_regno, GET_MODE (reg), | |
1196 | SUBREG_BYTE (operand), mode) < 0) | |
1197 | || CONSTANT_P (reg) || GET_CODE (reg) == PLUS || MEM_P (reg)) | |
1198 | { | |
1199 | enum op_type type = curr_static_id->operand[nop].type; | |
1200 | /* The class will be defined later in curr_insn_transform. */ | |
1201 | enum reg_class rclass | |
1202 | = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS); | |
1203 | ||
1204 | new_reg = lra_create_new_reg_with_unique_value (reg_mode, reg, rclass, | |
1205 | "subreg reg"); | |
1206 | bitmap_set_bit (&lra_optional_reload_pseudos, REGNO (new_reg)); | |
1207 | if (type != OP_OUT | |
1208 | || GET_MODE_SIZE (GET_MODE (reg)) > GET_MODE_SIZE (mode)) | |
1209 | { | |
1210 | push_to_sequence (before); | |
1211 | lra_emit_move (new_reg, reg); | |
1212 | before = get_insns (); | |
1213 | end_sequence (); | |
1214 | } | |
1215 | if (type != OP_IN) | |
1216 | { | |
1217 | start_sequence (); | |
1218 | lra_emit_move (reg, new_reg); | |
1219 | emit_insn (after); | |
1220 | after = get_insns (); | |
1221 | end_sequence (); | |
1222 | } | |
1223 | SUBREG_REG (operand) = new_reg; | |
1224 | lra_process_new_insns (curr_insn, before, after, | |
1225 | "Inserting subreg reload"); | |
1226 | return true; | |
1227 | } | |
1228 | return false; | |
1229 | } | |
1230 | ||
1231 | /* Return TRUE if X refers for a hard register from SET. */ | |
1232 | static bool | |
1233 | uses_hard_regs_p (rtx x, HARD_REG_SET set) | |
1234 | { | |
1235 | int i, j, x_hard_regno; | |
1236 | enum machine_mode mode; | |
1237 | const char *fmt; | |
1238 | enum rtx_code code; | |
1239 | ||
1240 | if (x == NULL_RTX) | |
1241 | return false; | |
1242 | code = GET_CODE (x); | |
1243 | mode = GET_MODE (x); | |
1244 | if (code == SUBREG) | |
1245 | { | |
1246 | x = SUBREG_REG (x); | |
1247 | code = GET_CODE (x); | |
1248 | if (GET_MODE_SIZE (GET_MODE (x)) > GET_MODE_SIZE (mode)) | |
1249 | mode = GET_MODE (x); | |
1250 | } | |
f4eafc30 | 1251 | |
55a2c322 VM |
1252 | if (REG_P (x)) |
1253 | { | |
1254 | x_hard_regno = get_hard_regno (x); | |
1255 | return (x_hard_regno >= 0 | |
1256 | && overlaps_hard_reg_set_p (set, mode, x_hard_regno)); | |
1257 | } | |
1258 | if (MEM_P (x)) | |
1259 | { | |
277f65de | 1260 | struct address_info ad; |
55a2c322 | 1261 | |
277f65de RS |
1262 | decompose_mem_address (&ad, x); |
1263 | if (ad.base_term != NULL && uses_hard_regs_p (*ad.base_term, set)) | |
1264 | return true; | |
1265 | if (ad.index_term != NULL && uses_hard_regs_p (*ad.index_term, set)) | |
1266 | return true; | |
55a2c322 VM |
1267 | } |
1268 | fmt = GET_RTX_FORMAT (code); | |
1269 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
1270 | { | |
1271 | if (fmt[i] == 'e') | |
1272 | { | |
1273 | if (uses_hard_regs_p (XEXP (x, i), set)) | |
1274 | return true; | |
1275 | } | |
1276 | else if (fmt[i] == 'E') | |
1277 | { | |
1278 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
1279 | if (uses_hard_regs_p (XVECEXP (x, i, j), set)) | |
1280 | return true; | |
1281 | } | |
1282 | } | |
1283 | return false; | |
1284 | } | |
1285 | ||
1286 | /* Return true if OP is a spilled pseudo. */ | |
1287 | static inline bool | |
1288 | spilled_pseudo_p (rtx op) | |
1289 | { | |
1290 | return (REG_P (op) | |
1291 | && REGNO (op) >= FIRST_PSEUDO_REGISTER && in_mem_p (REGNO (op))); | |
1292 | } | |
1293 | ||
1294 | /* Return true if X is a general constant. */ | |
1295 | static inline bool | |
1296 | general_constant_p (rtx x) | |
1297 | { | |
1298 | return CONSTANT_P (x) && (! flag_pic || LEGITIMATE_PIC_OPERAND_P (x)); | |
1299 | } | |
1300 | ||
55a2c322 VM |
1301 | /* Major function to choose the current insn alternative and what |
1302 | operands should be reloaded and how. If ONLY_ALTERNATIVE is not | |
1303 | negative we should consider only this alternative. Return false if | |
1304 | we can not choose the alternative or find how to reload the | |
1305 | operands. */ | |
1306 | static bool | |
1307 | process_alt_operands (int only_alternative) | |
1308 | { | |
1309 | bool ok_p = false; | |
1310 | int nop, small_class_operands_num, overall, nalt; | |
1311 | int n_alternatives = curr_static_id->n_alternatives; | |
1312 | int n_operands = curr_static_id->n_operands; | |
1313 | /* LOSERS counts the operands that don't fit this alternative and | |
1314 | would require loading. */ | |
1315 | int losers; | |
1316 | /* REJECT is a count of how undesirable this alternative says it is | |
1317 | if any reloading is required. If the alternative matches exactly | |
1318 | then REJECT is ignored, but otherwise it gets this much counted | |
1319 | against it in addition to the reloading needed. */ | |
1320 | int reject; | |
1321 | /* The number of elements in the following array. */ | |
1322 | int early_clobbered_regs_num; | |
1323 | /* Numbers of operands which are early clobber registers. */ | |
1324 | int early_clobbered_nops[MAX_RECOG_OPERANDS]; | |
1325 | enum reg_class curr_alt[MAX_RECOG_OPERANDS]; | |
1326 | HARD_REG_SET curr_alt_set[MAX_RECOG_OPERANDS]; | |
1327 | bool curr_alt_match_win[MAX_RECOG_OPERANDS]; | |
1328 | bool curr_alt_win[MAX_RECOG_OPERANDS]; | |
1329 | bool curr_alt_offmemok[MAX_RECOG_OPERANDS]; | |
1330 | int curr_alt_matches[MAX_RECOG_OPERANDS]; | |
1331 | /* The number of elements in the following array. */ | |
1332 | int curr_alt_dont_inherit_ops_num; | |
1333 | /* Numbers of operands whose reload pseudos should not be inherited. */ | |
1334 | int curr_alt_dont_inherit_ops[MAX_RECOG_OPERANDS]; | |
1335 | rtx op; | |
1336 | /* The register when the operand is a subreg of register, otherwise the | |
1337 | operand itself. */ | |
1338 | rtx no_subreg_reg_operand[MAX_RECOG_OPERANDS]; | |
1339 | /* The register if the operand is a register or subreg of register, | |
1340 | otherwise NULL. */ | |
1341 | rtx operand_reg[MAX_RECOG_OPERANDS]; | |
1342 | int hard_regno[MAX_RECOG_OPERANDS]; | |
1343 | enum machine_mode biggest_mode[MAX_RECOG_OPERANDS]; | |
1344 | int reload_nregs, reload_sum; | |
1345 | bool costly_p; | |
1346 | enum reg_class cl; | |
1347 | ||
1348 | /* Calculate some data common for all alternatives to speed up the | |
1349 | function. */ | |
1350 | for (nop = 0; nop < n_operands; nop++) | |
1351 | { | |
1352 | op = no_subreg_reg_operand[nop] = *curr_id->operand_loc[nop]; | |
1353 | /* The real hard regno of the operand after the allocation. */ | |
1354 | hard_regno[nop] = get_hard_regno (op); | |
f4eafc30 | 1355 | |
55a2c322 VM |
1356 | operand_reg[nop] = op; |
1357 | biggest_mode[nop] = GET_MODE (operand_reg[nop]); | |
1358 | if (GET_CODE (operand_reg[nop]) == SUBREG) | |
1359 | { | |
1360 | operand_reg[nop] = SUBREG_REG (operand_reg[nop]); | |
1361 | if (GET_MODE_SIZE (biggest_mode[nop]) | |
1362 | < GET_MODE_SIZE (GET_MODE (operand_reg[nop]))) | |
1363 | biggest_mode[nop] = GET_MODE (operand_reg[nop]); | |
1364 | } | |
1365 | if (REG_P (operand_reg[nop])) | |
1366 | no_subreg_reg_operand[nop] = operand_reg[nop]; | |
1367 | else | |
1368 | operand_reg[nop] = NULL_RTX; | |
1369 | } | |
1370 | ||
1371 | /* The constraints are made of several alternatives. Each operand's | |
1372 | constraint looks like foo,bar,... with commas separating the | |
1373 | alternatives. The first alternatives for all operands go | |
1374 | together, the second alternatives go together, etc. | |
1375 | ||
1376 | First loop over alternatives. */ | |
1377 | for (nalt = 0; nalt < n_alternatives; nalt++) | |
1378 | { | |
1379 | /* Loop over operands for one constraint alternative. */ | |
1380 | #ifdef HAVE_ATTR_enabled | |
1381 | if (curr_id->alternative_enabled_p != NULL | |
1382 | && ! curr_id->alternative_enabled_p[nalt]) | |
1383 | continue; | |
1384 | #endif | |
1385 | ||
1386 | if (only_alternative >= 0 && nalt != only_alternative) | |
1387 | continue; | |
1388 | ||
1389 | overall = losers = reject = reload_nregs = reload_sum = 0; | |
1390 | for (nop = 0; nop < n_operands; nop++) | |
1391 | reject += (curr_static_id | |
1392 | ->operand_alternative[nalt * n_operands + nop].reject); | |
1393 | early_clobbered_regs_num = 0; | |
1394 | ||
1395 | for (nop = 0; nop < n_operands; nop++) | |
1396 | { | |
1397 | const char *p; | |
1398 | char *end; | |
1399 | int len, c, m, i, opalt_num, this_alternative_matches; | |
1400 | bool win, did_match, offmemok, early_clobber_p; | |
1401 | /* false => this operand can be reloaded somehow for this | |
1402 | alternative. */ | |
1403 | bool badop; | |
1404 | /* true => this operand can be reloaded if the alternative | |
1405 | allows regs. */ | |
1406 | bool winreg; | |
1407 | /* True if a constant forced into memory would be OK for | |
1408 | this operand. */ | |
1409 | bool constmemok; | |
1410 | enum reg_class this_alternative, this_costly_alternative; | |
1411 | HARD_REG_SET this_alternative_set, this_costly_alternative_set; | |
1412 | bool this_alternative_match_win, this_alternative_win; | |
1413 | bool this_alternative_offmemok; | |
1414 | enum machine_mode mode; | |
1415 | ||
1416 | opalt_num = nalt * n_operands + nop; | |
1417 | if (curr_static_id->operand_alternative[opalt_num].anything_ok) | |
1418 | { | |
1419 | /* Fast track for no constraints at all. */ | |
1420 | curr_alt[nop] = NO_REGS; | |
1421 | CLEAR_HARD_REG_SET (curr_alt_set[nop]); | |
1422 | curr_alt_win[nop] = true; | |
1423 | curr_alt_match_win[nop] = false; | |
1424 | curr_alt_offmemok[nop] = false; | |
1425 | curr_alt_matches[nop] = -1; | |
1426 | continue; | |
1427 | } | |
f4eafc30 | 1428 | |
55a2c322 VM |
1429 | op = no_subreg_reg_operand[nop]; |
1430 | mode = curr_operand_mode[nop]; | |
1431 | ||
1432 | win = did_match = winreg = offmemok = constmemok = false; | |
1433 | badop = true; | |
f4eafc30 | 1434 | |
55a2c322 VM |
1435 | early_clobber_p = false; |
1436 | p = curr_static_id->operand_alternative[opalt_num].constraint; | |
f4eafc30 | 1437 | |
55a2c322 VM |
1438 | this_costly_alternative = this_alternative = NO_REGS; |
1439 | /* We update set of possible hard regs besides its class | |
1440 | because reg class might be inaccurate. For example, | |
1441 | union of LO_REGS (l), HI_REGS(h), and STACK_REG(k) in ARM | |
1442 | is translated in HI_REGS because classes are merged by | |
1443 | pairs and there is no accurate intermediate class. */ | |
1444 | CLEAR_HARD_REG_SET (this_alternative_set); | |
1445 | CLEAR_HARD_REG_SET (this_costly_alternative_set); | |
1446 | this_alternative_win = false; | |
1447 | this_alternative_match_win = false; | |
1448 | this_alternative_offmemok = false; | |
1449 | this_alternative_matches = -1; | |
f4eafc30 | 1450 | |
55a2c322 VM |
1451 | /* An empty constraint should be excluded by the fast |
1452 | track. */ | |
1453 | lra_assert (*p != 0 && *p != ','); | |
f4eafc30 | 1454 | |
55a2c322 VM |
1455 | /* Scan this alternative's specs for this operand; set WIN |
1456 | if the operand fits any letter in this alternative. | |
1457 | Otherwise, clear BADOP if this operand could fit some | |
1458 | letter after reloads, or set WINREG if this operand could | |
1459 | fit after reloads provided the constraint allows some | |
1460 | registers. */ | |
1461 | costly_p = false; | |
1462 | do | |
1463 | { | |
1464 | switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c) | |
1465 | { | |
1466 | case '\0': | |
1467 | len = 0; | |
1468 | break; | |
1469 | case ',': | |
1470 | c = '\0'; | |
1471 | break; | |
f4eafc30 | 1472 | |
55a2c322 VM |
1473 | case '=': case '+': case '?': case '*': case '!': |
1474 | case ' ': case '\t': | |
1475 | break; | |
f4eafc30 | 1476 | |
55a2c322 VM |
1477 | case '%': |
1478 | /* We only support one commutative marker, the first | |
1479 | one. We already set commutative above. */ | |
1480 | break; | |
f4eafc30 | 1481 | |
55a2c322 VM |
1482 | case '&': |
1483 | early_clobber_p = true; | |
1484 | break; | |
f4eafc30 | 1485 | |
55a2c322 VM |
1486 | case '#': |
1487 | /* Ignore rest of this alternative. */ | |
1488 | c = '\0'; | |
1489 | break; | |
f4eafc30 | 1490 | |
55a2c322 VM |
1491 | case '0': case '1': case '2': case '3': case '4': |
1492 | case '5': case '6': case '7': case '8': case '9': | |
1493 | { | |
1494 | int m_hregno; | |
1495 | bool match_p; | |
f4eafc30 | 1496 | |
55a2c322 VM |
1497 | m = strtoul (p, &end, 10); |
1498 | p = end; | |
1499 | len = 0; | |
1500 | lra_assert (nop > m); | |
f4eafc30 | 1501 | |
55a2c322 VM |
1502 | this_alternative_matches = m; |
1503 | m_hregno = get_hard_regno (*curr_id->operand_loc[m]); | |
1504 | /* We are supposed to match a previous operand. | |
1505 | If we do, we win if that one did. If we do | |
1506 | not, count both of the operands as losers. | |
1507 | (This is too conservative, since most of the | |
1508 | time only a single reload insn will be needed | |
1509 | to make the two operands win. As a result, | |
1510 | this alternative may be rejected when it is | |
1511 | actually desirable.) */ | |
1512 | match_p = false; | |
1513 | if (operands_match_p (*curr_id->operand_loc[nop], | |
1514 | *curr_id->operand_loc[m], m_hregno)) | |
1515 | { | |
1516 | /* We should reject matching of an early | |
1517 | clobber operand if the matching operand is | |
1518 | not dying in the insn. */ | |
1519 | if (! curr_static_id->operand[m].early_clobber | |
1520 | || operand_reg[nop] == NULL_RTX | |
1521 | || (find_regno_note (curr_insn, REG_DEAD, | |
1522 | REGNO (operand_reg[nop])) | |
1523 | != NULL_RTX)) | |
1524 | match_p = true; | |
1525 | } | |
1526 | if (match_p) | |
1527 | { | |
1528 | /* If we are matching a non-offsettable | |
1529 | address where an offsettable address was | |
1530 | expected, then we must reject this | |
1531 | combination, because we can't reload | |
1532 | it. */ | |
1533 | if (curr_alt_offmemok[m] | |
1534 | && MEM_P (*curr_id->operand_loc[m]) | |
1535 | && curr_alt[m] == NO_REGS && ! curr_alt_win[m]) | |
1536 | continue; | |
f4eafc30 | 1537 | |
55a2c322 VM |
1538 | } |
1539 | else | |
1540 | { | |
1541 | /* Operands don't match. Both operands must | |
1542 | allow a reload register, otherwise we | |
1543 | cannot make them match. */ | |
1544 | if (curr_alt[m] == NO_REGS) | |
1545 | break; | |
1546 | /* Retroactively mark the operand we had to | |
1547 | match as a loser, if it wasn't already and | |
1548 | it wasn't matched to a register constraint | |
1549 | (e.g it might be matched by memory). */ | |
1550 | if (curr_alt_win[m] | |
1551 | && (operand_reg[m] == NULL_RTX | |
1552 | || hard_regno[m] < 0)) | |
1553 | { | |
1554 | losers++; | |
1555 | reload_nregs | |
1556 | += (ira_reg_class_max_nregs[curr_alt[m]] | |
1557 | [GET_MODE (*curr_id->operand_loc[m])]); | |
1558 | } | |
f4eafc30 | 1559 | |
55a2c322 VM |
1560 | /* We prefer no matching alternatives because |
1561 | it gives more freedom in RA. */ | |
1562 | if (operand_reg[nop] == NULL_RTX | |
1563 | || (find_regno_note (curr_insn, REG_DEAD, | |
1564 | REGNO (operand_reg[nop])) | |
1565 | == NULL_RTX)) | |
1566 | reject += 2; | |
1567 | } | |
1568 | /* If we have to reload this operand and some | |
1569 | previous operand also had to match the same | |
1570 | thing as this operand, we don't know how to do | |
1571 | that. */ | |
1572 | if (!match_p || !curr_alt_win[m]) | |
1573 | { | |
1574 | for (i = 0; i < nop; i++) | |
1575 | if (curr_alt_matches[i] == m) | |
1576 | break; | |
1577 | if (i < nop) | |
1578 | break; | |
1579 | } | |
1580 | else | |
1581 | did_match = true; | |
f4eafc30 | 1582 | |
55a2c322 VM |
1583 | /* This can be fixed with reloads if the operand |
1584 | we are supposed to match can be fixed with | |
1585 | reloads. */ | |
1586 | badop = false; | |
1587 | this_alternative = curr_alt[m]; | |
1588 | COPY_HARD_REG_SET (this_alternative_set, curr_alt_set[m]); | |
821b7577 | 1589 | winreg = this_alternative != NO_REGS; |
55a2c322 VM |
1590 | break; |
1591 | } | |
f4eafc30 | 1592 | |
55a2c322 VM |
1593 | case 'p': |
1594 | cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC, | |
1595 | ADDRESS, SCRATCH); | |
1596 | this_alternative = reg_class_subunion[this_alternative][cl]; | |
1597 | IOR_HARD_REG_SET (this_alternative_set, | |
1598 | reg_class_contents[cl]); | |
1599 | if (costly_p) | |
1600 | { | |
1601 | this_costly_alternative | |
1602 | = reg_class_subunion[this_costly_alternative][cl]; | |
1603 | IOR_HARD_REG_SET (this_costly_alternative_set, | |
1604 | reg_class_contents[cl]); | |
1605 | } | |
1606 | win = true; | |
1607 | badop = false; | |
1608 | break; | |
f4eafc30 | 1609 | |
55a2c322 VM |
1610 | case TARGET_MEM_CONSTRAINT: |
1611 | if (MEM_P (op) || spilled_pseudo_p (op)) | |
1612 | win = true; | |
1bdc4b11 VM |
1613 | /* We can put constant or pseudo value into memory |
1614 | to satisfy the constraint. */ | |
1615 | if (CONST_POOL_OK_P (mode, op) || REG_P (op)) | |
55a2c322 VM |
1616 | badop = false; |
1617 | constmemok = true; | |
1618 | break; | |
f4eafc30 | 1619 | |
55a2c322 VM |
1620 | case '<': |
1621 | if (MEM_P (op) | |
1622 | && (GET_CODE (XEXP (op, 0)) == PRE_DEC | |
1623 | || GET_CODE (XEXP (op, 0)) == POST_DEC)) | |
1624 | win = true; | |
1625 | break; | |
f4eafc30 | 1626 | |
55a2c322 VM |
1627 | case '>': |
1628 | if (MEM_P (op) | |
1629 | && (GET_CODE (XEXP (op, 0)) == PRE_INC | |
1630 | || GET_CODE (XEXP (op, 0)) == POST_INC)) | |
1631 | win = true; | |
1632 | break; | |
f4eafc30 | 1633 | |
55a2c322 VM |
1634 | /* Memory op whose address is not offsettable. */ |
1635 | case 'V': | |
1636 | if (MEM_P (op) | |
1637 | && ! offsettable_nonstrict_memref_p (op)) | |
1638 | win = true; | |
1639 | break; | |
f4eafc30 | 1640 | |
55a2c322 VM |
1641 | /* Memory operand whose address is offsettable. */ |
1642 | case 'o': | |
1643 | if ((MEM_P (op) | |
1644 | && offsettable_nonstrict_memref_p (op)) | |
1645 | || spilled_pseudo_p (op)) | |
1646 | win = true; | |
1bdc4b11 VM |
1647 | /* We can put constant or pseudo value into memory |
1648 | or make memory address offsetable to satisfy the | |
1649 | constraint. */ | |
1650 | if (CONST_POOL_OK_P (mode, op) || MEM_P (op) || REG_P (op)) | |
55a2c322 VM |
1651 | badop = false; |
1652 | constmemok = true; | |
1653 | offmemok = true; | |
1654 | break; | |
f4eafc30 | 1655 | |
55a2c322 VM |
1656 | case 'E': |
1657 | case 'F': | |
1658 | if (GET_CODE (op) == CONST_DOUBLE | |
1659 | || (GET_CODE (op) == CONST_VECTOR | |
1660 | && (GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT))) | |
1661 | win = true; | |
1662 | break; | |
f4eafc30 | 1663 | |
55a2c322 VM |
1664 | case 'G': |
1665 | case 'H': | |
1666 | if (GET_CODE (op) == CONST_DOUBLE | |
1667 | && CONST_DOUBLE_OK_FOR_CONSTRAINT_P (op, c, p)) | |
1668 | win = true; | |
1669 | break; | |
f4eafc30 | 1670 | |
55a2c322 VM |
1671 | case 's': |
1672 | if (CONST_INT_P (op) | |
1673 | || (GET_CODE (op) == CONST_DOUBLE && mode == VOIDmode)) | |
1674 | break; | |
1bdc4b11 | 1675 | |
55a2c322 VM |
1676 | case 'i': |
1677 | if (general_constant_p (op)) | |
1678 | win = true; | |
1679 | break; | |
f4eafc30 | 1680 | |
55a2c322 VM |
1681 | case 'n': |
1682 | if (CONST_INT_P (op) | |
1683 | || (GET_CODE (op) == CONST_DOUBLE && mode == VOIDmode)) | |
1684 | win = true; | |
1685 | break; | |
f4eafc30 | 1686 | |
55a2c322 VM |
1687 | case 'I': |
1688 | case 'J': | |
1689 | case 'K': | |
1690 | case 'L': | |
1691 | case 'M': | |
1692 | case 'N': | |
1693 | case 'O': | |
1694 | case 'P': | |
1695 | if (CONST_INT_P (op) | |
1696 | && CONST_OK_FOR_CONSTRAINT_P (INTVAL (op), c, p)) | |
1697 | win = true; | |
1698 | break; | |
f4eafc30 | 1699 | |
55a2c322 VM |
1700 | case 'X': |
1701 | /* This constraint should be excluded by the fast | |
1702 | track. */ | |
1703 | gcc_unreachable (); | |
1704 | break; | |
f4eafc30 | 1705 | |
55a2c322 VM |
1706 | case 'g': |
1707 | if (MEM_P (op) | |
