re PR rtl-optimization/87305 (Segfault in end_hard_regno in setup_live_pseudos_and_sp...
[gcc.git] / gcc / lra-constraints.c
1 /* Code for RTL transformations to satisfy insn constraints.
2 Copyright (C) 2010-2019 Free Software Foundation, Inc.
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)
89 ... =>
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 "backend.h"
113 #include "target.h"
114 #include "rtl.h"
115 #include "tree.h"
116 #include "predict.h"
117 #include "df.h"
118 #include "memmodel.h"
119 #include "tm_p.h"
120 #include "expmed.h"
121 #include "optabs.h"
122 #include "regs.h"
123 #include "ira.h"
124 #include "recog.h"
125 #include "output.h"
126 #include "addresses.h"
127 #include "expr.h"
128 #include "cfgrtl.h"
129 #include "rtl-error.h"
130 #include "params.h"
131 #include "lra.h"
132 #include "lra-int.h"
133 #include "print-rtl.h"
134
135 /* Value of LRA_CURR_RELOAD_NUM at the beginning of BB of the current
136 insn. Remember that LRA_CURR_RELOAD_NUM is the number of emitted
137 reload insns. */
138 static int bb_reload_num;
139
140 /* The current insn being processed and corresponding its single set
141 (NULL otherwise), its data (basic block, the insn data, the insn
142 static data, and the mode of each operand). */
143 static rtx_insn *curr_insn;
144 static rtx curr_insn_set;
145 static basic_block curr_bb;
146 static lra_insn_recog_data_t curr_id;
147 static struct lra_static_insn_data *curr_static_id;
148 static machine_mode curr_operand_mode[MAX_RECOG_OPERANDS];
149 /* Mode of the register substituted by its equivalence with VOIDmode
150 (e.g. constant) and whose subreg is given operand of the current
151 insn. VOIDmode in all other cases. */
152 static machine_mode original_subreg_reg_mode[MAX_RECOG_OPERANDS];
153
154 \f
155
156 /* Start numbers for new registers and insns at the current constraints
157 pass start. */
158 static int new_regno_start;
159 static int new_insn_uid_start;
160
161 /* If LOC is nonnull, strip any outer subreg from it. */
162 static inline rtx *
163 strip_subreg (rtx *loc)
164 {
165 return loc && GET_CODE (*loc) == SUBREG ? &SUBREG_REG (*loc) : loc;
166 }
167
168 /* Return hard regno of REGNO or if it is was not assigned to a hard
169 register, use a hard register from its allocno class. */
170 static int
171 get_try_hard_regno (int regno)
172 {
173 int hard_regno;
174 enum reg_class rclass;
175
176 if ((hard_regno = regno) >= FIRST_PSEUDO_REGISTER)
177 hard_regno = lra_get_regno_hard_regno (regno);
178 if (hard_regno >= 0)
179 return hard_regno;
180 rclass = lra_get_allocno_class (regno);
181 if (rclass == NO_REGS)
182 return -1;
183 return ira_class_hard_regs[rclass][0];
184 }
185
186 /* Return the hard regno of X after removing its subreg. If X is not
187 a register or a subreg of a register, return -1. If X is a pseudo,
188 use its assignment. If FINAL_P return the final hard regno which will
189 be after elimination. */
190 static int
191 get_hard_regno (rtx x, bool final_p)
192 {
193 rtx reg;
194 int hard_regno;
195
196 reg = x;
197 if (SUBREG_P (x))
198 reg = SUBREG_REG (x);
199 if (! REG_P (reg))
200 return -1;
201 if (! HARD_REGISTER_NUM_P (hard_regno = REGNO (reg)))
202 hard_regno = lra_get_regno_hard_regno (hard_regno);
203 if (hard_regno < 0)
204 return -1;
205 if (final_p)
206 hard_regno = lra_get_elimination_hard_regno (hard_regno);
207 if (SUBREG_P (x))
208 hard_regno += subreg_regno_offset (hard_regno, GET_MODE (reg),
209 SUBREG_BYTE (x), GET_MODE (x));
210 return hard_regno;
211 }
212
213 /* If REGNO is a hard register or has been allocated a hard register,
214 return the class of that register. If REGNO is a reload pseudo
215 created by the current constraints pass, return its allocno class.
216 Return NO_REGS otherwise. */
217 static enum reg_class
218 get_reg_class (int regno)
219 {
220 int hard_regno;
221
222 if (! HARD_REGISTER_NUM_P (hard_regno = regno))
223 hard_regno = lra_get_regno_hard_regno (regno);
224 if (hard_regno >= 0)
225 {
226 hard_regno = lra_get_elimination_hard_regno (hard_regno);
227 return REGNO_REG_CLASS (hard_regno);
228 }
229 if (regno >= new_regno_start)
230 return lra_get_allocno_class (regno);
231 return NO_REGS;
232 }
233
234 /* Return true if REG satisfies (or will satisfy) reg class constraint
235 CL. Use elimination first if REG is a hard register. If REG is a
236 reload pseudo created by this constraints pass, assume that it will
237 be allocated a hard register from its allocno class, but allow that
238 class to be narrowed to CL if it is currently a superset of CL.
239
240 If NEW_CLASS is nonnull, set *NEW_CLASS to the new allocno class of
241 REGNO (reg), or NO_REGS if no change in its class was needed. */
242 static bool
243 in_class_p (rtx reg, enum reg_class cl, enum reg_class *new_class)
244 {
245 enum reg_class rclass, common_class;
246 machine_mode reg_mode;
247 int class_size, hard_regno, nregs, i, j;
248 int regno = REGNO (reg);
249
250 if (new_class != NULL)
251 *new_class = NO_REGS;
252 if (regno < FIRST_PSEUDO_REGISTER)
253 {
254 rtx final_reg = reg;
255 rtx *final_loc = &final_reg;
256
257 lra_eliminate_reg_if_possible (final_loc);
258 return TEST_HARD_REG_BIT (reg_class_contents[cl], REGNO (*final_loc));
259 }
260 reg_mode = GET_MODE (reg);
261 rclass = get_reg_class (regno);
262 if (regno < new_regno_start
263 /* Do not allow the constraints for reload instructions to
264 influence the classes of new pseudos. These reloads are
265 typically moves that have many alternatives, and restricting
266 reload pseudos for one alternative may lead to situations
267 where other reload pseudos are no longer allocatable. */
268 || (INSN_UID (curr_insn) >= new_insn_uid_start
269 && curr_insn_set != NULL
270 && ((OBJECT_P (SET_SRC (curr_insn_set))
271 && ! CONSTANT_P (SET_SRC (curr_insn_set)))
272 || (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
273 && OBJECT_P (SUBREG_REG (SET_SRC (curr_insn_set)))
274 && ! CONSTANT_P (SUBREG_REG (SET_SRC (curr_insn_set)))))))
275 /* When we don't know what class will be used finally for reload
276 pseudos, we use ALL_REGS. */
277 return ((regno >= new_regno_start && rclass == ALL_REGS)
278 || (rclass != NO_REGS && ira_class_subset_p[rclass][cl]
279 && ! hard_reg_set_subset_p (reg_class_contents[cl],
280 lra_no_alloc_regs)));
281 else
282 {
283 common_class = ira_reg_class_subset[rclass][cl];
284 if (new_class != NULL)
285 *new_class = common_class;
286 if (hard_reg_set_subset_p (reg_class_contents[common_class],
287 lra_no_alloc_regs))
288 return false;
289 /* Check that there are enough allocatable regs. */
290 class_size = ira_class_hard_regs_num[common_class];
291 for (i = 0; i < class_size; i++)
292 {
293 hard_regno = ira_class_hard_regs[common_class][i];
294 nregs = hard_regno_nregs (hard_regno, reg_mode);
295 if (nregs == 1)
296 return true;
297 for (j = 0; j < nregs; j++)
298 if (TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno + j)
299 || ! TEST_HARD_REG_BIT (reg_class_contents[common_class],
300 hard_regno + j))
301 break;
302 if (j >= nregs)
303 return true;
304 }
305 return false;
306 }
307 }
308
309 /* Return true if REGNO satisfies a memory constraint. */
310 static bool
311 in_mem_p (int regno)
312 {
313 return get_reg_class (regno) == NO_REGS;
314 }
315
316 /* Return 1 if ADDR is a valid memory address for mode MODE in address
317 space AS, and check that each pseudo has the proper kind of hard
318 reg. */
319 static int
320 valid_address_p (machine_mode mode ATTRIBUTE_UNUSED,
321 rtx addr, addr_space_t as)
322 {
323 #ifdef GO_IF_LEGITIMATE_ADDRESS
324 lra_assert (ADDR_SPACE_GENERIC_P (as));
325 GO_IF_LEGITIMATE_ADDRESS (mode, addr, win);
326 return 0;
327
328 win:
329 return 1;
330 #else
331 return targetm.addr_space.legitimate_address_p (mode, addr, 0, as);
332 #endif
333 }
334
335 namespace {
336 /* Temporarily eliminates registers in an address (for the lifetime of
337 the object). */
338 class address_eliminator {
339 public:
340 address_eliminator (struct address_info *ad);
341 ~address_eliminator ();
342
343 private:
344 struct address_info *m_ad;
345 rtx *m_base_loc;
346 rtx m_base_reg;
347 rtx *m_index_loc;
348 rtx m_index_reg;
349 };
350 }
351
352 address_eliminator::address_eliminator (struct address_info *ad)
353 : m_ad (ad),
354 m_base_loc (strip_subreg (ad->base_term)),
355 m_base_reg (NULL_RTX),
356 m_index_loc (strip_subreg (ad->index_term)),
357 m_index_reg (NULL_RTX)
358 {
359 if (m_base_loc != NULL)
360 {
361 m_base_reg = *m_base_loc;
362 /* If we have non-legitimate address which is decomposed not in
363 the way we expected, don't do elimination here. In such case
364 the address will be reloaded and elimination will be done in
365 reload insn finally. */
366 if (REG_P (m_base_reg))
367 lra_eliminate_reg_if_possible (m_base_loc);
368 if (m_ad->base_term2 != NULL)
369 *m_ad->base_term2 = *m_ad->base_term;
370 }
371 if (m_index_loc != NULL)
372 {
373 m_index_reg = *m_index_loc;
374 if (REG_P (m_index_reg))
375 lra_eliminate_reg_if_possible (m_index_loc);
376 }
377 }
378
379 address_eliminator::~address_eliminator ()
380 {
381 if (m_base_loc && *m_base_loc != m_base_reg)
382 {
383 *m_base_loc = m_base_reg;
384 if (m_ad->base_term2 != NULL)
385 *m_ad->base_term2 = *m_ad->base_term;
386 }
387 if (m_index_loc && *m_index_loc != m_index_reg)
388 *m_index_loc = m_index_reg;
389 }
390
391 /* Return true if the eliminated form of AD is a legitimate target address. */
392 static bool
393 valid_address_p (struct address_info *ad)
394 {
395 address_eliminator eliminator (ad);
396 return valid_address_p (ad->mode, *ad->outer, ad->as);
397 }
398
399 /* Return true if the eliminated form of memory reference OP satisfies
400 extra (special) memory constraint CONSTRAINT. */
401 static bool
402 satisfies_memory_constraint_p (rtx op, enum constraint_num constraint)
403 {
404 struct address_info ad;
405
406 decompose_mem_address (&ad, op);
407 address_eliminator eliminator (&ad);
408 return constraint_satisfied_p (op, constraint);
409 }
410
411 /* Return true if the eliminated form of address AD satisfies extra
412 address constraint CONSTRAINT. */
413 static bool
414 satisfies_address_constraint_p (struct address_info *ad,
415 enum constraint_num constraint)
416 {
417 address_eliminator eliminator (ad);
418 return constraint_satisfied_p (*ad->outer, constraint);
419 }
420
421 /* Return true if the eliminated form of address OP satisfies extra
422 address constraint CONSTRAINT. */
423 static bool
424 satisfies_address_constraint_p (rtx op, enum constraint_num constraint)
425 {
426 struct address_info ad;
427
428 decompose_lea_address (&ad, &op);
429 return satisfies_address_constraint_p (&ad, constraint);
430 }
431
432 /* Initiate equivalences for LRA. As we keep original equivalences
433 before any elimination, we need to make copies otherwise any change
434 in insns might change the equivalences. */
435 void
436 lra_init_equiv (void)
437 {
438 ira_expand_reg_equiv ();
439 for (int i = FIRST_PSEUDO_REGISTER; i < max_reg_num (); i++)
440 {
441 rtx res;
442
443 if ((res = ira_reg_equiv[i].memory) != NULL_RTX)
444 ira_reg_equiv[i].memory = copy_rtx (res);
445 if ((res = ira_reg_equiv[i].invariant) != NULL_RTX)
446 ira_reg_equiv[i].invariant = copy_rtx (res);
447 }
448 }
449
450 static rtx loc_equivalence_callback (rtx, const_rtx, void *);
451
452 /* Update equivalence for REGNO. We need to this as the equivalence
453 might contain other pseudos which are changed by their
454 equivalences. */
455 static void
456 update_equiv (int regno)
457 {
458 rtx x;
459
460 if ((x = ira_reg_equiv[regno].memory) != NULL_RTX)
461 ira_reg_equiv[regno].memory
462 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
463 NULL_RTX);
464 if ((x = ira_reg_equiv[regno].invariant) != NULL_RTX)
465 ira_reg_equiv[regno].invariant
466 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
467 NULL_RTX);
468 }
469
470 /* If we have decided to substitute X with another value, return that
471 value, otherwise return X. */
472 static rtx
473 get_equiv (rtx x)
474 {
475 int regno;
476 rtx res;
477
478 if (! REG_P (x) || (regno = REGNO (x)) < FIRST_PSEUDO_REGISTER
479 || ! ira_reg_equiv[regno].defined_p
480 || ! ira_reg_equiv[regno].profitable_p
481 || lra_get_regno_hard_regno (regno) >= 0)
482 return x;
483 if ((res = ira_reg_equiv[regno].memory) != NULL_RTX)
484 {
485 if (targetm.cannot_substitute_mem_equiv_p (res))
486 return x;
487 return res;
488 }
489 if ((res = ira_reg_equiv[regno].constant) != NULL_RTX)
490 return res;
491 if ((res = ira_reg_equiv[regno].invariant) != NULL_RTX)
492 return res;
493 gcc_unreachable ();
494 }
495
496 /* If we have decided to substitute X with the equivalent value,
497 return that value after elimination for INSN, otherwise return
498 X. */
499 static rtx
500 get_equiv_with_elimination (rtx x, rtx_insn *insn)
501 {
502 rtx res = get_equiv (x);
503
504 if (x == res || CONSTANT_P (res))
505 return res;
506 return lra_eliminate_regs_1 (insn, res, GET_MODE (res),
507 false, false, 0, true);
508 }
509
510 /* Set up curr_operand_mode. */
511 static void
512 init_curr_operand_mode (void)
513 {
514 int nop = curr_static_id->n_operands;
515 for (int i = 0; i < nop; i++)
516 {
517 machine_mode mode = GET_MODE (*curr_id->operand_loc[i]);
518 if (mode == VOIDmode)
519 {
520 /* The .md mode for address operands is the mode of the
521 addressed value rather than the mode of the address itself. */
522 if (curr_id->icode >= 0 && curr_static_id->operand[i].is_address)
523 mode = Pmode;
524 else
525 mode = curr_static_id->operand[i].mode;
526 }
527 curr_operand_mode[i] = mode;
528 }
529 }
530
531 \f
532
533 /* The page contains code to reuse input reloads. */
534
535 /* Structure describes input reload of the current insns. */
536 struct input_reload
537 {
538 /* True for input reload of matched operands. */
539 bool match_p;
540 /* Reloaded value. */
541 rtx input;
542 /* Reload pseudo used. */
543 rtx reg;
544 };
545
546 /* The number of elements in the following array. */
547 static int curr_insn_input_reloads_num;
548 /* Array containing info about input reloads. It is used to find the
549 same input reload and reuse the reload pseudo in this case. */
550 static struct input_reload curr_insn_input_reloads[LRA_MAX_INSN_RELOADS];
551
552 /* Initiate data concerning reuse of input reloads for the current
553 insn. */
554 static void
555 init_curr_insn_input_reloads (void)
556 {
557 curr_insn_input_reloads_num = 0;
558 }
559
560 /* Create a new pseudo using MODE, RCLASS, ORIGINAL or reuse already
561 created input reload pseudo (only if TYPE is not OP_OUT). Don't
562 reuse pseudo if IN_SUBREG_P is true and the reused pseudo should be
563 wrapped up in SUBREG. The result pseudo is returned through
564 RESULT_REG. Return TRUE if we created a new pseudo, FALSE if we
565 reused the already created input reload pseudo. Use TITLE to
566 describe new registers for debug purposes. */
567 static bool
568 get_reload_reg (enum op_type type, machine_mode mode, rtx original,
569 enum reg_class rclass, bool in_subreg_p,
570 const char *title, rtx *result_reg)
571 {
572 int i, regno;
573 enum reg_class new_class;
574 bool unique_p = false;
575
576 if (type == OP_OUT)
577 {
578 *result_reg
579 = lra_create_new_reg_with_unique_value (mode, original, rclass, title);
580 return true;
581 }
582 /* Prevent reuse value of expression with side effects,
583 e.g. volatile memory. */
584 if (! side_effects_p (original))
585 for (i = 0; i < curr_insn_input_reloads_num; i++)
586 {
587 if (! curr_insn_input_reloads[i].match_p
588 && rtx_equal_p (curr_insn_input_reloads[i].input, original)
589 && in_class_p (curr_insn_input_reloads[i].reg, rclass, &new_class))
590 {
591 rtx reg = curr_insn_input_reloads[i].reg;
592 regno = REGNO (reg);
593 /* If input is equal to original and both are VOIDmode,
594 GET_MODE (reg) might be still different from mode.
595 Ensure we don't return *result_reg with wrong mode. */
596 if (GET_MODE (reg) != mode)
597 {
598 if (in_subreg_p)
599 continue;
600 if (maybe_lt (GET_MODE_SIZE (GET_MODE (reg)),
601 GET_MODE_SIZE (mode)))
602 continue;
603 reg = lowpart_subreg (mode, reg, GET_MODE (reg));
604 if (reg == NULL_RTX || GET_CODE (reg) != SUBREG)
605 continue;
606 }
607 *result_reg = reg;
608 if (lra_dump_file != NULL)
609 {
610 fprintf (lra_dump_file, " Reuse r%d for reload ", regno);
611 dump_value_slim (lra_dump_file, original, 1);
612 }
613 if (new_class != lra_get_allocno_class (regno))
614 lra_change_class (regno, new_class, ", change to", false);
615 if (lra_dump_file != NULL)
616 fprintf (lra_dump_file, "\n");
617 return false;
618 }
619 /* If we have an input reload with a different mode, make sure it
620 will get a different hard reg. */
621 else if (REG_P (original)
622 && REG_P (curr_insn_input_reloads[i].input)
623 && REGNO (original) == REGNO (curr_insn_input_reloads[i].input)
624 && (GET_MODE (original)
625 != GET_MODE (curr_insn_input_reloads[i].input)))
626 unique_p = true;
627 }
628 *result_reg = (unique_p
629 ? lra_create_new_reg_with_unique_value
630 : lra_create_new_reg) (mode, original, rclass, title);
631 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
632 curr_insn_input_reloads[curr_insn_input_reloads_num].input = original;
633 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = false;
634 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = *result_reg;
635 return true;
636 }
637
638 \f
639 /* The page contains major code to choose the current insn alternative
640 and generate reloads for it. */
641
642 /* Return the offset from REGNO of the least significant register
643 in (reg:MODE REGNO).
644
645 This function is used to tell whether two registers satisfy
646 a matching constraint. (reg:MODE1 REGNO1) matches (reg:MODE2 REGNO2) if:
647
648 REGNO1 + lra_constraint_offset (REGNO1, MODE1)
649 == REGNO2 + lra_constraint_offset (REGNO2, MODE2) */
650 int
651 lra_constraint_offset (int regno, machine_mode mode)
652 {
653 lra_assert (regno < FIRST_PSEUDO_REGISTER);
654
655 scalar_int_mode int_mode;
656 if (WORDS_BIG_ENDIAN
657 && is_a <scalar_int_mode> (mode, &int_mode)
658 && GET_MODE_SIZE (int_mode) > UNITS_PER_WORD)
659 return hard_regno_nregs (regno, mode) - 1;
660 return 0;
661 }
662
663 /* Like rtx_equal_p except that it allows a REG and a SUBREG to match
664 if they are the same hard reg, and has special hacks for
665 auto-increment and auto-decrement. This is specifically intended for
666 process_alt_operands to use in determining whether two operands
667 match. X is the operand whose number is the lower of the two.
668
669 It is supposed that X is the output operand and Y is the input
670 operand. Y_HARD_REGNO is the final hard regno of register Y or
671 register in subreg Y as we know it now. Otherwise, it is a
672 negative value. */
673 static bool
674 operands_match_p (rtx x, rtx y, int y_hard_regno)
675 {
676 int i;
677 RTX_CODE code = GET_CODE (x);
678 const char *fmt;
679
680 if (x == y)
681 return true;
682 if ((code == REG || (code == SUBREG && REG_P (SUBREG_REG (x))))
683 && (REG_P (y) || (GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y)))))
684 {
685 int j;
686
687 i = get_hard_regno (x, false);
688 if (i < 0)
689 goto slow;
690
691 if ((j = y_hard_regno) < 0)
692 goto slow;
693
694 i += lra_constraint_offset (i, GET_MODE (x));
695 j += lra_constraint_offset (j, GET_MODE (y));
696
697 return i == j;
698 }
699
700 /* If two operands must match, because they are really a single
701 operand of an assembler insn, then two post-increments are invalid
702 because the assembler insn would increment only once. On the
703 other hand, a post-increment matches ordinary indexing if the
704 post-increment is the output operand. */
705 if (code == POST_DEC || code == POST_INC || code == POST_MODIFY)
706 return operands_match_p (XEXP (x, 0), y, y_hard_regno);
707
708 /* Two pre-increments are invalid because the assembler insn would
709 increment only once. On the other hand, a pre-increment matches
710 ordinary indexing if the pre-increment is the input operand. */
711 if (GET_CODE (y) == PRE_DEC || GET_CODE (y) == PRE_INC
712 || GET_CODE (y) == PRE_MODIFY)
713 return operands_match_p (x, XEXP (y, 0), -1);
714
715 slow:
716
717 if (code == REG && REG_P (y))
718 return REGNO (x) == REGNO (y);
719
720 if (code == REG && GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y))
721 && x == SUBREG_REG (y))
722 return true;
723 if (GET_CODE (y) == REG && code == SUBREG && REG_P (SUBREG_REG (x))
724 && SUBREG_REG (x) == y)
725 return true;
726
727 /* Now we have disposed of all the cases in which different rtx
728 codes can match. */
729 if (code != GET_CODE (y))
730 return false;
731
732 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
733 if (GET_MODE (x) != GET_MODE (y))
734 return false;
735
736 switch (code)
737 {
738 CASE_CONST_UNIQUE:
739 return false;
740
741 case LABEL_REF:
742 return label_ref_label (x) == label_ref_label (y);
743 case SYMBOL_REF:
744 return XSTR (x, 0) == XSTR (y, 0);
745
746 default:
747 break;
748 }
749
750 /* Compare the elements. If any pair of corresponding elements fail
751 to match, return false for the whole things. */
752
753 fmt = GET_RTX_FORMAT (code);
754 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
755 {
756 int val, j;
757 switch (fmt[i])
758 {
759 case 'w':
760 if (XWINT (x, i) != XWINT (y, i))
761 return false;
762 break;
763
764 case 'i':
765 if (XINT (x, i) != XINT (y, i))
766 return false;
767 break;
768
769 case 'p':
770 if (maybe_ne (SUBREG_BYTE (x), SUBREG_BYTE (y)))
771 return false;
772 break;
773
774 case 'e':
775 val = operands_match_p (XEXP (x, i), XEXP (y, i), -1);
776 if (val == 0)
777 return false;
778 break;
779
780 case '0':
781 break;
782
783 case 'E':
784 if (XVECLEN (x, i) != XVECLEN (y, i))
785 return false;
786 for (j = XVECLEN (x, i) - 1; j >= 0; --j)
787 {
788 val = operands_match_p (XVECEXP (x, i, j), XVECEXP (y, i, j), -1);
789 if (val == 0)
790 return false;
791 }
792 break;
793
794 /* It is believed that rtx's at this level will never
795 contain anything but integers and other rtx's, except for
796 within LABEL_REFs and SYMBOL_REFs. */
797 default:
798 gcc_unreachable ();
799 }
800 }
801 return true;
802 }
803
804 /* True if X is a constant that can be forced into the constant pool.
805 MODE is the mode of the operand, or VOIDmode if not known. */
806 #define CONST_POOL_OK_P(MODE, X) \
807 ((MODE) != VOIDmode \
808 && CONSTANT_P (X) \
809 && GET_CODE (X) != HIGH \
810 && GET_MODE_SIZE (MODE).is_constant () \
811 && !targetm.cannot_force_const_mem (MODE, X))
812
813 /* True if C is a non-empty register class that has too few registers
814 to be safely used as a reload target class. */
815 #define SMALL_REGISTER_CLASS_P(C) \
816 (ira_class_hard_regs_num [(C)] == 1 \
817 || (ira_class_hard_regs_num [(C)] >= 1 \
818 && targetm.class_likely_spilled_p (C)))
819
820 /* If REG is a reload pseudo, try to make its class satisfying CL. */
821 static void
822 narrow_reload_pseudo_class (rtx reg, enum reg_class cl)
823 {
824 enum reg_class rclass;
825
826 /* Do not make more accurate class from reloads generated. They are
827 mostly moves with a lot of constraints. Making more accurate
828 class may results in very narrow class and impossibility of find
829 registers for several reloads of one insn. */
830 if (INSN_UID (curr_insn) >= new_insn_uid_start)
831 return;
832 if (GET_CODE (reg) == SUBREG)
833 reg = SUBREG_REG (reg);
834 if (! REG_P (reg) || (int) REGNO (reg) < new_regno_start)
835 return;
836 if (in_class_p (reg, cl, &rclass) && rclass != cl)
837 lra_change_class (REGNO (reg), rclass, " Change to", true);
838 }
839
840 /* Searches X for any reference to a reg with the same value as REGNO,
841 returning the rtx of the reference found if any. Otherwise,
842 returns NULL_RTX. */
843 static rtx
844 regno_val_use_in (unsigned int regno, rtx x)
845 {
846 const char *fmt;
847 int i, j;
848 rtx tem;
849
850 if (REG_P (x) && lra_reg_info[REGNO (x)].val == lra_reg_info[regno].val)
851 return x;
852
853 fmt = GET_RTX_FORMAT (GET_CODE (x));
854 for (i = GET_RTX_LENGTH (GET_CODE (x)) - 1; i >= 0; i--)
855 {
856 if (fmt[i] == 'e')
857 {
858 if ((tem = regno_val_use_in (regno, XEXP (x, i))))
859 return tem;
860 }
861 else if (fmt[i] == 'E')
862 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
863 if ((tem = regno_val_use_in (regno , XVECEXP (x, i, j))))
864 return tem;
865 }
866
867 return NULL_RTX;
868 }
869
870 /* Return true if all current insn non-output operands except INS (it
871 has a negaitve end marker) do not use pseudos with the same value
872 as REGNO. */
873 static bool
874 check_conflict_input_operands (int regno, signed char *ins)
875 {
876 int in;
877 int n_operands = curr_static_id->n_operands;
878
879 for (int nop = 0; nop < n_operands; nop++)
880 if (! curr_static_id->operand[nop].is_operator
881 && curr_static_id->operand[nop].type != OP_OUT)
882 {
883 for (int i = 0; (in = ins[i]) >= 0; i++)
884 if (in == nop)
885 break;
886 if (in < 0
887 && regno_val_use_in (regno, *curr_id->operand_loc[nop]) != NULL_RTX)
888 return false;
889 }
890 return true;
891 }
892
893 /* Generate reloads for matching OUT and INS (array of input operand
894 numbers with end marker -1) with reg class GOAL_CLASS, considering
895 output operands OUTS (similar array to INS) needing to be in different
896 registers. Add input and output reloads correspondingly to the lists
897 *BEFORE and *AFTER. OUT might be negative. In this case we generate
898 input reloads for matched input operands INS. EARLY_CLOBBER_P is a flag
899 that the output operand is early clobbered for chosen alternative. */
900 static void
901 match_reload (signed char out, signed char *ins, signed char *outs,
902 enum reg_class goal_class, rtx_insn **before,
903 rtx_insn **after, bool early_clobber_p)
904 {
905 bool out_conflict;
906 int i, in;
907 rtx new_in_reg, new_out_reg, reg;
908 machine_mode inmode, outmode;
909 rtx in_rtx = *curr_id->operand_loc[ins[0]];
910 rtx out_rtx = out < 0 ? in_rtx : *curr_id->operand_loc[out];
911
912 inmode = curr_operand_mode[ins[0]];
913 outmode = out < 0 ? inmode : curr_operand_mode[out];
914 push_to_sequence (*before);
915 if (inmode != outmode)
916 {
917 /* process_alt_operands has already checked that the mode sizes
918 are ordered. */
919 if (partial_subreg_p (outmode, inmode))
920 {
921 reg = new_in_reg
922 = lra_create_new_reg_with_unique_value (inmode, in_rtx,
923 goal_class, "");
924 new_out_reg = gen_lowpart_SUBREG (outmode, reg);
925 LRA_SUBREG_P (new_out_reg) = 1;
926 /* If the input reg is dying here, we can use the same hard
927 register for REG and IN_RTX. We do it only for original
928 pseudos as reload pseudos can die although original
929 pseudos still live where reload pseudos dies. */
930 if (REG_P (in_rtx) && (int) REGNO (in_rtx) < lra_new_regno_start
931 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
932 && (!early_clobber_p
933 || check_conflict_input_operands(REGNO (in_rtx), ins)))
934 lra_assign_reg_val (REGNO (in_rtx), REGNO (reg));
935 }
936 else
937 {
938 reg = new_out_reg
939 = lra_create_new_reg_with_unique_value (outmode, out_rtx,
940 goal_class, "");
941 new_in_reg = gen_lowpart_SUBREG (inmode, reg);
942 /* NEW_IN_REG is non-paradoxical subreg. We don't want
943 NEW_OUT_REG living above. We add clobber clause for
944 this. This is just a temporary clobber. We can remove
945 it at the end of LRA work. */
946 rtx_insn *clobber = emit_clobber (new_out_reg);
947 LRA_TEMP_CLOBBER_P (PATTERN (clobber)) = 1;
948 LRA_SUBREG_P (new_in_reg) = 1;
949 if (GET_CODE (in_rtx) == SUBREG)
950 {
951 rtx subreg_reg = SUBREG_REG (in_rtx);
952
953 /* If SUBREG_REG is dying here and sub-registers IN_RTX
954 and NEW_IN_REG are similar, we can use the same hard
955 register for REG and SUBREG_REG. */
956 if (REG_P (subreg_reg)
957 && (int) REGNO (subreg_reg) < lra_new_regno_start
958 && GET_MODE (subreg_reg) == outmode
959 && known_eq (SUBREG_BYTE (in_rtx), SUBREG_BYTE (new_in_reg))
960 && find_regno_note (curr_insn, REG_DEAD, REGNO (subreg_reg))
961 && (! early_clobber_p
962 || check_conflict_input_operands (REGNO (subreg_reg),
963 ins)))
964 lra_assign_reg_val (REGNO (subreg_reg), REGNO (reg));
965 }
966 }
967 }
968 else
969 {
970 /* Pseudos have values -- see comments for lra_reg_info.
971 Different pseudos with the same value do not conflict even if
972 they live in the same place. When we create a pseudo we
973 assign value of original pseudo (if any) from which we
974 created the new pseudo. If we create the pseudo from the
975 input pseudo, the new pseudo will have no conflict with the
976 input pseudo which is wrong when the input pseudo lives after
977 the insn and as the new pseudo value is changed by the insn
978 output. Therefore we create the new pseudo from the output
979 except the case when we have single matched dying input
980 pseudo.
981
982 We cannot reuse the current output register because we might
983 have a situation like "a <- a op b", where the constraints
984 force the second input operand ("b") to match the output
985 operand ("a"). "b" must then be copied into a new register
986 so that it doesn't clobber the current value of "a".
987
988 We cannot use the same value if the output pseudo is
989 early clobbered or the input pseudo is mentioned in the
990 output, e.g. as an address part in memory, because
991 output reload will actually extend the pseudo liveness.
992 We don't care about eliminable hard regs here as we are
993 interesting only in pseudos. */
994
995 /* Matching input's register value is the same as one of the other
996 output operand. Output operands in a parallel insn must be in
997 different registers. */
998 out_conflict = false;
999 if (REG_P (in_rtx))
1000 {
1001 for (i = 0; outs[i] >= 0; i++)
1002 {
1003 rtx other_out_rtx = *curr_id->operand_loc[outs[i]];
1004 if (REG_P (other_out_rtx)
1005 && (regno_val_use_in (REGNO (in_rtx), other_out_rtx)
1006 != NULL_RTX))
1007 {
1008 out_conflict = true;
1009 break;
1010 }
1011 }
1012 }
1013
1014 new_in_reg = new_out_reg
1015 = (! early_clobber_p && ins[1] < 0 && REG_P (in_rtx)
1016 && (int) REGNO (in_rtx) < lra_new_regno_start
1017 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx))
1018 && (! early_clobber_p
1019 || check_conflict_input_operands (REGNO (in_rtx), ins))
1020 && (out < 0
1021 || regno_val_use_in (REGNO (in_rtx), out_rtx) == NULL_RTX)
1022 && !out_conflict
1023 ? lra_create_new_reg (inmode, in_rtx, goal_class, "")
1024 : lra_create_new_reg_with_unique_value (outmode, out_rtx,
1025 goal_class, ""));
1026 }
1027 /* In operand can be got from transformations before processing insn
1028 constraints. One example of such transformations is subreg
1029 reloading (see function simplify_operand_subreg). The new
1030 pseudos created by the transformations might have inaccurate
1031 class (ALL_REGS) and we should make their classes more
1032 accurate. */
1033 narrow_reload_pseudo_class (in_rtx, goal_class);
1034 lra_emit_move (copy_rtx (new_in_reg), in_rtx);
1035 *before = get_insns ();
1036 end_sequence ();
1037 /* Add the new pseudo to consider values of subsequent input reload
1038 pseudos. */
1039 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
1040 curr_insn_input_reloads[curr_insn_input_reloads_num].input = in_rtx;
1041 curr_insn_input_reloads[curr_insn_input_reloads_num].match_p = true;
1042 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = new_in_reg;
1043 for (i = 0; (in = ins[i]) >= 0; i++)
1044 {
1045 lra_assert
1046 (GET_MODE (*curr_id->operand_loc[in]) == VOIDmode
1047 || GET_MODE (new_in_reg) == GET_MODE (*curr_id->operand_loc[in]));
1048 *curr_id->operand_loc[in] = new_in_reg;
1049 }
1050 lra_update_dups (curr_id, ins);
1051 if (out < 0)
1052 return;
1053 /* See a comment for the input operand above. */
1054 narrow_reload_pseudo_class (out_rtx, goal_class);
1055 if (find_reg_note (curr_insn, REG_UNUSED, out_rtx) == NULL_RTX)
1056 {
1057 start_sequence ();
1058 lra_emit_move (out_rtx, copy_rtx (new_out_reg));
1059 emit_insn (*after);
1060 *after = get_insns ();
1061 end_sequence ();
1062 }
1063 *curr_id->operand_loc[out] = new_out_reg;
1064 lra_update_dup (curr_id, out);
1065 }
1066
1067 /* Return register class which is union of all reg classes in insn
1068 constraint alternative string starting with P. */
1069 static enum reg_class
1070 reg_class_from_constraints (const char *p)
1071 {
1072 int c, len;
1073 enum reg_class op_class = NO_REGS;
1074
1075 do
1076 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
1077 {
1078 case '#':
1079 case ',':
1080 return op_class;
1081
1082 case 'g':
1083 op_class = reg_class_subunion[op_class][GENERAL_REGS];
1084 break;
1085
1086 default:
1087 enum constraint_num cn = lookup_constraint (p);
1088 enum reg_class cl = reg_class_for_constraint (cn);
1089 if (cl == NO_REGS)
1090 {
1091 if (insn_extra_address_constraint (cn))
1092 op_class
1093 = (reg_class_subunion
1094 [op_class][base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
1095 ADDRESS, SCRATCH)]);
1096 break;
1097 }
1098
1099 op_class = reg_class_subunion[op_class][cl];
1100 break;
1101 }
1102 while ((p += len), c);
1103 return op_class;
1104 }
1105
1106 /* If OP is a register, return the class of the register as per
1107 get_reg_class, otherwise return NO_REGS. */
1108 static inline enum reg_class
1109 get_op_class (rtx op)
1110 {
1111 return REG_P (op) ? get_reg_class (REGNO (op)) : NO_REGS;
1112 }
1113
1114 /* Return generated insn mem_pseudo:=val if TO_P or val:=mem_pseudo
1115 otherwise. If modes of MEM_PSEUDO and VAL are different, use
1116 SUBREG for VAL to make them equal. */
1117 static rtx_insn *
1118 emit_spill_move (bool to_p, rtx mem_pseudo, rtx val)
1119 {
1120 if (GET_MODE (mem_pseudo) != GET_MODE (val))
1121 {
1122 /* Usually size of mem_pseudo is greater than val size but in
1123 rare cases it can be less as it can be defined by target
1124 dependent macro HARD_REGNO_CALLER_SAVE_MODE. */
1125 if (! MEM_P (val))
1126 {
1127 val = gen_lowpart_SUBREG (GET_MODE (mem_pseudo),
1128 GET_CODE (val) == SUBREG
1129 ? SUBREG_REG (val) : val);
1130 LRA_SUBREG_P (val) = 1;
1131 }
1132 else
1133 {
1134 mem_pseudo = gen_lowpart_SUBREG (GET_MODE (val), mem_pseudo);
1135 LRA_SUBREG_P (mem_pseudo) = 1;
1136 }
1137 }
1138 return to_p ? gen_move_insn (mem_pseudo, val)
1139 : gen_move_insn (val, mem_pseudo);
1140 }
1141
1142 /* Process a special case insn (register move), return true if we
1143 don't need to process it anymore. INSN should be a single set
1144 insn. Set up that RTL was changed through CHANGE_P and that hook
1145 TARGET_SECONDARY_MEMORY_NEEDED says to use secondary memory through
1146 SEC_MEM_P. */
1147 static bool
1148 check_and_process_move (bool *change_p, bool *sec_mem_p ATTRIBUTE_UNUSED)
1149 {
1150 int sregno, dregno;
1151 rtx dest, src, dreg, sreg, new_reg, scratch_reg;
1152 rtx_insn *before;
1153 enum reg_class dclass, sclass, secondary_class;
1154 secondary_reload_info sri;
1155
1156 lra_assert (curr_insn_set != NULL_RTX);
1157 dreg = dest = SET_DEST (curr_insn_set);
1158 sreg = src = SET_SRC (curr_insn_set);
1159 if (GET_CODE (dest) == SUBREG)
1160 dreg = SUBREG_REG (dest);
1161 if (GET_CODE (src) == SUBREG)
1162 sreg = SUBREG_REG (src);
1163 if (! (REG_P (dreg) || MEM_P (dreg)) || ! (REG_P (sreg) || MEM_P (sreg)))
1164 return false;
1165 sclass = dclass = NO_REGS;
1166 if (REG_P (dreg))
1167 dclass = get_reg_class (REGNO (dreg));
1168 gcc_assert (dclass < LIM_REG_CLASSES);
1169 if (dclass == ALL_REGS)
1170 /* ALL_REGS is used for new pseudos created by transformations
1171 like reload of SUBREG_REG (see function
1172 simplify_operand_subreg). We don't know their class yet. We
1173 should figure out the class from processing the insn
1174 constraints not in this fast path function. Even if ALL_REGS
1175 were a right class for the pseudo, secondary_... hooks usually
1176 are not define for ALL_REGS. */
1177 return false;
1178 if (REG_P (sreg))
1179 sclass = get_reg_class (REGNO (sreg));
1180 gcc_assert (sclass < LIM_REG_CLASSES);
1181 if (sclass == ALL_REGS)
1182 /* See comments above. */
1183 return false;
1184 if (sclass == NO_REGS && dclass == NO_REGS)
1185 return false;
1186 if (targetm.secondary_memory_needed (GET_MODE (src), sclass, dclass)
1187 && ((sclass != NO_REGS && dclass != NO_REGS)
1188 || (GET_MODE (src)
1189 != targetm.secondary_memory_needed_mode (GET_MODE (src)))))
1190 {
1191 *sec_mem_p = true;
1192 return false;
1193 }
1194 if (! REG_P (dreg) || ! REG_P (sreg))
1195 return false;
1196 sri.prev_sri = NULL;
1197 sri.icode = CODE_FOR_nothing;
1198 sri.extra_cost = 0;
1199 secondary_class = NO_REGS;
1200 /* Set up hard register for a reload pseudo for hook
1201 secondary_reload because some targets just ignore unassigned
1202 pseudos in the hook. */
1203 if (dclass != NO_REGS && lra_get_regno_hard_regno (REGNO (dreg)) < 0)
1204 {
1205 dregno = REGNO (dreg);
1206 reg_renumber[dregno] = ira_class_hard_regs[dclass][0];
1207 }
1208 else
1209 dregno = -1;
1210 if (sclass != NO_REGS && lra_get_regno_hard_regno (REGNO (sreg)) < 0)
1211 {
1212 sregno = REGNO (sreg);
1213 reg_renumber[sregno] = ira_class_hard_regs[sclass][0];
1214 }
1215 else
1216 sregno = -1;
1217 if (sclass != NO_REGS)
1218 secondary_class
1219 = (enum reg_class) targetm.secondary_reload (false, dest,
1220 (reg_class_t) sclass,
1221 GET_MODE (src), &sri);
1222 if (sclass == NO_REGS
1223 || ((secondary_class != NO_REGS || sri.icode != CODE_FOR_nothing)
1224 && dclass != NO_REGS))
1225 {
1226 enum reg_class old_sclass = secondary_class;
1227 secondary_reload_info old_sri = sri;
1228
1229 sri.prev_sri = NULL;
1230 sri.icode = CODE_FOR_nothing;
1231 sri.extra_cost = 0;
1232 secondary_class
1233 = (enum reg_class) targetm.secondary_reload (true, src,
1234 (reg_class_t) dclass,
1235 GET_MODE (src), &sri);
1236 /* Check the target hook consistency. */
1237 lra_assert
1238 ((secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1239 || (old_sclass == NO_REGS && old_sri.icode == CODE_FOR_nothing)
1240 || (secondary_class == old_sclass && sri.icode == old_sri.icode));
1241 }
1242 if (sregno >= 0)
1243 reg_renumber [sregno] = -1;
1244 if (dregno >= 0)
1245 reg_renumber [dregno] = -1;
1246 if (secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1247 return false;
1248 *change_p = true;
1249 new_reg = NULL_RTX;
1250 if (secondary_class != NO_REGS)
1251 new_reg = lra_create_new_reg_with_unique_value (GET_MODE (src), NULL_RTX,
1252 secondary_class,
1253 "secondary");
1254 start_sequence ();
1255 if (sri.icode == CODE_FOR_nothing)
1256 lra_emit_move (new_reg, src);
1257 else
1258 {
1259 enum reg_class scratch_class;
1260
1261 scratch_class = (reg_class_from_constraints
1262 (insn_data[sri.icode].operand[2].constraint));
1263 scratch_reg = (lra_create_new_reg_with_unique_value
1264 (insn_data[sri.icode].operand[2].mode, NULL_RTX,
1265 scratch_class, "scratch"));
1266 emit_insn (GEN_FCN (sri.icode) (new_reg != NULL_RTX ? new_reg : dest,
1267 src, scratch_reg));
1268 }
1269 before = get_insns ();
1270 end_sequence ();
1271 lra_process_new_insns (curr_insn, before, NULL, "Inserting the move");
1272 if (new_reg != NULL_RTX)
1273 SET_SRC (curr_insn_set) = new_reg;
1274 else
1275 {
1276 if (lra_dump_file != NULL)
1277 {
1278 fprintf (lra_dump_file, "Deleting move %u\n", INSN_UID (curr_insn));
1279 dump_insn_slim (lra_dump_file, curr_insn);
1280 }
1281 lra_set_insn_deleted (curr_insn);
1282 return true;
1283 }
1284 return false;
1285 }
1286
1287 /* The following data describe the result of process_alt_operands.