1708 | || general_constant_p (op) | |
1709 | || spilled_pseudo_p (op)) | |
1710 | win = true; | |
1711 | /* Drop through into 'r' case. */ | |
f4eafc30 | 1712 | |
55a2c322 VM |
1713 | case 'r': |
1714 | this_alternative | |
1715 | = reg_class_subunion[this_alternative][GENERAL_REGS]; | |
1716 | IOR_HARD_REG_SET (this_alternative_set, | |
1717 | reg_class_contents[GENERAL_REGS]); | |
1718 | if (costly_p) | |
1719 | { | |
1720 | this_costly_alternative | |
1721 | = (reg_class_subunion | |
1722 | [this_costly_alternative][GENERAL_REGS]); | |
1723 | IOR_HARD_REG_SET (this_costly_alternative_set, | |
1724 | reg_class_contents[GENERAL_REGS]); | |
1725 | } | |
1726 | goto reg; | |
f4eafc30 | 1727 | |
55a2c322 VM |
1728 | default: |
1729 | if (REG_CLASS_FROM_CONSTRAINT (c, p) == NO_REGS) | |
1730 | { | |
1731 | #ifdef EXTRA_CONSTRAINT_STR | |
1732 | if (EXTRA_MEMORY_CONSTRAINT (c, p)) | |
1733 | { | |
1734 | if (EXTRA_CONSTRAINT_STR (op, c, p)) | |
1735 | win = true; | |
1736 | else if (spilled_pseudo_p (op)) | |
1737 | win = true; | |
f4eafc30 | 1738 | |
55a2c322 | 1739 | /* If we didn't already win, we can reload |
1bdc4b11 VM |
1740 | constants via force_const_mem or put the |
1741 | pseudo value into memory, or make other | |
1742 | memory by reloading the address like for | |
55a2c322 | 1743 | 'o'. */ |
1bdc4b11 VM |
1744 | if (CONST_POOL_OK_P (mode, op) |
1745 | || MEM_P (op) || REG_P (op)) | |
55a2c322 VM |
1746 | badop = false; |
1747 | constmemok = true; | |
1748 | offmemok = true; | |
1749 | break; | |
1750 | } | |
1751 | if (EXTRA_ADDRESS_CONSTRAINT (c, p)) | |
1752 | { | |
1753 | if (EXTRA_CONSTRAINT_STR (op, c, p)) | |
1754 | win = true; | |
f4eafc30 | 1755 | |
55a2c322 VM |
1756 | /* If we didn't already win, we can reload |
1757 | the address into a base register. */ | |
1758 | cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC, | |
1759 | ADDRESS, SCRATCH); | |
1760 | this_alternative | |
1761 | = reg_class_subunion[this_alternative][cl]; | |
1762 | IOR_HARD_REG_SET (this_alternative_set, | |
1763 | reg_class_contents[cl]); | |
1764 | if (costly_p) | |
1765 | { | |
1766 | this_costly_alternative | |
1767 | = (reg_class_subunion | |
1768 | [this_costly_alternative][cl]); | |
1769 | IOR_HARD_REG_SET (this_costly_alternative_set, | |
1770 | reg_class_contents[cl]); | |
1771 | } | |
1772 | badop = false; | |
1773 | break; | |
1774 | } | |
f4eafc30 | 1775 | |
55a2c322 VM |
1776 | if (EXTRA_CONSTRAINT_STR (op, c, p)) |
1777 | win = true; | |
1778 | #endif | |
1779 | break; | |
1780 | } | |
f4eafc30 | 1781 | |
55a2c322 VM |
1782 | cl = REG_CLASS_FROM_CONSTRAINT (c, p); |
1783 | this_alternative = reg_class_subunion[this_alternative][cl]; | |
1784 | IOR_HARD_REG_SET (this_alternative_set, | |
1785 | reg_class_contents[cl]); | |
1786 | if (costly_p) | |
1787 | { | |
1788 | this_costly_alternative | |
1789 | = reg_class_subunion[this_costly_alternative][cl]; | |
1790 | IOR_HARD_REG_SET (this_costly_alternative_set, | |
1791 | reg_class_contents[cl]); | |
1792 | } | |
1793 | reg: | |
1794 | if (mode == BLKmode) | |
1795 | break; | |
1796 | winreg = true; | |
1797 | if (REG_P (op)) | |
1798 | { | |
1799 | if (hard_regno[nop] >= 0 | |
1800 | && in_hard_reg_set_p (this_alternative_set, | |
1801 | mode, hard_regno[nop])) | |
1802 | win = true; | |
1803 | else if (hard_regno[nop] < 0 | |
1804 | && in_class_p (op, this_alternative, NULL)) | |
1805 | win = true; | |
1806 | } | |
1807 | break; | |
1808 | } | |
1809 | if (c != ' ' && c != '\t') | |
1810 | costly_p = c == '*'; | |
1811 | } | |
1812 | while ((p += len), c); | |
f4eafc30 | 1813 | |
55a2c322 VM |
1814 | /* Record which operands fit this alternative. */ |
1815 | if (win) | |
1816 | { | |
1817 | this_alternative_win = true; | |
1818 | if (operand_reg[nop] != NULL_RTX) | |
1819 | { | |
1820 | if (hard_regno[nop] >= 0) | |
1821 | { | |
1822 | if (in_hard_reg_set_p (this_costly_alternative_set, | |
1823 | mode, hard_regno[nop])) | |
1824 | reject++; | |
1825 | } | |
1826 | else | |
1827 | { | |
1828 | /* Prefer won reg to spilled pseudo under other equal | |
1829 | conditions. */ | |
1830 | reject++; | |
1831 | if (in_class_p (operand_reg[nop], | |
1832 | this_costly_alternative, NULL)) | |
1833 | reject++; | |
1834 | } | |
1835 | /* We simulate the behaviour of old reload here. | |
1836 | Although scratches need hard registers and it | |
1837 | might result in spilling other pseudos, no reload | |
1838 | insns are generated for the scratches. So it | |
1839 | might cost something but probably less than old | |
1840 | reload pass believes. */ | |
1841 | if (lra_former_scratch_p (REGNO (operand_reg[nop]))) | |
821b7577 | 1842 | reject += LRA_LOSER_COST_FACTOR; |
55a2c322 VM |
1843 | } |
1844 | } | |
1845 | else if (did_match) | |
1846 | this_alternative_match_win = true; | |
1847 | else | |
1848 | { | |
1849 | int const_to_mem = 0; | |
1850 | bool no_regs_p; | |
1851 | ||
e86c0101 SB |
1852 | /* If this alternative asks for a specific reg class, see if there |
1853 | is at least one allocatable register in that class. */ | |
55a2c322 VM |
1854 | no_regs_p |
1855 | = (this_alternative == NO_REGS | |
1856 | || (hard_reg_set_subset_p | |
1857 | (reg_class_contents[this_alternative], | |
1858 | lra_no_alloc_regs))); | |
e86c0101 SB |
1859 | |
1860 | /* For asms, verify that the class for this alternative is possible | |
1861 | for the mode that is specified. */ | |
1862 | if (!no_regs_p && REG_P (op) && INSN_CODE (curr_insn) < 0) | |
1863 | { | |
1864 | int i; | |
1865 | for (i = 0; i < FIRST_PSEUDO_REGISTER; i++) | |
1866 | if (HARD_REGNO_MODE_OK (i, mode) | |
1867 | && in_hard_reg_set_p (reg_class_contents[this_alternative], mode, i)) | |
1868 | break; | |
1869 | if (i == FIRST_PSEUDO_REGISTER) | |
1870 | winreg = false; | |
1871 | } | |
1872 | ||
55a2c322 VM |
1873 | /* If this operand accepts a register, and if the |
1874 | register class has at least one allocatable register, | |
1875 | then this operand can be reloaded. */ | |
1876 | if (winreg && !no_regs_p) | |
1877 | badop = false; | |
f4eafc30 | 1878 | |
55a2c322 VM |
1879 | if (badop) |
1880 | goto fail; | |
1881 | ||
1882 | this_alternative_offmemok = offmemok; | |
1883 | if (this_costly_alternative != NO_REGS) | |
1884 | reject++; | |
1885 | /* If the operand is dying, has a matching constraint, | |
1886 | and satisfies constraints of the matched operand | |
1887 | which failed to satisfy the own constraints, we do | |
1888 | not need to generate a reload insn for this | |
1889 | operand. */ | |
1890 | if (!(this_alternative_matches >= 0 | |
1891 | && !curr_alt_win[this_alternative_matches] | |
1892 | && REG_P (op) | |
1893 | && find_regno_note (curr_insn, REG_DEAD, REGNO (op)) | |
1894 | && (hard_regno[nop] >= 0 | |
1895 | ? in_hard_reg_set_p (this_alternative_set, | |
1896 | mode, hard_regno[nop]) | |
1897 | : in_class_p (op, this_alternative, NULL)))) | |
027ece11 VM |
1898 | { |
1899 | /* Strict_low_part requires reload the register not | |
1900 | the sub-register. In this case we should check | |
1901 | that a final reload hard reg can hold the | |
1902 | value. */ | |
1903 | if (curr_static_id->operand[nop].strict_low | |
1904 | && REG_P (op) | |
1905 | && hard_regno[nop] < 0 | |
1906 | && GET_CODE (*curr_id->operand_loc[nop]) == SUBREG | |
1907 | && ira_class_hard_regs_num[this_alternative] > 0 | |
1908 | && ! HARD_REGNO_MODE_OK (ira_class_hard_regs | |
1909 | [this_alternative][0], | |
1910 | GET_MODE (op))) | |
1911 | goto fail; | |
1912 | losers++; | |
1913 | } | |
55a2c322 VM |
1914 | if (operand_reg[nop] != NULL_RTX |
1915 | /* Output operands and matched input operands are | |
1916 | not inherited. The following conditions do not | |
1917 | exactly describe the previous statement but they | |
1918 | are pretty close. */ | |
1919 | && curr_static_id->operand[nop].type != OP_OUT | |
1920 | && (this_alternative_matches < 0 | |
1921 | || curr_static_id->operand[nop].type != OP_IN)) | |
1922 | { | |
1923 | int last_reload = (lra_reg_info[ORIGINAL_REGNO | |
1924 | (operand_reg[nop])] | |
1925 | .last_reload); | |
1926 | ||
1927 | if (last_reload > bb_reload_num) | |
1928 | reload_sum += last_reload - bb_reload_num; | |
1929 | } | |
1930 | /* If this is a constant that is reloaded into the | |
1931 | desired class by copying it to memory first, count | |
1932 | that as another reload. This is consistent with | |
1933 | other code and is required to avoid choosing another | |
1934 | alternative when the constant is moved into memory. | |
1935 | Note that the test here is precisely the same as in | |
1936 | the code below that calls force_const_mem. */ | |
1937 | if (CONST_POOL_OK_P (mode, op) | |
1938 | && ((targetm.preferred_reload_class | |
1939 | (op, this_alternative) == NO_REGS) | |
1940 | || no_input_reloads_p)) | |
1941 | { | |
1942 | const_to_mem = 1; | |
1943 | if (! no_regs_p) | |
1944 | losers++; | |
1945 | } | |
f4eafc30 | 1946 | |
55a2c322 VM |
1947 | /* Alternative loses if it requires a type of reload not |
1948 | permitted for this insn. We can always reload | |
1949 | objects with a REG_UNUSED note. */ | |
1950 | if ((curr_static_id->operand[nop].type != OP_IN | |
1951 | && no_output_reloads_p | |
1952 | && ! find_reg_note (curr_insn, REG_UNUSED, op)) | |
1953 | || (curr_static_id->operand[nop].type != OP_OUT | |
1954 | && no_input_reloads_p && ! const_to_mem)) | |
1955 | goto fail; | |
f4eafc30 | 1956 | |
821b7577 VM |
1957 | /* Check strong discouragement of reload of non-constant |
1958 | into class THIS_ALTERNATIVE. */ | |
1959 | if (! CONSTANT_P (op) && ! no_regs_p | |
1960 | && (targetm.preferred_reload_class | |
1961 | (op, this_alternative) == NO_REGS | |
1962 | || (curr_static_id->operand[nop].type == OP_OUT | |
1963 | && (targetm.preferred_output_reload_class | |
1964 | (op, this_alternative) == NO_REGS)))) | |
1965 | reject += LRA_MAX_REJECT; | |
f4eafc30 | 1966 | |
55a2c322 VM |
1967 | if (! ((const_to_mem && constmemok) |
1968 | || (MEM_P (op) && offmemok))) | |
1969 | { | |
1970 | /* We prefer to reload pseudos over reloading other | |
1971 | things, since such reloads may be able to be | |
1972 | eliminated later. So bump REJECT in other cases. | |
1973 | Don't do this in the case where we are forcing a | |
1974 | constant into memory and it will then win since | |
1975 | we don't want to have a different alternative | |
1976 | match then. */ | |
1977 | if (! (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER)) | |
1978 | reject += 2; | |
f4eafc30 | 1979 | |
55a2c322 VM |
1980 | if (! no_regs_p) |
1981 | reload_nregs | |
1982 | += ira_reg_class_max_nregs[this_alternative][mode]; | |
1983 | } | |
1984 | ||
1bdc4b11 VM |
1985 | /* We are trying to spill pseudo into memory. It is |
1986 | usually more costly than moving to a hard register | |
1987 | although it might takes the same number of | |
1988 | reloads. */ | |
1989 | if (no_regs_p && REG_P (op)) | |
1990 | reject++; | |
1991 | ||
7100b561 UB |
1992 | #ifdef SECONDARY_MEMORY_NEEDED |
1993 | /* If reload requires moving value through secondary | |
1994 | memory, it will need one more insn at least. */ | |
1995 | if (this_alternative != NO_REGS | |
1996 | && REG_P (op) && (cl = get_reg_class (REGNO (op))) != NO_REGS | |
1997 | && ((curr_static_id->operand[nop].type != OP_OUT | |
1998 | && SECONDARY_MEMORY_NEEDED (cl, this_alternative, | |
1999 | GET_MODE (op))) | |
2000 | || (curr_static_id->operand[nop].type != OP_IN | |
2001 | && SECONDARY_MEMORY_NEEDED (this_alternative, cl, | |
2002 | GET_MODE (op))))) | |
2003 | losers++; | |
2004 | #endif | |
55a2c322 VM |
2005 | /* Input reloads can be inherited more often than output |
2006 | reloads can be removed, so penalize output | |
2007 | reloads. */ | |
2008 | if (!REG_P (op) || curr_static_id->operand[nop].type != OP_IN) | |
2009 | reject++; | |
2010 | } | |
f4eafc30 | 2011 | |
55a2c322 VM |
2012 | if (early_clobber_p) |
2013 | reject++; | |
2014 | /* ??? We check early clobbers after processing all operands | |
2015 | (see loop below) and there we update the costs more. | |
2016 | Should we update the cost (may be approximately) here | |
2017 | because of early clobber register reloads or it is a rare | |
2018 | or non-important thing to be worth to do it. */ | |
821b7577 | 2019 | overall = losers * LRA_LOSER_COST_FACTOR + reject; |
55a2c322 VM |
2020 | if ((best_losers == 0 || losers != 0) && best_overall < overall) |
2021 | goto fail; | |
2022 | ||
2023 | curr_alt[nop] = this_alternative; | |
2024 | COPY_HARD_REG_SET (curr_alt_set[nop], this_alternative_set); | |
2025 | curr_alt_win[nop] = this_alternative_win; | |
2026 | curr_alt_match_win[nop] = this_alternative_match_win; | |
2027 | curr_alt_offmemok[nop] = this_alternative_offmemok; | |
2028 | curr_alt_matches[nop] = this_alternative_matches; | |
f4eafc30 | 2029 | |
55a2c322 VM |
2030 | if (this_alternative_matches >= 0 |
2031 | && !did_match && !this_alternative_win) | |
2032 | curr_alt_win[this_alternative_matches] = false; | |
f4eafc30 | 2033 | |
55a2c322 VM |
2034 | if (early_clobber_p && operand_reg[nop] != NULL_RTX) |
2035 | early_clobbered_nops[early_clobbered_regs_num++] = nop; | |
2036 | } | |
2037 | ok_p = true; | |
2038 | curr_alt_dont_inherit_ops_num = 0; | |
2039 | for (nop = 0; nop < early_clobbered_regs_num; nop++) | |
2040 | { | |
2041 | int i, j, clobbered_hard_regno; | |
2042 | HARD_REG_SET temp_set; | |
2043 | ||
2044 | i = early_clobbered_nops[nop]; | |
2045 | if ((! curr_alt_win[i] && ! curr_alt_match_win[i]) | |
2046 | || hard_regno[i] < 0) | |
2047 | continue; | |
2048 | clobbered_hard_regno = hard_regno[i]; | |
2049 | CLEAR_HARD_REG_SET (temp_set); | |
2050 | add_to_hard_reg_set (&temp_set, biggest_mode[i], clobbered_hard_regno); | |
2051 | for (j = 0; j < n_operands; j++) | |
2052 | if (j == i | |
2053 | /* We don't want process insides of match_operator and | |
2054 | match_parallel because otherwise we would process | |
2055 | their operands once again generating a wrong | |
2056 | code. */ | |
2057 | || curr_static_id->operand[j].is_operator) | |
2058 | continue; | |
2059 | else if ((curr_alt_matches[j] == i && curr_alt_match_win[j]) | |
2060 | || (curr_alt_matches[i] == j && curr_alt_match_win[i])) | |
2061 | continue; | |
2062 | else if (uses_hard_regs_p (*curr_id->operand_loc[j], temp_set)) | |
2063 | break; | |
2064 | if (j >= n_operands) | |
2065 | continue; | |
2066 | /* We need to reload early clobbered register. */ | |
2067 | for (j = 0; j < n_operands; j++) | |
2068 | if (curr_alt_matches[j] == i) | |
2069 | { | |
2070 | curr_alt_match_win[j] = false; | |
2071 | losers++; | |
821b7577 | 2072 | overall += LRA_LOSER_COST_FACTOR; |
55a2c322 VM |
2073 | } |
2074 | if (! curr_alt_match_win[i]) | |
2075 | curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++] = i; | |
2076 | else | |
2077 | { | |
2078 | /* Remember pseudos used for match reloads are never | |
2079 | inherited. */ | |
2080 | lra_assert (curr_alt_matches[i] >= 0); | |
2081 | curr_alt_win[curr_alt_matches[i]] = false; | |
2082 | } | |
2083 | curr_alt_win[i] = curr_alt_match_win[i] = false; | |
2084 | losers++; | |
821b7577 | 2085 | overall += LRA_LOSER_COST_FACTOR; |
55a2c322 VM |
2086 | } |
2087 | small_class_operands_num = 0; | |
2088 | for (nop = 0; nop < n_operands; nop++) | |
2089 | small_class_operands_num | |
2090 | += SMALL_REGISTER_CLASS_P (curr_alt[nop]) ? 1 : 0; | |
2091 | ||
2092 | /* If this alternative can be made to work by reloading, and it | |
2093 | needs less reloading than the others checked so far, record | |
2094 | it as the chosen goal for reloading. */ | |
2095 | if ((best_losers != 0 && losers == 0) | |
2096 | || (((best_losers == 0 && losers == 0) | |
2097 | || (best_losers != 0 && losers != 0)) | |
2098 | && (best_overall > overall | |
2099 | || (best_overall == overall | |
2100 | /* If the cost of the reloads is the same, | |
2101 | prefer alternative which requires minimal | |
2102 | number of small register classes for the | |
2103 | operands. This improves chances of reloads | |
2104 | for insn requiring small register | |
2105 | classes. */ | |
2106 | && (small_class_operands_num | |
2107 | < best_small_class_operands_num | |
2108 | || (small_class_operands_num | |
2109 | == best_small_class_operands_num | |
2110 | && (reload_nregs < best_reload_nregs | |
2111 | || (reload_nregs == best_reload_nregs | |
2112 | && best_reload_sum < reload_sum)))))))) | |
2113 | { | |
2114 | for (nop = 0; nop < n_operands; nop++) | |
2115 | { | |
2116 | goal_alt_win[nop] = curr_alt_win[nop]; | |
2117 | goal_alt_match_win[nop] = curr_alt_match_win[nop]; | |
2118 | goal_alt_matches[nop] = curr_alt_matches[nop]; | |
2119 | goal_alt[nop] = curr_alt[nop]; | |
2120 | goal_alt_offmemok[nop] = curr_alt_offmemok[nop]; | |
2121 | } | |
2122 | goal_alt_dont_inherit_ops_num = curr_alt_dont_inherit_ops_num; | |
2123 | for (nop = 0; nop < curr_alt_dont_inherit_ops_num; nop++) | |
2124 | goal_alt_dont_inherit_ops[nop] = curr_alt_dont_inherit_ops[nop]; | |
2125 | goal_alt_swapped = curr_swapped; | |
2126 | best_overall = overall; | |
2127 | best_losers = losers; | |
2128 | best_small_class_operands_num = small_class_operands_num; | |
2129 | best_reload_nregs = reload_nregs; | |
2130 | best_reload_sum = reload_sum; | |
2131 | goal_alt_number = nalt; | |
2132 | } | |
2133 | if (losers == 0) | |
2134 | /* Everything is satisfied. Do not process alternatives | |
f4eafc30 | 2135 | anymore. */ |
55a2c322 VM |
2136 | break; |
2137 | fail: | |
2138 | ; | |
2139 | } | |
2140 | return ok_p; | |
2141 | } | |
2142 | ||
2143 | /* Return 1 if ADDR is a valid memory address for mode MODE in address | |
2144 | space AS, and check that each pseudo has the proper kind of hard | |
2145 | reg. */ | |
2146 | static int | |
2147 | valid_address_p (enum machine_mode mode ATTRIBUTE_UNUSED, | |
2148 | rtx addr, addr_space_t as) | |
2149 | { | |
2150 | #ifdef GO_IF_LEGITIMATE_ADDRESS | |
2151 | lra_assert (ADDR_SPACE_GENERIC_P (as)); | |
2152 | GO_IF_LEGITIMATE_ADDRESS (mode, addr, win); | |
2153 | return 0; | |
f4eafc30 | 2154 | |
55a2c322 VM |
2155 | win: |
2156 | return 1; | |
2157 | #else | |
2158 | return targetm.addr_space.legitimate_address_p (mode, addr, 0, as); | |
2159 | #endif | |
2160 | } | |
2161 | ||
277f65de | 2162 | /* Return whether address AD is valid. */ |
8bf9b489 RS |
2163 | |
2164 | static bool | |
277f65de | 2165 | valid_address_p (struct address_info *ad) |
8bf9b489 RS |
2166 | { |
2167 | /* Some ports do not check displacements for eliminable registers, | |
2168 | so we replace them temporarily with the elimination target. */ | |
2169 | rtx saved_base_reg = NULL_RTX; | |
2170 | rtx saved_index_reg = NULL_RTX; | |
277f65de RS |
2171 | rtx *base_term = strip_subreg (ad->base_term); |
2172 | rtx *index_term = strip_subreg (ad->index_term); | |
2173 | if (base_term != NULL) | |
8bf9b489 | 2174 | { |
277f65de RS |
2175 | saved_base_reg = *base_term; |
2176 | lra_eliminate_reg_if_possible (base_term); | |
2177 | if (ad->base_term2 != NULL) | |
2178 | *ad->base_term2 = *ad->base_term; | |
8bf9b489 | 2179 | } |
277f65de | 2180 | if (index_term != NULL) |
8bf9b489 | 2181 | { |
277f65de RS |
2182 | saved_index_reg = *index_term; |
2183 | lra_eliminate_reg_if_possible (index_term); | |
8bf9b489 | 2184 | } |
277f65de | 2185 | bool ok_p = valid_address_p (ad->mode, *ad->outer, ad->as); |
8bf9b489 RS |
2186 | if (saved_base_reg != NULL_RTX) |
2187 | { | |
277f65de RS |
2188 | *base_term = saved_base_reg; |
2189 | if (ad->base_term2 != NULL) | |
2190 | *ad->base_term2 = *ad->base_term; | |
8bf9b489 RS |
2191 | } |
2192 | if (saved_index_reg != NULL_RTX) | |
277f65de | 2193 | *index_term = saved_index_reg; |
8bf9b489 RS |
2194 | return ok_p; |
2195 | } | |
2196 | ||
277f65de | 2197 | /* Make reload base reg + disp from address AD. Return the new pseudo. */ |
55a2c322 | 2198 | static rtx |
277f65de | 2199 | base_plus_disp_to_reg (struct address_info *ad) |
55a2c322 VM |
2200 | { |
2201 | enum reg_class cl; | |
2202 | rtx new_reg; | |
2203 | ||
277f65de RS |
2204 | lra_assert (ad->base == ad->base_term && ad->disp == ad->disp_term); |
2205 | cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code, | |
2206 | get_index_code (ad)); | |
2207 | new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX, | |
2208 | cl, "base + disp"); | |
2209 | lra_emit_add (new_reg, *ad->base_term, *ad->disp_term); | |
55a2c322 VM |
2210 | return new_reg; |
2211 | } | |
2212 | ||
277f65de RS |
2213 | /* Return true if we can add a displacement to address AD, even if that |
2214 | makes the address invalid. The fix-up code requires any new address | |
2215 | to be the sum of the BASE_TERM, INDEX and DISP_TERM fields. */ | |
02ea4bf4 | 2216 | static bool |
277f65de | 2217 | can_add_disp_p (struct address_info *ad) |
02ea4bf4 | 2218 | { |
277f65de RS |
2219 | return (!ad->autoinc_p |
2220 | && ad->segment == NULL | |
2221 | && ad->base == ad->base_term | |
2222 | && ad->disp == ad->disp_term); | |
02ea4bf4 RS |
2223 | } |
2224 | ||
277f65de RS |
2225 | /* Make equiv substitution in address AD. Return true if a substitution |
2226 | was made. */ | |
55a2c322 | 2227 | static bool |
277f65de | 2228 | equiv_address_substitution (struct address_info *ad) |
55a2c322 | 2229 | { |
277f65de | 2230 | rtx base_reg, new_base_reg, index_reg, new_index_reg, *base_term, *index_term; |
55a2c322 VM |
2231 | HOST_WIDE_INT disp, scale; |
2232 | bool change_p; | |
2233 | ||
277f65de RS |
2234 | base_term = strip_subreg (ad->base_term); |
2235 | if (base_term == NULL) | |
55a2c322 VM |
2236 | base_reg = new_base_reg = NULL_RTX; |
2237 | else | |
2238 | { | |
277f65de | 2239 | base_reg = *base_term; |
55a2c322 VM |
2240 | new_base_reg = get_equiv_substitution (base_reg); |
2241 | } | |
277f65de RS |
2242 | index_term = strip_subreg (ad->index_term); |
2243 | if (index_term == NULL) | |
55a2c322 VM |
2244 | index_reg = new_index_reg = NULL_RTX; |
2245 | else | |
2246 | { | |
277f65de | 2247 | index_reg = *index_term; |
55a2c322 VM |
2248 | new_index_reg = get_equiv_substitution (index_reg); |
2249 | } | |
2250 | if (base_reg == new_base_reg && index_reg == new_index_reg) | |
2251 | return false; | |
2252 | disp = 0; | |
2253 | change_p = false; | |
2254 | if (lra_dump_file != NULL) | |
2255 | { | |
2256 | fprintf (lra_dump_file, "Changing address in insn %d ", | |
2257 | INSN_UID (curr_insn)); | |
cfbeaedf | 2258 | dump_value_slim (lra_dump_file, *ad->outer, 1); |
55a2c322 VM |
2259 | } |
2260 | if (base_reg != new_base_reg) | |
2261 | { | |
2262 | if (REG_P (new_base_reg)) | |
2263 | { | |
277f65de | 2264 | *base_term = new_base_reg; |
55a2c322 VM |
2265 | change_p = true; |
2266 | } | |
2267 | else if (GET_CODE (new_base_reg) == PLUS | |
2268 | && REG_P (XEXP (new_base_reg, 0)) | |
02ea4bf4 | 2269 | && CONST_INT_P (XEXP (new_base_reg, 1)) |
277f65de | 2270 | && can_add_disp_p (ad)) |
55a2c322 VM |
2271 | { |
2272 | disp += INTVAL (XEXP (new_base_reg, 1)); | |
277f65de | 2273 | *base_term = XEXP (new_base_reg, 0); |
55a2c322 VM |
2274 | change_p = true; |
2275 | } | |
277f65de RS |
2276 | if (ad->base_term2 != NULL) |
2277 | *ad->base_term2 = *ad->base_term; | |
55a2c322 | 2278 | } |
55a2c322 VM |
2279 | if (index_reg != new_index_reg) |
2280 | { | |
2281 | if (REG_P (new_index_reg)) | |
2282 | { | |
277f65de | 2283 | *index_term = new_index_reg; |
55a2c322 VM |
2284 | change_p = true; |
2285 | } | |
2286 | else if (GET_CODE (new_index_reg) == PLUS | |
2287 | && REG_P (XEXP (new_index_reg, 0)) | |
02ea4bf4 | 2288 | && CONST_INT_P (XEXP (new_index_reg, 1)) |
277f65de | 2289 | && can_add_disp_p (ad) |
02ea4bf4 | 2290 | && (scale = get_index_scale (ad))) |
55a2c322 VM |
2291 | { |
2292 | disp += INTVAL (XEXP (new_index_reg, 1)) * scale; | |
277f65de | 2293 | *index_term = XEXP (new_index_reg, 0); |
55a2c322 VM |
2294 | change_p = true; |
2295 | } | |
2296 | } | |
2297 | if (disp != 0) | |
2298 | { | |
277f65de RS |
2299 | if (ad->disp != NULL) |
2300 | *ad->disp = plus_constant (GET_MODE (*ad->inner), *ad->disp, disp); | |
55a2c322 VM |
2301 | else |
2302 | { | |
277f65de RS |
2303 | *ad->inner = plus_constant (GET_MODE (*ad->inner), *ad->inner, disp); |
2304 | update_address (ad); | |
55a2c322 VM |
2305 | } |
2306 | change_p = true; | |
2307 | } | |
2308 | if (lra_dump_file != NULL) | |
2309 | { | |
2310 | if (! change_p) | |
2311 | fprintf (lra_dump_file, " -- no change\n"); | |
2312 | else | |
2313 | { | |
2314 | fprintf (lra_dump_file, " on equiv "); | |
cfbeaedf | 2315 | dump_value_slim (lra_dump_file, *ad->outer, 1); |
55a2c322 VM |
2316 | fprintf (lra_dump_file, "\n"); |
2317 | } | |
2318 | } | |
2319 | return change_p; | |
2320 | } | |
2321 | ||
bd3d34d4 RS |
2322 | /* Major function to make reloads for an address in operand NOP. |
2323 | The supported cases are: | |
2324 | ||
2325 | 1) an address that existed before LRA started, at which point it must | |
2326 | have been valid. These addresses are subject to elimination and | |
2327 | may have become invalid due to the elimination offset being out | |
2328 | of range. | |
2329 | ||
2330 | 2) an address created by forcing a constant to memory (force_const_to_mem). | |
2331 | The initial form of these addresses might not be valid, and it is this | |
2332 | function's job to make them valid. | |
2333 | ||
2334 | 3) a frame address formed from a register and a (possibly zero) | |
2335 | constant offset. As above, these addresses might not be valid | |
2336 | and this function must make them so. | |
2337 | ||
2338 | Add reloads to the lists *BEFORE and *AFTER. We might need to add | |
55a2c322 VM |
2339 | reloads to *AFTER because of inc/dec, {pre, post} modify in the |
2340 | address. Return true for any RTL change. */ | |
2341 | static bool | |
2342 | process_address (int nop, rtx *before, rtx *after) | |
2343 | { | |
277f65de RS |
2344 | struct address_info ad; |
2345 | rtx new_reg; | |
55a2c322 VM |
2346 | rtx op = *curr_id->operand_loc[nop]; |
2347 | const char *constraint = curr_static_id->operand[nop].constraint; | |
2348 | bool change_p; | |
55a2c322 VM |
2349 | |
2350 | if (constraint[0] == 'p' | |
2351 | || EXTRA_ADDRESS_CONSTRAINT (constraint[0], constraint)) | |
277f65de | 2352 | decompose_lea_address (&ad, curr_id->operand_loc[nop]); |
55a2c322 | 2353 | else if (MEM_P (op)) |
277f65de | 2354 | decompose_mem_address (&ad, op); |
55a2c322 VM |
2355 | else if (GET_CODE (op) == SUBREG |
2356 | && MEM_P (SUBREG_REG (op))) | |
277f65de | 2357 | decompose_mem_address (&ad, SUBREG_REG (op)); |
55a2c322 VM |
2358 | else |
2359 | return false; | |
277f65de RS |
2360 | change_p = equiv_address_substitution (&ad); |
2361 | if (ad.base_term != NULL | |
55a2c322 | 2362 | && (process_addr_reg |
277f65de RS |
2363 | (ad.base_term, before, |
2364 | (ad.autoinc_p | |
2365 | && !(REG_P (*ad.base_term) | |
2366 | && find_regno_note (curr_insn, REG_DEAD, | |
2367 | REGNO (*ad.base_term)) != NULL_RTX) | |
55a2c322 | 2368 | ? after : NULL), |
277f65de RS |
2369 | base_reg_class (ad.mode, ad.as, ad.base_outer_code, |
2370 | get_index_code (&ad))))) | |
55a2c322 VM |
2371 | { |
2372 | change_p = true; | |
277f65de RS |
2373 | if (ad.base_term2 != NULL) |
2374 | *ad.base_term2 = *ad.base_term; | |
55a2c322 | 2375 | } |
277f65de RS |
2376 | if (ad.index_term != NULL |
2377 | && process_addr_reg (ad.index_term, before, NULL, INDEX_REG_CLASS)) | |
55a2c322 VM |
2378 | change_p = true; |
2379 | ||
277f65de | 2380 | /* There are three cases where the shape of *AD.INNER may now be invalid: |
bd3d34d4 RS |
2381 | |
2382 | 1) the original address was valid, but either elimination or | |
2383 | equiv_address_substitution applied a displacement that made | |
2384 | it invalid. | |
2385 | ||
2386 | 2) the address is an invalid symbolic address created by | |
2387 | force_const_to_mem. | |
2388 | ||
2389 | 3) the address is a frame address with an invalid offset. | |
2390 | ||
277f65de RS |
2391 | All these cases involve a displacement and a non-autoinc address, |
2392 | so there is no point revalidating other types. */ | |
2393 | if (ad.disp == NULL || ad.autoinc_p || valid_address_p (&ad)) | |
55a2c322 VM |
2394 | return change_p; |
2395 | ||
bd3d34d4 RS |
2396 | /* Any index existed before LRA started, so we can assume that the |
2397 | presence and shape of the index is valid. */ | |
55a2c322 | 2398 | push_to_sequence (*before); |
277f65de RS |
2399 | gcc_assert (ad.segment == NULL); |
2400 | gcc_assert (ad.disp == ad.disp_term); | |
2401 | if (ad.base == NULL) | |
55a2c322 | 2402 | { |
277f65de | 2403 | if (ad.index == NULL) |
55a2c322 VM |
2404 | { |
2405 | int code = -1; | |
277f65de RS |
2406 | enum reg_class cl = base_reg_class (ad.mode, ad.as, |
2407 | SCRATCH, SCRATCH); | |
2408 | rtx disp = *ad.disp; | |
2409 | ||
55a2c322 VM |
2410 | new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp"); |
2411 | #ifdef HAVE_lo_sum | |
2412 | { | |
2413 | rtx insn; | |
2414 | rtx last = get_last_insn (); | |
2415 | ||
bd3d34d4 | 2416 | /* disp => lo_sum (new_base, disp), case (2) above. */ |
55a2c322 VM |
2417 | insn = emit_insn (gen_rtx_SET |
2418 | (VOIDmode, new_reg, | |
277f65de | 2419 | gen_rtx_HIGH (Pmode, copy_rtx (disp)))); |
55a2c322 VM |
2420 | code = recog_memoized (insn); |
2421 | if (code >= 0) | |
2422 | { | |
277f65de RS |
2423 | *ad.disp = gen_rtx_LO_SUM (Pmode, new_reg, disp); |
2424 | if (! valid_address_p (ad.mode, *ad.outer, ad.as)) | |
55a2c322 | 2425 | { |
277f65de | 2426 | *ad.disp = disp; |
55a2c322 VM |
2427 | code = -1; |
2428 | } | |
2429 | } | |
2430 | if (code < 0) | |
2431 | delete_insns_since (last); | |
2432 | } | |
2433 | #endif | |
2434 | if (code < 0) | |
2435 | { | |
bd3d34d4 | 2436 | /* disp => new_base, case (2) above. */ |
277f65de RS |
2437 | lra_emit_move (new_reg, disp); |
2438 | *ad.disp = new_reg; | |
55a2c322 VM |
2439 | } |
2440 | } | |
2441 | else | |
2442 | { | |
bd3d34d4 RS |
2443 | /* index * scale + disp => new base + index * scale, |
2444 | case (1) above. */ | |
277f65de RS |
2445 | enum reg_class cl = base_reg_class (ad.mode, ad.as, PLUS, |
2446 | GET_CODE (*ad.index)); | |
55a2c322 VM |
2447 | |
2448 | lra_assert (INDEX_REG_CLASS != NO_REGS); | |
2449 | new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp"); | |
277f65de RS |
2450 | lra_emit_move (new_reg, *ad.disp); |
2451 | *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg), | |
2452 | new_reg, *ad.index); | |
55a2c322 VM |
2453 | } |
2454 | } | |
277f65de | 2455 | else if (ad.index == NULL) |
55a2c322 | 2456 | { |
bd3d34d4 | 2457 | /* base + disp => new base, cases (1) and (3) above. */ |
55a2c322 VM |
2458 | /* Another option would be to reload the displacement into an |
2459 | index register. However, postreload has code to optimize | |
2460 | address reloads that have the same base and different | |
2461 | displacements, so reloading into an index register would | |
2462 | not necessarily be a win. */ | |
277f65de RS |
2463 | new_reg = base_plus_disp_to_reg (&ad); |
2464 | *ad.inner = new_reg; | |
55a2c322 VM |
2465 | } |
2466 | else | |
2467 | { | |
bd3d34d4 RS |
2468 | /* base + scale * index + disp => new base + scale * index, |
2469 | case (1) above. */ | |
277f65de RS |
2470 | new_reg = base_plus_disp_to_reg (&ad); |
2471 | *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg), | |
2472 | new_reg, *ad.index); | |
55a2c322 VM |
2473 | } |
2474 | *before = get_insns (); | |
2475 | end_sequence (); | |
2476 | return true; | |
2477 | } | |
2478 | ||
2479 | /* Emit insns to reload VALUE into a new register. VALUE is an | |
2480 | auto-increment or auto-decrement RTX whose operand is a register or | |
2481 | memory location; so reloading involves incrementing that location. | |
2482 | IN is either identical to VALUE, or some cheaper place to reload | |
2483 | value being incremented/decremented from. | |
2484 | ||
2485 | INC_AMOUNT is the number to increment or decrement by (always | |
2486 | positive and ignored for POST_MODIFY/PRE_MODIFY). | |
2487 | ||
2488 | Return pseudo containing the result. */ | |
2489 | static rtx | |
2490 | emit_inc (enum reg_class new_rclass, rtx in, rtx value, int inc_amount) | |
2491 | { | |
2492 | /* REG or MEM to be copied and incremented. */ | |
2493 | rtx incloc = XEXP (value, 0); | |
2494 | /* Nonzero if increment after copying. */ | |
2495 | int post = (GET_CODE (value) == POST_DEC || GET_CODE (value) == POST_INC | |
2496 | || GET_CODE (value) == POST_MODIFY); | |
2497 | rtx last; | |
2498 | rtx inc; | |
2499 | rtx add_insn; | |
2500 | int code; | |
2501 | rtx real_in = in == value ? incloc : in; | |
2502 | rtx result; | |
2503 | bool plus_p = true; | |
2504 | ||
2505 | if (GET_CODE (value) == PRE_MODIFY || GET_CODE (value) == POST_MODIFY) | |
2506 | { | |
2507 | lra_assert (GET_CODE (XEXP (value, 1)) == PLUS | |
2508 | || GET_CODE (XEXP (value, 1)) == MINUS); | |
2509 | lra_assert (rtx_equal_p (XEXP (XEXP (value, 1), 0), XEXP (value, 0))); | |
2510 | plus_p = GET_CODE (XEXP (value, 1)) == PLUS; | |
2511 | inc = XEXP (XEXP (value, 1), 1); | |
2512 | } | |
2513 | else | |
2514 | { | |
2515 | if (GET_CODE (value) == PRE_DEC || GET_CODE (value) == POST_DEC) | |
2516 | inc_amount = -inc_amount; | |
2517 | ||
2518 | inc = GEN_INT (inc_amount); | |
2519 | } | |
2520 | ||
2521 | if (! post && REG_P (incloc)) | |
2522 | result = incloc; | |
2523 | else | |
2524 | result = lra_create_new_reg (GET_MODE (value), value, new_rclass, | |
2525 | "INC/DEC result"); | |
2526 | ||
2527 | if (real_in != result) | |
2528 | { | |
2529 | /* First copy the location to the result register. */ | |
2530 | lra_assert (REG_P (result)); | |
2531 | emit_insn (gen_move_insn (result, real_in)); | |
2532 | } | |
2533 | ||
2534 | /* We suppose that there are insns to add/sub with the constant | |
2535 | increment permitted in {PRE/POST)_{DEC/INC/MODIFY}. At least the | |
2536 | old reload worked with this assumption. If the assumption | |
2537 | becomes wrong, we should use approach in function | |
2538 | base_plus_disp_to_reg. */ | |
2539 | if (in == value) | |
2540 | { | |
2541 | /* See if we can directly increment INCLOC. */ | |
2542 | last = get_last_insn (); | |
2543 | add_insn = emit_insn (plus_p | |
2544 | ? gen_add2_insn (incloc, inc) | |
2545 | : gen_sub2_insn (incloc, inc)); | |
2546 | ||
2547 | code = recog_memoized (add_insn); | |
2548 | if (code >= 0) | |
2549 | { | |
2550 | if (! post && result != incloc) | |
2551 | emit_insn (gen_move_insn (result, incloc)); | |
2552 | return result; | |
2553 | } | |
2554 | delete_insns_since (last); | |
2555 | } | |
2556 | ||
2557 | /* If couldn't do the increment directly, must increment in RESULT. | |
2558 | The way we do this depends on whether this is pre- or | |
2559 | post-increment. For pre-increment, copy INCLOC to the reload | |
2560 | register, increment it there, then save back. */ | |
2561 | if (! post) | |
2562 | { | |
2563 | if (real_in != result) | |
2564 | emit_insn (gen_move_insn (result, real_in)); | |
2565 | if (plus_p) | |
2566 | emit_insn (gen_add2_insn (result, inc)); | |
2567 | else | |
2568 | emit_insn (gen_sub2_insn (result, inc)); | |
2569 | if (result != incloc) | |
2570 | emit_insn (gen_move_insn (incloc, result)); | |
2571 | } | |
2572 | else | |
2573 | { | |
2574 | /* Post-increment. | |
2575 | ||
2576 | Because this might be a jump insn or a compare, and because | |
2577 | RESULT may not be available after the insn in an input | |
2578 | reload, we must do the incrementing before the insn being | |
2579 | reloaded for. | |
2580 | ||
2581 | We have already copied IN to RESULT. Increment the copy in | |
2582 | RESULT, save that back, then decrement RESULT so it has | |
2583 | the original value. */ | |
2584 | if (plus_p) | |
2585 | emit_insn (gen_add2_insn (result, inc)); | |
2586 | else | |
2587 | emit_insn (gen_sub2_insn (result, inc)); | |
2588 | emit_insn (gen_move_insn (incloc, result)); | |
2589 | /* Restore non-modified value for the result. We prefer this | |
2590 | way because it does not require an additional hard | |
2591 | register. */ | |
2592 | if (plus_p) | |
2593 | { | |
2594 | if (CONST_INT_P (inc)) | |
2595 | emit_insn (gen_add2_insn (result, GEN_INT (-INTVAL (inc)))); | |
2596 | else | |
2597 | emit_insn (gen_sub2_insn (result, inc)); | |
2598 | } | |
2599 | else | |
2600 | emit_insn (gen_add2_insn (result, inc)); | |
2601 | } | |
2602 | return result; | |
2603 | } | |
2604 | ||
2605 | /* Swap operands NOP and NOP + 1. */ | |
2606 | static inline void | |
2607 | swap_operands (int nop) | |
2608 | { | |
2609 | enum machine_mode mode = curr_operand_mode[nop]; | |
2610 | curr_operand_mode[nop] = curr_operand_mode[nop + 1]; | |
2611 | curr_operand_mode[nop + 1] = mode; | |
2612 | rtx x = *curr_id->operand_loc[nop]; | |
2613 | *curr_id->operand_loc[nop] = *curr_id->operand_loc[nop + 1]; | |
2614 | *curr_id->operand_loc[nop + 1] = x; | |
2615 | /* Swap the duplicates too. */ | |
2616 | lra_update_dup (curr_id, nop); | |
2617 | lra_update_dup (curr_id, nop + 1); | |
2618 | } | |
2619 | ||
2620 | /* Main entry point of the constraint code: search the body of the | |
2621 | current insn to choose the best alternative. It is mimicking insn | |
2622 | alternative cost calculation model of former reload pass. That is | |
2623 | because machine descriptions were written to use this model. This | |
2624 | model can be changed in future. Make commutative operand exchange | |
2625 | if it is chosen. | |
2626 | ||
2627 | Return true if some RTL changes happened during function call. */ | |
2628 | static bool | |
2629 | curr_insn_transform (void) | |
2630 | { | |
2631 | int i, j, k; | |
2632 | int n_operands; | |
2633 | int n_alternatives; | |
2634 | int commutative; | |
2635 | signed char goal_alt_matched[MAX_RECOG_OPERANDS][MAX_RECOG_OPERANDS]; | |
511dcace | 2636 | signed char match_inputs[MAX_RECOG_OPERANDS + 1]; |
55a2c322 VM |
2637 | rtx before, after; |
2638 | bool alt_p = false; | |
2639 | /* Flag that the insn has been changed through a transformation. */ | |
2640 | bool change_p; | |
2641 | bool sec_mem_p; | |
2642 | #ifdef SECONDARY_MEMORY_NEEDED | |
2643 | bool use_sec_mem_p; | |
2644 | #endif | |
2645 | int max_regno_before; | |
2646 | int reused_alternative_num; | |
2647 | ||
2648 | no_input_reloads_p = no_output_reloads_p = false; | |
2649 | goal_alt_number = -1; | |
2650 | ||
2651 | if (check_and_process_move (&change_p, &sec_mem_p)) | |
2652 | return change_p; | |
2653 | ||
2654 | /* JUMP_INSNs and CALL_INSNs are not allowed to have any output | |
2655 | reloads; neither are insns that SET cc0. Insns that use CC0 are | |
2656 | not allowed to have any input reloads. */ | |
2657 | if (JUMP_P (curr_insn) || CALL_P (curr_insn)) | |
2658 | no_output_reloads_p = true; | |
2659 | ||
2660 | #ifdef HAVE_cc0 | |
2661 | if (reg_referenced_p (cc0_rtx, PATTERN (curr_insn))) | |
2662 | no_input_reloads_p = true; | |
2663 | if (reg_set_p (cc0_rtx, PATTERN (curr_insn))) | |
2664 | no_output_reloads_p = true; | |
2665 | #endif | |
2666 | ||
2667 | n_operands = curr_static_id->n_operands; | |
2668 | n_alternatives = curr_static_id->n_alternatives; | |
2669 | ||
2670 | /* Just return "no reloads" if insn has no operands with | |
2671 | constraints. */ | |
2672 | if (n_operands == 0 || n_alternatives == 0) | |
2673 | return false; | |
2674 | ||