1288 The data are used in curr_insn_transform to generate reloads. */
1289
1290 /* The chosen reg classes which should be used for the corresponding
1291 operands. */
1292 static enum reg_class goal_alt[MAX_RECOG_OPERANDS];
1293 /* True if the operand should be the same as another operand and that
1294 other operand does not need a reload. */
1295 static bool goal_alt_match_win[MAX_RECOG_OPERANDS];
1296 /* True if the operand does not need a reload. */
1297 static bool goal_alt_win[MAX_RECOG_OPERANDS];
1298 /* True if the operand can be offsetable memory. */
1299 static bool goal_alt_offmemok[MAX_RECOG_OPERANDS];
1300 /* The number of an operand to which given operand can be matched to. */
1301 static int goal_alt_matches[MAX_RECOG_OPERANDS];
1302 /* The number of elements in the following array. */
1303 static int goal_alt_dont_inherit_ops_num;
1304 /* Numbers of operands whose reload pseudos should not be inherited. */
1305 static int goal_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1306 /* True if the insn commutative operands should be swapped. */
1307 static bool goal_alt_swapped;
1308 /* The chosen insn alternative. */
1309 static int goal_alt_number;
1310
1311 /* True if the corresponding operand is the result of an equivalence
1312 substitution. */
1313 static bool equiv_substition_p[MAX_RECOG_OPERANDS];
1314
1315 /* The following five variables are used to choose the best insn
1316 alternative. They reflect final characteristics of the best
1317 alternative. */
1318
1319 /* Number of necessary reloads and overall cost reflecting the
1320 previous value and other unpleasantness of the best alternative. */
1321 static int best_losers, best_overall;
1322 /* Overall number hard registers used for reloads. For example, on
1323 some targets we need 2 general registers to reload DFmode and only
1324 one floating point register. */
1325 static int best_reload_nregs;
1326 /* Overall number reflecting distances of previous reloading the same
1327 value. The distances are counted from the current BB start. It is
1328 used to improve inheritance chances. */
1329 static int best_reload_sum;
1330
1331 /* True if the current insn should have no correspondingly input or
1332 output reloads. */
1333 static bool no_input_reloads_p, no_output_reloads_p;
1334
1335 /* True if we swapped the commutative operands in the current
1336 insn. */
1337 static int curr_swapped;
1338
1339 /* if CHECK_ONLY_P is false, arrange for address element *LOC to be a
1340 register of class CL. Add any input reloads to list BEFORE. AFTER
1341 is nonnull if *LOC is an automodified value; handle that case by
1342 adding the required output reloads to list AFTER. Return true if
1343 the RTL was changed.
1344
1345 if CHECK_ONLY_P is true, check that the *LOC is a correct address
1346 register. Return false if the address register is correct. */
1347 static bool
1348 process_addr_reg (rtx *loc, bool check_only_p, rtx_insn **before, rtx_insn **after,
1349 enum reg_class cl)
1350 {
1351 int regno;
1352 enum reg_class rclass, new_class;
1353 rtx reg;
1354 rtx new_reg;
1355 machine_mode mode;
1356 bool subreg_p, before_p = false;
1357
1358 subreg_p = GET_CODE (*loc) == SUBREG;
1359 if (subreg_p)
1360 {
1361 reg = SUBREG_REG (*loc);
1362 mode = GET_MODE (reg);
1363
1364 /* For mode with size bigger than ptr_mode, there unlikely to be "mov"
1365 between two registers with different classes, but there normally will
1366 be "mov" which transfers element of vector register into the general
1367 register, and this normally will be a subreg which should be reloaded
1368 as a whole. This is particularly likely to be triggered when
1369 -fno-split-wide-types specified. */
1370 if (!REG_P (reg)
1371 || in_class_p (reg, cl, &new_class)
1372 || known_le (GET_MODE_SIZE (mode), GET_MODE_SIZE (ptr_mode)))
1373 loc = &SUBREG_REG (*loc);
1374 }
1375
1376 reg = *loc;
1377 mode = GET_MODE (reg);
1378 if (! REG_P (reg))
1379 {
1380 if (check_only_p)
1381 return true;
1382 /* Always reload memory in an address even if the target supports
1383 such addresses. */
1384 new_reg = lra_create_new_reg_with_unique_value (mode, reg, cl, "address");
1385 before_p = true;
1386 }
1387 else
1388 {
1389 regno = REGNO (reg);
1390 rclass = get_reg_class (regno);
1391 if (! check_only_p
1392 && (*loc = get_equiv_with_elimination (reg, curr_insn)) != reg)
1393 {
1394 if (lra_dump_file != NULL)
1395 {
1396 fprintf (lra_dump_file,
1397 "Changing pseudo %d in address of insn %u on equiv ",
1398 REGNO (reg), INSN_UID (curr_insn));
1399 dump_value_slim (lra_dump_file, *loc, 1);
1400 fprintf (lra_dump_file, "\n");
1401 }
1402 *loc = copy_rtx (*loc);
1403 }
1404 if (*loc != reg || ! in_class_p (reg, cl, &new_class))
1405 {
1406 if (check_only_p)
1407 return true;
1408 reg = *loc;
1409 if (get_reload_reg (after == NULL ? OP_IN : OP_INOUT,
1410 mode, reg, cl, subreg_p, "address", &new_reg))
1411 before_p = true;
1412 }
1413 else if (new_class != NO_REGS && rclass != new_class)
1414 {
1415 if (check_only_p)
1416 return true;
1417 lra_change_class (regno, new_class, " Change to", true);
1418 return false;
1419 }
1420 else
1421 return false;
1422 }
1423 if (before_p)
1424 {
1425 push_to_sequence (*before);
1426 lra_emit_move (new_reg, reg);
1427 *before = get_insns ();
1428 end_sequence ();
1429 }
1430 *loc = new_reg;
1431 if (after != NULL)
1432 {
1433 start_sequence ();
1434 lra_emit_move (before_p ? copy_rtx (reg) : reg, new_reg);
1435 emit_insn (*after);
1436 *after = get_insns ();
1437 end_sequence ();
1438 }
1439 return true;
1440 }
1441
1442 /* Insert move insn in simplify_operand_subreg. BEFORE returns
1443 the insn to be inserted before curr insn. AFTER returns the
1444 the insn to be inserted after curr insn. ORIGREG and NEWREG
1445 are the original reg and new reg for reload. */
1446 static void
1447 insert_move_for_subreg (rtx_insn **before, rtx_insn **after, rtx origreg,
1448 rtx newreg)
1449 {
1450 if (before)
1451 {
1452 push_to_sequence (*before);
1453 lra_emit_move (newreg, origreg);
1454 *before = get_insns ();
1455 end_sequence ();
1456 }
1457 if (after)
1458 {
1459 start_sequence ();
1460 lra_emit_move (origreg, newreg);
1461 emit_insn (*after);
1462 *after = get_insns ();
1463 end_sequence ();
1464 }
1465 }
1466
1467 static int valid_address_p (machine_mode mode, rtx addr, addr_space_t as);
1468 static bool process_address (int, bool, rtx_insn **, rtx_insn **);
1469
1470 /* Make reloads for subreg in operand NOP with internal subreg mode
1471 REG_MODE, add new reloads for further processing. Return true if
1472 any change was done. */
1473 static bool
1474 simplify_operand_subreg (int nop, machine_mode reg_mode)
1475 {
1476 int hard_regno;
1477 rtx_insn *before, *after;
1478 machine_mode mode, innermode;
1479 rtx reg, new_reg;
1480 rtx operand = *curr_id->operand_loc[nop];
1481 enum reg_class regclass;
1482 enum op_type type;
1483
1484 before = after = NULL;
1485
1486 if (GET_CODE (operand) != SUBREG)
1487 return false;
1488
1489 mode = GET_MODE (operand);
1490 reg = SUBREG_REG (operand);
1491 innermode = GET_MODE (reg);
1492 type = curr_static_id->operand[nop].type;
1493 if (MEM_P (reg))
1494 {
1495 const bool addr_was_valid
1496 = valid_address_p (innermode, XEXP (reg, 0), MEM_ADDR_SPACE (reg));
1497 alter_subreg (curr_id->operand_loc[nop], false);
1498 rtx subst = *curr_id->operand_loc[nop];
1499 lra_assert (MEM_P (subst));
1500
1501 if (!addr_was_valid
1502 || valid_address_p (GET_MODE (subst), XEXP (subst, 0),
1503 MEM_ADDR_SPACE (subst))
1504 || ((get_constraint_type (lookup_constraint
1505 (curr_static_id->operand[nop].constraint))
1506 != CT_SPECIAL_MEMORY)
1507 /* We still can reload address and if the address is
1508 valid, we can remove subreg without reloading its
1509 inner memory. */
1510 && valid_address_p (GET_MODE (subst),
1511 regno_reg_rtx
1512 [ira_class_hard_regs
1513 [base_reg_class (GET_MODE (subst),
1514 MEM_ADDR_SPACE (subst),
1515 ADDRESS, SCRATCH)][0]],
1516 MEM_ADDR_SPACE (subst))))
1517 {
1518 /* If we change the address for a paradoxical subreg of memory, the
1519 new address might violate the necessary alignment or the access
1520 might be slow; take this into consideration. We need not worry
1521 about accesses beyond allocated memory for paradoxical memory
1522 subregs as we don't substitute such equiv memory (see processing
1523 equivalences in function lra_constraints) and because for spilled
1524 pseudos we allocate stack memory enough for the biggest
1525 corresponding paradoxical subreg.
1526
1527 However, do not blindly simplify a (subreg (mem ...)) for
1528 WORD_REGISTER_OPERATIONS targets as this may lead to loading junk
1529 data into a register when the inner is narrower than outer or
1530 missing important data from memory when the inner is wider than
1531 outer. This rule only applies to modes that are no wider than
1532 a word. */
1533 if (!(maybe_ne (GET_MODE_PRECISION (mode),
1534 GET_MODE_PRECISION (innermode))
1535 && known_le (GET_MODE_SIZE (mode), UNITS_PER_WORD)
1536 && known_le (GET_MODE_SIZE (innermode), UNITS_PER_WORD)
1537 && WORD_REGISTER_OPERATIONS)
1538 && (!(MEM_ALIGN (subst) < GET_MODE_ALIGNMENT (mode)
1539 && targetm.slow_unaligned_access (mode, MEM_ALIGN (subst)))
1540 || (MEM_ALIGN (reg) < GET_MODE_ALIGNMENT (innermode)
1541 && targetm.slow_unaligned_access (innermode,
1542 MEM_ALIGN (reg)))))
1543 return true;
1544
1545 *curr_id->operand_loc[nop] = operand;
1546
1547 /* But if the address was not valid, we cannot reload the MEM without
1548 reloading the address first. */
1549 if (!addr_was_valid)
1550 process_address (nop, false, &before, &after);
1551
1552 /* INNERMODE is fast, MODE slow. Reload the mem in INNERMODE. */
1553 enum reg_class rclass
1554 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1555 if (get_reload_reg (curr_static_id->operand[nop].type, innermode,
1556 reg, rclass, TRUE, "slow mem", &new_reg))
1557 {
1558 bool insert_before, insert_after;
1559 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1560
1561 insert_before = (type != OP_OUT
1562 || partial_subreg_p (mode, innermode));
1563 insert_after = type != OP_IN;
1564 insert_move_for_subreg (insert_before ? &before : NULL,
1565 insert_after ? &after : NULL,
1566 reg, new_reg);
1567 }
1568 SUBREG_REG (operand) = new_reg;
1569
1570 /* Convert to MODE. */
1571 reg = operand;
1572 rclass
1573 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1574 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1575 rclass, TRUE, "slow mem", &new_reg))
1576 {
1577 bool insert_before, insert_after;
1578 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1579
1580 insert_before = type != OP_OUT;
1581 insert_after = type != OP_IN;
1582 insert_move_for_subreg (insert_before ? &before : NULL,
1583 insert_after ? &after : NULL,
1584 reg, new_reg);
1585 }
1586 *curr_id->operand_loc[nop] = new_reg;
1587 lra_process_new_insns (curr_insn, before, after,
1588 "Inserting slow mem reload");
1589 return true;
1590 }
1591
1592 /* If the address was valid and became invalid, prefer to reload
1593 the memory. Typical case is when the index scale should
1594 correspond the memory. */
1595 *curr_id->operand_loc[nop] = operand;
1596 /* Do not return false here as the MEM_P (reg) will be processed
1597 later in this function. */
1598 }
1599 else if (REG_P (reg) && REGNO (reg) < FIRST_PSEUDO_REGISTER)
1600 {
1601 alter_subreg (curr_id->operand_loc[nop], false);
1602 return true;
1603 }
1604 else if (CONSTANT_P (reg))
1605 {
1606 /* Try to simplify subreg of constant. It is usually result of
1607 equivalence substitution. */
1608 if (innermode == VOIDmode
1609 && (innermode = original_subreg_reg_mode[nop]) == VOIDmode)
1610 innermode = curr_static_id->operand[nop].mode;
1611 if ((new_reg = simplify_subreg (mode, reg, innermode,
1612 SUBREG_BYTE (operand))) != NULL_RTX)
1613 {
1614 *curr_id->operand_loc[nop] = new_reg;
1615 return true;
1616 }
1617 }
1618 /* Put constant into memory when we have mixed modes. It generates
1619 a better code in most cases as it does not need a secondary
1620 reload memory. It also prevents LRA looping when LRA is using
1621 secondary reload memory again and again. */
1622 if (CONSTANT_P (reg) && CONST_POOL_OK_P (reg_mode, reg)
1623 && SCALAR_INT_MODE_P (reg_mode) != SCALAR_INT_MODE_P (mode))
1624 {
1625 SUBREG_REG (operand) = force_const_mem (reg_mode, reg);
1626 alter_subreg (curr_id->operand_loc[nop], false);
1627 return true;
1628 }
1629 /* Force a reload of the SUBREG_REG if this is a constant or PLUS or
1630 if there may be a problem accessing OPERAND in the outer
1631 mode. */
1632 if ((REG_P (reg)
1633 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1634 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1635 /* Don't reload paradoxical subregs because we could be looping
1636 having repeatedly final regno out of hard regs range. */
1637 && (hard_regno_nregs (hard_regno, innermode)
1638 >= hard_regno_nregs (hard_regno, mode))
1639 && simplify_subreg_regno (hard_regno, innermode,
1640 SUBREG_BYTE (operand), mode) < 0
1641 /* Don't reload subreg for matching reload. It is actually
1642 valid subreg in LRA. */
1643 && ! LRA_SUBREG_P (operand))
1644 || CONSTANT_P (reg) || GET_CODE (reg) == PLUS || MEM_P (reg))
1645 {
1646 enum reg_class rclass;
1647
1648 if (REG_P (reg))
1649 /* There is a big probability that we will get the same class
1650 for the new pseudo and we will get the same insn which
1651 means infinite looping. So spill the new pseudo. */
1652 rclass = NO_REGS;
1653 else
1654 /* The class will be defined later in curr_insn_transform. */
1655 rclass
1656 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1657
1658 if (get_reload_reg (curr_static_id->operand[nop].type, reg_mode, reg,
1659 rclass, TRUE, "subreg reg", &new_reg))
1660 {
1661 bool insert_before, insert_after;
1662 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1663
1664 insert_before = (type != OP_OUT
1665 || read_modify_subreg_p (operand));
1666 insert_after = (type != OP_IN);
1667 insert_move_for_subreg (insert_before ? &before : NULL,
1668 insert_after ? &after : NULL,
1669 reg, new_reg);
1670 }
1671 SUBREG_REG (operand) = new_reg;
1672 lra_process_new_insns (curr_insn, before, after,
1673 "Inserting subreg reload");
1674 return true;
1675 }
1676 /* Force a reload for a paradoxical subreg. For paradoxical subreg,
1677 IRA allocates hardreg to the inner pseudo reg according to its mode
1678 instead of the outermode, so the size of the hardreg may not be enough
1679 to contain the outermode operand, in that case we may need to insert
1680 reload for the reg. For the following two types of paradoxical subreg,
1681 we need to insert reload:
1682 1. If the op_type is OP_IN, and the hardreg could not be paired with
1683 other hardreg to contain the outermode operand
1684 (checked by in_hard_reg_set_p), we need to insert the reload.
1685 2. If the op_type is OP_OUT or OP_INOUT.
1686
1687 Here is a paradoxical subreg example showing how the reload is generated:
1688
1689 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1690 (subreg:TI (reg:DI 107 [ __comp ]) 0)) {*movti_internal_rex64}
1691
1692 In IRA, reg107 is allocated to a DImode hardreg. We use x86-64 as example
1693 here, if reg107 is assigned to hardreg R15, because R15 is the last
1694 hardreg, compiler cannot find another hardreg to pair with R15 to
1695 contain TImode data. So we insert a TImode reload reg180 for it.
1696 After reload is inserted:
1697
1698 (insn 283 0 0 (set (subreg:DI (reg:TI 180 [orig:107 __comp ] [107]) 0)
1699 (reg:DI 107 [ __comp ])) -1
1700 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1701 (subreg:TI (reg:TI 180 [orig:107 __comp ] [107]) 0)) {*movti_internal_rex64}
1702
1703 Two reload hard registers will be allocated to reg180 to save TImode data
1704 in LRA_assign.