2675 | max_regno_before = max_reg_num (); | |
2676 | ||
2677 | for (i = 0; i < n_operands; i++) | |
2678 | { | |
2679 | goal_alt_matched[i][0] = -1; | |
2680 | goal_alt_matches[i] = -1; | |
2681 | } | |
2682 | ||
2683 | commutative = curr_static_id->commutative; | |
2684 | ||
2685 | /* Now see what we need for pseudos that didn't get hard regs or got | |
2686 | the wrong kind of hard reg. For this, we must consider all the | |
2687 | operands together against the register constraints. */ | |
2688 | ||
821b7577 | 2689 | best_losers = best_overall = INT_MAX; |
55a2c322 VM |
2690 | best_small_class_operands_num = best_reload_sum = 0; |
2691 | ||
2692 | curr_swapped = false; | |
2693 | goal_alt_swapped = false; | |
2694 | ||
2695 | /* Make equivalence substitution and memory subreg elimination | |
2696 | before address processing because an address legitimacy can | |
2697 | depend on memory mode. */ | |
2698 | for (i = 0; i < n_operands; i++) | |
2699 | { | |
2700 | rtx op = *curr_id->operand_loc[i]; | |
2701 | rtx subst, old = op; | |
2702 | bool op_change_p = false; | |
2703 | ||
2704 | if (GET_CODE (old) == SUBREG) | |
2705 | old = SUBREG_REG (old); | |
2706 | subst = get_equiv_substitution (old); | |
2707 | if (subst != old) | |
2708 | { | |
2709 | subst = copy_rtx (subst); | |
2710 | lra_assert (REG_P (old)); | |
2711 | if (GET_CODE (op) == SUBREG) | |
2712 | SUBREG_REG (op) = subst; | |
2713 | else | |
2714 | *curr_id->operand_loc[i] = subst; | |
2715 | if (lra_dump_file != NULL) | |
2716 | { | |
2717 | fprintf (lra_dump_file, | |
2718 | "Changing pseudo %d in operand %i of insn %u on equiv ", | |
2719 | REGNO (old), i, INSN_UID (curr_insn)); | |
cfbeaedf | 2720 | dump_value_slim (lra_dump_file, subst, 1); |
55a2c322 VM |
2721 | fprintf (lra_dump_file, "\n"); |
2722 | } | |
2723 | op_change_p = change_p = true; | |
2724 | } | |
2725 | if (simplify_operand_subreg (i, GET_MODE (old)) || op_change_p) | |
2726 | { | |
2727 | change_p = true; | |
2728 | lra_update_dup (curr_id, i); | |
2729 | } | |
2730 | } | |
2731 | ||
2732 | /* Reload address registers and displacements. We do it before | |
2733 | finding an alternative because of memory constraints. */ | |
2734 | before = after = NULL_RTX; | |
2735 | for (i = 0; i < n_operands; i++) | |
2736 | if (! curr_static_id->operand[i].is_operator | |
2737 | && process_address (i, &before, &after)) | |
2738 | { | |
2739 | change_p = true; | |
2740 | lra_update_dup (curr_id, i); | |
2741 | } | |
f4eafc30 | 2742 | |
55a2c322 VM |
2743 | if (change_p) |
2744 | /* If we've changed the instruction then any alternative that | |
2745 | we chose previously may no longer be valid. */ | |
2746 | lra_set_used_insn_alternative (curr_insn, -1); | |
2747 | ||
2748 | try_swapped: | |
2749 | ||
2750 | reused_alternative_num = curr_id->used_insn_alternative; | |
2751 | if (lra_dump_file != NULL && reused_alternative_num >= 0) | |
2752 | fprintf (lra_dump_file, "Reusing alternative %d for insn #%u\n", | |
2753 | reused_alternative_num, INSN_UID (curr_insn)); | |
2754 | ||
2755 | if (process_alt_operands (reused_alternative_num)) | |
2756 | alt_p = true; | |
2757 | ||
2758 | /* If insn is commutative (it's safe to exchange a certain pair of | |
2759 | operands) then we need to try each alternative twice, the second | |
2760 | time matching those two operands as if we had exchanged them. To | |
2761 | do this, really exchange them in operands. | |
2762 | ||
2763 | If we have just tried the alternatives the second time, return | |
2764 | operands to normal and drop through. */ | |
2765 | ||
2766 | if (reused_alternative_num < 0 && commutative >= 0) | |
2767 | { | |
2768 | curr_swapped = !curr_swapped; | |
2769 | if (curr_swapped) | |
2770 | { | |
2771 | swap_operands (commutative); | |
2772 | goto try_swapped; | |
2773 | } | |
2774 | else | |
2775 | swap_operands (commutative); | |
2776 | } | |
2777 | ||
55a2c322 VM |
2778 | if (! alt_p && ! sec_mem_p) |
2779 | { | |
2780 | /* No alternative works with reloads?? */ | |
2781 | if (INSN_CODE (curr_insn) >= 0) | |
2782 | fatal_insn ("unable to generate reloads for:", curr_insn); | |
2783 | error_for_asm (curr_insn, | |
2784 | "inconsistent operand constraints in an %<asm%>"); | |
2785 | /* Avoid further trouble with this insn. */ | |
2786 | PATTERN (curr_insn) = gen_rtx_USE (VOIDmode, const0_rtx); | |
2787 | lra_invalidate_insn_data (curr_insn); | |
2788 | return true; | |
2789 | } | |
2790 | ||
2791 | /* If the best alternative is with operands 1 and 2 swapped, swap | |
2792 | them. Update the operand numbers of any reloads already | |
2793 | pushed. */ | |
2794 | ||
2795 | if (goal_alt_swapped) | |
2796 | { | |
2797 | if (lra_dump_file != NULL) | |
2798 | fprintf (lra_dump_file, " Commutative operand exchange in insn %u\n", | |
2799 | INSN_UID (curr_insn)); | |
2800 | ||
2801 | /* Swap the duplicates too. */ | |
2802 | swap_operands (commutative); | |
2803 | change_p = true; | |
2804 | } | |
2805 | ||
2806 | #ifdef SECONDARY_MEMORY_NEEDED | |
2807 | /* Some target macros SECONDARY_MEMORY_NEEDED (e.g. x86) are defined | |
2808 | too conservatively. So we use the secondary memory only if there | |
2809 | is no any alternative without reloads. */ | |
2810 | use_sec_mem_p = false; | |
2811 | if (! alt_p) | |
2812 | use_sec_mem_p = true; | |
2813 | else if (sec_mem_p) | |
2814 | { | |
2815 | for (i = 0; i < n_operands; i++) | |
2816 | if (! goal_alt_win[i] && ! goal_alt_match_win[i]) | |
2817 | break; | |
2818 | use_sec_mem_p = i < n_operands; | |
2819 | } | |
2820 | ||
2821 | if (use_sec_mem_p) | |
2822 | { | |
89d56d79 | 2823 | rtx new_reg, src, dest, rld; |
66aa7879 | 2824 | enum machine_mode sec_mode, rld_mode; |
55a2c322 VM |
2825 | |
2826 | lra_assert (sec_mem_p); | |
66aa7879 VM |
2827 | lra_assert (curr_static_id->operand[0].type == OP_OUT |
2828 | && curr_static_id->operand[1].type == OP_IN); | |
2829 | dest = *curr_id->operand_loc[0]; | |
2830 | src = *curr_id->operand_loc[1]; | |
2831 | rld = (GET_MODE_SIZE (GET_MODE (dest)) <= GET_MODE_SIZE (GET_MODE (src)) | |
2832 | ? dest : src); | |
2833 | rld_mode = GET_MODE (rld); | |
55a2c322 | 2834 | #ifdef SECONDARY_MEMORY_NEEDED_MODE |
66aa7879 | 2835 | sec_mode = SECONDARY_MEMORY_NEEDED_MODE (rld_mode); |
55a2c322 | 2836 | #else |
66aa7879 | 2837 | sec_mode = rld_mode; |
55a2c322 VM |
2838 | #endif |
2839 | new_reg = lra_create_new_reg (sec_mode, NULL_RTX, | |
2840 | NO_REGS, "secondary"); | |
2841 | /* If the mode is changed, it should be wider. */ | |
66aa7879 | 2842 | lra_assert (GET_MODE_SIZE (sec_mode) >= GET_MODE_SIZE (rld_mode)); |
89d56d79 VM |
2843 | if (sec_mode != rld_mode) |
2844 | { | |
2845 | /* If the target says specifically to use another mode for | |
2846 | secondary memory moves we can not reuse the original | |
2847 | insn. */ | |
2848 | after = emit_spill_move (false, new_reg, dest); | |
2849 | lra_process_new_insns (curr_insn, NULL_RTX, after, | |
2850 | "Inserting the sec. move"); | |
2851 | before = emit_spill_move (true, new_reg, src); | |
2852 | lra_process_new_insns (curr_insn, before, NULL_RTX, "Changing on"); | |
2853 | lra_set_insn_deleted (curr_insn); | |
2854 | } | |
2855 | else if (dest == rld) | |
2856 | { | |
2857 | *curr_id->operand_loc[0] = new_reg; | |
66aa7879 VM |
2858 | after = emit_spill_move (false, new_reg, dest); |
2859 | lra_process_new_insns (curr_insn, NULL_RTX, after, | |
2860 | "Inserting the sec. move"); | |
2861 | } | |
2862 | else | |
2863 | { | |
89d56d79 | 2864 | *curr_id->operand_loc[1] = new_reg; |
66aa7879 VM |
2865 | before = emit_spill_move (true, new_reg, src); |
2866 | lra_process_new_insns (curr_insn, before, NULL_RTX, | |
2867 | "Inserting the sec. move"); | |
2868 | } | |
2869 | lra_update_insn_regno_info (curr_insn); | |
55a2c322 VM |
2870 | return true; |
2871 | } | |
2872 | #endif | |
2873 | ||
2874 | lra_assert (goal_alt_number >= 0); | |
2875 | lra_set_used_insn_alternative (curr_insn, goal_alt_number); | |
2876 | ||
2877 | if (lra_dump_file != NULL) | |
2878 | { | |
2879 | const char *p; | |
2880 | ||
2881 | fprintf (lra_dump_file, " Choosing alt %d in insn %u:", | |
2882 | goal_alt_number, INSN_UID (curr_insn)); | |
2883 | for (i = 0; i < n_operands; i++) | |
2884 | { | |
2885 | p = (curr_static_id->operand_alternative | |
2886 | [goal_alt_number * n_operands + i].constraint); | |
2887 | if (*p == '\0') | |
2888 | continue; | |
2889 | fprintf (lra_dump_file, " (%d) ", i); | |
2890 | for (; *p != '\0' && *p != ',' && *p != '#'; p++) | |
2891 | fputc (*p, lra_dump_file); | |
2892 | } | |
2893 | fprintf (lra_dump_file, "\n"); | |
2894 | } | |
2895 | ||
2896 | /* Right now, for any pair of operands I and J that are required to | |
2897 | match, with J < I, goal_alt_matches[I] is J. Add I to | |
2898 | goal_alt_matched[J]. */ | |
f4eafc30 | 2899 | |
55a2c322 VM |
2900 | for (i = 0; i < n_operands; i++) |
2901 | if ((j = goal_alt_matches[i]) >= 0) | |
2902 | { | |
2903 | for (k = 0; goal_alt_matched[j][k] >= 0; k++) | |
2904 | ; | |
2905 | /* We allow matching one output operand and several input | |
2906 | operands. */ | |
2907 | lra_assert (k == 0 | |
2908 | || (curr_static_id->operand[j].type == OP_OUT | |
2909 | && curr_static_id->operand[i].type == OP_IN | |
2910 | && (curr_static_id->operand | |
2911 | [goal_alt_matched[j][0]].type == OP_IN))); | |
2912 | goal_alt_matched[j][k] = i; | |
2913 | goal_alt_matched[j][k + 1] = -1; | |
2914 | } | |
f4eafc30 | 2915 | |
55a2c322 VM |
2916 | for (i = 0; i < n_operands; i++) |
2917 | goal_alt_win[i] |= goal_alt_match_win[i]; | |
f4eafc30 | 2918 | |
55a2c322 VM |
2919 | /* Any constants that aren't allowed and can't be reloaded into |
2920 | registers are here changed into memory references. */ | |
2921 | for (i = 0; i < n_operands; i++) | |
2922 | if (goal_alt_win[i]) | |
2923 | { | |
2924 | int regno; | |
2925 | enum reg_class new_class; | |
2926 | rtx reg = *curr_id->operand_loc[i]; | |
2927 | ||
2928 | if (GET_CODE (reg) == SUBREG) | |
2929 | reg = SUBREG_REG (reg); | |
f4eafc30 | 2930 | |
55a2c322 VM |
2931 | if (REG_P (reg) && (regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER) |
2932 | { | |
2933 | bool ok_p = in_class_p (reg, goal_alt[i], &new_class); | |
2934 | ||
2935 | if (new_class != NO_REGS && get_reg_class (regno) != new_class) | |
2936 | { | |
2937 | lra_assert (ok_p); | |
2938 | change_class (regno, new_class, " Change", true); | |
2939 | } | |
2940 | } | |
2941 | } | |
2942 | else | |
2943 | { | |
2944 | const char *constraint; | |
2945 | char c; | |
2946 | rtx op = *curr_id->operand_loc[i]; | |
2947 | rtx subreg = NULL_RTX; | |
2948 | enum machine_mode mode = curr_operand_mode[i]; | |
f4eafc30 | 2949 | |
55a2c322 VM |
2950 | if (GET_CODE (op) == SUBREG) |
2951 | { | |
2952 | subreg = op; | |
2953 | op = SUBREG_REG (op); | |
2954 | mode = GET_MODE (op); | |
2955 | } | |
f4eafc30 | 2956 | |
55a2c322 VM |
2957 | if (CONST_POOL_OK_P (mode, op) |
2958 | && ((targetm.preferred_reload_class | |
2959 | (op, (enum reg_class) goal_alt[i]) == NO_REGS) | |
2960 | || no_input_reloads_p)) | |
2961 | { | |
2962 | rtx tem = force_const_mem (mode, op); | |
f4eafc30 | 2963 | |
55a2c322 VM |
2964 | change_p = true; |
2965 | if (subreg != NULL_RTX) | |
2966 | tem = gen_rtx_SUBREG (mode, tem, SUBREG_BYTE (subreg)); | |
f4eafc30 | 2967 | |
55a2c322 VM |
2968 | *curr_id->operand_loc[i] = tem; |
2969 | lra_update_dup (curr_id, i); | |
2970 | process_address (i, &before, &after); | |
f4eafc30 | 2971 | |
55a2c322 VM |
2972 | /* If the alternative accepts constant pool refs directly |
2973 | there will be no reload needed at all. */ | |
2974 | if (subreg != NULL_RTX) | |
2975 | continue; | |
2976 | /* Skip alternatives before the one requested. */ | |
2977 | constraint = (curr_static_id->operand_alternative | |
2978 | [goal_alt_number * n_operands + i].constraint); | |
2979 | for (; | |
2980 | (c = *constraint) && c != ',' && c != '#'; | |
2981 | constraint += CONSTRAINT_LEN (c, constraint)) | |
2982 | { | |
2983 | if (c == TARGET_MEM_CONSTRAINT || c == 'o') | |
2984 | break; | |
2985 | #ifdef EXTRA_CONSTRAINT_STR | |
2986 | if (EXTRA_MEMORY_CONSTRAINT (c, constraint) | |
2987 | && EXTRA_CONSTRAINT_STR (tem, c, constraint)) | |
2988 | break; | |
2989 | #endif | |
2990 | } | |
2991 | if (c == '\0' || c == ',' || c == '#') | |
2992 | continue; | |
f4eafc30 | 2993 | |
55a2c322 VM |
2994 | goal_alt_win[i] = true; |
2995 | } | |
2996 | } | |
f4eafc30 | 2997 | |
55a2c322 VM |
2998 | for (i = 0; i < n_operands; i++) |
2999 | { | |
3000 | rtx old, new_reg; | |
3001 | rtx op = *curr_id->operand_loc[i]; | |
3002 | ||
3003 | if (goal_alt_win[i]) | |
3004 | { | |
3005 | if (goal_alt[i] == NO_REGS | |
3006 | && REG_P (op) | |
3007 | /* When we assign NO_REGS it means that we will not | |
3008 | assign a hard register to the scratch pseudo by | |
3009 | assigment pass and the scratch pseudo will be | |
3010 | spilled. Spilled scratch pseudos are transformed | |
3011 | back to scratches at the LRA end. */ | |
3012 | && lra_former_scratch_operand_p (curr_insn, i)) | |
3013 | change_class (REGNO (op), NO_REGS, " Change", true); | |
3014 | continue; | |
3015 | } | |
f4eafc30 | 3016 | |
55a2c322 VM |
3017 | /* Operands that match previous ones have already been handled. */ |
3018 | if (goal_alt_matches[i] >= 0) | |
3019 | continue; | |
3020 | ||
3021 | /* We should not have an operand with a non-offsettable address | |
3022 | appearing where an offsettable address will do. It also may | |
3023 | be a case when the address should be special in other words | |
3024 | not a general one (e.g. it needs no index reg). */ | |
3025 | if (goal_alt_matched[i][0] == -1 && goal_alt_offmemok[i] && MEM_P (op)) | |
3026 | { | |
3027 | enum reg_class rclass; | |
3028 | rtx *loc = &XEXP (op, 0); | |
3029 | enum rtx_code code = GET_CODE (*loc); | |
3030 | ||
3031 | push_to_sequence (before); | |
3032 | rclass = base_reg_class (GET_MODE (op), MEM_ADDR_SPACE (op), | |
3033 | MEM, SCRATCH); | |
3034 | if (GET_RTX_CLASS (code) == RTX_AUTOINC) | |
3035 | new_reg = emit_inc (rclass, *loc, *loc, | |
3036 | /* This value does not matter for MODIFY. */ | |
3037 | GET_MODE_SIZE (GET_MODE (op))); | |
3038 | else if (get_reload_reg (OP_IN, Pmode, *loc, rclass, | |
3039 | "offsetable address", &new_reg)) | |
3040 | lra_emit_move (new_reg, *loc); | |
3041 | before = get_insns (); | |
3042 | end_sequence (); | |
3043 | *loc = new_reg; | |
3044 | lra_update_dup (curr_id, i); | |
3045 | } | |
3046 | else if (goal_alt_matched[i][0] == -1) | |
3047 | { | |
3048 | enum machine_mode mode; | |
3049 | rtx reg, *loc; | |
3050 | int hard_regno, byte; | |
3051 | enum op_type type = curr_static_id->operand[i].type; | |
3052 | ||
3053 | loc = curr_id->operand_loc[i]; | |
3054 | mode = curr_operand_mode[i]; | |
3055 | if (GET_CODE (*loc) == SUBREG) | |
3056 | { | |
3057 | reg = SUBREG_REG (*loc); | |
3058 | byte = SUBREG_BYTE (*loc); | |
3059 | if (REG_P (reg) | |
3060 | /* Strict_low_part requires reload the register not | |
3061 | the sub-register. */ | |
3062 | && (curr_static_id->operand[i].strict_low | |
3063 | || (GET_MODE_SIZE (mode) | |
3064 | <= GET_MODE_SIZE (GET_MODE (reg)) | |
3065 | && (hard_regno | |
3066 | = get_try_hard_regno (REGNO (reg))) >= 0 | |
3067 | && (simplify_subreg_regno | |
3068 | (hard_regno, | |
3069 | GET_MODE (reg), byte, mode) < 0) | |
3070 | && (goal_alt[i] == NO_REGS | |
3071 | || (simplify_subreg_regno | |
3072 | (ira_class_hard_regs[goal_alt[i]][0], | |
3073 | GET_MODE (reg), byte, mode) >= 0))))) | |
3074 | { | |
3075 | loc = &SUBREG_REG (*loc); | |
3076 | mode = GET_MODE (*loc); | |
3077 | } | |
3078 | } | |
3079 | old = *loc; | |
3080 | if (get_reload_reg (type, mode, old, goal_alt[i], "", &new_reg) | |
3081 | && type != OP_OUT) | |
3082 | { | |
3083 | push_to_sequence (before); | |
3084 | lra_emit_move (new_reg, old); | |
3085 | before = get_insns (); | |
3086 | end_sequence (); | |
3087 | } | |
3088 | *loc = new_reg; | |
3089 | if (type != OP_IN | |
3090 | && find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX) | |
3091 | { | |
3092 | start_sequence (); | |
3093 | lra_emit_move (type == OP_INOUT ? copy_rtx (old) : old, new_reg); | |
3094 | emit_insn (after); | |
3095 | after = get_insns (); | |
3096 | end_sequence (); | |
3097 | *loc = new_reg; | |
3098 | } | |
3099 | for (j = 0; j < goal_alt_dont_inherit_ops_num; j++) | |
3100 | if (goal_alt_dont_inherit_ops[j] == i) | |
3101 | { | |
3102 | lra_set_regno_unique_value (REGNO (new_reg)); | |
3103 | break; | |
3104 | } | |
3105 | lra_update_dup (curr_id, i); | |
3106 | } | |
3107 | else if (curr_static_id->operand[i].type == OP_IN | |
3108 | && (curr_static_id->operand[goal_alt_matched[i][0]].type | |
3109 | == OP_OUT)) | |
3110 | { | |
511dcace VM |
3111 | /* generate reloads for input and matched outputs. */ |
3112 | match_inputs[0] = i; | |
3113 | match_inputs[1] = -1; | |
3114 | match_reload (goal_alt_matched[i][0], match_inputs, | |
55a2c322 VM |
3115 | goal_alt[i], &before, &after); |
3116 | } | |
3117 | else if (curr_static_id->operand[i].type == OP_OUT | |
3118 | && (curr_static_id->operand[goal_alt_matched[i][0]].type | |
3119 | == OP_IN)) | |
511dcace | 3120 | /* Generate reloads for output and matched inputs. */ |
55a2c322 | 3121 | match_reload (i, goal_alt_matched[i], goal_alt[i], &before, &after); |
511dcace VM |
3122 | else if (curr_static_id->operand[i].type == OP_IN |
3123 | && (curr_static_id->operand[goal_alt_matched[i][0]].type | |
3124 | == OP_IN)) | |
3125 | { | |
3126 | /* Generate reloads for matched inputs. */ | |
3127 | match_inputs[0] = i; | |
3128 | for (j = 0; (k = goal_alt_matched[i][j]) >= 0; j++) | |
3129 | match_inputs[j + 1] = k; | |
3130 | match_inputs[j + 1] = -1; | |
3131 | match_reload (-1, match_inputs, goal_alt[i], &before, &after); | |
3132 | } | |
55a2c322 VM |
3133 | else |
3134 | /* We must generate code in any case when function | |
3135 | process_alt_operands decides that it is possible. */ | |
3136 | gcc_unreachable (); | |
3137 | } | |
3138 | if (before != NULL_RTX || after != NULL_RTX | |
3139 | || max_regno_before != max_reg_num ()) | |
3140 | change_p = true; | |
3141 | if (change_p) | |
3142 | { | |
3143 | lra_update_operator_dups (curr_id); | |
3144 | /* Something changes -- process the insn. */ | |
3145 | lra_update_insn_regno_info (curr_insn); | |
3146 | } | |
3147 | lra_process_new_insns (curr_insn, before, after, "Inserting insn reload"); | |
3148 | return change_p; | |
3149 | } | |
3150 | ||
3151 | /* Return true if X is in LIST. */ | |
3152 | static bool | |
3153 | in_list_p (rtx x, rtx list) | |
3154 | { | |
3155 | for (; list != NULL_RTX; list = XEXP (list, 1)) | |
3156 | if (XEXP (list, 0) == x) | |
3157 | return true; | |
3158 | return false; | |
3159 | } | |
3160 | ||
3161 | /* Return true if X contains an allocatable hard register (if | |
3162 | HARD_REG_P) or a (spilled if SPILLED_P) pseudo. */ | |
3163 | static bool | |
3164 | contains_reg_p (rtx x, bool hard_reg_p, bool spilled_p) | |
3165 | { | |
3166 | int i, j; | |
3167 | const char *fmt; | |
3168 | enum rtx_code code; | |
3169 | ||
3170 | code = GET_CODE (x); | |
3171 | if (REG_P (x)) | |
3172 | { | |
3173 | int regno = REGNO (x); | |
3174 | HARD_REG_SET alloc_regs; | |
3175 | ||
3176 | if (hard_reg_p) | |
3177 | { | |
3178 | if (regno >= FIRST_PSEUDO_REGISTER) | |
3179 | regno = lra_get_regno_hard_regno (regno); | |
3180 | if (regno < 0) | |
3181 | return false; | |
3182 | COMPL_HARD_REG_SET (alloc_regs, lra_no_alloc_regs); | |
3183 | return overlaps_hard_reg_set_p (alloc_regs, GET_MODE (x), regno); | |
3184 | } | |
3185 | else | |
3186 | { | |
3187 | if (regno < FIRST_PSEUDO_REGISTER) | |
3188 | return false; | |
3189 | if (! spilled_p) | |
3190 | return true; | |
3191 | return lra_get_regno_hard_regno (regno) < 0; | |
3192 | } | |
3193 | } | |
3194 | fmt = GET_RTX_FORMAT (code); | |
3195 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
3196 | { | |
3197 | if (fmt[i] == 'e') | |
3198 | { | |
3199 | if (contains_reg_p (XEXP (x, i), hard_reg_p, spilled_p)) | |
3200 | return true; | |
3201 | } | |
3202 | else if (fmt[i] == 'E') | |
3203 | { | |
3204 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
3205 | if (contains_reg_p (XVECEXP (x, i, j), hard_reg_p, spilled_p)) | |
3206 | return true; | |
3207 | } | |
3208 | } | |
3209 | return false; | |
3210 | } | |
3211 | ||
28430b2e VM |
3212 | /* Process all regs in location *LOC and change them on equivalent |
3213 | substitution. Return true if any change was done. */ | |
55a2c322 | 3214 | static bool |
28430b2e | 3215 | loc_equivalence_change_p (rtx *loc) |
55a2c322 VM |
3216 | { |
3217 | rtx subst, reg, x = *loc; | |
3218 | bool result = false; | |
3219 | enum rtx_code code = GET_CODE (x); | |
3220 | const char *fmt; | |
3221 | int i, j; | |
3222 | ||
3223 | if (code == SUBREG) | |
3224 | { | |
3225 | reg = SUBREG_REG (x); | |
3226 | if ((subst = get_equiv_substitution (reg)) != reg | |
3227 | && GET_MODE (subst) == VOIDmode) | |
3228 | { | |
3229 | /* We cannot reload debug location. Simplify subreg here | |
3230 | while we know the inner mode. */ | |
3231 | *loc = simplify_gen_subreg (GET_MODE (x), subst, | |
3232 | GET_MODE (reg), SUBREG_BYTE (x)); | |
3233 | return true; | |
3234 | } | |
3235 | } | |
3236 | if (code == REG && (subst = get_equiv_substitution (x)) != x) | |
3237 | { | |
3238 | *loc = subst; | |
3239 | return true; | |
3240 | } | |
3241 | ||
3242 | /* Scan all the operand sub-expressions. */ | |
3243 | fmt = GET_RTX_FORMAT (code); | |
3244 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
3245 | { | |
3246 | if (fmt[i] == 'e') | |
28430b2e | 3247 | result = loc_equivalence_change_p (&XEXP (x, i)) || result; |
55a2c322 VM |
3248 | else if (fmt[i] == 'E') |
3249 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
3250 | result | |
28430b2e | 3251 | = loc_equivalence_change_p (&XVECEXP (x, i, j)) || result; |
55a2c322 VM |
3252 | } |
3253 | return result; | |
3254 | } | |
3255 | ||
d0608e59 JJ |
3256 | /* Similar to loc_equivalence_change_p, but for use as |
3257 | simplify_replace_fn_rtx callback. */ | |
3258 | static rtx | |
3259 | loc_equivalence_callback (rtx loc, const_rtx, void *) | |
3260 | { | |
3261 | if (!REG_P (loc)) | |
3262 | return NULL_RTX; | |
3263 | ||
3264 | rtx subst = get_equiv_substitution (loc); | |
3265 | if (subst != loc) | |
3266 | return subst; | |
3267 | ||
3268 | return NULL_RTX; | |
3269 | } | |
3270 | ||
55a2c322 VM |
3271 | /* Maximum number of generated reload insns per an insn. It is for |
3272 | preventing this pass cycling in a bug case. */ | |
3273 | #define MAX_RELOAD_INSNS_NUMBER LRA_MAX_INSN_RELOADS | |
3274 | ||
3275 | /* The current iteration number of this LRA pass. */ | |
3276 | int lra_constraint_iter; | |
3277 | ||
3278 | /* The current iteration number of this LRA pass after the last spill | |
3279 | pass. */ | |
3280 | int lra_constraint_iter_after_spill; | |
3281 | ||
3282 | /* True if we substituted equiv which needs checking register | |
3283 | allocation correctness because the equivalent value contains | |
3284 | allocatable hard registers or when we restore multi-register | |
3285 | pseudo. */ | |
3286 | bool lra_risky_transformations_p; | |
3287 | ||
3288 | /* Return true if REGNO is referenced in more than one block. */ | |
3289 | static bool | |
3290 | multi_block_pseudo_p (int regno) | |
3291 | { | |
3292 | basic_block bb = NULL; | |
3293 | unsigned int uid; | |
3294 | bitmap_iterator bi; | |
f4eafc30 | 3295 | |
55a2c322 VM |
3296 | if (regno < FIRST_PSEUDO_REGISTER) |
3297 | return false; | |
f4eafc30 | 3298 | |
55a2c322 VM |
3299 | EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi) |
3300 | if (bb == NULL) | |
3301 | bb = BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn); | |
3302 | else if (BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn) != bb) | |
3303 | return true; | |
3304 | return false; | |
3305 | } | |
3306 | ||
1966c91b VM |
3307 | /* Return true if LIST contains a deleted insn. */ |
3308 | static bool | |
3309 | contains_deleted_insn_p (rtx list) | |
3310 | { | |
3311 | for (; list != NULL_RTX; list = XEXP (list, 1)) | |
3312 | if (NOTE_P (XEXP (list, 0)) | |
3313 | && NOTE_KIND (XEXP (list, 0)) == NOTE_INSN_DELETED) | |
3314 | return true; | |
3315 | return false; | |
3316 | } | |
3317 | ||
55a2c322 VM |
3318 | /* Return true if X contains a pseudo dying in INSN. */ |
3319 | static bool | |
3320 | dead_pseudo_p (rtx x, rtx insn) | |
3321 | { | |
3322 | int i, j; | |
3323 | const char *fmt; | |
3324 | enum rtx_code code; | |
3325 | ||
3326 | if (REG_P (x)) | |
3327 | return (insn != NULL_RTX | |
3328 | && find_regno_note (insn, REG_DEAD, REGNO (x)) != NULL_RTX); | |
3329 | code = GET_CODE (x); | |
3330 | fmt = GET_RTX_FORMAT (code); | |
3331 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
3332 | { | |
3333 | if (fmt[i] == 'e') | |
3334 | { | |
3335 | if (dead_pseudo_p (XEXP (x, i), insn)) | |
3336 | return true; | |
3337 | } | |
3338 | else if (fmt[i] == 'E') | |
3339 | { | |
3340 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
3341 | if (dead_pseudo_p (XVECEXP (x, i, j), insn)) | |
3342 | return true; | |
3343 | } | |
3344 | } | |
3345 | return false; | |
3346 | } | |
3347 | ||
3348 | /* Return true if INSN contains a dying pseudo in INSN right hand | |
3349 | side. */ | |
3350 | static bool | |
3351 | insn_rhs_dead_pseudo_p (rtx insn) | |
3352 | { | |
3353 | rtx set = single_set (insn); | |
3354 | ||
3355 | gcc_assert (set != NULL); | |
3356 | return dead_pseudo_p (SET_SRC (set), insn); | |
3357 | } | |
3358 | ||
3359 | /* Return true if any init insn of REGNO contains a dying pseudo in | |
3360 | insn right hand side. */ | |
3361 | static bool | |
3362 | init_insn_rhs_dead_pseudo_p (int regno) | |
3363 | { | |
3364 | rtx insns = ira_reg_equiv[regno].init_insns; | |
3365 | ||
3366 | if (insns == NULL) | |
3367 | return false; | |
3368 | if (INSN_P (insns)) | |
3369 | return insn_rhs_dead_pseudo_p (insns); | |
3370 | for (; insns != NULL_RTX; insns = XEXP (insns, 1)) | |
3371 | if (insn_rhs_dead_pseudo_p (XEXP (insns, 0))) | |
3372 | return true; | |
3373 | return false; | |
3374 | } | |
3375 | ||
3376 | /* Entry function of LRA constraint pass. Return true if the | |
3377 | constraint pass did change the code. */ | |
3378 | bool | |
3379 | lra_constraints (bool first_p) | |
3380 | { | |
3381 | bool changed_p; | |
3382 | int i, hard_regno, new_insns_num; | |
6cd1dd26 VM |
3383 | unsigned int min_len, new_min_len, uid; |
3384 | rtx set, x, reg, dest_reg; | |
55a2c322 | 3385 | basic_block last_bb; |
6cd1dd26 VM |
3386 | bitmap_head equiv_insn_bitmap; |
3387 | bitmap_iterator bi; | |
55a2c322 VM |
3388 | |
3389 | lra_constraint_iter++; | |
3390 | if (lra_dump_file != NULL) | |
3391 | fprintf (lra_dump_file, "\n********** Local #%d: **********\n\n", | |
3392 | lra_constraint_iter); | |
3393 | lra_constraint_iter_after_spill++; | |
8e3a4869 | 3394 | if (lra_constraint_iter_after_spill > LRA_MAX_CONSTRAINT_ITERATION_NUMBER) |
55a2c322 VM |
3395 | internal_error |
3396 | ("Maximum number of LRA constraint passes is achieved (%d)\n", | |
8e3a4869 | 3397 | LRA_MAX_CONSTRAINT_ITERATION_NUMBER); |
55a2c322 VM |
3398 | changed_p = false; |
3399 | lra_risky_transformations_p = false; | |
3400 | new_insn_uid_start = get_max_uid (); | |
3401 | new_regno_start = first_p ? lra_constraint_new_regno_start : max_reg_num (); | |
6cd1dd26 | 3402 | bitmap_initialize (&equiv_insn_bitmap, ®_obstack); |
55a2c322 VM |
3403 | for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++) |
3404 | if (lra_reg_info[i].nrefs != 0) | |
3405 | { | |
3406 | ira_reg_equiv[i].profitable_p = true; | |
6cd1dd26 | 3407 | reg = regno_reg_rtx[i]; |
55a2c322 VM |
3408 | if ((hard_regno = lra_get_regno_hard_regno (i)) >= 0) |
3409 | { | |
9011b0f6 | 3410 | int j, nregs; |
f4eafc30 | 3411 | |
9011b0f6 | 3412 | nregs = hard_regno_nregs[hard_regno][lra_reg_info[i].biggest_mode]; |
55a2c322 VM |
3413 | for (j = 0; j < nregs; j++) |
3414 | df_set_regs_ever_live (hard_regno + j, true); | |
3415 | } | |
6cd1dd26 | 3416 | else if ((x = get_equiv_substitution (reg)) != reg) |
55a2c322 VM |
3417 | { |
3418 | bool pseudo_p = contains_reg_p (x, false, false); | |
3419 | rtx set, insn; | |
3420 | ||
1966c91b VM |
3421 | /* After RTL transformation, we can not guarantee that |
3422 | pseudo in the substitution was not reloaded which might | |
3423 | make equivalence invalid. For example, in reverse | |
3424 | equiv of p0 | |
3425 | ||
3426 | p0 <- ... | |
3427 | ... | |
3428 | equiv_mem <- p0 | |
3429 | ||
3430 | the memory address register was reloaded before the 2nd | |
3431 | insn. */ | |
3432 | if ((! first_p && pseudo_p) | |
3433 | /* We don't use DF for compilation speed sake. So it | |
3434 | is problematic to update live info when we use an | |
3435 | equivalence containing pseudos in more than one | |
3436 | BB. */ | |
3437 | || (pseudo_p && multi_block_pseudo_p (i)) | |
3438 | /* If an init insn was deleted for some reason, cancel | |
3439 | the equiv. We could update the equiv insns after | |
3440 | transformations including an equiv insn deletion | |
3441 | but it is not worthy as such cases are extremely | |
3442 | rare. */ | |
3443 | || contains_deleted_insn_p (ira_reg_equiv[i].init_insns) | |
55a2c322 VM |
3444 | /* If it is not a reverse equivalence, we check that a |
3445 | pseudo in rhs of the init insn is not dying in the | |
3446 | insn. Otherwise, the live info at the beginning of | |
3447 | the corresponding BB might be wrong after we | |
3448 | removed the insn. When the equiv can be a | |
3449 | constant, the right hand side of the init insn can | |
3450 | be a pseudo. */ | |
3451 | || (! ((insn = ira_reg_equiv[i].init_insns) != NULL_RTX | |
3452 | && INSN_P (insn) | |
3453 | && (set = single_set (insn)) != NULL_RTX | |
3454 | && REG_P (SET_DEST (set)) | |
3455 | && (int) REGNO (SET_DEST (set)) == i) | |
b28ece32 VM |
3456 | && init_insn_rhs_dead_pseudo_p (i)) |
3457 | /* Prevent access beyond equivalent memory for | |
3458 | paradoxical subregs. */ | |
3459 | || (MEM_P (x) | |
3460 | && (GET_MODE_SIZE (lra_reg_info[i].biggest_mode) | |
3461 | > GET_MODE_SIZE (GET_MODE (x))))) | |
55a2c322 | 3462 | ira_reg_equiv[i].defined_p = false; |
55a2c322 VM |
3463 | if (contains_reg_p (x, false, true)) |
3464 | ira_reg_equiv[i].profitable_p = false; | |
6cd1dd26 VM |
3465 | if (get_equiv_substitution (reg) != reg) |
3466 | bitmap_ior_into (&equiv_insn_bitmap, &lra_reg_info[i].insn_bitmap); | |
55a2c322 VM |
3467 | } |
3468 | } | |
6cd1dd26 VM |
3469 | /* We should add all insns containing pseudos which should be |
3470 | substituted by their equivalences. */ | |
3471 | EXECUTE_IF_SET_IN_BITMAP (&equiv_insn_bitmap, 0, uid, bi) | |
3472 | lra_push_insn_by_uid (uid); | |
55a2c322 VM |
3473 | lra_eliminate (false); |
3474 | min_len = lra_insn_stack_length (); | |
3475 | new_insns_num = 0; | |
3476 | last_bb = NULL; | |
3477 | changed_p = false; | |
3478 | while ((new_min_len = lra_insn_stack_length ()) != 0) | |
3479 | { | |
3480 | curr_insn = lra_pop_insn (); | |
3481 | --new_min_len; | |
f4eafc30 | 3482 | curr_bb = BLOCK_FOR_INSN (curr_insn); |
55a2c322 VM |
3483 | if (curr_bb != last_bb) |
3484 | { | |
3485 | last_bb = curr_bb; | |
3486 | bb_reload_num = lra_curr_reload_num; | |
3487 | } | |
3488 | if (min_len > new_min_len) | |
3489 | { | |
3490 | min_len = new_min_len; | |
3491 | new_insns_num = 0; | |
3492 | } | |
3493 | if (new_insns_num > MAX_RELOAD_INSNS_NUMBER) | |
3494 | internal_error | |
3495 | ("Max. number of generated reload insns per insn is achieved (%d)\n", | |
3496 | MAX_RELOAD_INSNS_NUMBER); | |
3497 | new_insns_num++; | |
3498 | if (DEBUG_INSN_P (curr_insn)) | |
3499 | { | |
3500 | /* We need to check equivalence in debug insn and change | |
3501 | pseudo to the equivalent value if necessary. */ | |
3502 | curr_id = lra_get_insn_recog_data (curr_insn); | |
d0608e59 | 3503 | if (bitmap_bit_p (&equiv_insn_bitmap, INSN_UID (curr_insn))) |
4d64ce5c | 3504 | { |
d0608e59 JJ |
3505 | rtx old = *curr_id->operand_loc[0]; |
3506 | *curr_id->operand_loc[0] | |
3507 | = simplify_replace_fn_rtx (old, NULL_RTX, | |
3508 | loc_equivalence_callback, NULL); | |
3509 | if (old != *curr_id->operand_loc[0]) | |
3510 | { | |
3511 | lra_update_insn_regno_info (curr_insn); | |
3512 | changed_p = true; | |
3513 | } | |
4d64ce5c | 3514 | } |
55a2c322 VM |
3515 | } |
3516 | else if (INSN_P (curr_insn)) | |
3517 | { | |
3518 | if ((set = single_set (curr_insn)) != NULL_RTX) | |
3519 | { | |
3520 | dest_reg = SET_DEST (set); | |
3521 | /* The equivalence pseudo could be set up as SUBREG in a | |
3522 | case when it is a call restore insn in a mode | |
3523 | different from the pseudo mode. */ | |
3524 | if (GET_CODE (dest_reg) == SUBREG) | |
3525 | dest_reg = SUBREG_REG (dest_reg); | |
3526 | if ((REG_P (dest_reg) | |
3527 | && (x = get_equiv_substitution (dest_reg)) != dest_reg | |
3528 | /* Remove insns which set up a pseudo whose value | |
3529 | can not be changed. Such insns might be not in | |
3530 | init_insns because we don't update equiv data | |
3531 | during insn transformations. | |
f4eafc30 | 3532 | |
55a2c322 VM |
3533 | As an example, let suppose that a pseudo got |
3534 | hard register and on the 1st pass was not | |
3535 | changed to equivalent constant. We generate an | |
3536 | additional insn setting up the pseudo because of | |
3537 | secondary memory movement. Then the pseudo is | |
3538 | spilled and we use the equiv constant. In this | |
3539 | case we should remove the additional insn and | |
3540 | this insn is not init_insns list. */ | |
3541 | && (! MEM_P (x) || MEM_READONLY_P (x) | |
3542 | || in_list_p (curr_insn, | |
3543 | ira_reg_equiv | |
3544 | [REGNO (dest_reg)].init_insns))) | |
3545 | || (((x = get_equiv_substitution (SET_SRC (set))) | |
3546 | != SET_SRC (set)) | |
3547 | && in_list_p (curr_insn, | |
3548 | ira_reg_equiv | |
3549 | [REGNO (SET_SRC (set))].init_insns))) | |
3550 | { | |
3551 | /* This is equiv init insn of pseudo which did not get a | |
3552 | hard register -- remove the insn. */ | |
3553 | if (lra_dump_file != NULL) | |
3554 | { | |
3555 | fprintf (lra_dump_file, | |
3556 | " Removing equiv init insn %i (freq=%d)\n", | |
3557 | INSN_UID (curr_insn), | |
3558 | BLOCK_FOR_INSN (curr_insn)->frequency); | |
cfbeaedf | 3559 | dump_insn_slim (lra_dump_file, curr_insn); |
55a2c322 VM |
3560 | } |
3561 | if (contains_reg_p (x, true, false)) | |
3562 | lra_risky_transformations_p = true; | |
3563 | lra_set_insn_deleted (curr_insn); | |
3564 | continue; | |
3565 | } | |
3566 | } | |
3567 | curr_id = lra_get_insn_recog_data (curr_insn); | |
3568 | curr_static_id = curr_id->insn_static_data; | |
3569 | init_curr_insn_input_reloads (); | |
3570 | init_curr_operand_mode (); | |
3571 | if (curr_insn_transform ()) | |
3572 | changed_p = true; | |
28430b2e VM |
3573 | /* Check non-transformed insns too for equiv change as USE |
3574 | or CLOBBER don't need reloads but can contain pseudos | |
3575 | being changed on their equivalences. */ | |
3576 | else if (bitmap_bit_p (&equiv_insn_bitmap, INSN_UID (curr_insn)) | |
3577 | && loc_equivalence_change_p (&PATTERN (curr_insn))) | |
3578 | { | |
3579 | lra_update_insn_regno_info (curr_insn); | |
3580 | changed_p = true; | |
3581 | } | |
55a2c322 VM |
3582 | } |
3583 | } | |
28430b2e | 3584 | bitmap_clear (&equiv_insn_bitmap); |
55a2c322 VM |
3585 | /* If we used a new hard regno, changed_p should be true because the |
3586 | hard reg is assigned to a new pseudo. */ | |
3587 | #ifdef ENABLE_CHECKING | |
3588 | if (! changed_p) | |
3589 | { | |
3590 | for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++) | |
3591 | if (lra_reg_info[i].nrefs != 0 | |
3592 | && (hard_regno = lra_get_regno_hard_regno (i)) >= 0) | |
3593 | { | |
3594 | int j, nregs = hard_regno_nregs[hard_regno][PSEUDO_REGNO_MODE (i)]; | |
f4eafc30 | 3595 | |
55a2c322 VM |
3596 | for (j = 0; j < nregs; j++) |
3597 | lra_assert (df_regs_ever_live_p (hard_regno + j)); | |
3598 | } | |
3599 | } | |
3600 | #endif | |
3601 | return changed_p; | |
3602 | } | |
3603 | ||
3604 | /* Initiate the LRA constraint pass. It is done once per | |
3605 | function. */ | |
3606 | void | |
3607 | lra_constraints_init (void) | |
3608 | { | |
3609 | } | |
3610 | ||
3611 | /* Finalize the LRA constraint pass. It is done once per | |
3612 | function. */ | |
3613 | void | |
3614 | lra_constraints_finish (void) | |
3615 | { | |
3616 | } | |
3617 | ||
3618 | \f | |
3619 | ||
3620 | /* This page contains code to do inheritance/split | |
3621 | transformations. */ | |
3622 | ||
3623 | /* Number of reloads passed so far in current EBB. */ | |
3624 | static int reloads_num; | |
3625 | ||
3626 | /* Number of calls passed so far in current EBB. */ | |
3627 | static int calls_num; | |
3628 | ||
3629 | /* Current reload pseudo check for validity of elements in | |
3630 | USAGE_INSNS. */ | |
3631 | static int curr_usage_insns_check; | |
3632 | ||
3633 | /* Info about last usage of registers in EBB to do inheritance/split | |
3634 | transformation. Inheritance transformation is done from a spilled | |
3635 | pseudo and split transformations from a hard register or a pseudo | |
3636 | assigned to a hard register. */ | |
3637 | struct usage_insns | |
3638 | { | |
3639 | /* If the value is equal to CURR_USAGE_INSNS_CHECK, then the member | |
3640 | value INSNS is valid. The insns is chain of optional debug insns | |
3641 | and a finishing non-debug insn using the corresponding reg. */ | |
3642 | int check; | |
3643 | /* Value of global reloads_num at the last insn in INSNS. */ | |
3644 | int reloads_num; | |
3645 | /* Value of global reloads_nums at the last insn in INSNS. */ | |
3646 | int calls_num; | |
3647 | /* It can be true only for splitting. And it means that the restore | |
3648 | insn should be put after insn given by the following member. */ | |
3649 | bool after_p; | |
3650 | /* Next insns in the current EBB which use the original reg and the | |
3651 | original reg value is not changed between the current insn and | |
3652 | the next insns. In order words, e.g. for inheritance, if we need | |
3653 | to use the original reg value again in the next insns we can try | |
3654 | to use the value in a hard register from a reload insn of the | |
3655 | current insn. */ | |
3656 | rtx insns; | |
3657 | }; | |
3658 | ||
3659 | /* Map: regno -> corresponding pseudo usage insns. */ | |
3660 | static struct usage_insns *usage_insns; | |
3661 | ||
3662 | static void | |
3663 | setup_next_usage_insn (int regno, rtx insn, int reloads_num, bool after_p) | |
3664 | { | |
3665 | usage_insns[regno].check = curr_usage_insns_check; | |
3666 | usage_insns[regno].insns = insn; | |
3667 | usage_insns[regno].reloads_num = reloads_num; | |
3668 | usage_insns[regno].calls_num = calls_num; | |
3669 | usage_insns[regno].after_p = after_p; | |
3670 | } | |
3671 | ||
3672 | /* The function is used to form list REGNO usages which consists of | |
3673 | optional debug insns finished by a non-debug insn using REGNO. | |
3674 | RELOADS_NUM is current number of reload insns processed so far. */ | |
3675 | static void | |
3676 | add_next_usage_insn (int regno, rtx insn, int reloads_num) | |
3677 | { | |
3678 | rtx next_usage_insns; | |
f4eafc30 | 3679 | |