1705
1706 For LRA pseudos this should normally be handled by the biggest_mode
1707 mechanism. However, it's possible for new uses of an LRA pseudo
1708 to be introduced after we've allocated it, such as when undoing
1709 inheritance, and the allocated register might not then be appropriate
1710 for the new uses. */
1711 else if (REG_P (reg)
1712 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1713 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1714 && (hard_regno_nregs (hard_regno, innermode)
1715 < hard_regno_nregs (hard_regno, mode))
1716 && (regclass = lra_get_allocno_class (REGNO (reg)))
1717 && (type != OP_IN
1718 || !in_hard_reg_set_p (reg_class_contents[regclass],
1719 mode, hard_regno)
1720 || overlaps_hard_reg_set_p (lra_no_alloc_regs,
1721 mode, hard_regno)))
1722 {
1723 /* The class will be defined later in curr_insn_transform. */
1724 enum reg_class rclass
1725 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1726
1727 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1728 rclass, TRUE, "paradoxical subreg", &new_reg))
1729 {
1730 rtx subreg;
1731 bool insert_before, insert_after;
1732
1733 PUT_MODE (new_reg, mode);
1734 subreg = gen_lowpart_SUBREG (innermode, new_reg);
1735 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1736
1737 insert_before = (type != OP_OUT);
1738 insert_after = (type != OP_IN);
1739 insert_move_for_subreg (insert_before ? &before : NULL,
1740 insert_after ? &after : NULL,
1741 reg, subreg);
1742 }
1743 SUBREG_REG (operand) = new_reg;
1744 lra_process_new_insns (curr_insn, before, after,
1745 "Inserting paradoxical subreg reload");
1746 return true;
1747 }
1748 return false;
1749 }
1750
1751 /* Return TRUE if X refers for a hard register from SET. */
1752 static bool
1753 uses_hard_regs_p (rtx x, HARD_REG_SET set)
1754 {
1755 int i, j, x_hard_regno;
1756 machine_mode mode;
1757 const char *fmt;
1758 enum rtx_code code;
1759
1760 if (x == NULL_RTX)
1761 return false;
1762 code = GET_CODE (x);
1763 mode = GET_MODE (x);
1764 if (code == SUBREG)
1765 {
1766 mode = wider_subreg_mode (x);
1767 x = SUBREG_REG (x);
1768 code = GET_CODE (x);
1769 }
1770
1771 if (REG_P (x))
1772 {
1773 x_hard_regno = get_hard_regno (x, true);
1774 return (x_hard_regno >= 0
1775 && overlaps_hard_reg_set_p (set, mode, x_hard_regno));
1776 }
1777 if (MEM_P (x))
1778 {
1779 struct address_info ad;
1780
1781 decompose_mem_address (&ad, x);
1782 if (ad.base_term != NULL && uses_hard_regs_p (*ad.base_term, set))
1783 return true;
1784 if (ad.index_term != NULL && uses_hard_regs_p (*ad.index_term, set))
1785 return true;
1786 }
1787 fmt = GET_RTX_FORMAT (code);
1788 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1789 {
1790 if (fmt[i] == 'e')
1791 {
1792 if (uses_hard_regs_p (XEXP (x, i), set))
1793 return true;
1794 }
1795 else if (fmt[i] == 'E')
1796 {
1797 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1798 if (uses_hard_regs_p (XVECEXP (x, i, j), set))
1799 return true;
1800 }
1801 }
1802 return false;
1803 }
1804
1805 /* Return true if OP is a spilled pseudo. */
1806 static inline bool
1807 spilled_pseudo_p (rtx op)
1808 {
1809 return (REG_P (op)
1810 && REGNO (op) >= FIRST_PSEUDO_REGISTER && in_mem_p (REGNO (op)));
1811 }
1812
1813 /* Return true if X is a general constant. */
1814 static inline bool
1815 general_constant_p (rtx x)
1816 {
1817 return CONSTANT_P (x) && (! flag_pic || LEGITIMATE_PIC_OPERAND_P (x));
1818 }
1819
1820 static bool
1821 reg_in_class_p (rtx reg, enum reg_class cl)
1822 {
1823 if (cl == NO_REGS)
1824 return get_reg_class (REGNO (reg)) == NO_REGS;
1825 return in_class_p (reg, cl, NULL);
1826 }
1827
1828 /* Return true if SET of RCLASS contains no hard regs which can be
1829 used in MODE. */
1830 static bool
1831 prohibited_class_reg_set_mode_p (enum reg_class rclass,
1832 HARD_REG_SET &set,
1833 machine_mode mode)
1834 {
1835 HARD_REG_SET temp;
1836
1837 lra_assert (hard_reg_set_subset_p (reg_class_contents[rclass], set));
1838 COPY_HARD_REG_SET (temp, set);
1839 AND_COMPL_HARD_REG_SET (temp, lra_no_alloc_regs);
1840 return (hard_reg_set_subset_p
1841 (temp, ira_prohibited_class_mode_regs[rclass][mode]));
1842 }
1843
1844
1845 /* Used to check validity info about small class input operands. It
1846 should be incremented at start of processing an insn
1847 alternative. */
1848 static unsigned int curr_small_class_check = 0;
1849
1850 /* Update number of used inputs of class OP_CLASS for operand NOP.
1851 Return true if we have more such class operands than the number of
1852 available regs. */
1853 static bool
1854 update_and_check_small_class_inputs (int nop, enum reg_class op_class)
1855 {
1856 static unsigned int small_class_check[LIM_REG_CLASSES];
1857 static int small_class_input_nums[LIM_REG_CLASSES];
1858
1859 if (SMALL_REGISTER_CLASS_P (op_class)
1860 /* We are interesting in classes became small because of fixing
1861 some hard regs, e.g. by an user through GCC options. */
1862 && hard_reg_set_intersect_p (reg_class_contents[op_class],
1863 ira_no_alloc_regs)
1864 && (curr_static_id->operand[nop].type != OP_OUT
1865 || curr_static_id->operand[nop].early_clobber))
1866 {
1867 if (small_class_check[op_class] == curr_small_class_check)
1868 small_class_input_nums[op_class]++;
1869 else
1870 {
1871 small_class_check[op_class] = curr_small_class_check;
1872 small_class_input_nums[op_class] = 1;
1873 }
1874 if (small_class_input_nums[op_class] > ira_class_hard_regs_num[op_class])
1875 return true;
1876 }
1877 return false;
1878 }
1879
1880 /* Major function to choose the current insn alternative and what
1881 operands should be reloaded and how. If ONLY_ALTERNATIVE is not
1882 negative we should consider only this alternative. Return false if
1883 we cannot choose the alternative or find how to reload the
1884 operands. */
1885 static bool
1886 process_alt_operands (int only_alternative)
1887 {
1888 bool ok_p = false;
1889 int nop, overall, nalt;
1890 int n_alternatives = curr_static_id->n_alternatives;
1891 int n_operands = curr_static_id->n_operands;
1892 /* LOSERS counts the operands that don't fit this alternative and
1893 would require loading. */
1894 int losers;
1895 int addr_losers;
1896 /* REJECT is a count of how undesirable this alternative says it is
1897 if any reloading is required. If the alternative matches exactly
1898 then REJECT is ignored, but otherwise it gets this much counted
1899 against it in addition to the reloading needed. */
1900 int reject;
1901 /* This is defined by '!' or '?' alternative constraint and added to
1902 reject. But in some cases it can be ignored. */
1903 int static_reject;
1904 int op_reject;
1905 /* The number of elements in the following array. */
1906 int early_clobbered_regs_num;
1907 /* Numbers of operands which are early clobber registers. */
1908 int early_clobbered_nops[MAX_RECOG_OPERANDS];
1909 enum reg_class curr_alt[MAX_RECOG_OPERANDS];
1910 HARD_REG_SET curr_alt_set[MAX_RECOG_OPERANDS];
1911 bool curr_alt_match_win[MAX_RECOG_OPERANDS];
1912 bool curr_alt_win[MAX_RECOG_OPERANDS];
1913 bool curr_alt_offmemok[MAX_RECOG_OPERANDS];
1914 int curr_alt_matches[MAX_RECOG_OPERANDS];
1915 /* The number of elements in the following array. */
1916 int curr_alt_dont_inherit_ops_num;
1917 /* Numbers of operands whose reload pseudos should not be inherited. */
1918 int curr_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1919 rtx op;
1920 /* The register when the operand is a subreg of register, otherwise the
1921 operand itself. */
1922 rtx no_subreg_reg_operand[MAX_RECOG_OPERANDS];
1923 /* The register if the operand is a register or subreg of register,
1924 otherwise NULL. */
1925 rtx operand_reg[MAX_RECOG_OPERANDS];
1926 int hard_regno[MAX_RECOG_OPERANDS];
1927 machine_mode biggest_mode[MAX_RECOG_OPERANDS];
1928 int reload_nregs, reload_sum;
1929 bool costly_p;
1930 enum reg_class cl;
1931
1932 /* Calculate some data common for all alternatives to speed up the
1933 function. */
1934 for (nop = 0; nop < n_operands; nop++)
1935 {
1936 rtx reg;
1937
1938 op = no_subreg_reg_operand[nop] = *curr_id->operand_loc[nop];
1939 /* The real hard regno of the operand after the allocation. */
1940 hard_regno[nop] = get_hard_regno (op, true);
1941
1942 operand_reg[nop] = reg = op;
1943 biggest_mode[nop] = GET_MODE (op);
1944 if (GET_CODE (op) == SUBREG)
1945 {
1946 biggest_mode[nop] = wider_subreg_mode (op);
1947 operand_reg[nop] = reg = SUBREG_REG (op);
1948 }
1949 if (! REG_P (reg))
1950 operand_reg[nop] = NULL_RTX;
1951 else if (REGNO (reg) >= FIRST_PSEUDO_REGISTER
1952 || ((int) REGNO (reg)
1953 == lra_get_elimination_hard_regno (REGNO (reg))))
1954 no_subreg_reg_operand[nop] = reg;
1955 else
1956 operand_reg[nop] = no_subreg_reg_operand[nop]
1957 /* Just use natural mode for elimination result. It should
1958 be enough for extra constraints hooks. */
1959 = regno_reg_rtx[hard_regno[nop]];
1960 }
1961
1962 /* The constraints are made of several alternatives. Each operand's
1963 constraint looks like foo,bar,... with commas separating the
1964 alternatives. The first alternatives for all operands go
1965 together, the second alternatives go together, etc.
1966
1967 First loop over alternatives. */
1968 alternative_mask preferred = curr_id->preferred_alternatives;
1969 if (only_alternative >= 0)
1970 preferred &= ALTERNATIVE_BIT (only_alternative);
1971
1972 for (nalt = 0; nalt < n_alternatives; nalt++)
1973 {
1974 /* Loop over operands for one constraint alternative. */
1975 if (!TEST_BIT (preferred, nalt))
1976 continue;
1977
1978 bool matching_early_clobber[MAX_RECOG_OPERANDS];
1979 curr_small_class_check++;
1980 overall = losers = addr_losers = 0;
1981 static_reject = reject = reload_nregs = reload_sum = 0;
1982 for (nop = 0; nop < n_operands; nop++)
1983 {
1984 int inc = (curr_static_id
1985 ->operand_alternative[nalt * n_operands + nop].reject);
1986 if (lra_dump_file != NULL && inc != 0)
1987 fprintf (lra_dump_file,
1988 " Staticly defined alt reject+=%d\n", inc);
1989 static_reject += inc;
1990 matching_early_clobber[nop] = 0;
1991 }
1992 reject += static_reject;
1993 early_clobbered_regs_num = 0;
1994
1995 for (nop = 0; nop < n_operands; nop++)
1996 {
1997 const char *p;
1998 char *end;
1999 int len, c, m, i, opalt_num, this_alternative_matches;
2000 bool win, did_match, offmemok, early_clobber_p;
2001 /* false => this operand can be reloaded somehow for this
2002 alternative. */
2003 bool badop;
2004 /* true => this operand can be reloaded if the alternative
2005 allows regs. */
2006 bool winreg;
2007 /* True if a constant forced into memory would be OK for
2008 this operand. */
2009 bool constmemok;
2010 enum reg_class this_alternative, this_costly_alternative;
2011 HARD_REG_SET this_alternative_set, this_costly_alternative_set;
2012 bool this_alternative_match_win, this_alternative_win;
2013 bool this_alternative_offmemok;
2014 bool scratch_p;
2015 machine_mode mode;
2016 enum constraint_num cn;
2017
2018 opalt_num = nalt * n_operands + nop;
2019 if (curr_static_id->operand_alternative[opalt_num].anything_ok)
2020 {
2021 /* Fast track for no constraints at all. */
2022 curr_alt[nop] = NO_REGS;
2023 CLEAR_HARD_REG_SET (curr_alt_set[nop]);
2024 curr_alt_win[nop] = true;
2025 curr_alt_match_win[nop] = false;
2026 curr_alt_offmemok[nop] = false;
2027 curr_alt_matches[nop] = -1;
2028 continue;
2029 }
2030
2031 op = no_subreg_reg_operand[nop];
2032 mode = curr_operand_mode[nop];
2033
2034 win = did_match = winreg = offmemok = constmemok = false;
2035 badop = true;
2036
2037 early_clobber_p = false;
2038 p = curr_static_id->operand_alternative[opalt_num].constraint;
2039
2040 this_costly_alternative = this_alternative = NO_REGS;
2041 /* We update set of possible hard regs besides its class
2042 because reg class might be inaccurate. For example,
2043 union of LO_REGS (l), HI_REGS(h), and STACK_REG(k) in ARM
2044 is translated in HI_REGS because classes are merged by
2045 pairs and there is no accurate intermediate class. */
2046 CLEAR_HARD_REG_SET (this_alternative_set);
2047 CLEAR_HARD_REG_SET (this_costly_alternative_set);
2048 this_alternative_win = false;
2049 this_alternative_match_win = false;
2050 this_alternative_offmemok = false;
2051 this_alternative_matches = -1;
2052
2053 /* An empty constraint should be excluded by the fast
2054 track. */
2055 lra_assert (*p != 0 && *p != ',');
2056
2057 op_reject = 0;
2058 /* Scan this alternative's specs for this operand; set WIN
2059 if the operand fits any letter in this alternative.
2060 Otherwise, clear BADOP if this operand could fit some
2061 letter after reloads, or set WINREG if this operand could
2062 fit after reloads provided the constraint allows some
2063 registers. */
2064 costly_p = false;
2065 do
2066 {
2067 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
2068 {
2069 case '\0':
2070 len = 0;
2071 break;
2072 case ',':
2073 c = '\0';
2074 break;
2075
2076 case '&':
2077 early_clobber_p = true;
2078 break;
2079
2080 case '$':
2081 op_reject += LRA_MAX_REJECT;
2082 break;
2083 case '^':
2084 op_reject += LRA_LOSER_COST_FACTOR;
2085 break;
2086
2087 case '#':
2088 /* Ignore rest of this alternative. */
2089 c = '\0';
2090 break;
2091
2092 case '0': case '1': case '2': case '3': case '4':
2093 case '5': case '6': case '7': case '8': case '9':
2094 {
2095 int m_hregno;
2096 bool match_p;
2097
2098 m = strtoul (p, &end, 10);
2099 p = end;
2100 len = 0;
2101 lra_assert (nop > m);
2102
2103 /* Reject matches if we don't know which operand is
2104 bigger. This situation would arguably be a bug in
2105 an .md pattern, but could also occur in a user asm. */
2106 if (!ordered_p (GET_MODE_SIZE (biggest_mode[m]),
2107 GET_MODE_SIZE (biggest_mode[nop])))
2108 break;
2109
2110 /* Don't match wrong asm insn operands for proper
2111 diagnostic later. */
2112 if (INSN_CODE (curr_insn) < 0
2113 && (curr_operand_mode[m] == BLKmode
2114 || curr_operand_mode[nop] == BLKmode)
2115 && curr_operand_mode[m] != curr_operand_mode[nop])
2116 break;
2117
2118 m_hregno = get_hard_regno (*curr_id->operand_loc[m], false);
2119 /* We are supposed to match a previous operand.
2120 If we do, we win if that one did. If we do
2121 not, count both of the operands as losers.
2122 (This is too conservative, since most of the
2123 time only a single reload insn will be needed
2124 to make the two operands win. As a result,
2125 this alternative may be rejected when it is
2126 actually desirable.) */
2127 match_p = false;
2128 if (operands_match_p (*curr_id->operand_loc[nop],
2129 *curr_id->operand_loc[m], m_hregno))
2130 {
2131 /* We should reject matching of an early
2132 clobber operand if the matching operand is
2133 not dying in the insn. */
2134 if (! curr_static_id->operand[m].early_clobber
2135 || operand_reg[nop] == NULL_RTX
2136 || (find_regno_note (curr_insn, REG_DEAD,
2137 REGNO (op))
2138 || REGNO (op) == REGNO (operand_reg[m])))
2139 match_p = true;
2140 }
2141 if (match_p)
2142 {
2143 /* If we are matching a non-offsettable
2144 address where an offsettable address was
2145 expected, then we must reject this
2146 combination, because we can't reload
2147 it. */
2148 if (curr_alt_offmemok[m]
2149 && MEM_P (*curr_id->operand_loc[m])
2150 && curr_alt[m] == NO_REGS && ! curr_alt_win[m])
2151 continue;
2152 }
2153 else
2154 {
2155 /* Operands don't match. If the operands are
2156 different user defined explicit hard registers,
2157 then we cannot make them match. */
2158 if ((REG_P (*curr_id->operand_loc[nop])
2159 || SUBREG_P (*curr_id->operand_loc[nop]))
2160 && (REG_P (*curr_id->operand_loc[m])
2161 || SUBREG_P (*curr_id->operand_loc[m])))
2162 {
2163 rtx nop_reg = *curr_id->operand_loc[nop];
2164 if (SUBREG_P (nop_reg))
2165 nop_reg = SUBREG_REG (nop_reg);
2166 rtx m_reg = *curr_id->operand_loc[m];
2167 if (SUBREG_P (m_reg))
2168 m_reg = SUBREG_REG (m_reg);
2169
2170 if (REG_P (nop_reg)
2171 && HARD_REGISTER_P (nop_reg)
2172 && REG_USERVAR_P (nop_reg)
2173 && REG_P (m_reg)
2174 && HARD_REGISTER_P (m_reg)
2175 && REG_USERVAR_P (m_reg))
2176 break;
2177 }
2178
2179 /* Both operands must allow a reload register,
2180 otherwise we cannot make them match. */
2181 if (curr_alt[m] == NO_REGS)
2182 break;
2183 /* Retroactively mark the operand we had to
2184 match as a loser, if it wasn't already and
2185 it wasn't matched to a register constraint
2186 (e.g it might be matched by memory). */
2187 if (curr_alt_win[m]
2188 && (operand_reg[m] == NULL_RTX
2189 || hard_regno[m] < 0))
2190 {
2191 losers++;
2192 reload_nregs
2193 += (ira_reg_class_max_nregs[curr_alt[m]]
2194 [GET_MODE (*curr_id->operand_loc[m])]);
2195 }
2196
2197 /* Prefer matching earlyclobber alternative as
2198 it results in less hard regs required for
2199 the insn than a non-matching earlyclobber
2200 alternative. */
2201 if (curr_static_id->operand[m].early_clobber)
2202 {
2203 if (lra_dump_file != NULL)
2204 fprintf
2205 (lra_dump_file,
2206 " %d Matching earlyclobber alt:"
2207 " reject--\n",
2208 nop);
2209 if (!matching_early_clobber[m])
2210 {
2211 reject--;
2212 matching_early_clobber[m] = 1;
2213 }
2214 }
2215 /* Otherwise we prefer no matching
2216 alternatives because it gives more freedom
2217 in RA. */
2218 else if (operand_reg[nop] == NULL_RTX
2219 || (find_regno_note (curr_insn, REG_DEAD,
2220 REGNO (operand_reg[nop]))
2221 == NULL_RTX))
2222 {
2223 if (lra_dump_file != NULL)
2224 fprintf
2225 (lra_dump_file,
2226 " %d Matching alt: reject+=2\n",
2227 nop);
2228 reject += 2;
2229 }
2230 }
2231 /* If we have to reload this operand and some
2232 previous operand also had to match the same
2233 thing as this operand, we don't know how to do
2234 that. */
2235 if (!match_p || !curr_alt_win[m])
2236 {
2237 for (i = 0; i < nop; i++)
2238 if (curr_alt_matches[i] == m)
2239 break;
2240 if (i < nop)
2241 break;
2242 }
2243 else
2244 did_match = true;
2245
2246 this_alternative_matches = m;
2247 /* This can be fixed with reloads if the operand
2248 we are supposed to match can be fixed with
2249 reloads. */
2250 badop = false;
2251 this_alternative = curr_alt[m];
2252 COPY_HARD_REG_SET (this_alternative_set, curr_alt_set[m]);
2253 winreg = this_alternative != NO_REGS;
2254 break;
2255 }
2256
2257 case 'g':
2258 if (MEM_P (op)
2259 || general_constant_p (op)
2260 || spilled_pseudo_p (op))
2261 win = true;
2262 cl = GENERAL_REGS;
2263 goto reg;
2264
2265 default:
2266 cn = lookup_constraint (p);
2267 switch (get_constraint_type (cn))
2268 {
2269 case CT_REGISTER:
2270 cl = reg_class_for_constraint (cn);
2271 if (cl != NO_REGS)
2272 goto reg;
2273 break;
2274
2275 case CT_CONST_INT:
2276 if (CONST_INT_P (op)
2277 && insn_const_int_ok_for_constraint (INTVAL (op), cn))
2278 win = true;
2279 break;
2280
2281 case CT_MEMORY:
2282 if (MEM_P (op)
2283 && satisfies_memory_constraint_p (op, cn))
2284 win = true;
2285 else if (spilled_pseudo_p (op))
2286 win = true;
2287
2288 /* If we didn't already win, we can reload constants
2289 via force_const_mem or put the pseudo value into
2290 memory, or make other memory by reloading the
2291 address like for 'o'. */
2292 if (CONST_POOL_OK_P (mode, op)
2293 || MEM_P (op) || REG_P (op)
2294 /* We can restore the equiv insn by a
2295 reload. */
2296 || equiv_substition_p[nop])
2297 badop = false;
2298 constmemok = true;
2299 offmemok = true;
2300 break;
2301
2302 case CT_ADDRESS:
2303 /* An asm operand with an address constraint
2304 that doesn't satisfy address_operand has
2305 is_address cleared, so that we don't try to
2306 make a non-address fit. */
2307 if (!curr_static_id->operand[nop].is_address)
2308 break;
2309 /* If we didn't already win, we can reload the address
2310 into a base register. */
2311 if (satisfies_address_constraint_p (op, cn))
2312 win = true;
2313 cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
2314 ADDRESS, SCRATCH);
2315 badop = false;
2316 goto reg;
2317
2318 case CT_FIXED_FORM:
2319 if (constraint_satisfied_p (op, cn))
2320 win = true;
2321 break;
2322
2323 case CT_SPECIAL_MEMORY:
2324 if (MEM_P (op)
2325 && satisfies_memory_constraint_p (op, cn))
2326 win = true;
2327 else if (spilled_pseudo_p (op))
2328 win = true;
2329 break;
2330 }
2331 break;
2332
2333 reg:
2334 this_alternative = reg_class_subunion[this_alternative][cl];
2335 IOR_HARD_REG_SET (this_alternative_set,
2336 reg_class_contents[cl]);
2337 if (costly_p)
2338 {
2339 this_costly_alternative
2340 = reg_class_subunion[this_costly_alternative][cl];
2341 IOR_HARD_REG_SET (this_costly_alternative_set,
2342 reg_class_contents[cl]);
2343 }
2344 if (mode == BLKmode)
2345 break;
2346 winreg = true;
2347 if (REG_P (op))
2348 {
2349 if (hard_regno[nop] >= 0
2350 && in_hard_reg_set_p (this_alternative_set,
2351 mode, hard_regno[nop]))
2352 win = true;
2353 else if (hard_regno[nop] < 0
2354 && in_class_p (op, this_alternative, NULL))
2355 win = true;
2356 }
2357 break;
2358 }
2359 if (c != ' ' && c != '\t')
2360 costly_p = c == '*';
2361 }
2362 while ((p += len), c);
2363
2364 scratch_p = (operand_reg[nop] != NULL_RTX
2365 && lra_former_scratch_p (REGNO (operand_reg[nop])));
2366 /* Record which operands fit this alternative. */
2367 if (win)
2368 {
2369 this_alternative_win = true;
2370 if (operand_reg[nop] != NULL_RTX)
2371 {
2372 if (hard_regno[nop] >= 0)
2373 {
2374 if (in_hard_reg_set_p (this_costly_alternative_set,
2375 mode, hard_regno[nop]))
2376 {
2377 if (lra_dump_file != NULL)
2378 fprintf (lra_dump_file,
2379 " %d Costly set: reject++\n",
2380 nop);
2381 reject++;
2382 }
2383 }
2384 else
2385 {
2386 /* Prefer won reg to spilled pseudo under other
2387 equal conditions for possibe inheritance. */
2388 if (! scratch_p)
2389 {
2390 if (lra_dump_file != NULL)
2391 fprintf
2392 (lra_dump_file,
2393 " %d Non pseudo reload: reject++\n",
2394 nop);
2395 reject++;
2396 }
2397 if (in_class_p (operand_reg[nop],
2398 this_costly_alternative, NULL))
2399 {
2400 if (lra_dump_file != NULL)
2401 fprintf
2402 (lra_dump_file,
2403 " %d Non pseudo costly reload:"
2404 " reject++\n",
2405 nop);
2406 reject++;
2407 }
2408 }
2409 /* We simulate the behavior of old reload here.