55a2c322 VM |
3680 | if (usage_insns[regno].check == curr_usage_insns_check |
3681 | && (next_usage_insns = usage_insns[regno].insns) != NULL_RTX | |
3682 | && DEBUG_INSN_P (insn)) | |
3683 | { | |
3684 | /* Check that we did not add the debug insn yet. */ | |
3685 | if (next_usage_insns != insn | |
3686 | && (GET_CODE (next_usage_insns) != INSN_LIST | |
3687 | || XEXP (next_usage_insns, 0) != insn)) | |
3688 | usage_insns[regno].insns = gen_rtx_INSN_LIST (VOIDmode, insn, | |
3689 | next_usage_insns); | |
3690 | } | |
3691 | else if (NONDEBUG_INSN_P (insn)) | |
3692 | setup_next_usage_insn (regno, insn, reloads_num, false); | |
3693 | else | |
3694 | usage_insns[regno].check = 0; | |
3695 | } | |
f4eafc30 | 3696 | |
55a2c322 VM |
3697 | /* Replace all references to register OLD_REGNO in *LOC with pseudo |
3698 | register NEW_REG. Return true if any change was made. */ | |
3699 | static bool | |
3700 | substitute_pseudo (rtx *loc, int old_regno, rtx new_reg) | |
3701 | { | |
3702 | rtx x = *loc; | |
3703 | bool result = false; | |
3704 | enum rtx_code code; | |
3705 | const char *fmt; | |
3706 | int i, j; | |
3707 | ||
3708 | if (x == NULL_RTX) | |
3709 | return false; | |
3710 | ||
3711 | code = GET_CODE (x); | |
3712 | if (code == REG && (int) REGNO (x) == old_regno) | |
3713 | { | |
3714 | enum machine_mode mode = GET_MODE (*loc); | |
3715 | enum machine_mode inner_mode = GET_MODE (new_reg); | |
3716 | ||
3717 | if (mode != inner_mode) | |
3718 | { | |
3719 | if (GET_MODE_SIZE (mode) >= GET_MODE_SIZE (inner_mode) | |
3720 | || ! SCALAR_INT_MODE_P (inner_mode)) | |
3721 | new_reg = gen_rtx_SUBREG (mode, new_reg, 0); | |
3722 | else | |
3723 | new_reg = gen_lowpart_SUBREG (mode, new_reg); | |
3724 | } | |
3725 | *loc = new_reg; | |
3726 | return true; | |
3727 | } | |
3728 | ||
3729 | /* Scan all the operand sub-expressions. */ | |
3730 | fmt = GET_RTX_FORMAT (code); | |
3731 | for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--) | |
3732 | { | |
3733 | if (fmt[i] == 'e') | |
3734 | { | |
3735 | if (substitute_pseudo (&XEXP (x, i), old_regno, new_reg)) | |
3736 | result = true; | |
3737 | } | |
3738 | else if (fmt[i] == 'E') | |
3739 | { | |
3740 | for (j = XVECLEN (x, i) - 1; j >= 0; j--) | |
3741 | if (substitute_pseudo (&XVECEXP (x, i, j), old_regno, new_reg)) | |
3742 | result = true; | |
3743 | } | |
3744 | } | |
3745 | return result; | |
3746 | } | |
3747 | ||
bc3591eb VM |
3748 | /* Return first non-debug insn in list USAGE_INSNS. */ |
3749 | static rtx | |
3750 | skip_usage_debug_insns (rtx usage_insns) | |
3751 | { | |
3752 | rtx insn; | |
3753 | ||
3754 | /* Skip debug insns. */ | |
3755 | for (insn = usage_insns; | |
3756 | insn != NULL_RTX && GET_CODE (insn) == INSN_LIST; | |
3757 | insn = XEXP (insn, 1)) | |
3758 | ; | |
3759 | return insn; | |
3760 | } | |
3761 | ||
3762 | /* Return true if we need secondary memory moves for insn in | |
3763 | USAGE_INSNS after inserting inherited pseudo of class INHER_CL | |
3764 | into the insn. */ | |
3765 | static bool | |
fbebbadd JR |
3766 | check_secondary_memory_needed_p (enum reg_class inher_cl ATTRIBUTE_UNUSED, |
3767 | rtx usage_insns ATTRIBUTE_UNUSED) | |
bc3591eb VM |
3768 | { |
3769 | #ifndef SECONDARY_MEMORY_NEEDED | |
3770 | return false; | |
3771 | #else | |
3772 | rtx insn, set, dest; | |
3773 | enum reg_class cl; | |
3774 | ||
3775 | if (inher_cl == ALL_REGS | |
3776 | || (insn = skip_usage_debug_insns (usage_insns)) == NULL_RTX) | |
3777 | return false; | |
3778 | lra_assert (INSN_P (insn)); | |
3779 | if ((set = single_set (insn)) == NULL_RTX || ! REG_P (SET_DEST (set))) | |
3780 | return false; | |
3781 | dest = SET_DEST (set); | |
3782 | if (! REG_P (dest)) | |
3783 | return false; | |
3784 | lra_assert (inher_cl != NO_REGS); | |
3785 | cl = get_reg_class (REGNO (dest)); | |
3786 | return (cl != NO_REGS && cl != ALL_REGS | |
3787 | && SECONDARY_MEMORY_NEEDED (inher_cl, cl, GET_MODE (dest))); | |
3788 | #endif | |
3789 | } | |
3790 | ||
55a2c322 VM |
3791 | /* Registers involved in inheritance/split in the current EBB |
3792 | (inheritance/split pseudos and original registers). */ | |
3793 | static bitmap_head check_only_regs; | |
3794 | ||
3795 | /* Do inheritance transformations for insn INSN, which defines (if | |
3796 | DEF_P) or uses ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which | |
3797 | instruction in the EBB next uses ORIGINAL_REGNO; it has the same | |
3798 | form as the "insns" field of usage_insns. Return true if we | |
3799 | succeed in such transformation. | |
3800 | ||
3801 | The transformations look like: | |
3802 | ||
3803 | p <- ... i <- ... | |
3804 | ... p <- i (new insn) | |
3805 | ... => | |
3806 | <- ... p ... <- ... i ... | |
3807 | or | |
3808 | ... i <- p (new insn) | |
3809 | <- ... p ... <- ... i ... | |
3810 | ... => | |
3811 | <- ... p ... <- ... i ... | |
3812 | where p is a spilled original pseudo and i is a new inheritance pseudo. | |
f4eafc30 L |
3813 | |
3814 | ||
55a2c322 VM |
3815 | The inheritance pseudo has the smallest class of two classes CL and |
3816 | class of ORIGINAL REGNO. */ | |
3817 | static bool | |
3818 | inherit_reload_reg (bool def_p, int original_regno, | |
3819 | enum reg_class cl, rtx insn, rtx next_usage_insns) | |
3820 | { | |
3821 | enum reg_class rclass = lra_get_allocno_class (original_regno); | |
3822 | rtx original_reg = regno_reg_rtx[original_regno]; | |
3823 | rtx new_reg, new_insns, usage_insn; | |
3824 | ||
3825 | lra_assert (! usage_insns[original_regno].after_p); | |
3826 | if (lra_dump_file != NULL) | |
3827 | fprintf (lra_dump_file, | |
bc3591eb | 3828 | " <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n"); |
55a2c322 VM |
3829 | if (! ira_reg_classes_intersect_p[cl][rclass]) |
3830 | { | |
3831 | if (lra_dump_file != NULL) | |
3832 | { | |
3833 | fprintf (lra_dump_file, | |
bc3591eb | 3834 | " Rejecting inheritance for %d " |
55a2c322 VM |
3835 | "because of disjoint classes %s and %s\n", |
3836 | original_regno, reg_class_names[cl], | |
3837 | reg_class_names[rclass]); | |
3838 | fprintf (lra_dump_file, | |
bc3591eb | 3839 | " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n"); |
55a2c322 VM |
3840 | } |
3841 | return false; | |
3842 | } | |
3843 | if ((ira_class_subset_p[cl][rclass] && cl != rclass) | |
3844 | /* We don't use a subset of two classes because it can be | |
3845 | NO_REGS. This transformation is still profitable in most | |
3846 | cases even if the classes are not intersected as register | |
3847 | move is probably cheaper than a memory load. */ | |
3848 | || ira_class_hard_regs_num[cl] < ira_class_hard_regs_num[rclass]) | |
3849 | { | |
3850 | if (lra_dump_file != NULL) | |
3851 | fprintf (lra_dump_file, " Use smallest class of %s and %s\n", | |
3852 | reg_class_names[cl], reg_class_names[rclass]); | |
f4eafc30 | 3853 | |
55a2c322 VM |
3854 | rclass = cl; |
3855 | } | |
66aa7879 | 3856 | if (check_secondary_memory_needed_p (rclass, next_usage_insns)) |
bc3591eb VM |
3857 | { |
3858 | /* Reject inheritance resulting in secondary memory moves. | |
3859 | Otherwise, there is a danger in LRA cycling. Also such | |
3860 | transformation will be unprofitable. */ | |
3861 | if (lra_dump_file != NULL) | |
3862 | { | |
3863 | rtx insn = skip_usage_debug_insns (next_usage_insns); | |
3864 | rtx set = single_set (insn); | |
3865 | ||
3866 | lra_assert (set != NULL_RTX); | |
3867 | ||
3868 | rtx dest = SET_DEST (set); | |
3869 | ||
3870 | lra_assert (REG_P (dest)); | |
3871 | fprintf (lra_dump_file, | |
3872 | " Rejecting inheritance for insn %d(%s)<-%d(%s) " | |
3873 | "as secondary mem is needed\n", | |
3874 | REGNO (dest), reg_class_names[get_reg_class (REGNO (dest))], | |
66aa7879 | 3875 | original_regno, reg_class_names[rclass]); |
bc3591eb VM |
3876 | fprintf (lra_dump_file, |
3877 | " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n"); | |
3878 | } | |
3879 | return false; | |
3880 | } | |
55a2c322 VM |
3881 | new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg, |
3882 | rclass, "inheritance"); | |
3883 | start_sequence (); | |
3884 | if (def_p) | |
3885 | emit_move_insn (original_reg, new_reg); | |
3886 | else | |
3887 | emit_move_insn (new_reg, original_reg); | |
3888 | new_insns = get_insns (); | |
3889 | end_sequence (); | |
3890 | if (NEXT_INSN (new_insns) != NULL_RTX) | |
3891 | { | |
3892 | if (lra_dump_file != NULL) | |
3893 | { | |
3894 | fprintf (lra_dump_file, | |
bc3591eb | 3895 | " Rejecting inheritance %d->%d " |
55a2c322 VM |
3896 | "as it results in 2 or more insns:\n", |
3897 | original_regno, REGNO (new_reg)); | |
cfbeaedf | 3898 | dump_rtl_slim (lra_dump_file, new_insns, NULL_RTX, -1, 0); |
55a2c322 VM |
3899 | fprintf (lra_dump_file, |
3900 | " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n"); | |
3901 | } | |
3902 | return false; | |
3903 | } | |
3904 | substitute_pseudo (&insn, original_regno, new_reg); | |
3905 | lra_update_insn_regno_info (insn); | |
3906 | if (! def_p) | |
3907 | /* We now have a new usage insn for original regno. */ | |
3908 | setup_next_usage_insn (original_regno, new_insns, reloads_num, false); | |
3909 | if (lra_dump_file != NULL) | |
bc3591eb | 3910 | fprintf (lra_dump_file, " Original reg change %d->%d (bb%d):\n", |
55a2c322 VM |
3911 | original_regno, REGNO (new_reg), BLOCK_FOR_INSN (insn)->index); |
3912 | lra_reg_info[REGNO (new_reg)].restore_regno = original_regno; | |
3913 | bitmap_set_bit (&check_only_regs, REGNO (new_reg)); | |
3914 | bitmap_set_bit (&check_only_regs, original_regno); | |
3915 | bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg)); | |
3916 | if (def_p) | |
3917 | lra_process_new_insns (insn, NULL_RTX, new_insns, | |
3918 | "Add original<-inheritance"); | |
3919 | else | |
3920 | lra_process_new_insns (insn, new_insns, NULL_RTX, | |
3921 | "Add inheritance<-original"); | |
3922 | while (next_usage_insns != NULL_RTX) | |
3923 | { | |
3924 | if (GET_CODE (next_usage_insns) != INSN_LIST) | |
3925 | { | |
3926 | usage_insn = next_usage_insns; | |
3927 | lra_assert (NONDEBUG_INSN_P (usage_insn)); | |
3928 | next_usage_insns = NULL; | |
3929 | } | |
3930 | else | |
3931 | { | |
3932 | usage_insn = XEXP (next_usage_insns, 0); | |
3933 | lra_assert (DEBUG_INSN_P (usage_insn)); | |
3934 | next_usage_insns = XEXP (next_usage_insns, 1); | |
3935 | } | |
3936 | substitute_pseudo (&usage_insn, original_regno, new_reg); | |
3937 | lra_update_insn_regno_info (usage_insn); | |
3938 | if (lra_dump_file != NULL) | |
3939 | { | |
3940 | fprintf (lra_dump_file, | |
3941 | " Inheritance reuse change %d->%d (bb%d):\n", | |
3942 | original_regno, REGNO (new_reg), | |
3943 | BLOCK_FOR_INSN (usage_insn)->index); | |
cfbeaedf | 3944 | dump_insn_slim (lra_dump_file, usage_insn); |
55a2c322 VM |
3945 | } |
3946 | } | |
3947 | if (lra_dump_file != NULL) | |
3948 | fprintf (lra_dump_file, | |
3949 | " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n"); | |
3950 | return true; | |
3951 | } | |
3952 | ||
3953 | /* Return true if we need a caller save/restore for pseudo REGNO which | |
3954 | was assigned to a hard register. */ | |
3955 | static inline bool | |
3956 | need_for_call_save_p (int regno) | |
3957 | { | |
3958 | lra_assert (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] >= 0); | |
3959 | return (usage_insns[regno].calls_num < calls_num | |
3960 | && (overlaps_hard_reg_set_p | |
3961 | (call_used_reg_set, | |
3962 | PSEUDO_REGNO_MODE (regno), reg_renumber[regno]))); | |
3963 | } | |
3964 | ||
3965 | /* Global registers occuring in the current EBB. */ | |
3966 | static bitmap_head ebb_global_regs; | |
3967 | ||
3968 | /* Return true if we need a split for hard register REGNO or pseudo | |
3969 | REGNO which was assigned to a hard register. | |
3970 | POTENTIAL_RELOAD_HARD_REGS contains hard registers which might be | |
3971 | used for reloads since the EBB end. It is an approximation of the | |
3972 | used hard registers in the split range. The exact value would | |
3973 | require expensive calculations. If we were aggressive with | |
3974 | splitting because of the approximation, the split pseudo will save | |
3975 | the same hard register assignment and will be removed in the undo | |
3976 | pass. We still need the approximation because too aggressive | |
3977 | splitting would result in too inaccurate cost calculation in the | |
3978 | assignment pass because of too many generated moves which will be | |
3979 | probably removed in the undo pass. */ | |
3980 | static inline bool | |
3981 | need_for_split_p (HARD_REG_SET potential_reload_hard_regs, int regno) | |
3982 | { | |
3983 | int hard_regno = regno < FIRST_PSEUDO_REGISTER ? regno : reg_renumber[regno]; | |
3984 | ||
3985 | lra_assert (hard_regno >= 0); | |
3986 | return ((TEST_HARD_REG_BIT (potential_reload_hard_regs, hard_regno) | |
3987 | /* Don't split eliminable hard registers, otherwise we can | |
3988 | split hard registers like hard frame pointer, which | |
3989 | lives on BB start/end according to DF-infrastructure, | |
3990 | when there is a pseudo assigned to the register and | |
3991 | living in the same BB. */ | |
3992 | && (regno >= FIRST_PSEUDO_REGISTER | |
3993 | || ! TEST_HARD_REG_BIT (eliminable_regset, hard_regno)) | |
3994 | && ! TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno) | |
3995 | /* We need at least 2 reloads to make pseudo splitting | |
3996 | profitable. We should provide hard regno splitting in | |
3997 | any case to solve 1st insn scheduling problem when | |
3998 | moving hard register definition up might result in | |
3999 | impossibility to find hard register for reload pseudo of | |
4000 | small register class. */ | |
4001 | && (usage_insns[regno].reloads_num | |
4002 | + (regno < FIRST_PSEUDO_REGISTER ? 0 : 2) < reloads_num) | |
4003 | && (regno < FIRST_PSEUDO_REGISTER | |
4004 | /* For short living pseudos, spilling + inheritance can | |
4005 | be considered a substitution for splitting. | |
4006 | Therefore we do not splitting for local pseudos. It | |
4007 | decreases also aggressiveness of splitting. The | |
4008 | minimal number of references is chosen taking into | |
4009 | account that for 2 references splitting has no sense | |
4010 | as we can just spill the pseudo. */ | |
4011 | || (regno >= FIRST_PSEUDO_REGISTER | |
4012 | && lra_reg_info[regno].nrefs > 3 | |
4013 | && bitmap_bit_p (&ebb_global_regs, regno)))) | |
4014 | || (regno >= FIRST_PSEUDO_REGISTER && need_for_call_save_p (regno))); | |
4015 | } | |
4016 | ||
4017 | /* Return class for the split pseudo created from original pseudo with | |
4018 | ALLOCNO_CLASS and MODE which got a hard register HARD_REGNO. We | |
4019 | choose subclass of ALLOCNO_CLASS which contains HARD_REGNO and | |
4020 | results in no secondary memory movements. */ | |
4021 | static enum reg_class | |
4022 | choose_split_class (enum reg_class allocno_class, | |
4023 | int hard_regno ATTRIBUTE_UNUSED, | |
4024 | enum machine_mode mode ATTRIBUTE_UNUSED) | |
4025 | { | |
4026 | #ifndef SECONDARY_MEMORY_NEEDED | |
4027 | return allocno_class; | |
4028 | #else | |
4029 | int i; | |
4030 | enum reg_class cl, best_cl = NO_REGS; | |
ef4dbe49 JR |
4031 | enum reg_class hard_reg_class ATTRIBUTE_UNUSED |
4032 | = REGNO_REG_CLASS (hard_regno); | |
f4eafc30 | 4033 | |
55a2c322 VM |
4034 | if (! SECONDARY_MEMORY_NEEDED (allocno_class, allocno_class, mode) |
4035 | && TEST_HARD_REG_BIT (reg_class_contents[allocno_class], hard_regno)) | |
4036 | return allocno_class; | |
4037 | for (i = 0; | |
4038 | (cl = reg_class_subclasses[allocno_class][i]) != LIM_REG_CLASSES; | |
4039 | i++) | |
4040 | if (! SECONDARY_MEMORY_NEEDED (cl, hard_reg_class, mode) | |
4041 | && ! SECONDARY_MEMORY_NEEDED (hard_reg_class, cl, mode) | |
4042 | && TEST_HARD_REG_BIT (reg_class_contents[cl], hard_regno) | |
4043 | && (best_cl == NO_REGS | |
4044 | || ira_class_hard_regs_num[best_cl] < ira_class_hard_regs_num[cl])) | |
4045 | best_cl = cl; | |
4046 | return best_cl; | |
4047 | #endif | |
4048 | } | |
4049 | ||
4050 | /* Do split transformations for insn INSN, which defines or uses | |
4051 | ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which instruction in | |
4052 | the EBB next uses ORIGINAL_REGNO; it has the same form as the | |
4053 | "insns" field of usage_insns. | |
4054 | ||
4055 | The transformations look like: | |
4056 | ||
4057 | p <- ... p <- ... | |
4058 | ... s <- p (new insn -- save) | |
4059 | ... => | |
4060 | ... p <- s (new insn -- restore) | |
4061 | <- ... p ... <- ... p ... | |
4062 | or | |
4063 | <- ... p ... <- ... p ... | |
4064 | ... s <- p (new insn -- save) | |
4065 | ... => | |
4066 | ... p <- s (new insn -- restore) | |
4067 | <- ... p ... <- ... p ... | |
4068 | ||
4069 | where p is an original pseudo got a hard register or a hard | |
4070 | register and s is a new split pseudo. The save is put before INSN | |
4071 | if BEFORE_P is true. Return true if we succeed in such | |
4072 | transformation. */ | |
4073 | static bool | |
4074 | split_reg (bool before_p, int original_regno, rtx insn, rtx next_usage_insns) | |
4075 | { | |
4076 | enum reg_class rclass; | |
4077 | rtx original_reg; | |
4078 | int hard_regno; | |
4079 | rtx new_reg, save, restore, usage_insn; | |
4080 | bool after_p; | |
4081 | bool call_save_p; | |
4082 | ||
4083 | if (original_regno < FIRST_PSEUDO_REGISTER) | |
4084 | { | |
4085 | rclass = ira_allocno_class_translate[REGNO_REG_CLASS (original_regno)]; | |
4086 | hard_regno = original_regno; | |
4087 | call_save_p = false; | |
4088 | } | |
4089 | else | |
4090 | { | |
4091 | hard_regno = reg_renumber[original_regno]; | |
4092 | rclass = lra_get_allocno_class (original_regno); | |
4093 | original_reg = regno_reg_rtx[original_regno]; | |
4094 | call_save_p = need_for_call_save_p (original_regno); | |
4095 | } | |
4096 | original_reg = regno_reg_rtx[original_regno]; | |
4097 | lra_assert (hard_regno >= 0); | |
4098 | if (lra_dump_file != NULL) | |
4099 | fprintf (lra_dump_file, | |
4100 | " ((((((((((((((((((((((((((((((((((((((((((((((((\n"); | |
4101 | if (call_save_p) | |
4102 | { | |
4103 | enum machine_mode sec_mode; | |
f4eafc30 | 4104 | |
55a2c322 VM |
4105 | #ifdef SECONDARY_MEMORY_NEEDED_MODE |
4106 | sec_mode = SECONDARY_MEMORY_NEEDED_MODE (GET_MODE (original_reg)); | |
4107 | #else | |
4108 | sec_mode = GET_MODE (original_reg); | |
4109 | #endif | |
4110 | new_reg = lra_create_new_reg (sec_mode, NULL_RTX, | |
4111 | NO_REGS, "save"); | |
4112 | } | |
4113 | else | |
4114 | { | |
4115 | rclass = choose_split_class (rclass, hard_regno, | |
4116 | GET_MODE (original_reg)); | |
4117 | if (rclass == NO_REGS) | |
4118 | { | |
4119 | if (lra_dump_file != NULL) | |
4120 | { | |
4121 | fprintf (lra_dump_file, | |
4122 | " Rejecting split of %d(%s): " | |
4123 | "no good reg class for %d(%s)\n", | |
4124 | original_regno, | |
4125 | reg_class_names[lra_get_allocno_class (original_regno)], | |
4126 | hard_regno, | |
4127 | reg_class_names[REGNO_REG_CLASS (hard_regno)]); | |
4128 | fprintf | |
4129 | (lra_dump_file, | |
4130 | " ))))))))))))))))))))))))))))))))))))))))))))))))\n"); | |
4131 | } | |
4132 | return false; | |
4133 | } | |
4134 | new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg, | |
4135 | rclass, "split"); | |
4136 | reg_renumber[REGNO (new_reg)] = hard_regno; | |
4137 | } | |
4138 | save = emit_spill_move (true, new_reg, original_reg); | |
4139 | if (NEXT_INSN (save) != NULL_RTX) | |
4140 | { | |
4141 | lra_assert (! call_save_p); | |
4142 | if (lra_dump_file != NULL) | |
4143 | { | |
4144 | fprintf | |
4145 | (lra_dump_file, | |
4146 | " Rejecting split %d->%d resulting in > 2 %s save insns:\n", | |
4147 | original_regno, REGNO (new_reg), call_save_p ? "call" : ""); | |
cfbeaedf | 4148 | dump_rtl_slim (lra_dump_file, save, NULL_RTX, -1, 0); |
55a2c322 VM |
4149 | fprintf (lra_dump_file, |
4150 | " ))))))))))))))))))))))))))))))))))))))))))))))))\n"); | |
4151 | } | |
4152 | return false; | |
4153 | } | |
4154 | restore = emit_spill_move (false, new_reg, original_reg); | |
4155 | if (NEXT_INSN (restore) != NULL_RTX) | |
4156 | { | |
4157 | lra_assert (! call_save_p); | |
4158 | if (lra_dump_file != NULL) | |
4159 | { | |
4160 | fprintf (lra_dump_file, | |
4161 | " Rejecting split %d->%d " | |
4162 | "resulting in > 2 %s restore insns:\n", | |
4163 | original_regno, REGNO (new_reg), call_save_p ? "call" : ""); | |
cfbeaedf | 4164 | dump_rtl_slim (lra_dump_file, restore, NULL_RTX, -1, 0); |
55a2c322 VM |
4165 | fprintf (lra_dump_file, |
4166 | " ))))))))))))))))))))))))))))))))))))))))))))))))\n"); | |
4167 | } | |
4168 | return false; | |
4169 | } | |
4170 | after_p = usage_insns[original_regno].after_p; | |
4171 | lra_reg_info[REGNO (new_reg)].restore_regno = original_regno; | |
4172 | bitmap_set_bit (&check_only_regs, REGNO (new_reg)); | |
4173 | bitmap_set_bit (&check_only_regs, original_regno); | |
4174 | bitmap_set_bit (&lra_split_regs, REGNO (new_reg)); | |
4175 | for (;;) | |
4176 | { | |
4177 | if (GET_CODE (next_usage_insns) != INSN_LIST) | |
4178 | { | |
4179 | usage_insn = next_usage_insns; | |
4180 | break; | |
4181 | } | |
4182 | usage_insn = XEXP (next_usage_insns, 0); | |
4183 | lra_assert (DEBUG_INSN_P (usage_insn)); | |
4184 | next_usage_insns = XEXP (next_usage_insns, 1); | |
4185 | substitute_pseudo (&usage_insn, original_regno, new_reg); | |
4186 | lra_update_insn_regno_info (usage_insn); | |
4187 | if (lra_dump_file != NULL) | |
4188 | { | |
4189 | fprintf (lra_dump_file, " Split reuse change %d->%d:\n", | |
4190 | original_regno, REGNO (new_reg)); | |
cfbeaedf | 4191 | dump_insn_slim (lra_dump_file, usage_insn); |
55a2c322 VM |
4192 | } |
4193 | } | |
4194 | lra_assert (NOTE_P (usage_insn) || NONDEBUG_INSN_P (usage_insn)); | |
4195 | lra_assert (usage_insn != insn || (after_p && before_p)); | |
4196 | lra_process_new_insns (usage_insn, after_p ? NULL_RTX : restore, | |
4197 | after_p ? restore : NULL_RTX, | |
4198 | call_save_p | |
4199 | ? "Add reg<-save" : "Add reg<-split"); | |
4200 | lra_process_new_insns (insn, before_p ? save : NULL_RTX, | |
4201 | before_p ? NULL_RTX : save, | |
4202 | call_save_p | |
4203 | ? "Add save<-reg" : "Add split<-reg"); | |
4204 | if (lra_dump_file != NULL) | |
4205 | fprintf (lra_dump_file, | |
4206 | " ))))))))))))))))))))))))))))))))))))))))))))))))\n"); | |
4207 | return true; | |
4208 | } | |
4209 | ||
4210 | /* Recognize that we need a split transformation for insn INSN, which | |
4211 | defines or uses REGNO in its insn biggest MODE (we use it only if | |
4212 | REGNO is a hard register). POTENTIAL_RELOAD_HARD_REGS contains | |
4213 | hard registers which might be used for reloads since the EBB end. | |
4214 | Put the save before INSN if BEFORE_P is true. MAX_UID is maximla | |
4215 | uid before starting INSN processing. Return true if we succeed in | |
4216 | such transformation. */ | |
4217 | static bool | |
4218 | split_if_necessary (int regno, enum machine_mode mode, | |
4219 | HARD_REG_SET potential_reload_hard_regs, | |
4220 | bool before_p, rtx insn, int max_uid) | |
4221 | { | |
4222 | bool res = false; | |
4223 | int i, nregs = 1; | |
4224 | rtx next_usage_insns; | |
4225 | ||
4226 | if (regno < FIRST_PSEUDO_REGISTER) | |
4227 | nregs = hard_regno_nregs[regno][mode]; | |
4228 | for (i = 0; i < nregs; i++) | |
4229 | if (usage_insns[regno + i].check == curr_usage_insns_check | |
4230 | && (next_usage_insns = usage_insns[regno + i].insns) != NULL_RTX | |
4231 | /* To avoid processing the register twice or more. */ | |
4232 | && ((GET_CODE (next_usage_insns) != INSN_LIST | |
4233 | && INSN_UID (next_usage_insns) < max_uid) | |
4234 | || (GET_CODE (next_usage_insns) == INSN_LIST | |
4235 | && (INSN_UID (XEXP (next_usage_insns, 0)) < max_uid))) | |
4236 | && need_for_split_p (potential_reload_hard_regs, regno + i) | |
4237 | && split_reg (before_p, regno + i, insn, next_usage_insns)) | |
4238 | res = true; | |
4239 | return res; | |
4240 | } | |
4241 | ||
4242 | /* Check only registers living at the current program point in the | |
4243 | current EBB. */ | |
4244 | static bitmap_head live_regs; | |
4245 | ||
4246 | /* Update live info in EBB given by its HEAD and TAIL insns after | |
4247 | inheritance/split transformation. The function removes dead moves | |
4248 | too. */ | |
4249 | static void | |
4250 | update_ebb_live_info (rtx head, rtx tail) | |
4251 | { | |
4252 | unsigned int j; | |
4253 | int regno; | |
4254 | bool live_p; | |
4255 | rtx prev_insn, set; | |
4256 | bool remove_p; | |
4257 | basic_block last_bb, prev_bb, curr_bb; | |
4258 | bitmap_iterator bi; | |
4259 | struct lra_insn_reg *reg; | |
4260 | edge e; | |
4261 | edge_iterator ei; | |
4262 | ||
f4eafc30 | 4263 | last_bb = BLOCK_FOR_INSN (tail); |
55a2c322 VM |
4264 | prev_bb = NULL; |
4265 | for (curr_insn = tail; | |
4266 | curr_insn != PREV_INSN (head); | |
4267 | curr_insn = prev_insn) | |
4268 | { | |
4269 | prev_insn = PREV_INSN (curr_insn); | |
911598e3 VM |
4270 | /* We need to process empty blocks too. They contain |
4271 | NOTE_INSN_BASIC_BLOCK referring for the basic block. */ | |
4272 | if (NOTE_P (curr_insn) && NOTE_KIND (curr_insn) != NOTE_INSN_BASIC_BLOCK) | |
4273 | continue; | |
55a2c322 VM |
4274 | curr_bb = BLOCK_FOR_INSN (curr_insn); |
4275 | if (curr_bb != prev_bb) | |
4276 | { | |
4277 | if (prev_bb != NULL) | |
4278 | { | |
4279 | /* Update df_get_live_in (prev_bb): */ | |
4280 | EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi) | |
4281 | if (bitmap_bit_p (&live_regs, j)) | |
4282 | bitmap_set_bit (df_get_live_in (prev_bb), j); | |
4283 | else | |
4284 | bitmap_clear_bit (df_get_live_in (prev_bb), j); | |
4285 | } | |
4286 | if (curr_bb != last_bb) | |
4287 | { | |
4288 | /* Update df_get_live_out (curr_bb): */ | |
4289 | EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi) | |
4290 | { | |
4291 | live_p = bitmap_bit_p (&live_regs, j); | |
4292 | if (! live_p) | |
4293 | FOR_EACH_EDGE (e, ei, curr_bb->succs) | |
4294 | if (bitmap_bit_p (df_get_live_in (e->dest), j)) | |
4295 | { | |
4296 | live_p = true; | |
4297 | break; | |
4298 | } | |
4299 | if (live_p) | |
4300 | bitmap_set_bit (df_get_live_out (curr_bb), j); | |
4301 | else | |
4302 | bitmap_clear_bit (df_get_live_out (curr_bb), j); | |
4303 | } | |
4304 | } | |
4305 | prev_bb = curr_bb; | |
4306 | bitmap_and (&live_regs, &check_only_regs, df_get_live_out (curr_bb)); | |
4307 | } | |
44b94bdb | 4308 | if (! NONDEBUG_INSN_P (curr_insn)) |
55a2c322 VM |
4309 | continue; |
4310 | curr_id = lra_get_insn_recog_data (curr_insn); | |
4311 | remove_p = false; | |
4312 | if ((set = single_set (curr_insn)) != NULL_RTX && REG_P (SET_DEST (set)) | |
4313 | && (regno = REGNO (SET_DEST (set))) >= FIRST_PSEUDO_REGISTER | |
4314 | && bitmap_bit_p (&check_only_regs, regno) | |
4315 | && ! bitmap_bit_p (&live_regs, regno)) | |
4316 | remove_p = true; | |
4317 | /* See which defined values die here. */ | |
4318 | for (reg = curr_id->regs; reg != NULL; reg = reg->next) | |
4319 | if (reg->type == OP_OUT && ! reg->subreg_p) | |
4320 | bitmap_clear_bit (&live_regs, reg->regno); | |
4321 | /* Mark each used value as live. */ | |
4322 | for (reg = curr_id->regs; reg != NULL; reg = reg->next) | |
4323 | if (reg->type == OP_IN | |
4324 | && bitmap_bit_p (&check_only_regs, reg->regno)) | |
4325 | bitmap_set_bit (&live_regs, reg->regno); | |
4326 | /* It is quite important to remove dead move insns because it | |
4327 | means removing dead store. We don't need to process them for | |
4328 | constraints. */ | |
4329 | if (remove_p) | |
4330 | { | |
4331 | if (lra_dump_file != NULL) | |
4332 | { | |
4333 | fprintf (lra_dump_file, " Removing dead insn:\n "); | |
cfbeaedf | 4334 | dump_insn_slim (lra_dump_file, curr_insn); |
55a2c322 VM |
4335 | } |
4336 | lra_set_insn_deleted (curr_insn); | |
4337 | } | |
4338 | } | |
4339 | } | |
4340 | ||
4341 | /* The structure describes info to do an inheritance for the current | |
4342 | insn. We need to collect such info first before doing the | |
4343 | transformations because the transformations change the insn | |
4344 | internal representation. */ | |
4345 | struct to_inherit | |
4346 | { | |
4347 | /* Original regno. */ | |
4348 | int regno; | |
4349 | /* Subsequent insns which can inherit original reg value. */ | |
4350 | rtx insns; | |
4351 | }; | |
4352 | ||
4353 | /* Array containing all info for doing inheritance from the current | |
4354 | insn. */ | |
4355 | static struct to_inherit to_inherit[LRA_MAX_INSN_RELOADS]; | |
4356 | ||
4357 | /* Number elements in the previous array. */ | |
4358 | static int to_inherit_num; | |
4359 | ||
4360 | /* Add inheritance info REGNO and INSNS. Their meaning is described in | |
4361 | structure to_inherit. */ | |
4362 | static void | |
4363 | add_to_inherit (int regno, rtx insns) | |
4364 | { | |
4365 | int i; | |
4366 | ||
4367 | for (i = 0; i < to_inherit_num; i++) | |
4368 | if (to_inherit[i].regno == regno) | |
4369 | return; | |
4370 | lra_assert (to_inherit_num < LRA_MAX_INSN_RELOADS); | |
4371 | to_inherit[to_inherit_num].regno = regno; | |
4372 | to_inherit[to_inherit_num++].insns = insns; | |
4373 | } | |
4374 | ||
4375 | /* Return the last non-debug insn in basic block BB, or the block begin | |
4376 | note if none. */ | |
4377 | static rtx | |
4378 | get_last_insertion_point (basic_block bb) | |
4379 | { | |
4380 | rtx insn; | |
4381 | ||
4382 | FOR_BB_INSNS_REVERSE (bb, insn) | |
4383 | if (NONDEBUG_INSN_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn)) | |
4384 | return insn; | |
4385 | gcc_unreachable (); | |
4386 | } | |
4387 | ||
4388 | /* Set up RES by registers living on edges FROM except the edge (FROM, | |
4389 | TO) or by registers set up in a jump insn in BB FROM. */ | |
4390 | static void | |
4391 | get_live_on_other_edges (basic_block from, basic_block to, bitmap res) | |
4392 | { | |
4393 | rtx last; | |
4394 | struct lra_insn_reg *reg; | |
4395 | edge e; | |
4396 | edge_iterator ei; | |
4397 | ||
4398 | lra_assert (to != NULL); | |
4399 | bitmap_clear (res); | |
4400 | FOR_EACH_EDGE (e, ei, from->succs) | |
4401 | if (e->dest != to) | |
4402 | bitmap_ior_into (res, df_get_live_in (e->dest)); | |
4403 | last = get_last_insertion_point (from); | |
4404 | if (! JUMP_P (last)) | |
4405 | return; | |
4406 | curr_id = lra_get_insn_recog_data (last); | |
4407 | for (reg = curr_id->regs; reg != NULL; reg = reg->next) | |
4408 | if (reg->type != OP_IN) | |
4409 | bitmap_set_bit (res, reg->regno); | |
4410 | } | |
f4eafc30 | 4411 | |
55a2c322 VM |
4412 | /* Used as a temporary results of some bitmap calculations. */ |
4413 | static bitmap_head temp_bitmap; | |
4414 | ||
4415 | /* Do inheritance/split transformations in EBB starting with HEAD and | |
4416 | finishing on TAIL. We process EBB insns in the reverse order. | |
4417 | Return true if we did any inheritance/split transformation in the | |
4418 | EBB. | |
4419 | ||
4420 | We should avoid excessive splitting which results in worse code | |
4421 | because of inaccurate cost calculations for spilling new split | |
4422 | pseudos in such case. To achieve this we do splitting only if | |
4423 | register pressure is high in given basic block and there are reload | |
4424 | pseudos requiring hard registers. We could do more register | |
4425 | pressure calculations at any given program point to avoid necessary | |
4426 | splitting even more but it is to expensive and the current approach | |
4427 | works well enough. */ | |
4428 | static bool | |
4429 | inherit_in_ebb (rtx head, rtx tail) | |
4430 | { | |
4431 | int i, src_regno, dst_regno, nregs; | |
4432 | bool change_p, succ_p; | |
4433 | rtx prev_insn, next_usage_insns, set, last_insn; | |
4434 | enum reg_class cl; | |
4435 | struct lra_insn_reg *reg; | |
4436 | basic_block last_processed_bb, curr_bb = NULL; | |
4437 | HARD_REG_SET potential_reload_hard_regs, live_hard_regs; | |
4438 | bitmap to_process; | |
4439 | unsigned int j; | |
4440 | bitmap_iterator bi; | |
4441 | bool head_p, after_p; | |
4442 | ||
4443 | change_p = false; | |
4444 | curr_usage_insns_check++; | |
4445 | reloads_num = calls_num = 0; | |
4446 | bitmap_clear (&check_only_regs); | |
4447 | last_processed_bb = NULL; | |
4448 | CLEAR_HARD_REG_SET (potential_reload_hard_regs); | |
4449 | CLEAR_HARD_REG_SET (live_hard_regs); | |
4450 | /* We don't process new insns generated in the loop. */ | |
4451 | for (curr_insn = tail; curr_insn != PREV_INSN (head); curr_insn = prev_insn) | |
4452 | { | |
4453 | prev_insn = PREV_INSN (curr_insn); | |
4454 | if (BLOCK_FOR_INSN (curr_insn) != NULL) | |
4455 | curr_bb = BLOCK_FOR_INSN (curr_insn); | |
4456 | if (last_processed_bb != curr_bb) | |
4457 | { | |
4458 | /* We are at the end of BB. Add qualified living | |
4459 | pseudos for potential splitting. */ | |
4460 | to_process = df_get_live_out (curr_bb); | |
4461 | if (last_processed_bb != NULL) | |
f4eafc30 | 4462 | { |
55a2c322 VM |
4463 | /* We are somewhere in the middle of EBB. */ |
4464 | get_live_on_other_edges (curr_bb, last_processed_bb, | |
4465 | &temp_bitmap); | |
4466 | to_process = &temp_bitmap; | |
4467 | } | |
4468 | last_processed_bb = curr_bb; | |
4469 | last_insn = get_last_insertion_point (curr_bb); | |
4470 | after_p = (! JUMP_P (last_insn) | |
4471 | && (! CALL_P (last_insn) | |
4472 | || (find_reg_note (last_insn, | |
4473 | REG_NORETURN, NULL_RTX) == NULL_RTX | |
4474 | && ! SIBLING_CALL_P (last_insn)))); | |
4475 | REG_SET_TO_HARD_REG_SET (live_hard_regs, df_get_live_out (curr_bb)); | |
4476 | IOR_HARD_REG_SET (live_hard_regs, eliminable_regset); | |
4477 | IOR_HARD_REG_SET (live_hard_regs, lra_no_alloc_regs); | |
4478 | CLEAR_HARD_REG_SET (potential_reload_hard_regs); | |
4479 | EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi) | |
4480 | { | |
4481 | if ((int) j >= lra_constraint_new_regno_start) | |
4482 | break; | |
4483 | if (j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0) | |
4484 | { | |
4485 | if (j < FIRST_PSEUDO_REGISTER) | |
4486 | SET_HARD_REG_BIT (live_hard_regs, j); | |
4487 | else | |
4488 | add_to_hard_reg_set (&live_hard_regs, | |
4489 | PSEUDO_REGNO_MODE (j), | |
4490 | reg_renumber[j]); | |
4491 | setup_next_usage_insn (j, last_insn, reloads_num, after_p); | |
4492 | } | |
4493 | } | |
4494 | } | |
4495 | src_regno = dst_regno = -1; | |
4496 | if (NONDEBUG_INSN_P (curr_insn) | |
4497 | && (set = single_set (curr_insn)) != NULL_RTX | |
4498 | && REG_P (SET_DEST (set)) && REG_P (SET_SRC (set))) | |
4499 | { | |
4500 | src_regno = REGNO (SET_SRC (set)); | |
4501 | dst_regno = REGNO (SET_DEST (set)); | |
4502 | } | |
4503 | if (src_regno < lra_constraint_new_regno_start | |
4504 | && src_regno >= FIRST_PSEUDO_REGISTER | |
4505 | && reg_renumber[src_regno] < 0 | |
4506 | && dst_regno >= lra_constraint_new_regno_start | |
4507 | && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS) | |
4508 | { | |
4509 | /* 'reload_pseudo <- original_pseudo'. */ | |
4510 | reloads_num++; | |
4511 | succ_p = false; | |
4512 | if (usage_insns[src_regno].check == curr_usage_insns_check | |
4513 | && (next_usage_insns = usage_insns[src_regno].insns) != NULL_RTX) | |
4514 | succ_p = inherit_reload_reg (false, src_regno, cl, | |
4515 | curr_insn, next_usage_insns); | |
4516 | if (succ_p) | |
4517 | change_p = true; | |
4518 | else | |
4519 | setup_next_usage_insn (src_regno, curr_insn, reloads_num, false); | |
4520 | if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs)) | |
4521 | IOR_HARD_REG_SET (potential_reload_hard_regs, | |
4522 | reg_class_contents[cl]); | |
4523 | } | |
4524 | else if (src_regno >= lra_constraint_new_regno_start | |
4525 | && dst_regno < lra_constraint_new_regno_start | |
4526 | && dst_regno >= FIRST_PSEUDO_REGISTER | |
4527 | && reg_renumber[dst_regno] < 0 | |
4528 | && (cl = lra_get_allocno_class (src_regno)) != NO_REGS | |
4529 | && usage_insns[dst_regno].check == curr_usage_insns_check | |
4530 | && (next_usage_insns | |
4531 | = usage_insns[dst_regno].insns) != NULL_RTX) | |
4532 | { | |
4533 | reloads_num++; | |
4534 | /* 'original_pseudo <- reload_pseudo'. */ | |
4535 | if (! JUMP_P (curr_insn) | |
4536 | && inherit_reload_reg (true, dst_regno, cl, | |
4537 | curr_insn, next_usage_insns)) | |
4538 | change_p = true; | |
4539 | /* Invalidate. */ | |
4540 | usage_insns[dst_regno].check = 0; | |
4541 | if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs)) | |
4542 | IOR_HARD_REG_SET (potential_reload_hard_regs, | |
4543 | reg_class_contents[cl]); | |
4544 | } | |
4545 | else if (INSN_P (curr_insn)) | |
4546 | { | |
4547 | int max_uid = get_max_uid (); | |
4548 | ||
4549 | curr_id = lra_get_insn_recog_data (curr_insn); | |
4550 | to_inherit_num = 0; | |
4551 | /* Process insn definitions. */ | |
4552 | for (reg = curr_id->regs; reg != NULL; reg = reg->next) | |
4553 | if (reg->type != OP_IN | |
4554 | && (dst_regno = reg->regno) < lra_constraint_new_regno_start) | |
4555 | { | |
4556 | if (dst_regno >= FIRST_PSEUDO_REGISTER && reg->type == OP_OUT | |
4557 | && reg_renumber[dst_regno] < 0 && ! reg->subreg_p | |
4558 | && usage_insns[dst_regno].check == curr_usage_insns_check | |
4559 | && (next_usage_insns | |
4560 | = usage_insns[dst_regno].insns) != NULL_RTX) | |
4561 | { | |
4562 | struct lra_insn_reg *r; | |
f4eafc30 | 4563 | |
55a2c322 VM |
4564 | for (r = curr_id->regs; r != NULL; r = r->next) |
4565 | if (r->type != OP_OUT && r->regno == dst_regno) | |
4566 | break; | |
4567 | /* Don't do inheritance if the pseudo is also | |
4568 | used in the insn. */ | |
4569 | if (r == NULL) | |
4570 | /* We can not do inheritance right now | |
4571 | because the current insn reg info (chain | |
4572 | regs) can change after that. */ | |
4573 | add_to_inherit (dst_regno, next_usage_insns); | |
4574 | } | |
4575 | /* We can not process one reg twice here because of | |
4576 | usage_insns invalidation. */ | |
4577 | if ((dst_regno < FIRST_PSEUDO_REGISTER | |
4578 | || reg_renumber[dst_regno] >= 0) | |
4579 | && ! reg->subreg_p && reg->type == OP_OUT) | |
4580 | { | |
4581 | HARD_REG_SET s; | |
f4eafc30 | 4582 | |
55a2c322 VM |
4583 | if (split_if_necessary (dst_regno, reg->biggest_mode, |
4584 | potential_reload_hard_regs, | |
4585 | false, curr_insn, max_uid)) | |
4586 | change_p = true; | |
4587 | CLEAR_HARD_REG_SET (s); | |
4588 | if (dst_regno < FIRST_PSEUDO_REGISTER) | |
4589 | add_to_hard_reg_set (&s, reg->biggest_mode, dst_regno); | |
4590 | else | |
4591 | add_to_hard_reg_set (&s, PSEUDO_REGNO_MODE (dst_regno), | |
4592 | reg_renumber[dst_regno]); | |
4593 | AND_COMPL_HARD_REG_SET (live_hard_regs, s); | |
4594 | } | |
4595 | /* We should invalidate potential inheritance or | |
4596 | splitting for the current insn usages to the next | |
4597 | usage insns (see code below) as the output pseudo | |
4598 | prevents this. */ | |
4599 | if ((dst_regno >= FIRST_PSEUDO_REGISTER | |
4600 | && reg_renumber[dst_regno] < 0) | |
4601 | || (reg->type == OP_OUT && ! reg->subreg_p | |
4602 | && (dst_regno < FIRST_PSEUDO_REGISTER | |
4603 | || reg_renumber[dst_regno] >= 0))) | |
4604 | { | |
4605 | /* Invalidate. */ | |
4606 | if (dst_regno >= FIRST_PSEUDO_REGISTER) | |
4607 | usage_insns[dst_regno].check = 0; | |
4608 | else | |
4609 | { | |
4610 | nregs = hard_regno_nregs[dst_regno][reg->biggest_mode]; | |
4611 | for (i = 0; i < nregs; i++) | |