2410 Although scratches need hard registers and it
2411 might result in spilling other pseudos, no reload
2412 insns are generated for the scratches. So it
2413 might cost something but probably less than old
2414 reload pass believes. */
2415 if (scratch_p)
2416 {
2417 if (lra_dump_file != NULL)
2418 fprintf (lra_dump_file,
2419 " %d Scratch win: reject+=2\n",
2420 nop);
2421 reject += 2;
2422 }
2423 }
2424 }
2425 else if (did_match)
2426 this_alternative_match_win = true;
2427 else
2428 {
2429 int const_to_mem = 0;
2430 bool no_regs_p;
2431
2432 reject += op_reject;
2433 /* Never do output reload of stack pointer. It makes
2434 impossible to do elimination when SP is changed in
2435 RTL. */
2436 if (op == stack_pointer_rtx && ! frame_pointer_needed
2437 && curr_static_id->operand[nop].type != OP_IN)
2438 goto fail;
2439
2440 /* If this alternative asks for a specific reg class, see if there
2441 is at least one allocatable register in that class. */
2442 no_regs_p
2443 = (this_alternative == NO_REGS
2444 || (hard_reg_set_subset_p
2445 (reg_class_contents[this_alternative],
2446 lra_no_alloc_regs)));
2447
2448 /* For asms, verify that the class for this alternative is possible
2449 for the mode that is specified. */
2450 if (!no_regs_p && INSN_CODE (curr_insn) < 0)
2451 {
2452 int i;
2453 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2454 if (targetm.hard_regno_mode_ok (i, mode)
2455 && in_hard_reg_set_p (reg_class_contents[this_alternative],
2456 mode, i))
2457 break;
2458 if (i == FIRST_PSEUDO_REGISTER)
2459 winreg = false;
2460 }
2461
2462 /* If this operand accepts a register, and if the
2463 register class has at least one allocatable register,
2464 then this operand can be reloaded. */
2465 if (winreg && !no_regs_p)
2466 badop = false;
2467
2468 if (badop)
2469 {
2470 if (lra_dump_file != NULL)
2471 fprintf (lra_dump_file,
2472 " alt=%d: Bad operand -- refuse\n",
2473 nalt);
2474 goto fail;
2475 }
2476
2477 if (this_alternative != NO_REGS)
2478 {
2479 HARD_REG_SET available_regs;
2480
2481 COPY_HARD_REG_SET (available_regs,
2482 reg_class_contents[this_alternative]);
2483 AND_COMPL_HARD_REG_SET
2484 (available_regs,
2485 ira_prohibited_class_mode_regs[this_alternative][mode]);
2486 AND_COMPL_HARD_REG_SET (available_regs, lra_no_alloc_regs);
2487 if (hard_reg_set_empty_p (available_regs))
2488 {
2489 /* There are no hard regs holding a value of given
2490 mode. */
2491 if (offmemok)
2492 {
2493 this_alternative = NO_REGS;
2494 if (lra_dump_file != NULL)
2495 fprintf (lra_dump_file,
2496 " %d Using memory because of"
2497 " a bad mode: reject+=2\n",
2498 nop);
2499 reject += 2;
2500 }
2501 else
2502 {
2503 if (lra_dump_file != NULL)
2504 fprintf (lra_dump_file,
2505 " alt=%d: Wrong mode -- refuse\n",
2506 nalt);
2507 goto fail;
2508 }
2509 }
2510 }
2511
2512 /* If not assigned pseudo has a class which a subset of
2513 required reg class, it is a less costly alternative
2514 as the pseudo still can get a hard reg of necessary
2515 class. */
2516 if (! no_regs_p && REG_P (op) && hard_regno[nop] < 0
2517 && (cl = get_reg_class (REGNO (op))) != NO_REGS
2518 && ira_class_subset_p[this_alternative][cl])
2519 {
2520 if (lra_dump_file != NULL)
2521 fprintf
2522 (lra_dump_file,
2523 " %d Super set class reg: reject-=3\n", nop);
2524 reject -= 3;
2525 }
2526
2527 this_alternative_offmemok = offmemok;
2528 if (this_costly_alternative != NO_REGS)
2529 {
2530 if (lra_dump_file != NULL)
2531 fprintf (lra_dump_file,
2532 " %d Costly loser: reject++\n", nop);
2533 reject++;
2534 }
2535 /* If the operand is dying, has a matching constraint,
2536 and satisfies constraints of the matched operand
2537 which failed to satisfy the own constraints, most probably
2538 the reload for this operand will be gone. */
2539 if (this_alternative_matches >= 0
2540 && !curr_alt_win[this_alternative_matches]
2541 && REG_P (op)
2542 && find_regno_note (curr_insn, REG_DEAD, REGNO (op))
2543 && (hard_regno[nop] >= 0
2544 ? in_hard_reg_set_p (this_alternative_set,
2545 mode, hard_regno[nop])
2546 : in_class_p (op, this_alternative, NULL)))
2547 {
2548 if (lra_dump_file != NULL)
2549 fprintf
2550 (lra_dump_file,
2551 " %d Dying matched operand reload: reject++\n",
2552 nop);
2553 reject++;
2554 }
2555 else
2556 {
2557 /* Strict_low_part requires to reload the register
2558 not the sub-register. In this case we should
2559 check that a final reload hard reg can hold the
2560 value mode. */
2561 if (curr_static_id->operand[nop].strict_low
2562 && REG_P (op)
2563 && hard_regno[nop] < 0
2564 && GET_CODE (*curr_id->operand_loc[nop]) == SUBREG
2565 && ira_class_hard_regs_num[this_alternative] > 0
2566 && (!targetm.hard_regno_mode_ok
2567 (ira_class_hard_regs[this_alternative][0],
2568 GET_MODE (*curr_id->operand_loc[nop]))))
2569 {
2570 if (lra_dump_file != NULL)
2571 fprintf
2572 (lra_dump_file,
2573 " alt=%d: Strict low subreg reload -- refuse\n",
2574 nalt);
2575 goto fail;
2576 }
2577 losers++;
2578 }
2579 if (operand_reg[nop] != NULL_RTX
2580 /* Output operands and matched input operands are
2581 not inherited. The following conditions do not
2582 exactly describe the previous statement but they
2583 are pretty close. */
2584 && curr_static_id->operand[nop].type != OP_OUT
2585 && (this_alternative_matches < 0
2586 || curr_static_id->operand[nop].type != OP_IN))
2587 {
2588 int last_reload = (lra_reg_info[ORIGINAL_REGNO
2589 (operand_reg[nop])]
2590 .last_reload);
2591
2592 /* The value of reload_sum has sense only if we
2593 process insns in their order. It happens only on
2594 the first constraints sub-pass when we do most of
2595 reload work. */
2596 if (lra_constraint_iter == 1 && last_reload > bb_reload_num)
2597 reload_sum += last_reload - bb_reload_num;
2598 }
2599 /* If this is a constant that is reloaded into the
2600 desired class by copying it to memory first, count
2601 that as another reload. This is consistent with
2602 other code and is required to avoid choosing another
2603 alternative when the constant is moved into memory.
2604 Note that the test here is precisely the same as in
2605 the code below that calls force_const_mem. */
2606 if (CONST_POOL_OK_P (mode, op)
2607 && ((targetm.preferred_reload_class
2608 (op, this_alternative) == NO_REGS)
2609 || no_input_reloads_p))
2610 {
2611 const_to_mem = 1;
2612 if (! no_regs_p)
2613 losers++;
2614 }
2615
2616 /* Alternative loses if it requires a type of reload not
2617 permitted for this insn. We can always reload
2618 objects with a REG_UNUSED note. */
2619 if ((curr_static_id->operand[nop].type != OP_IN
2620 && no_output_reloads_p
2621 && ! find_reg_note (curr_insn, REG_UNUSED, op))
2622 || (curr_static_id->operand[nop].type != OP_OUT
2623 && no_input_reloads_p && ! const_to_mem)
2624 || (this_alternative_matches >= 0
2625 && (no_input_reloads_p
2626 || (no_output_reloads_p
2627 && (curr_static_id->operand
2628 [this_alternative_matches].type != OP_IN)
2629 && ! find_reg_note (curr_insn, REG_UNUSED,
2630 no_subreg_reg_operand
2631 [this_alternative_matches])))))
2632 {
2633 if (lra_dump_file != NULL)
2634 fprintf
2635 (lra_dump_file,
2636 " alt=%d: No input/otput reload -- refuse\n",
2637 nalt);
2638 goto fail;
2639 }
2640
2641 /* Alternative loses if it required class pseudo cannot
2642 hold value of required mode. Such insns can be
2643 described by insn definitions with mode iterators. */
2644 if (GET_MODE (*curr_id->operand_loc[nop]) != VOIDmode
2645 && ! hard_reg_set_empty_p (this_alternative_set)
2646 /* It is common practice for constraints to use a
2647 class which does not have actually enough regs to
2648 hold the value (e.g. x86 AREG for mode requiring
2649 more one general reg). Therefore we have 2
2650 conditions to check that the reload pseudo cannot
2651 hold the mode value. */
2652 && (!targetm.hard_regno_mode_ok
2653 (ira_class_hard_regs[this_alternative][0],
2654 GET_MODE (*curr_id->operand_loc[nop])))
2655 /* The above condition is not enough as the first
2656 reg in ira_class_hard_regs can be not aligned for
2657 multi-words mode values. */
2658 && (prohibited_class_reg_set_mode_p
2659 (this_alternative, this_alternative_set,
2660 GET_MODE (*curr_id->operand_loc[nop]))))
2661 {
2662 if (lra_dump_file != NULL)
2663 fprintf (lra_dump_file,
2664 " alt=%d: reload pseudo for op %d "
2665 " cannot hold the mode value -- refuse\n",
2666 nalt, nop);
2667 goto fail;
2668 }
2669
2670 /* Check strong discouragement of reload of non-constant
2671 into class THIS_ALTERNATIVE. */
2672 if (! CONSTANT_P (op) && ! no_regs_p
2673 && (targetm.preferred_reload_class
2674 (op, this_alternative) == NO_REGS
2675 || (curr_static_id->operand[nop].type == OP_OUT
2676 && (targetm.preferred_output_reload_class
2677 (op, this_alternative) == NO_REGS))))
2678 {
2679 if (lra_dump_file != NULL)
2680 fprintf (lra_dump_file,
2681 " %d Non-prefered reload: reject+=%d\n",
2682 nop, LRA_MAX_REJECT);
2683 reject += LRA_MAX_REJECT;
2684 }
2685
2686 if (! (MEM_P (op) && offmemok)
2687 && ! (const_to_mem && constmemok))
2688 {
2689 /* We prefer to reload pseudos over reloading other
2690 things, since such reloads may be able to be
2691 eliminated later. So bump REJECT in other cases.
2692 Don't do this in the case where we are forcing a
2693 constant into memory and it will then win since
2694 we don't want to have a different alternative
2695 match then. */
2696 if (! (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER))
2697 {
2698 if (lra_dump_file != NULL)
2699 fprintf
2700 (lra_dump_file,
2701 " %d Non-pseudo reload: reject+=2\n",
2702 nop);
2703 reject += 2;
2704 }
2705
2706 if (! no_regs_p)
2707 reload_nregs
2708 += ira_reg_class_max_nregs[this_alternative][mode];
2709
2710 if (SMALL_REGISTER_CLASS_P (this_alternative))
2711 {
2712 if (lra_dump_file != NULL)
2713 fprintf
2714 (lra_dump_file,
2715 " %d Small class reload: reject+=%d\n",
2716 nop, LRA_LOSER_COST_FACTOR / 2);
2717 reject += LRA_LOSER_COST_FACTOR / 2;
2718 }
2719 }
2720
2721 /* We are trying to spill pseudo into memory. It is
2722 usually more costly than moving to a hard register
2723 although it might takes the same number of
2724 reloads.
2725
2726 Non-pseudo spill may happen also. Suppose a target allows both
2727 register and memory in the operand constraint alternatives,
2728 then it's typical that an eliminable register has a substition
2729 of "base + offset" which can either be reloaded by a simple
2730 "new_reg <= base + offset" which will match the register
2731 constraint, or a similar reg addition followed by further spill
2732 to and reload from memory which will match the memory
2733 constraint, but this memory spill will be much more costly
2734 usually.
2735
2736 Code below increases the reject for both pseudo and non-pseudo
2737 spill. */
2738 if (no_regs_p
2739 && !(MEM_P (op) && offmemok)
2740 && !(REG_P (op) && hard_regno[nop] < 0))
2741 {
2742 if (lra_dump_file != NULL)
2743 fprintf
2744 (lra_dump_file,
2745 " %d Spill %spseudo into memory: reject+=3\n",
2746 nop, REG_P (op) ? "" : "Non-");
2747 reject += 3;
2748 if (VECTOR_MODE_P (mode))
2749 {
2750 /* Spilling vectors into memory is usually more
2751 costly as they contain big values. */
2752 if (lra_dump_file != NULL)
2753 fprintf
2754 (lra_dump_file,
2755 " %d Spill vector pseudo: reject+=2\n",
2756 nop);
2757 reject += 2;
2758 }
2759 }
2760
2761 /* When we use an operand requiring memory in given
2762 alternative, the insn should write *and* read the
2763 value to/from memory it is costly in comparison with
2764 an insn alternative which does not use memory
2765 (e.g. register or immediate operand). We exclude
2766 memory operand for such case as we can satisfy the
2767 memory constraints by reloading address. */
2768 if (no_regs_p && offmemok && !MEM_P (op))
2769 {
2770 if (lra_dump_file != NULL)
2771 fprintf
2772 (lra_dump_file,
2773 " Using memory insn operand %d: reject+=3\n",
2774 nop);
2775 reject += 3;
2776 }
2777
2778 /* If reload requires moving value through secondary
2779 memory, it will need one more insn at least. */
2780 if (this_alternative != NO_REGS
2781 && REG_P (op) && (cl = get_reg_class (REGNO (op))) != NO_REGS
2782 && ((curr_static_id->operand[nop].type != OP_OUT
2783 && targetm.secondary_memory_needed (GET_MODE (op), cl,
2784 this_alternative))
2785 || (curr_static_id->operand[nop].type != OP_IN
2786 && (targetm.secondary_memory_needed
2787 (GET_MODE (op), this_alternative, cl)))))
2788 losers++;
2789
2790 /* Input reloads can be inherited more often than output
2791 reloads can be removed, so penalize output
2792 reloads. */
2793 if (!REG_P (op) || curr_static_id->operand[nop].type != OP_IN)
2794 {
2795 if (lra_dump_file != NULL)
2796 fprintf
2797 (lra_dump_file,
2798 " %d Non input pseudo reload: reject++\n",
2799 nop);
2800 reject++;
2801 }
2802
2803 if (MEM_P (op) && offmemok)
2804 addr_losers++;
2805 else if (curr_static_id->operand[nop].type == OP_INOUT)
2806 {
2807 if (lra_dump_file != NULL)
2808 fprintf
2809 (lra_dump_file,
2810 " %d Input/Output reload: reject+=%d\n",
2811 nop, LRA_LOSER_COST_FACTOR);
2812 reject += LRA_LOSER_COST_FACTOR;
2813 }
2814 }
2815
2816 if (early_clobber_p && ! scratch_p)
2817 {
2818 if (lra_dump_file != NULL)
2819 fprintf (lra_dump_file,
2820 " %d Early clobber: reject++\n", nop);
2821 reject++;
2822 }
2823 /* ??? We check early clobbers after processing all operands
2824 (see loop below) and there we update the costs more.
2825 Should we update the cost (may be approximately) here
2826 because of early clobber register reloads or it is a rare
2827 or non-important thing to be worth to do it. */
2828 overall = (losers * LRA_LOSER_COST_FACTOR + reject
2829 - (addr_losers == losers ? static_reject : 0));
2830 if ((best_losers == 0 || losers != 0) && best_overall < overall)
2831 {
2832 if (lra_dump_file != NULL)
2833 fprintf (lra_dump_file,
2834 " alt=%d,overall=%d,losers=%d -- refuse\n",
2835 nalt, overall, losers);
2836 goto fail;
2837 }
2838
2839 if (update_and_check_small_class_inputs (nop, this_alternative))
2840 {
2841 if (lra_dump_file != NULL)
2842 fprintf (lra_dump_file,
2843 " alt=%d, not enough small class regs -- refuse\n",
2844 nalt);
2845 goto fail;
2846 }
2847 curr_alt[nop] = this_alternative;
2848 COPY_HARD_REG_SET (curr_alt_set[nop], this_alternative_set);
2849 curr_alt_win[nop] = this_alternative_win;
2850 curr_alt_match_win[nop] = this_alternative_match_win;
2851 curr_alt_offmemok[nop] = this_alternative_offmemok;
2852 curr_alt_matches[nop] = this_alternative_matches;
2853
2854 if (this_alternative_matches >= 0
2855 && !did_match && !this_alternative_win)
2856 curr_alt_win[this_alternative_matches] = false;
2857
2858 if (early_clobber_p && operand_reg[nop] != NULL_RTX)
2859 early_clobbered_nops[early_clobbered_regs_num++] = nop;
2860 }
2861
2862 if (curr_insn_set != NULL_RTX && n_operands == 2
2863 /* Prevent processing non-move insns. */
2864 && (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
2865 || SET_SRC (curr_insn_set) == no_subreg_reg_operand[1])
2866 && ((! curr_alt_win[0] && ! curr_alt_win[1]
2867 && REG_P (no_subreg_reg_operand[0])
2868 && REG_P (no_subreg_reg_operand[1])
2869 && (reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2870 || reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0])))
2871 || (! curr_alt_win[0] && curr_alt_win[1]
2872 && REG_P (no_subreg_reg_operand[1])
2873 /* Check that we reload memory not the memory
2874 address. */
2875 && ! (curr_alt_offmemok[0]
2876 && MEM_P (no_subreg_reg_operand[0]))
2877 && reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0]))
2878 || (curr_alt_win[0] && ! curr_alt_win[1]
2879 && REG_P (no_subreg_reg_operand[0])
2880 /* Check that we reload memory not the memory
2881 address. */
2882 && ! (curr_alt_offmemok[1]
2883 && MEM_P (no_subreg_reg_operand[1]))
2884 && reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2885 && (! CONST_POOL_OK_P (curr_operand_mode[1],
2886 no_subreg_reg_operand[1])
2887 || (targetm.preferred_reload_class
2888 (no_subreg_reg_operand[1],
2889 (enum reg_class) curr_alt[1]) != NO_REGS))
2890 /* If it is a result of recent elimination in move
2891 insn we can transform it into an add still by
2892 using this alternative. */
2893 && GET_CODE (no_subreg_reg_operand[1]) != PLUS
2894 /* Likewise if the source has been replaced with an
2895 equivalent value. This only happens once -- the reload
2896 will use the equivalent value instead of the register it
2897 replaces -- so there should be no danger of cycling. */
2898 && !equiv_substition_p[1])))
2899 {
2900 /* We have a move insn and a new reload insn will be similar
2901 to the current insn. We should avoid such situation as
2902 it results in LRA cycling. */
2903 if (lra_dump_file != NULL)
2904 fprintf (lra_dump_file,
2905 " Cycle danger: overall += LRA_MAX_REJECT\n");
2906 overall += LRA_MAX_REJECT;
2907 }
2908 ok_p = true;
2909 curr_alt_dont_inherit_ops_num = 0;
2910 for (nop = 0; nop < early_clobbered_regs_num; nop++)
2911 {
2912 int i, j, clobbered_hard_regno, first_conflict_j, last_conflict_j;
2913 HARD_REG_SET temp_set;
2914
2915 i = early_clobbered_nops[nop];
2916 if ((! curr_alt_win[i] && ! curr_alt_match_win[i])
2917 || hard_regno[i] < 0)
2918 continue;
2919 lra_assert (operand_reg[i] != NULL_RTX);
2920 clobbered_hard_regno = hard_regno[i];
2921 CLEAR_HARD_REG_SET (temp_set);
2922 add_to_hard_reg_set (&temp_set, biggest_mode[i], clobbered_hard_regno);
2923 first_conflict_j = last_conflict_j = -1;
2924 for (j = 0; j < n_operands; j++)
2925 if (j == i
2926 /* We don't want process insides of match_operator and
2927 match_parallel because otherwise we would process
2928 their operands once again generating a wrong
2929 code. */
2930 || curr_static_id->operand[j].is_operator)
2931 continue;
2932 else if ((curr_alt_matches[j] == i && curr_alt_match_win[j])
2933 || (curr_alt_matches[i] == j && curr_alt_match_win[i]))
2934 continue;
2935 /* If we don't reload j-th operand, check conflicts. */
2936 else if ((curr_alt_win[j] || curr_alt_match_win[j])
2937 && uses_hard_regs_p (*curr_id->operand_loc[j], temp_set))
2938 {
2939 if (first_conflict_j < 0)
2940 first_conflict_j = j;
2941 last_conflict_j = j;
2942 /* Both the earlyclobber operand and conflicting operand
2943 cannot both be user defined hard registers. */
2944 if (HARD_REGISTER_P (operand_reg[i])
2945 && REG_USERVAR_P (operand_reg[i])
2946 && operand_reg[j] != NULL_RTX
2947 && HARD_REGISTER_P (operand_reg[j])
2948 && REG_USERVAR_P (operand_reg[j]))
2949 fatal_insn ("unable to generate reloads for "
2950 "impossible constraints:", curr_insn);
2951 }
2952 if (last_conflict_j < 0)
2953 continue;
2954
2955 /* If an earlyclobber operand conflicts with another non-matching
2956 operand (ie, they have been assigned the same hard register),
2957 then it is better to reload the other operand, as there may
2958 exist yet another operand with a matching constraint associated
2959 with the earlyclobber operand. However, if one of the operands
2960 is an explicit use of a hard register, then we must reload the
2961 other non-hard register operand. */
2962 if (HARD_REGISTER_P (operand_reg[i])
2963 || (first_conflict_j == last_conflict_j
2964 && operand_reg[last_conflict_j] != NULL_RTX
2965 && !curr_alt_match_win[last_conflict_j]
2966 && !HARD_REGISTER_P (operand_reg[last_conflict_j])))
2967 {
2968 curr_alt_win[last_conflict_j] = false;
2969 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++]
2970 = last_conflict_j;
2971 losers++;
2972 if (lra_dump_file != NULL)
2973 fprintf
2974 (lra_dump_file,
2975 " %d Conflict early clobber reload: reject--\n",
2976 i);
2977 }
2978 else
2979 {
2980 /* We need to reload early clobbered register and the
2981 matched registers. */
2982 for (j = 0; j < n_operands; j++)
2983 if (curr_alt_matches[j] == i)
2984 {
2985 curr_alt_match_win[j] = false;
2986 losers++;
2987 overall += LRA_LOSER_COST_FACTOR;
2988 }
2989 if (! curr_alt_match_win[i])
2990 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++] = i;
2991 else
2992 {
2993 /* Remember pseudos used for match reloads are never
2994 inherited. */
2995 lra_assert (curr_alt_matches[i] >= 0);
2996 curr_alt_win[curr_alt_matches[i]] = false;
2997 }
2998 curr_alt_win[i] = curr_alt_match_win[i] = false;
2999 losers++;
3000 if (lra_dump_file != NULL)
3001 fprintf
3002 (lra_dump_file,
3003 " %d Matched conflict early clobber reloads: "
3004 "reject--\n",
3005 i);
3006 }
3007 /* Early clobber was already reflected in REJECT. */
3008 if (!matching_early_clobber[i])
3009 {
3010 lra_assert (reject > 0);
3011 reject--;
3012 matching_early_clobber[i] = 1;
3013 }
3014 overall += LRA_LOSER_COST_FACTOR - 1;
3015 }
3016 if (lra_dump_file != NULL)
3017 fprintf (lra_dump_file, " alt=%d,overall=%d,losers=%d,rld_nregs=%d\n",
3018 nalt, overall, losers, reload_nregs);
3019
3020 /* If this alternative can be made to work by reloading, and it
3021 needs less reloading than the others checked so far, record
3022 it as the chosen goal for reloading. */
3023 if ((best_losers != 0 && losers == 0)
3024 || (((best_losers == 0 && losers == 0)
3025 || (best_losers != 0 && losers != 0))
3026 && (best_overall > overall
3027 || (best_overall == overall
3028 /* If the cost of the reloads is the same,
3029 prefer alternative which requires minimal
3030 number of reload regs. */
3031 && (reload_nregs < best_reload_nregs
3032 || (reload_nregs == best_reload_nregs
3033 && (best_reload_sum < reload_sum
3034 || (best_reload_sum == reload_sum
3035 && nalt < goal_alt_number))))))))
3036 {
3037 for (nop = 0; nop < n_operands; nop++)
3038 {
3039 goal_alt_win[nop] = curr_alt_win[nop];
3040 goal_alt_match_win[nop] = curr_alt_match_win[nop];
3041 goal_alt_matches[nop] = curr_alt_matches[nop];
3042 goal_alt[nop] = curr_alt[nop];
3043 goal_alt_offmemok[nop] = curr_alt_offmemok[nop];
3044 }
3045 goal_alt_dont_inherit_ops_num = curr_alt_dont_inherit_ops_num;
3046 for (nop = 0; nop < curr_alt_dont_inherit_ops_num; nop++)
3047 goal_alt_dont_inherit_ops[nop] = curr_alt_dont_inherit_ops[nop];
3048 goal_alt_swapped = curr_swapped;
3049 best_overall = overall;
3050 best_losers = losers;
3051 best_reload_nregs = reload_nregs;
3052 best_reload_sum = reload_sum;
3053 goal_alt_number = nalt;
3054 }
3055 if (losers == 0)
3056 /* Everything is satisfied. Do not process alternatives
3057 anymore. */
3058 break;
3059 fail:
3060 ;
3061 }
3062 return ok_p;
3063 }
3064
3065 /* Make reload base reg from address AD. */
3066 static rtx
3067 base_to_reg (struct address_info *ad)
3068 {
3069 enum reg_class cl;
3070 int code = -1;
3071 rtx new_inner = NULL_RTX;
3072 rtx new_reg = NULL_RTX;
3073 rtx_insn *insn;
3074 rtx_insn *last_insn = get_last_insn();
3075
3076 lra_assert (ad->disp == ad->disp_term);
3077 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
3078 get_index_code (ad));
3079 new_reg = lra_create_new_reg (GET_MODE (*ad->base), NULL_RTX,
3080 cl, "base");
3081 new_inner = simplify_gen_binary (PLUS, GET_MODE (new_reg), new_reg,
3082 ad->disp_term == NULL
3083 ? const0_rtx
3084 : *ad->disp_term);
3085 if (!valid_address_p (ad->mode, new_inner, ad->as))
3086 return NULL_RTX;
3087 insn = emit_insn (gen_rtx_SET (new_reg, *ad->base));
3088 code = recog_memoized (insn);
3089 if (code < 0)
3090 {
3091 delete_insns_since (last_insn);
3092 return NULL_RTX;
3093 }
3094
3095 return new_inner;
3096 }
3097
3098 /* Make reload base reg + DISP from address AD. Return the new pseudo. */
3099 static rtx
3100 base_plus_disp_to_reg (struct address_info *ad, rtx disp)
3101 {
3102 enum reg_class cl;
3103 rtx new_reg;
3104
3105 lra_assert (ad->base == ad->base_term);
3106 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
3107 get_index_code (ad));
3108 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
3109 cl, "base + disp");
3110 lra_emit_add (new_reg, *ad->base_term, disp);
3111 return new_reg;
3112 }
3113
3114 /* Make reload of index part of address AD. Return the new
3115 pseudo. */
3116 static rtx
3117 index_part_to_reg (struct address_info *ad)
3118 {
3119 rtx new_reg;
3120
3121 new_reg = lra_create_new_reg (GET_MODE (*ad->index), NULL_RTX,
3122 INDEX_REG_CLASS, "index term");
3123 expand_mult (GET_MODE (*ad->index), *ad->index_term,
3124 GEN_INT (get_index_scale (ad)), new_reg, 1);
3125 return new_reg;
3126 }
3127
3128 /* Return true if we can add a displacement to address AD, even if that
3129 makes the address invalid. The fix-up code requires any new address
3130 to be the sum of the BASE_TERM, INDEX and DISP_TERM fields. */
3131 static bool
3132 can_add_disp_p (struct address_info *ad)
3133 {
3134 return (!ad->autoinc_p
3135 && ad->segment == NULL
3136 && ad->base == ad->base_term
3137 && ad->disp == ad->disp_term);
3138 }
3139
3140 /* Make equiv substitution in address AD. Return true if a substitution
3141 was made. */
3142 static bool
3143 equiv_address_substitution (struct address_info *ad)
3144 {
3145 rtx base_reg, new_base_reg, index_reg, new_index_reg, *base_term, *index_term;
3146 poly_int64 disp;
3147 HOST_WIDE_INT scale;
3148 bool change_p;
3149
3150 base_term = strip_subreg (ad->base_term);
3151 if (base_term == NULL)
3152 base_reg = new_base_reg = NULL_RTX;
3153 else
3154 {
3155 base_reg = *base_term;
3156 new_base_reg = get_equiv_with_elimination (base_reg, curr_insn);
3157 }
3158 index_term = strip_subreg (ad->index_term);
3159 if (index_term == NULL)
3160 index_reg = new_index_reg = NULL_RTX;
3161 else
3162 {
3163 index_reg = *index_term;
3164 new_index_reg = get_equiv_with_elimination (index_reg, curr_insn);
3165 }
3166 if (base_reg == new_base_reg && index_reg == new_index_reg)
3167 return false;
3168 disp = 0;
3169 change_p = false;
3170 if (lra_dump_file != NULL)
3171 {
3172 fprintf (lra_dump_file, "Changing address in insn %d ",
3173 INSN_UID (curr_insn));
3174 dump_value_slim (lra_dump_file, *ad->outer, 1);
3175 }
3176 if (base_reg != new_base_reg)
3177 {
3178 poly_int64 offset;
3179 if (REG_P (new_base_reg))
3180 {
3181 *base_term = new_base_reg;
3182 change_p = true;
3183 }
3184 else if (GET_CODE (new_base_reg) == PLUS
3185 && REG_P (XEXP (new_base_reg, 0))
3186 && poly_int_rtx_p (XEXP (new_base_reg, 1), &offset)
3187 && can_add_disp_p (ad))
3188 {
3189 disp += offset;
3190 *base_term = XEXP (new_base_reg, 0);
3191 change_p = true;
3192 }
3193 if (ad->base_term2 != NULL)
3194 *ad->base_term2 = *ad->base_term;
3195 }
3196 if (index_reg != new_index_reg)
3197 {
3198 poly_int64 offset;
3199 if (REG_P (new_index_reg))
3200 {
3201 *index_term = new_index_reg;
3202 change_p = true;
3203 }
3204 else if (GET_CODE (new_index_reg) == PLUS
3205 && REG_P (XEXP (new_index_reg, 0))
3206 && poly_int_rtx_p (XEXP (new_index_reg, 1), &offset)
3207 && can_add_disp_p (ad)
3208 && (scale = get_index_scale (ad)))
3209 {
3210 disp += offset * scale;
3211 *index_term = XEXP (new_index_reg, 0);
3212 change_p = true;
3213 }
3214 }
3215 if (maybe_ne (disp, 0))
3216 {
3217 if (ad->disp != NULL)
3218 *ad->disp = plus_constant (GET_MODE (*ad->inner), *ad->disp, disp);
3219 else
3220 {
3221 *ad->inner = plus_constant (GET_MODE (*ad->inner), *ad->inner, disp);
3222 update_address (ad);
3223 }
3224 change_p = true;
3225 }
3226 if (lra_dump_file != NULL)
3227 {
3228 if (! change_p)
3229 fprintf (lra_dump_file, " -- no change\n");
3230 else
3231 {
3232 fprintf (lra_dump_file, " on equiv ");
3233 dump_value_slim (lra_dump_file, *ad->outer, 1);
3234 fprintf (lra_dump_file, "\n");
3235 }
3236 }
3237 return change_p;
3238 }
3239
3240 /* Major function to make reloads for an address in operand NOP or
3241 check its correctness (If CHECK_ONLY_P is true). The supported
3242 cases are:
3243
3244 1) an address that existed before LRA started, at which point it
3245 must have been valid. These addresses are subject to elimination
3246 and may have become invalid due to the elimination offset being out
3247 of range.
3248
3249 2) an address created by forcing a constant to memory
3250 (force_const_to_mem). The initial form of these addresses might
3251 not be valid, and it is this function's job to make them valid.
3252
3253 3) a frame address formed from a register and a (possibly zero)
3254 constant offset. As above, these addresses might not be valid and
3255 this function must make them so.
3256
3257 Add reloads to the lists *BEFORE and *AFTER. We might need to add
3258 reloads to *AFTER because of inc/dec, {pre, post} modify in the
3259 address. Return true for any RTL change.
3260
3261 The function is a helper function which does not produce all
3262 transformations (when CHECK_ONLY_P is false) which can be
3263 necessary. It does just basic steps. To do all necessary
3264 transformations use function process_address. */
3265 static bool
3266 process_address_1 (int nop, bool check_only_p,
3267 rtx_insn **before, rtx_insn **after)
3268 {
3269 struct address_info ad;
3270 rtx new_reg;
3271 HOST_WIDE_INT scale;
3272 rtx op = *curr_id->operand_loc[nop];
3273 const char *constraint = curr_static_id->operand[nop].constraint;
3274 enum constraint_num cn = lookup_constraint (constraint);
3275 bool change_p = false;
3276
3277 if (MEM_P (op)
3278 && GET_MODE (op) == BLKmode
3279 && GET_CODE (XEXP (op, 0)) == SCRATCH)
3280 return false;
3281
3282 if (insn_extra_address_constraint (cn)
3283 /* When we find an asm operand with an address constraint that
3284 doesn't satisfy address_operand to begin with, we clear
3285 is_address, so that we don't try to make a non-address fit.
3286 If the asm statement got this far, it's because other
3287 constraints are available, and we'll use them, disregarding
3288 the unsatisfiable address ones. */
3289 && curr_static_id->operand[nop].is_address)
3290 decompose_lea_address (&ad, curr_id->operand_loc[nop]);
3291 /* Do not attempt to decompose arbitrary addresses generated by combine
3292 for asm operands with loose constraints, e.g 'X'. */
3293 else if (MEM_P (op)
3294 && !(INSN_CODE (curr_insn) < 0
3295 && get_constraint_type (cn) == CT_FIXED_FORM
3296 && constraint_satisfied_p (op, cn)))
3297 decompose_mem_address (&ad, op);
3298 else if (GET_CODE (op) == SUBREG
3299 && MEM_P (SUBREG_REG (op)))
3300 decompose_mem_address (&ad, SUBREG_REG (op));
3301 else
3302 return false;
3303 /* If INDEX_REG_CLASS is assigned to base_term already and isn't to
3304 index_term, swap them so to avoid assigning INDEX_REG_CLASS to both
3305 when INDEX_REG_CLASS is a single register class. */
3306 if (ad.base_term != NULL
3307 && ad.index_term != NULL
3308 && ira_class_hard_regs_num[INDEX_REG_CLASS] == 1
3309 && REG_P (*ad.base_term)
3310 && REG_P (*ad.index_term)
3311 && in_class_p (*ad.base_term, INDEX_REG_CLASS, NULL)
3312 && ! in_class_p (*ad.index_term, INDEX_REG_CLASS, NULL))
3313 {
3314 std::swap (ad.base, ad.index);
3315 std::swap (ad.base_term, ad.index_term);
3316 }
3317 if (! check_only_p)
3318 change_p = equiv_address_substitution (&ad);
3319 if (ad.base_term != NULL
3320 && (process_addr_reg
3321 (ad.base_term, check_only_p, before,
3322 (ad.autoinc_p
3323 && !(REG_P (*ad.base_term)
3324 && find_regno_note (curr_insn, REG_DEAD,
3325 REGNO (*ad.base_term)) != NULL_RTX)
3326 ? after : NULL),
3327 base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3328 get_index_code (&ad)))))
3329 {
3330 change_p = true;
3331 if (ad.base_term2 != NULL)
3332 *ad.base_term2 = *ad.base_term;
3333 }
3334 if (ad.index_term != NULL
3335 && process_addr_reg (ad.index_term, check_only_p,
3336 before, NULL, INDEX_REG_CLASS))
3337 change_p = true;
3338
3339 /* Target hooks sometimes don't treat extra-constraint addresses as
3340 legitimate address_operands, so handle them specially. */
3341 if (insn_extra_address_constraint (cn)
3342 && satisfies_address_constraint_p (&ad, cn))
3343 return change_p;
3344
3345 if (check_only_p)
3346 return change_p;
3347
3348 /* There are three cases where the shape of *AD.INNER may now be invalid:
3349
3350 1) the original address was valid, but either elimination or
3351 equiv_address_substitution was applied and that made
3352 the address invalid.
3353
3354 2) the address is an invalid symbolic address created by
3355 force_const_to_mem.
3356
3357 3) the address is a frame address with an invalid offset.
3358
3359 4) the address is a frame address with an invalid base.
3360
3361 All these cases involve a non-autoinc address, so there is no
3362 point revalidating other types. */
3363 if (ad.autoinc_p || valid_address_p (&ad))
3364 return change_p;
3365
3366 /* Any index existed before LRA started, so we can assume that the
3367 presence and shape of the index is valid. */
3368 push_to_sequence (*before);
3369 lra_assert (ad.disp == ad.disp_term);
3370 if (ad.base == NULL)
3371 {
3372 if (ad.index == NULL)
3373 {
3374 rtx_insn *insn;
3375 rtx_insn *last = get_last_insn ();
3376 int code = -1;
3377 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3378 SCRATCH, SCRATCH);
3379 rtx addr = *ad.inner;
3380
3381 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3382 if (HAVE_lo_sum)
3383 {
3384 /* addr => lo_sum (new_base, addr), case (2) above. */
3385 insn = emit_insn (gen_rtx_SET
3386 (new_reg,
3387 gen_rtx_HIGH (Pmode, copy_rtx (addr))));
3388 code = recog_memoized (insn);
3389 if (code >= 0)
3390 {
3391 *ad.inner = gen_rtx_LO_SUM (Pmode, new_reg, addr);
3392 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3393 {
3394 /* Try to put lo_sum into register. */
3395 insn = emit_insn (gen_rtx_SET
3396 (new_reg,
3397 gen_rtx_LO_SUM (Pmode, new_reg, addr)));
3398 code = recog_memoized (insn);
3399 if (code >= 0)
3400 {
3401 *ad.inner = new_reg;
3402 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
3403 {
3404 *ad.inner = addr;
3405 code = -1;
3406 }
3407 }
3408
3409 }
3410 }
3411 if (code < 0)
3412 delete_insns_since (last);
3413 }
3414
3415 if (code < 0)
3416 {
3417 /* addr => new_base, case (2) above. */
3418 lra_emit_move (new_reg, addr);
3419
3420 for (insn = last == NULL_RTX ? get_insns () : NEXT_INSN (last);
3421 insn != NULL_RTX;
3422 insn = NEXT_INSN (insn))
3423 if (recog_memoized (insn) < 0)
3424 break;
3425 if (insn != NULL_RTX)
3426 {
3427 /* Do nothing if we cannot generate right insns.