4612 | usage_insns[dst_regno + i].check = 0; | |
4613 | } | |
4614 | } | |
4615 | } | |
4616 | if (! JUMP_P (curr_insn)) | |
4617 | for (i = 0; i < to_inherit_num; i++) | |
4618 | if (inherit_reload_reg (true, to_inherit[i].regno, | |
4619 | ALL_REGS, curr_insn, | |
4620 | to_inherit[i].insns)) | |
4621 | change_p = true; | |
4622 | if (CALL_P (curr_insn)) | |
4623 | { | |
4624 | rtx cheap, pat, dest, restore; | |
4625 | int regno, hard_regno; | |
4626 | ||
4627 | calls_num++; | |
4628 | if ((cheap = find_reg_note (curr_insn, | |
4629 | REG_RETURNED, NULL_RTX)) != NULL_RTX | |
4630 | && ((cheap = XEXP (cheap, 0)), true) | |
4631 | && (regno = REGNO (cheap)) >= FIRST_PSEUDO_REGISTER | |
4632 | && (hard_regno = reg_renumber[regno]) >= 0 | |
4633 | /* If there are pending saves/restores, the | |
4634 | optimization is not worth. */ | |
4635 | && usage_insns[regno].calls_num == calls_num - 1 | |
4636 | && TEST_HARD_REG_BIT (call_used_reg_set, hard_regno)) | |
4637 | { | |
4638 | /* Restore the pseudo from the call result as | |
4639 | REG_RETURNED note says that the pseudo value is | |
4640 | in the call result and the pseudo is an argument | |
4641 | of the call. */ | |
4642 | pat = PATTERN (curr_insn); | |
4643 | if (GET_CODE (pat) == PARALLEL) | |
4644 | pat = XVECEXP (pat, 0, 0); | |
4645 | dest = SET_DEST (pat); | |
4646 | start_sequence (); | |
4647 | emit_move_insn (cheap, copy_rtx (dest)); | |
4648 | restore = get_insns (); | |
4649 | end_sequence (); | |
4650 | lra_process_new_insns (curr_insn, NULL, restore, | |
4651 | "Inserting call parameter restore"); | |
4652 | /* We don't need to save/restore of the pseudo from | |
4653 | this call. */ | |
4654 | usage_insns[regno].calls_num = calls_num; | |
4655 | bitmap_set_bit (&check_only_regs, regno); | |
4656 | } | |
4657 | } | |
4658 | to_inherit_num = 0; | |
4659 | /* Process insn usages. */ | |
4660 | for (reg = curr_id->regs; reg != NULL; reg = reg->next) | |
4661 | if ((reg->type != OP_OUT | |
4662 | || (reg->type == OP_OUT && reg->subreg_p)) | |
4663 | && (src_regno = reg->regno) < lra_constraint_new_regno_start) | |
4664 | { | |
4665 | if (src_regno >= FIRST_PSEUDO_REGISTER | |
4666 | && reg_renumber[src_regno] < 0 && reg->type == OP_IN) | |
4667 | { | |
4668 | if (usage_insns[src_regno].check == curr_usage_insns_check | |
4669 | && (next_usage_insns | |
4670 | = usage_insns[src_regno].insns) != NULL_RTX | |
4671 | && NONDEBUG_INSN_P (curr_insn)) | |
4672 | add_to_inherit (src_regno, next_usage_insns); | |
4673 | else | |
4674 | /* Add usages. */ | |
4675 | add_next_usage_insn (src_regno, curr_insn, reloads_num); | |
4676 | } | |
4677 | else if (src_regno < FIRST_PSEUDO_REGISTER | |
4678 | || reg_renumber[src_regno] >= 0) | |
4679 | { | |
4680 | bool before_p; | |
4681 | rtx use_insn = curr_insn; | |
4682 | ||
4683 | before_p = (JUMP_P (curr_insn) | |
4684 | || (CALL_P (curr_insn) && reg->type == OP_IN)); | |
4685 | if (NONDEBUG_INSN_P (curr_insn) | |
4686 | && split_if_necessary (src_regno, reg->biggest_mode, | |
4687 | potential_reload_hard_regs, | |
4688 | before_p, curr_insn, max_uid)) | |
4689 | { | |
4690 | if (reg->subreg_p) | |
4691 | lra_risky_transformations_p = true; | |
4692 | change_p = true; | |
4693 | /* Invalidate. */ | |
4694 | usage_insns[src_regno].check = 0; | |
4695 | if (before_p) | |
4696 | use_insn = PREV_INSN (curr_insn); | |
4697 | } | |
4698 | if (NONDEBUG_INSN_P (curr_insn)) | |
4699 | { | |
4700 | if (src_regno < FIRST_PSEUDO_REGISTER) | |
4701 | add_to_hard_reg_set (&live_hard_regs, | |
4702 | reg->biggest_mode, src_regno); | |
4703 | else | |
4704 | add_to_hard_reg_set (&live_hard_regs, | |
4705 | PSEUDO_REGNO_MODE (src_regno), | |
4706 | reg_renumber[src_regno]); | |
4707 | } | |
4708 | add_next_usage_insn (src_regno, use_insn, reloads_num); | |
4709 | } | |
4710 | } | |
4711 | for (i = 0; i < to_inherit_num; i++) | |
4712 | { | |
4713 | src_regno = to_inherit[i].regno; | |
4714 | if (inherit_reload_reg (false, src_regno, ALL_REGS, | |
4715 | curr_insn, to_inherit[i].insns)) | |
4716 | change_p = true; | |
4717 | else | |
4718 | setup_next_usage_insn (src_regno, curr_insn, reloads_num, false); | |
4719 | } | |
4720 | } | |
4721 | /* We reached the start of the current basic block. */ | |
4722 | if (prev_insn == NULL_RTX || prev_insn == PREV_INSN (head) | |
4723 | || BLOCK_FOR_INSN (prev_insn) != curr_bb) | |
4724 | { | |
4725 | /* We reached the beginning of the current block -- do | |
4726 | rest of spliting in the current BB. */ | |
4727 | to_process = df_get_live_in (curr_bb); | |
4728 | if (BLOCK_FOR_INSN (head) != curr_bb) | |
f4eafc30 | 4729 | { |
55a2c322 VM |
4730 | /* We are somewhere in the middle of EBB. */ |
4731 | get_live_on_other_edges (EDGE_PRED (curr_bb, 0)->src, | |
4732 | curr_bb, &temp_bitmap); | |
4733 | to_process = &temp_bitmap; | |
4734 | } | |
4735 | head_p = true; | |
4736 | EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi) | |
4737 | { | |
4738 | if ((int) j >= lra_constraint_new_regno_start) | |
4739 | break; | |
4740 | if (((int) j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0) | |
4741 | && usage_insns[j].check == curr_usage_insns_check | |
4742 | && (next_usage_insns = usage_insns[j].insns) != NULL_RTX) | |
4743 | { | |
4744 | if (need_for_split_p (potential_reload_hard_regs, j)) | |
4745 | { | |
4746 | if (lra_dump_file != NULL && head_p) | |
4747 | { | |
4748 | fprintf (lra_dump_file, | |
4749 | " ----------------------------------\n"); | |
4750 | head_p = false; | |
4751 | } | |
4752 | if (split_reg (false, j, bb_note (curr_bb), | |
4753 | next_usage_insns)) | |
4754 | change_p = true; | |
4755 | } | |
4756 | usage_insns[j].check = 0; | |
4757 | } | |
4758 | } | |
4759 | } | |
4760 | } | |
4761 | return change_p; | |
4762 | } | |
4763 | ||
4764 | /* This value affects EBB forming. If probability of edge from EBB to | |
4765 | a BB is not greater than the following value, we don't add the BB | |
f4eafc30 | 4766 | to EBB. */ |
55a2c322 VM |
4767 | #define EBB_PROBABILITY_CUTOFF (REG_BR_PROB_BASE / 2) |
4768 | ||
4769 | /* Current number of inheritance/split iteration. */ | |
4770 | int lra_inheritance_iter; | |
4771 | ||
4772 | /* Entry function for inheritance/split pass. */ | |
4773 | void | |
4774 | lra_inheritance (void) | |
4775 | { | |
4776 | int i; | |
4777 | basic_block bb, start_bb; | |
4778 | edge e; | |
4779 | ||
55a2c322 | 4780 | lra_inheritance_iter++; |
8e3a4869 | 4781 | if (lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES) |
e731262b VM |
4782 | return; |
4783 | timevar_push (TV_LRA_INHERITANCE); | |
55a2c322 VM |
4784 | if (lra_dump_file != NULL) |
4785 | fprintf (lra_dump_file, "\n********** Inheritance #%d: **********\n\n", | |
4786 | lra_inheritance_iter); | |
4787 | curr_usage_insns_check = 0; | |
4788 | usage_insns = XNEWVEC (struct usage_insns, lra_constraint_new_regno_start); | |
4789 | for (i = 0; i < lra_constraint_new_regno_start; i++) | |
4790 | usage_insns[i].check = 0; | |
4791 | bitmap_initialize (&check_only_regs, ®_obstack); | |
4792 | bitmap_initialize (&live_regs, ®_obstack); | |
4793 | bitmap_initialize (&temp_bitmap, ®_obstack); | |
4794 | bitmap_initialize (&ebb_global_regs, ®_obstack); | |
4795 | FOR_EACH_BB (bb) | |
4796 | { | |
4797 | start_bb = bb; | |
4798 | if (lra_dump_file != NULL) | |
4799 | fprintf (lra_dump_file, "EBB"); | |
4800 | /* Form a EBB starting with BB. */ | |
4801 | bitmap_clear (&ebb_global_regs); | |
4802 | bitmap_ior_into (&ebb_global_regs, df_get_live_in (bb)); | |
4803 | for (;;) | |
4804 | { | |
4805 | if (lra_dump_file != NULL) | |
4806 | fprintf (lra_dump_file, " %d", bb->index); | |
4807 | if (bb->next_bb == EXIT_BLOCK_PTR || LABEL_P (BB_HEAD (bb->next_bb))) | |
4808 | break; | |
4809 | e = find_fallthru_edge (bb->succs); | |
4810 | if (! e) | |
4811 | break; | |
4812 | if (e->probability <= EBB_PROBABILITY_CUTOFF) | |
4813 | break; | |
4814 | bb = bb->next_bb; | |
4815 | } | |
4816 | bitmap_ior_into (&ebb_global_regs, df_get_live_out (bb)); | |
4817 | if (lra_dump_file != NULL) | |
4818 | fprintf (lra_dump_file, "\n"); | |
4819 | if (inherit_in_ebb (BB_HEAD (start_bb), BB_END (bb))) | |
4820 | /* Remember that the EBB head and tail can change in | |
4821 | inherit_in_ebb. */ | |
4822 | update_ebb_live_info (BB_HEAD (start_bb), BB_END (bb)); | |
4823 | } | |
4824 | bitmap_clear (&ebb_global_regs); | |
4825 | bitmap_clear (&temp_bitmap); | |
4826 | bitmap_clear (&live_regs); | |
4827 | bitmap_clear (&check_only_regs); | |
4828 | free (usage_insns); | |
4829 | ||
4830 | timevar_pop (TV_LRA_INHERITANCE); | |
4831 | } | |
4832 | ||
4833 | \f | |
4834 | ||
4835 | /* This page contains code to undo failed inheritance/split | |
4836 | transformations. */ | |
4837 | ||
4838 | /* Current number of iteration undoing inheritance/split. */ | |
4839 | int lra_undo_inheritance_iter; | |
4840 | ||
4841 | /* Fix BB live info LIVE after removing pseudos created on pass doing | |
4842 | inheritance/split which are REMOVED_PSEUDOS. */ | |
4843 | static void | |
4844 | fix_bb_live_info (bitmap live, bitmap removed_pseudos) | |
4845 | { | |
4846 | unsigned int regno; | |
4847 | bitmap_iterator bi; | |
4848 | ||
4849 | EXECUTE_IF_SET_IN_BITMAP (removed_pseudos, 0, regno, bi) | |
4850 | if (bitmap_clear_bit (live, regno)) | |
4851 | bitmap_set_bit (live, lra_reg_info[regno].restore_regno); | |
4852 | } | |
4853 | ||
4854 | /* Return regno of the (subreg of) REG. Otherwise, return a negative | |
4855 | number. */ | |
4856 | static int | |
4857 | get_regno (rtx reg) | |
4858 | { | |
4859 | if (GET_CODE (reg) == SUBREG) | |
4860 | reg = SUBREG_REG (reg); | |
4861 | if (REG_P (reg)) | |
4862 | return REGNO (reg); | |
4863 | return -1; | |
4864 | } | |
4865 | ||
4866 | /* Remove inheritance/split pseudos which are in REMOVE_PSEUDOS and | |
4867 | return true if we did any change. The undo transformations for | |
4868 | inheritance looks like | |
4869 | i <- i2 | |
4870 | p <- i => p <- i2 | |
4871 | or removing | |
4872 | p <- i, i <- p, and i <- i3 | |
4873 | where p is original pseudo from which inheritance pseudo i was | |
4874 | created, i and i3 are removed inheritance pseudos, i2 is another | |
4875 | not removed inheritance pseudo. All split pseudos or other | |
4876 | occurrences of removed inheritance pseudos are changed on the | |
4877 | corresponding original pseudos. | |
4878 | ||
4879 | The function also schedules insns changed and created during | |
4880 | inheritance/split pass for processing by the subsequent constraint | |
4881 | pass. */ | |
4882 | static bool | |
4883 | remove_inheritance_pseudos (bitmap remove_pseudos) | |
4884 | { | |
4885 | basic_block bb; | |
4886 | int regno, sregno, prev_sregno, dregno, restore_regno; | |
4887 | rtx set, prev_set, prev_insn; | |
4888 | bool change_p, done_p; | |
4889 | ||
4890 | change_p = ! bitmap_empty_p (remove_pseudos); | |
4891 | /* We can not finish the function right away if CHANGE_P is true | |
4892 | because we need to marks insns affected by previous | |
4893 | inheritance/split pass for processing by the subsequent | |
4894 | constraint pass. */ | |
4895 | FOR_EACH_BB (bb) | |
4896 | { | |
4897 | fix_bb_live_info (df_get_live_in (bb), remove_pseudos); | |
4898 | fix_bb_live_info (df_get_live_out (bb), remove_pseudos); | |
4899 | FOR_BB_INSNS_REVERSE (bb, curr_insn) | |
4900 | { | |
4901 | if (! INSN_P (curr_insn)) | |
4902 | continue; | |
4903 | done_p = false; | |
4904 | sregno = dregno = -1; | |
4905 | if (change_p && NONDEBUG_INSN_P (curr_insn) | |
4906 | && (set = single_set (curr_insn)) != NULL_RTX) | |
4907 | { | |
4908 | dregno = get_regno (SET_DEST (set)); | |
4909 | sregno = get_regno (SET_SRC (set)); | |
4910 | } | |
f4eafc30 | 4911 | |
55a2c322 VM |
4912 | if (sregno >= 0 && dregno >= 0) |
4913 | { | |
4914 | if ((bitmap_bit_p (remove_pseudos, sregno) | |
4915 | && (lra_reg_info[sregno].restore_regno == dregno | |
4916 | || (bitmap_bit_p (remove_pseudos, dregno) | |
4917 | && (lra_reg_info[sregno].restore_regno | |
4918 | == lra_reg_info[dregno].restore_regno)))) | |
4919 | || (bitmap_bit_p (remove_pseudos, dregno) | |
4920 | && lra_reg_info[dregno].restore_regno == sregno)) | |
4921 | /* One of the following cases: | |
4922 | original <- removed inheritance pseudo | |
4923 | removed inherit pseudo <- another removed inherit pseudo | |
4924 | removed inherit pseudo <- original pseudo | |
4925 | Or | |
4926 | removed_split_pseudo <- original_reg | |
4927 | original_reg <- removed_split_pseudo */ | |
4928 | { | |
4929 | if (lra_dump_file != NULL) | |
4930 | { | |
4931 | fprintf (lra_dump_file, " Removing %s:\n", | |
4932 | bitmap_bit_p (&lra_split_regs, sregno) | |
4933 | || bitmap_bit_p (&lra_split_regs, dregno) | |
4934 | ? "split" : "inheritance"); | |
cfbeaedf | 4935 | dump_insn_slim (lra_dump_file, curr_insn); |
55a2c322 VM |
4936 | } |
4937 | lra_set_insn_deleted (curr_insn); | |
4938 | done_p = true; | |
4939 | } | |
4940 | else if (bitmap_bit_p (remove_pseudos, sregno) | |
4941 | && bitmap_bit_p (&lra_inheritance_pseudos, sregno)) | |
4942 | { | |
4943 | /* Search the following pattern: | |
4944 | inherit_or_split_pseudo1 <- inherit_or_split_pseudo2 | |
4945 | original_pseudo <- inherit_or_split_pseudo1 | |
4946 | where the 2nd insn is the current insn and | |
4947 | inherit_or_split_pseudo2 is not removed. If it is found, | |
4948 | change the current insn onto: | |
4949 | original_pseudo <- inherit_or_split_pseudo2. */ | |
4950 | for (prev_insn = PREV_INSN (curr_insn); | |
4951 | prev_insn != NULL_RTX && ! NONDEBUG_INSN_P (prev_insn); | |
4952 | prev_insn = PREV_INSN (prev_insn)) | |
4953 | ; | |
4954 | if (prev_insn != NULL_RTX && BLOCK_FOR_INSN (prev_insn) == bb | |
4955 | && (prev_set = single_set (prev_insn)) != NULL_RTX | |
4956 | /* There should be no subregs in insn we are | |
4957 | searching because only the original reg might | |
4958 | be in subreg when we changed the mode of | |
4959 | load/store for splitting. */ | |
4960 | && REG_P (SET_DEST (prev_set)) | |
4961 | && REG_P (SET_SRC (prev_set)) | |
4962 | && (int) REGNO (SET_DEST (prev_set)) == sregno | |
4963 | && ((prev_sregno = REGNO (SET_SRC (prev_set))) | |
4964 | >= FIRST_PSEUDO_REGISTER) | |
4965 | /* As we consider chain of inheritance or | |
4966 | splitting described in above comment we should | |
4967 | check that sregno and prev_sregno were | |
4968 | inheritance/split pseudos created from the | |
4969 | same original regno. */ | |
4970 | && (lra_reg_info[sregno].restore_regno | |
4971 | == lra_reg_info[prev_sregno].restore_regno) | |
4972 | && ! bitmap_bit_p (remove_pseudos, prev_sregno)) | |
4973 | { | |
4974 | lra_assert (GET_MODE (SET_SRC (prev_set)) | |
4975 | == GET_MODE (regno_reg_rtx[sregno])); | |
4976 | if (GET_CODE (SET_SRC (set)) == SUBREG) | |
4977 | SUBREG_REG (SET_SRC (set)) = SET_SRC (prev_set); | |
4978 | else | |
4979 | SET_SRC (set) = SET_SRC (prev_set); | |
4980 | lra_push_insn_and_update_insn_regno_info (curr_insn); | |
4981 | lra_set_used_insn_alternative_by_uid | |
4982 | (INSN_UID (curr_insn), -1); | |
4983 | done_p = true; | |
4984 | if (lra_dump_file != NULL) | |
4985 | { | |
4986 | fprintf (lra_dump_file, " Change reload insn:\n"); | |
cfbeaedf | 4987 | dump_insn_slim (lra_dump_file, curr_insn); |
55a2c322 VM |
4988 | } |
4989 | } | |
4990 | } | |
4991 | } | |
4992 | if (! done_p) | |
4993 | { | |
4994 | struct lra_insn_reg *reg; | |
4995 | bool restored_regs_p = false; | |
4996 | bool kept_regs_p = false; | |
4997 | ||
4998 | curr_id = lra_get_insn_recog_data (curr_insn); | |
4999 | for (reg = curr_id->regs; reg != NULL; reg = reg->next) | |
5000 | { | |
5001 | regno = reg->regno; | |
5002 | restore_regno = lra_reg_info[regno].restore_regno; | |
5003 | if (restore_regno >= 0) | |
5004 | { | |
5005 | if (change_p && bitmap_bit_p (remove_pseudos, regno)) | |
5006 | { | |
5007 | substitute_pseudo (&curr_insn, regno, | |
5008 | regno_reg_rtx[restore_regno]); | |
5009 | restored_regs_p = true; | |
5010 | } | |
5011 | else | |
5012 | kept_regs_p = true; | |
5013 | } | |
5014 | } | |
5015 | if (NONDEBUG_INSN_P (curr_insn) && kept_regs_p) | |
5016 | { | |
5017 | /* The instruction has changed since the previous | |
5018 | constraints pass. */ | |
5019 | lra_push_insn_and_update_insn_regno_info (curr_insn); | |
5020 | lra_set_used_insn_alternative_by_uid | |
5021 | (INSN_UID (curr_insn), -1); | |
5022 | } | |
5023 | else if (restored_regs_p) | |
5024 | /* The instruction has been restored to the form that | |
5025 | it had during the previous constraints pass. */ | |
5026 | lra_update_insn_regno_info (curr_insn); | |
5027 | if (restored_regs_p && lra_dump_file != NULL) | |
5028 | { | |
5029 | fprintf (lra_dump_file, " Insn after restoring regs:\n"); | |
cfbeaedf | 5030 | dump_insn_slim (lra_dump_file, curr_insn); |
55a2c322 VM |
5031 | } |
5032 | } | |
5033 | } | |
5034 | } | |
5035 | return change_p; | |
5036 | } | |
5037 | ||
5038 | /* Entry function for undoing inheritance/split transformation. Return true | |
5039 | if we did any RTL change in this pass. */ | |
5040 | bool | |
5041 | lra_undo_inheritance (void) | |
5042 | { | |
5043 | unsigned int regno; | |
5044 | int restore_regno, hard_regno; | |
5045 | int n_all_inherit, n_inherit, n_all_split, n_split; | |
5046 | bitmap_head remove_pseudos; | |
5047 | bitmap_iterator bi; | |
5048 | bool change_p; | |
5049 | ||
5050 | lra_undo_inheritance_iter++; | |
8e3a4869 | 5051 | if (lra_undo_inheritance_iter > LRA_MAX_INHERITANCE_PASSES) |
e731262b | 5052 | return false; |
55a2c322 VM |
5053 | if (lra_dump_file != NULL) |
5054 | fprintf (lra_dump_file, | |
5055 | "\n********** Undoing inheritance #%d: **********\n\n", | |
5056 | lra_undo_inheritance_iter); | |
5057 | bitmap_initialize (&remove_pseudos, ®_obstack); | |
5058 | n_inherit = n_all_inherit = 0; | |
5059 | EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi) | |
5060 | if (lra_reg_info[regno].restore_regno >= 0) | |
5061 | { | |
5062 | n_all_inherit++; | |
5063 | if (reg_renumber[regno] < 0) | |
5064 | bitmap_set_bit (&remove_pseudos, regno); | |
5065 | else | |
5066 | n_inherit++; | |
5067 | } | |
5068 | if (lra_dump_file != NULL && n_all_inherit != 0) | |
5069 | fprintf (lra_dump_file, "Inherit %d out of %d (%.2f%%)\n", | |
5070 | n_inherit, n_all_inherit, | |
5071 | (double) n_inherit / n_all_inherit * 100); | |
5072 | n_split = n_all_split = 0; | |
5073 | EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi) | |
5074 | if ((restore_regno = lra_reg_info[regno].restore_regno) >= 0) | |
5075 | { | |
5076 | n_all_split++; | |
5077 | hard_regno = (restore_regno >= FIRST_PSEUDO_REGISTER | |
5078 | ? reg_renumber[restore_regno] : restore_regno); | |
5079 | if (hard_regno < 0 || reg_renumber[regno] == hard_regno) | |
5080 | bitmap_set_bit (&remove_pseudos, regno); | |
5081 | else | |
5082 | { | |
5083 | n_split++; | |
5084 | if (lra_dump_file != NULL) | |
5085 | fprintf (lra_dump_file, " Keep split r%d (orig=r%d)\n", | |
5086 | regno, restore_regno); | |
5087 | } | |
5088 | } | |
5089 | if (lra_dump_file != NULL && n_all_split != 0) | |
5090 | fprintf (lra_dump_file, "Split %d out of %d (%.2f%%)\n", | |
5091 | n_split, n_all_split, | |
5092 | (double) n_split / n_all_split * 100); | |
5093 | change_p = remove_inheritance_pseudos (&remove_pseudos); | |
5094 | bitmap_clear (&remove_pseudos); | |
5095 | /* Clear restore_regnos. */ | |
5096 | EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi) | |
5097 | lra_reg_info[regno].restore_regno = -1; | |
5098 | EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi) | |
5099 | lra_reg_info[regno].restore_regno = -1; | |
5100 | return change_p; | |
5101 | } |