3428 This is analogous to reload pass behavior. */
3429 delete_insns_since (last);
3430 end_sequence ();
3431 return false;
3432 }
3433 *ad.inner = new_reg;
3434 }
3435 }
3436 else
3437 {
3438 /* index * scale + disp => new base + index * scale,
3439 case (1) above. */
3440 enum reg_class cl = base_reg_class (ad.mode, ad.as, PLUS,
3441 GET_CODE (*ad.index));
3442
3443 lra_assert (INDEX_REG_CLASS != NO_REGS);
3444 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp");
3445 lra_emit_move (new_reg, *ad.disp);
3446 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3447 new_reg, *ad.index);
3448 }
3449 }
3450 else if (ad.index == NULL)
3451 {
3452 int regno;
3453 enum reg_class cl;
3454 rtx set;
3455 rtx_insn *insns, *last_insn;
3456 /* Try to reload base into register only if the base is invalid
3457 for the address but with valid offset, case (4) above. */
3458 start_sequence ();
3459 new_reg = base_to_reg (&ad);
3460
3461 /* base + disp => new base, cases (1) and (3) above. */
3462 /* Another option would be to reload the displacement into an
3463 index register. However, postreload has code to optimize
3464 address reloads that have the same base and different
3465 displacements, so reloading into an index register would
3466 not necessarily be a win. */
3467 if (new_reg == NULL_RTX)
3468 {
3469 /* See if the target can split the displacement into a
3470 legitimate new displacement from a local anchor. */
3471 gcc_assert (ad.disp == ad.disp_term);
3472 poly_int64 orig_offset;
3473 rtx offset1, offset2;
3474 if (poly_int_rtx_p (*ad.disp, &orig_offset)
3475 && targetm.legitimize_address_displacement (&offset1, &offset2,
3476 orig_offset,
3477 ad.mode))
3478 {
3479 new_reg = base_plus_disp_to_reg (&ad, offset1);
3480 new_reg = gen_rtx_PLUS (GET_MODE (new_reg), new_reg, offset2);
3481 }
3482 else
3483 new_reg = base_plus_disp_to_reg (&ad, *ad.disp);
3484 }
3485 insns = get_insns ();
3486 last_insn = get_last_insn ();
3487 /* If we generated at least two insns, try last insn source as
3488 an address. If we succeed, we generate one less insn. */
3489 if (REG_P (new_reg)
3490 && last_insn != insns
3491 && (set = single_set (last_insn)) != NULL_RTX
3492 && GET_CODE (SET_SRC (set)) == PLUS
3493 && REG_P (XEXP (SET_SRC (set), 0))
3494 && CONSTANT_P (XEXP (SET_SRC (set), 1)))
3495 {
3496 *ad.inner = SET_SRC (set);
3497 if (valid_address_p (ad.mode, *ad.outer, ad.as))
3498 {
3499 *ad.base_term = XEXP (SET_SRC (set), 0);
3500 *ad.disp_term = XEXP (SET_SRC (set), 1);
3501 cl = base_reg_class (ad.mode, ad.as, ad.base_outer_code,
3502 get_index_code (&ad));
3503 regno = REGNO (*ad.base_term);
3504 if (regno >= FIRST_PSEUDO_REGISTER
3505 && cl != lra_get_allocno_class (regno))
3506 lra_change_class (regno, cl, " Change to", true);
3507 new_reg = SET_SRC (set);
3508 delete_insns_since (PREV_INSN (last_insn));
3509 }
3510 }
3511 end_sequence ();
3512 emit_insn (insns);
3513 *ad.inner = new_reg;
3514 }
3515 else if (ad.disp_term != NULL)
3516 {
3517 /* base + scale * index + disp => new base + scale * index,
3518 case (1) above. */
3519 gcc_assert (ad.disp == ad.disp_term);
3520 new_reg = base_plus_disp_to_reg (&ad, *ad.disp);
3521 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3522 new_reg, *ad.index);
3523 }
3524 else if ((scale = get_index_scale (&ad)) == 1)
3525 {
3526 /* The last transformation to one reg will be made in
3527 curr_insn_transform function. */
3528 end_sequence ();
3529 return false;
3530 }
3531 else if (scale != 0)
3532 {
3533 /* base + scale * index => base + new_reg,
3534 case (1) above.
3535 Index part of address may become invalid. For example, we
3536 changed pseudo on the equivalent memory and a subreg of the
3537 pseudo onto the memory of different mode for which the scale is
3538 prohibitted. */
3539 new_reg = index_part_to_reg (&ad);
3540 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3541 *ad.base_term, new_reg);
3542 }
3543 else
3544 {
3545 enum reg_class cl = base_reg_class (ad.mode, ad.as,
3546 SCRATCH, SCRATCH);
3547 rtx addr = *ad.inner;
3548
3549 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
3550 /* addr => new_base. */
3551 lra_emit_move (new_reg, addr);
3552 *ad.inner = new_reg;
3553 }
3554 *before = get_insns ();
3555 end_sequence ();
3556 return true;
3557 }
3558
3559 /* If CHECK_ONLY_P is false, do address reloads until it is necessary.
3560 Use process_address_1 as a helper function. Return true for any
3561 RTL changes.
3562
3563 If CHECK_ONLY_P is true, just check address correctness. Return
3564 false if the address correct. */
3565 static bool
3566 process_address (int nop, bool check_only_p,
3567 rtx_insn **before, rtx_insn **after)
3568 {
3569 bool res = false;
3570
3571 while (process_address_1 (nop, check_only_p, before, after))
3572 {
3573 if (check_only_p)
3574 return true;
3575 res = true;
3576 }
3577 return res;
3578 }
3579
3580 /* Emit insns to reload VALUE into a new register. VALUE is an
3581 auto-increment or auto-decrement RTX whose operand is a register or
3582 memory location; so reloading involves incrementing that location.
3583 IN is either identical to VALUE, or some cheaper place to reload
3584 value being incremented/decremented from.
3585
3586 INC_AMOUNT is the number to increment or decrement by (always
3587 positive and ignored for POST_MODIFY/PRE_MODIFY).
3588
3589 Return pseudo containing the result. */
3590 static rtx
3591 emit_inc (enum reg_class new_rclass, rtx in, rtx value, poly_int64 inc_amount)
3592 {
3593 /* REG or MEM to be copied and incremented. */
3594 rtx incloc = XEXP (value, 0);
3595 /* Nonzero if increment after copying. */
3596 int post = (GET_CODE (value) == POST_DEC || GET_CODE (value) == POST_INC
3597 || GET_CODE (value) == POST_MODIFY);
3598 rtx_insn *last;
3599 rtx inc;
3600 rtx_insn *add_insn;
3601 int code;
3602 rtx real_in = in == value ? incloc : in;
3603 rtx result;
3604 bool plus_p = true;
3605
3606 if (GET_CODE (value) == PRE_MODIFY || GET_CODE (value) == POST_MODIFY)
3607 {
3608 lra_assert (GET_CODE (XEXP (value, 1)) == PLUS
3609 || GET_CODE (XEXP (value, 1)) == MINUS);
3610 lra_assert (rtx_equal_p (XEXP (XEXP (value, 1), 0), XEXP (value, 0)));
3611 plus_p = GET_CODE (XEXP (value, 1)) == PLUS;
3612 inc = XEXP (XEXP (value, 1), 1);
3613 }
3614 else
3615 {
3616 if (GET_CODE (value) == PRE_DEC || GET_CODE (value) == POST_DEC)
3617 inc_amount = -inc_amount;
3618
3619 inc = gen_int_mode (inc_amount, GET_MODE (value));
3620 }
3621
3622 if (! post && REG_P (incloc))
3623 result = incloc;
3624 else
3625 result = lra_create_new_reg (GET_MODE (value), value, new_rclass,
3626 "INC/DEC result");
3627
3628 if (real_in != result)
3629 {
3630 /* First copy the location to the result register. */
3631 lra_assert (REG_P (result));
3632 emit_insn (gen_move_insn (result, real_in));
3633 }
3634
3635 /* We suppose that there are insns to add/sub with the constant
3636 increment permitted in {PRE/POST)_{DEC/INC/MODIFY}. At least the
3637 old reload worked with this assumption. If the assumption
3638 becomes wrong, we should use approach in function
3639 base_plus_disp_to_reg. */
3640 if (in == value)
3641 {
3642 /* See if we can directly increment INCLOC. */
3643 last = get_last_insn ();
3644 add_insn = emit_insn (plus_p
3645 ? gen_add2_insn (incloc, inc)
3646 : gen_sub2_insn (incloc, inc));
3647
3648 code = recog_memoized (add_insn);
3649 if (code >= 0)
3650 {
3651 if (! post && result != incloc)
3652 emit_insn (gen_move_insn (result, incloc));
3653 return result;
3654 }
3655 delete_insns_since (last);
3656 }
3657
3658 /* If couldn't do the increment directly, must increment in RESULT.
3659 The way we do this depends on whether this is pre- or
3660 post-increment. For pre-increment, copy INCLOC to the reload
3661 register, increment it there, then save back. */
3662 if (! post)
3663 {
3664 if (real_in != result)
3665 emit_insn (gen_move_insn (result, real_in));
3666 if (plus_p)
3667 emit_insn (gen_add2_insn (result, inc));
3668 else
3669 emit_insn (gen_sub2_insn (result, inc));
3670 if (result != incloc)
3671 emit_insn (gen_move_insn (incloc, result));
3672 }
3673 else
3674 {
3675 /* Post-increment.
3676
3677 Because this might be a jump insn or a compare, and because
3678 RESULT may not be available after the insn in an input
3679 reload, we must do the incrementing before the insn being
3680 reloaded for.
3681
3682 We have already copied IN to RESULT. Increment the copy in
3683 RESULT, save that back, then decrement RESULT so it has
3684 the original value. */
3685 if (plus_p)
3686 emit_insn (gen_add2_insn (result, inc));
3687 else
3688 emit_insn (gen_sub2_insn (result, inc));
3689 emit_insn (gen_move_insn (incloc, result));
3690 /* Restore non-modified value for the result. We prefer this
3691 way because it does not require an additional hard
3692 register. */
3693 if (plus_p)
3694 {
3695 poly_int64 offset;
3696 if (poly_int_rtx_p (inc, &offset))
3697 emit_insn (gen_add2_insn (result,
3698 gen_int_mode (-offset,
3699 GET_MODE (result))));
3700 else
3701 emit_insn (gen_sub2_insn (result, inc));
3702 }
3703 else
3704 emit_insn (gen_add2_insn (result, inc));
3705 }
3706 return result;
3707 }
3708
3709 /* Return true if the current move insn does not need processing as we
3710 already know that it satisfies its constraints. */
3711 static bool
3712 simple_move_p (void)
3713 {
3714 rtx dest, src;
3715 enum reg_class dclass, sclass;
3716
3717 lra_assert (curr_insn_set != NULL_RTX);
3718 dest = SET_DEST (curr_insn_set);
3719 src = SET_SRC (curr_insn_set);
3720
3721 /* If the instruction has multiple sets we need to process it even if it
3722 is single_set. This can happen if one or more of the SETs are dead.
3723 See PR73650. */
3724 if (multiple_sets (curr_insn))
3725 return false;
3726
3727 return ((dclass = get_op_class (dest)) != NO_REGS
3728 && (sclass = get_op_class (src)) != NO_REGS
3729 /* The backend guarantees that register moves of cost 2
3730 never need reloads. */
3731 && targetm.register_move_cost (GET_MODE (src), sclass, dclass) == 2);
3732 }
3733
3734 /* Swap operands NOP and NOP + 1. */
3735 static inline void
3736 swap_operands (int nop)
3737 {
3738 std::swap (curr_operand_mode[nop], curr_operand_mode[nop + 1]);
3739 std::swap (original_subreg_reg_mode[nop], original_subreg_reg_mode[nop + 1]);
3740 std::swap (*curr_id->operand_loc[nop], *curr_id->operand_loc[nop + 1]);
3741 std::swap (equiv_substition_p[nop], equiv_substition_p[nop + 1]);
3742 /* Swap the duplicates too. */
3743 lra_update_dup (curr_id, nop);
3744 lra_update_dup (curr_id, nop + 1);
3745 }
3746
3747 /* Main entry point of the constraint code: search the body of the
3748 current insn to choose the best alternative. It is mimicking insn
3749 alternative cost calculation model of former reload pass. That is
3750 because machine descriptions were written to use this model. This
3751 model can be changed in future. Make commutative operand exchange
3752 if it is chosen.
3753
3754 if CHECK_ONLY_P is false, do RTL changes to satisfy the
3755 constraints. Return true if any change happened during function
3756 call.
3757
3758 If CHECK_ONLY_P is true then don't do any transformation. Just
3759 check that the insn satisfies all constraints. If the insn does
3760 not satisfy any constraint, return true. */
3761 static bool
3762 curr_insn_transform (bool check_only_p)
3763 {
3764 int i, j, k;
3765 int n_operands;
3766 int n_alternatives;
3767 int n_outputs;
3768 int commutative;
3769 signed char goal_alt_matched[MAX_RECOG_OPERANDS][MAX_RECOG_OPERANDS];
3770 signed char match_inputs[MAX_RECOG_OPERANDS + 1];
3771 signed char outputs[MAX_RECOG_OPERANDS + 1];
3772 rtx_insn *before, *after;
3773 bool alt_p = false;
3774 /* Flag that the insn has been changed through a transformation. */
3775 bool change_p;
3776 bool sec_mem_p;
3777 bool use_sec_mem_p;
3778 int max_regno_before;
3779 int reused_alternative_num;
3780
3781 curr_insn_set = single_set (curr_insn);
3782 if (curr_insn_set != NULL_RTX && simple_move_p ())
3783 {
3784 /* We assume that the corresponding insn alternative has no
3785 earlier clobbers. If it is not the case, don't define move
3786 cost equal to 2 for the corresponding register classes. */
3787 lra_set_used_insn_alternative (curr_insn, LRA_NON_CLOBBERED_ALT);
3788 return false;
3789 }
3790
3791 no_input_reloads_p = no_output_reloads_p = false;
3792 goal_alt_number = -1;
3793 change_p = sec_mem_p = false;
3794 /* JUMP_INSNs and CALL_INSNs are not allowed to have any output
3795 reloads; neither are insns that SET cc0. Insns that use CC0 are
3796 not allowed to have any input reloads. */
3797 if (JUMP_P (curr_insn) || CALL_P (curr_insn))
3798 no_output_reloads_p = true;
3799
3800 if (HAVE_cc0 && reg_referenced_p (cc0_rtx, PATTERN (curr_insn)))
3801 no_input_reloads_p = true;
3802 if (HAVE_cc0 && reg_set_p (cc0_rtx, PATTERN (curr_insn)))
3803 no_output_reloads_p = true;
3804
3805 n_operands = curr_static_id->n_operands;
3806 n_alternatives = curr_static_id->n_alternatives;
3807
3808 /* Just return "no reloads" if insn has no operands with
3809 constraints. */
3810 if (n_operands == 0 || n_alternatives == 0)
3811 return false;
3812
3813 max_regno_before = max_reg_num ();
3814
3815 for (i = 0; i < n_operands; i++)
3816 {
3817 goal_alt_matched[i][0] = -1;
3818 goal_alt_matches[i] = -1;
3819 }
3820
3821 commutative = curr_static_id->commutative;
3822
3823 /* Now see what we need for pseudos that didn't get hard regs or got
3824 the wrong kind of hard reg. For this, we must consider all the
3825 operands together against the register constraints. */
3826
3827 best_losers = best_overall = INT_MAX;
3828 best_reload_sum = 0;
3829
3830 curr_swapped = false;
3831 goal_alt_swapped = false;
3832
3833 if (! check_only_p)
3834 /* Make equivalence substitution and memory subreg elimination
3835 before address processing because an address legitimacy can
3836 depend on memory mode. */
3837 for (i = 0; i < n_operands; i++)
3838 {
3839 rtx op, subst, old;
3840 bool op_change_p = false;
3841
3842 if (curr_static_id->operand[i].is_operator)
3843 continue;
3844
3845 old = op = *curr_id->operand_loc[i];
3846 if (GET_CODE (old) == SUBREG)
3847 old = SUBREG_REG (old);
3848 subst = get_equiv_with_elimination (old, curr_insn);
3849 original_subreg_reg_mode[i] = VOIDmode;
3850 equiv_substition_p[i] = false;
3851 if (subst != old)
3852 {
3853 equiv_substition_p[i] = true;
3854 subst = copy_rtx (subst);
3855 lra_assert (REG_P (old));
3856 if (GET_CODE (op) != SUBREG)
3857 *curr_id->operand_loc[i] = subst;
3858 else
3859 {
3860 SUBREG_REG (op) = subst;
3861 if (GET_MODE (subst) == VOIDmode)
3862 original_subreg_reg_mode[i] = GET_MODE (old);
3863 }
3864 if (lra_dump_file != NULL)
3865 {
3866 fprintf (lra_dump_file,
3867 "Changing pseudo %d in operand %i of insn %u on equiv ",
3868 REGNO (old), i, INSN_UID (curr_insn));
3869 dump_value_slim (lra_dump_file, subst, 1);
3870 fprintf (lra_dump_file, "\n");
3871 }
3872 op_change_p = change_p = true;
3873 }
3874 if (simplify_operand_subreg (i, GET_MODE (old)) || op_change_p)
3875 {
3876 change_p = true;
3877 lra_update_dup (curr_id, i);
3878 }
3879 }
3880
3881 /* Reload address registers and displacements. We do it before
3882 finding an alternative because of memory constraints. */
3883 before = after = NULL;
3884 for (i = 0; i < n_operands; i++)
3885 if (! curr_static_id->operand[i].is_operator
3886 && process_address (i, check_only_p, &before, &after))
3887 {
3888 if (check_only_p)
3889 return true;
3890 change_p = true;
3891 lra_update_dup (curr_id, i);
3892 }
3893
3894 if (change_p)
3895 /* If we've changed the instruction then any alternative that
3896 we chose previously may no longer be valid. */
3897 lra_set_used_insn_alternative (curr_insn, LRA_UNKNOWN_ALT);
3898
3899 if (! check_only_p && curr_insn_set != NULL_RTX
3900 && check_and_process_move (&change_p, &sec_mem_p))
3901 return change_p;
3902
3903 try_swapped:
3904
3905 reused_alternative_num = check_only_p ? LRA_UNKNOWN_ALT : curr_id->used_insn_alternative;
3906 if (lra_dump_file != NULL && reused_alternative_num >= 0)
3907 fprintf (lra_dump_file, "Reusing alternative %d for insn #%u\n",
3908 reused_alternative_num, INSN_UID (curr_insn));
3909
3910 if (process_alt_operands (reused_alternative_num))
3911 alt_p = true;
3912
3913 if (check_only_p)
3914 return ! alt_p || best_losers != 0;
3915
3916 /* If insn is commutative (it's safe to exchange a certain pair of
3917 operands) then we need to try each alternative twice, the second
3918 time matching those two operands as if we had exchanged them. To
3919 do this, really exchange them in operands.
3920
3921 If we have just tried the alternatives the second time, return
3922 operands to normal and drop through. */
3923
3924 if (reused_alternative_num < 0 && commutative >= 0)
3925 {
3926 curr_swapped = !curr_swapped;
3927 if (curr_swapped)
3928 {
3929 swap_operands (commutative);
3930 goto try_swapped;
3931 }
3932 else
3933 swap_operands (commutative);
3934 }
3935
3936 if (! alt_p && ! sec_mem_p)
3937 {
3938 /* No alternative works with reloads?? */
3939 if (INSN_CODE (curr_insn) >= 0)
3940 fatal_insn ("unable to generate reloads for:", curr_insn);
3941 error_for_asm (curr_insn,
3942 "inconsistent operand constraints in an %<asm%>");
3943 lra_asm_error_p = true;
3944 /* Avoid further trouble with this insn. Don't generate use
3945 pattern here as we could use the insn SP offset. */
3946 lra_set_insn_deleted (curr_insn);
3947 return true;
3948 }
3949
3950 /* If the best alternative is with operands 1 and 2 swapped, swap
3951 them. Update the operand numbers of any reloads already
3952 pushed. */
3953
3954 if (goal_alt_swapped)
3955 {
3956 if (lra_dump_file != NULL)
3957 fprintf (lra_dump_file, " Commutative operand exchange in insn %u\n",
3958 INSN_UID (curr_insn));
3959
3960 /* Swap the duplicates too. */
3961 swap_operands (commutative);
3962 change_p = true;
3963 }
3964
3965 /* Some targets' TARGET_SECONDARY_MEMORY_NEEDED (e.g. x86) are defined
3966 too conservatively. So we use the secondary memory only if there
3967 is no any alternative without reloads. */
3968 use_sec_mem_p = false;
3969 if (! alt_p)
3970 use_sec_mem_p = true;
3971 else if (sec_mem_p)
3972 {
3973 for (i = 0; i < n_operands; i++)
3974 if (! goal_alt_win[i] && ! goal_alt_match_win[i])
3975 break;
3976 use_sec_mem_p = i < n_operands;
3977 }
3978
3979 if (use_sec_mem_p)
3980 {
3981 int in = -1, out = -1;
3982 rtx new_reg, src, dest, rld;
3983 machine_mode sec_mode, rld_mode;
3984
3985 lra_assert (curr_insn_set != NULL_RTX && sec_mem_p);
3986 dest = SET_DEST (curr_insn_set);
3987 src = SET_SRC (curr_insn_set);
3988 for (i = 0; i < n_operands; i++)
3989 if (*curr_id->operand_loc[i] == dest)
3990 out = i;
3991 else if (*curr_id->operand_loc[i] == src)
3992 in = i;
3993 for (i = 0; i < curr_static_id->n_dups; i++)
3994 if (out < 0 && *curr_id->dup_loc[i] == dest)
3995 out = curr_static_id->dup_num[i];
3996 else if (in < 0 && *curr_id->dup_loc[i] == src)
3997 in = curr_static_id->dup_num[i];
3998 lra_assert (out >= 0 && in >= 0
3999 && curr_static_id->operand[out].type == OP_OUT
4000 && curr_static_id->operand[in].type == OP_IN);
4001 rld = partial_subreg_p (GET_MODE (src), GET_MODE (dest)) ? src : dest;
4002 rld_mode = GET_MODE (rld);
4003 sec_mode = targetm.secondary_memory_needed_mode (rld_mode);
4004 new_reg = lra_create_new_reg (sec_mode, NULL_RTX,
4005 NO_REGS, "secondary");
4006 /* If the mode is changed, it should be wider. */
4007 lra_assert (!partial_subreg_p (sec_mode, rld_mode));
4008 if (sec_mode != rld_mode)
4009 {
4010 /* If the target says specifically to use another mode for
4011 secondary memory moves we cannot reuse the original
4012 insn. */
4013 after = emit_spill_move (false, new_reg, dest);
4014 lra_process_new_insns (curr_insn, NULL, after,
4015 "Inserting the sec. move");
4016 /* We may have non null BEFORE here (e.g. after address
4017 processing. */
4018 push_to_sequence (before);
4019 before = emit_spill_move (true, new_reg, src);
4020 emit_insn (before);
4021 before = get_insns ();
4022 end_sequence ();
4023 lra_process_new_insns (curr_insn, before, NULL, "Changing on");
4024 lra_set_insn_deleted (curr_insn);
4025 }
4026 else if (dest == rld)
4027 {
4028 *curr_id->operand_loc[out] = new_reg;
4029 lra_update_dup (curr_id, out);
4030 after = emit_spill_move (false, new_reg, dest);
4031 lra_process_new_insns (curr_insn, NULL, after,
4032 "Inserting the sec. move");
4033 }
4034 else
4035 {
4036 *curr_id->operand_loc[in] = new_reg;
4037 lra_update_dup (curr_id, in);
4038 /* See comments above. */
4039 push_to_sequence (before);
4040 before = emit_spill_move (true, new_reg, src);
4041 emit_insn (before);
4042 before = get_insns ();
4043 end_sequence ();
4044 lra_process_new_insns (curr_insn, before, NULL,
4045 "Inserting the sec. move");
4046 }
4047 lra_update_insn_regno_info (curr_insn);
4048 return true;
4049 }
4050
4051 lra_assert (goal_alt_number >= 0);
4052 lra_set_used_insn_alternative (curr_insn, goal_alt_number);
4053
4054 if (lra_dump_file != NULL)
4055 {
4056 const char *p;
4057
4058 fprintf (lra_dump_file, " Choosing alt %d in insn %u:",
4059 goal_alt_number, INSN_UID (curr_insn));
4060 for (i = 0; i < n_operands; i++)
4061 {
4062 p = (curr_static_id->operand_alternative
4063 [goal_alt_number * n_operands + i].constraint);
4064 if (*p == '\0')
4065 continue;
4066 fprintf (lra_dump_file, " (%d) ", i);
4067 for (; *p != '\0' && *p != ',' && *p != '#'; p++)
4068 fputc (*p, lra_dump_file);
4069 }
4070 if (INSN_CODE (curr_insn) >= 0
4071 && (p = get_insn_name (INSN_CODE (curr_insn))) != NULL)
4072 fprintf (lra_dump_file, " {%s}", p);
4073 if (maybe_ne (curr_id->sp_offset, 0))
4074 {
4075 fprintf (lra_dump_file, " (sp_off=");
4076 print_dec (curr_id->sp_offset, lra_dump_file);
4077 fprintf (lra_dump_file, ")");
4078 }
4079 fprintf (lra_dump_file, "\n");
4080 }
4081
4082 /* Right now, for any pair of operands I and J that are required to
4083 match, with J < I, goal_alt_matches[I] is J. Add I to
4084 goal_alt_matched[J]. */
4085
4086 for (i = 0; i < n_operands; i++)
4087 if ((j = goal_alt_matches[i]) >= 0)
4088 {
4089 for (k = 0; goal_alt_matched[j][k] >= 0; k++)
4090 ;
4091 /* We allow matching one output operand and several input
4092 operands. */
4093 lra_assert (k == 0
4094 || (curr_static_id->operand[j].type == OP_OUT
4095 && curr_static_id->operand[i].type == OP_IN
4096 && (curr_static_id->operand
4097 [goal_alt_matched[j][0]].type == OP_IN)));
4098 goal_alt_matched[j][k] = i;
4099 goal_alt_matched[j][k + 1] = -1;
4100 }
4101
4102 for (i = 0; i < n_operands; i++)
4103 goal_alt_win[i] |= goal_alt_match_win[i];
4104
4105 /* Any constants that aren't allowed and can't be reloaded into
4106 registers are here changed into memory references. */
4107 for (i = 0; i < n_operands; i++)
4108 if (goal_alt_win[i])
4109 {
4110 int regno;
4111 enum reg_class new_class;
4112 rtx reg = *curr_id->operand_loc[i];
4113
4114 if (GET_CODE (reg) == SUBREG)
4115 reg = SUBREG_REG (reg);
4116
4117 if (REG_P (reg) && (regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER)
4118 {
4119 bool ok_p = in_class_p (reg, goal_alt[i], &new_class);
4120
4121 if (new_class != NO_REGS && get_reg_class (regno) != new_class)
4122 {
4123 lra_assert (ok_p);
4124 lra_change_class (regno, new_class, " Change to", true);
4125 }
4126 }
4127 }
4128 else
4129 {
4130 const char *constraint;
4131 char c;
4132 rtx op = *curr_id->operand_loc[i];
4133 rtx subreg = NULL_RTX;
4134 machine_mode mode = curr_operand_mode[i];
4135
4136 if (GET_CODE (op) == SUBREG)
4137 {
4138 subreg = op;
4139 op = SUBREG_REG (op);
4140 mode = GET_MODE (op);
4141 }
4142
4143 if (CONST_POOL_OK_P (mode, op)
4144 && ((targetm.preferred_reload_class
4145 (op, (enum reg_class) goal_alt[i]) == NO_REGS)
4146 || no_input_reloads_p))
4147 {
4148 rtx tem = force_const_mem (mode, op);
4149
4150 change_p = true;
4151 if (subreg != NULL_RTX)
4152 tem = gen_rtx_SUBREG (mode, tem, SUBREG_BYTE (subreg));
4153
4154 *curr_id->operand_loc[i] = tem;
4155 lra_update_dup (curr_id, i);
4156 process_address (i, false, &before, &after);
4157
4158 /* If the alternative accepts constant pool refs directly
4159 there will be no reload needed at all. */
4160 if (subreg != NULL_RTX)
4161 continue;
4162 /* Skip alternatives before the one requested. */
4163 constraint = (curr_static_id->operand_alternative
4164 [goal_alt_number * n_operands + i].constraint);
4165 for (;
4166 (c = *constraint) && c != ',' && c != '#';
4167 constraint += CONSTRAINT_LEN (c, constraint))
4168 {
4169 enum constraint_num cn = lookup_constraint (constraint);
4170 if ((insn_extra_memory_constraint (cn)
4171 || insn_extra_special_memory_constraint (cn))
4172 && satisfies_memory_constraint_p (tem, cn))
4173 break;
4174 }
4175 if (c == '\0' || c == ',' || c == '#')
4176 continue;
4177
4178 goal_alt_win[i] = true;
4179 }
4180 }
4181
4182 n_outputs = 0;
4183 outputs[0] = -1;
4184 for (i = 0; i < n_operands; i++)
4185 {
4186 int regno;
4187 bool optional_p = false;
4188 rtx old, new_reg;
4189 rtx op = *curr_id->operand_loc[i];
4190
4191 if (goal_alt_win[i])
4192 {
4193 if (goal_alt[i] == NO_REGS
4194 && REG_P (op)
4195 /* When we assign NO_REGS it means that we will not
4196 assign a hard register to the scratch pseudo by
4197 assigment pass and the scratch pseudo will be
4198 spilled. Spilled scratch pseudos are transformed
4199 back to scratches at the LRA end. */
4200 && lra_former_scratch_operand_p (curr_insn, i)
4201 && lra_former_scratch_p (REGNO (op)))
4202 {
4203 int regno = REGNO (op);
4204 lra_change_class (regno, NO_REGS, " Change to", true);
4205 if (lra_get_regno_hard_regno (regno) >= 0)
4206 /* We don't have to mark all insn affected by the
4207 spilled pseudo as there is only one such insn, the
4208 current one. */
4209 reg_renumber[regno] = -1;
4210 lra_assert (bitmap_single_bit_set_p
4211 (&lra_reg_info[REGNO (op)].insn_bitmap));
4212 }
4213 /* We can do an optional reload. If the pseudo got a hard
4214 reg, we might improve the code through inheritance. If
4215 it does not get a hard register we coalesce memory/memory
4216 moves later. Ignore move insns to avoid cycling. */
4217 if (! lra_simple_p
4218 && lra_undo_inheritance_iter < LRA_MAX_INHERITANCE_PASSES
4219 && goal_alt[i] != NO_REGS && REG_P (op)
4220 && (regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER
4221 && regno < new_regno_start
4222 && ! lra_former_scratch_p (regno)
4223 && reg_renumber[regno] < 0
4224 /* Check that the optional reload pseudo will be able to
4225 hold given mode value. */
4226 && ! (prohibited_class_reg_set_mode_p
4227 (goal_alt[i], reg_class_contents[goal_alt[i]],
4228 PSEUDO_REGNO_MODE (regno)))
4229 && (curr_insn_set == NULL_RTX
4230 || !((REG_P (SET_SRC (curr_insn_set))
4231 || MEM_P (SET_SRC (curr_insn_set))
4232 || GET_CODE (SET_SRC (curr_insn_set)) == SUBREG)
4233 && (REG_P (SET_DEST (curr_insn_set))
4234 || MEM_P (SET_DEST (curr_insn_set))
4235 || GET_CODE (SET_DEST (curr_insn_set)) == SUBREG))))
4236 optional_p = true;
4237 else
4238 continue;
4239 }
4240
4241 /* Operands that match previous ones have already been handled. */
4242 if (goal_alt_matches[i] >= 0)
4243 continue;
4244
4245 /* We should not have an operand with a non-offsettable address
4246 appearing where an offsettable address will do. It also may
4247 be a case when the address should be special in other words
4248 not a general one (e.g. it needs no index reg). */
4249 if (goal_alt_matched[i][0] == -1 && goal_alt_offmemok[i] && MEM_P (op))
4250 {
4251 enum reg_class rclass;
4252 rtx *loc = &XEXP (op, 0);
4253 enum rtx_code code = GET_CODE (*loc);
4254
4255 push_to_sequence (before);
4256 rclass = base_reg_class (GET_MODE (op), MEM_ADDR_SPACE (op),
4257 MEM, SCRATCH);
4258 if (GET_RTX_CLASS (code) == RTX_AUTOINC)
4259 new_reg = emit_inc (rclass, *loc, *loc,
4260 /* This value does not matter for MODIFY. */
4261 GET_MODE_SIZE (GET_MODE (op)));
4262 else if (get_reload_reg (OP_IN, Pmode, *loc, rclass, FALSE,
4263 "offsetable address", &new_reg))
4264 {
4265 rtx addr = *loc;
4266 enum rtx_code code = GET_CODE (addr);
4267
4268 if (code == AND && CONST_INT_P (XEXP (addr, 1)))
4269 /* (and ... (const_int -X)) is used to align to X bytes. */
4270 addr = XEXP (*loc, 0);
4271 lra_emit_move (new_reg, addr);
4272 if (addr != *loc)
4273 emit_move_insn (new_reg, gen_rtx_AND (GET_MODE (new_reg), new_reg, XEXP (*loc, 1)));
4274 }
4275 before = get_insns ();
4276 end_sequence ();
4277 *loc = new_reg;
4278 lra_update_dup (curr_id, i);
4279 }
4280 else if (goal_alt_matched[i][0] == -1)
4281 {
4282 machine_mode mode;
4283 rtx reg, *loc;
4284 int hard_regno;
4285 enum op_type type = curr_static_id->operand[i].type;
4286
4287 loc = curr_id->operand_loc[i];
4288 mode = curr_operand_mode[i];
4289 if (GET_CODE (*loc) == SUBREG)
4290 {
4291 reg = SUBREG_REG (*loc);
4292 poly_int64 byte = SUBREG_BYTE (*loc);
4293 if (REG_P (reg)
4294 /* Strict_low_part requires reloading the register and not
4295 just the subreg. Likewise for a strict subreg no wider
4296 than a word for WORD_REGISTER_OPERATIONS targets. */
4297 && (curr_static_id->operand[i].strict_low
4298 || (!paradoxical_subreg_p (mode, GET_MODE (reg))
4299 && (hard_regno
4300 = get_try_hard_regno (REGNO (reg))) >= 0
4301 && (simplify_subreg_regno
4302 (hard_regno,
4303 GET_MODE (reg), byte, mode) < 0)
4304 && (goal_alt[i] == NO_REGS
4305 || (simplify_subreg_regno
4306 (ira_class_hard_regs[goal_alt[i]][0],
4307 GET_MODE (reg), byte, mode) >= 0)))
4308 || (partial_subreg_p (mode, GET_MODE (reg))
4309 && known_le (GET_MODE_SIZE (GET_MODE (reg)),
4310 UNITS_PER_WORD)
4311 && WORD_REGISTER_OPERATIONS)))
4312 {
4313 /* An OP_INOUT is required when reloading a subreg of a
4314 mode wider than a word to ensure that data beyond the
4315 word being reloaded is preserved. Also automatically
4316 ensure that strict_low_part reloads are made into
4317 OP_INOUT which should already be true from the backend
4318 constraints. */
4319 if (type == OP_OUT
4320 && (curr_static_id->operand[i].strict_low
4321 || read_modify_subreg_p (*loc)))
4322 type = OP_INOUT;
4323 loc = &SUBREG_REG (*loc);
4324 mode = GET_MODE (*loc);
4325 }
4326 }
4327 old = *loc;
4328 if (get_reload_reg (type, mode, old, goal_alt[i],
4329 loc != curr_id->operand_loc[i], "", &new_reg)
4330 && type != OP_OUT)
4331 {
4332 push_to_sequence (before);
4333 lra_emit_move (new_reg, old);
4334 before = get_insns ();
4335 end_sequence ();
4336 }
4337 *loc = new_reg;
4338 if (type != OP_IN
4339 && find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX)
4340 {
4341 start_sequence ();
4342 lra_emit_move (type == OP_INOUT ? copy_rtx (old) : old, new_reg);
4343 emit_insn (after);
4344 after = get_insns ();
4345 end_sequence ();
4346 *loc = new_reg;
4347 }
4348 for (j = 0; j < goal_alt_dont_inherit_ops_num; j++)
4349 if (goal_alt_dont_inherit_ops[j] == i)
4350 {
4351 lra_set_regno_unique_value (REGNO (new_reg));
4352 break;
4353 }
4354 lra_update_dup (curr_id, i);
4355 }
4356 else if (curr_static_id->operand[i].type == OP_IN
4357 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4358 == OP_OUT
4359 || (curr_static_id->operand[goal_alt_matched[i][0]].type
4360 == OP_INOUT
4361 && (operands_match_p
4362 (*curr_id->operand_loc[i],
4363 *curr_id->operand_loc[goal_alt_matched[i][0]],
4364 -1)))))
4365 {
4366 /* generate reloads for input and matched outputs. */
4367 match_inputs[0] = i;
4368 match_inputs[1] = -1;
4369 match_reload (goal_alt_matched[i][0], match_inputs, outputs,
4370 goal_alt[i], &before, &after,
4371 curr_static_id->operand_alternative
4372 [goal_alt_number * n_operands + goal_alt_matched[i][0]]
4373 .earlyclobber);
4374 }
4375 else if ((curr_static_id->operand[i].type == OP_OUT
4376 || (curr_static_id->operand[i].type == OP_INOUT
4377 && (operands_match_p
4378 (*curr_id->operand_loc[i],
4379 *curr_id->operand_loc[goal_alt_matched[i][0]],
4380 -1))))
4381 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4382 == OP_IN))
4383 /* Generate reloads for output and matched inputs. */
4384 match_reload (i, goal_alt_matched[i], outputs, goal_alt[i], &before,
4385 &after, curr_static_id->operand_alternative
4386 [goal_alt_number * n_operands + i].earlyclobber);
4387 else if (curr_static_id->operand[i].type == OP_IN
4388 && (curr_static_id->operand[goal_alt_matched[i][0]].type
4389 == OP_IN))
4390 {
4391 /* Generate reloads for matched inputs. */
4392 match_inputs[0] = i;
4393 for (j = 0; (k = goal_alt_matched[i][j]) >= 0; j++)
4394 match_inputs[j + 1] = k;
4395 match_inputs[j + 1] = -1;
4396 match_reload (-1, match_inputs, outputs, goal_alt[i], &before,
4397 &after, false);
4398 }
4399 else
4400 /* We must generate code in any case when function
4401 process_alt_operands decides that it is possible. */
4402 gcc_unreachable ();
4403
4404 /* Memorise processed outputs so that output remaining to be processed
4405 can avoid using the same register value (see match_reload). */
4406 if (curr_static_id->operand[i].type == OP_OUT)
4407 {
4408 outputs[n_outputs++] = i;
4409 outputs[n_outputs] = -1;
4410 }
4411
4412 if (optional_p)
4413 {
4414 rtx reg = op;
4415
4416 lra_assert (REG_P (reg));
4417 regno = REGNO (reg);
4418 op = *curr_id->operand_loc[i]; /* Substitution. */
4419 if (GET_CODE (op) == SUBREG)
4420 op = SUBREG_REG (op);
4421 gcc_assert (REG_P (op) && (int) REGNO (op) >= new_regno_start);
4422 bitmap_set_bit (&lra_optional_reload_pseudos, REGNO (op));
4423 lra_reg_info[REGNO (op)].restore_rtx = reg;
4424 if (lra_dump_file != NULL)
4425 fprintf (lra_dump_file,
4426 " Making reload reg %d for reg %d optional\n",
4427 REGNO (op), regno);
4428 }
4429 }
4430 if (before != NULL_RTX || after != NULL_RTX
4431 || max_regno_before != max_reg_num ())
4432 change_p = true;
4433 if (change_p)
4434 {
4435 lra_update_operator_dups (curr_id);
4436 /* Something changes -- process the insn. */
4437 lra_update_insn_regno_info (curr_insn);
4438 }
4439 lra_process_new_insns (curr_insn, before, after, "Inserting insn reload");
4440 return change_p;
4441 }
4442
4443 /* Return true if INSN satisfies all constraints. In other words, no
4444 reload insns are needed. */
4445 bool
4446 lra_constrain_insn (rtx_insn *insn)
4447 {
4448 int saved_new_regno_start = new_regno_start;
4449 int saved_new_insn_uid_start = new_insn_uid_start;
4450 bool change_p;
4451
4452 curr_insn = insn;
4453 curr_id = lra_get_insn_recog_data (curr_insn);
4454 curr_static_id = curr_id->insn_static_data;
4455 new_insn_uid_start = get_max_uid ();
4456 new_regno_start = max_reg_num ();
4457 change_p = curr_insn_transform (true);
4458 new_regno_start = saved_new_regno_start;
4459 new_insn_uid_start = saved_new_insn_uid_start;
4460 return ! change_p;
4461 }
4462
4463 /* Return true if X is in LIST. */
4464 static bool
4465 in_list_p (rtx x, rtx list)
4466 {
4467 for (; list != NULL_RTX; list = XEXP (list, 1))
4468 if (XEXP (list, 0) == x)
4469 return true;
4470 return false;
4471 }
4472
4473 /* Return true if X contains an allocatable hard register (if
4474 HARD_REG_P) or a (spilled if SPILLED_P) pseudo. */
4475 static bool
4476 contains_reg_p (rtx x, bool hard_reg_p, bool spilled_p)
4477 {
4478 int i, j;
4479 const char *fmt;
4480 enum rtx_code code;
4481
4482 code = GET_CODE (x);
4483 if (REG_P (x))
4484 {
4485 int regno = REGNO (x);
4486 HARD_REG_SET alloc_regs;
4487
4488 if (hard_reg_p)
4489 {
4490 if (regno >= FIRST_PSEUDO_REGISTER)
4491 regno = lra_get_regno_hard_regno (regno);
4492 if (regno < 0)
4493 return false;
4494 COMPL_HARD_REG_SET (alloc_regs, lra_no_alloc_regs);
4495 return overlaps_hard_reg_set_p (alloc_regs, GET_MODE (x), regno);
4496 }
4497 else
4498 {
4499 if (regno < FIRST_PSEUDO_REGISTER)
4500 return false;
4501 if (! spilled_p)
4502 return true;
4503 return lra_get_regno_hard_regno (regno) < 0;
4504 }
4505 }
4506 fmt = GET_RTX_FORMAT (code);
4507 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4508 {
4509 if (fmt[i] == 'e')
4510 {
4511 if (contains_reg_p (XEXP (x, i), hard_reg_p, spilled_p))
4512 return true;
4513 }
4514 else if (fmt[i] == 'E')
4515 {
4516 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4517 if (contains_reg_p (XVECEXP (x, i, j), hard_reg_p, spilled_p))
4518 return true;
4519 }
4520 }
4521 return false;
4522 }
4523
4524 /* Process all regs in location *LOC and change them on equivalent
4525 substitution. Return true if any change was done. */
4526 static bool
4527 loc_equivalence_change_p (rtx *loc)
4528 {
4529 rtx subst, reg, x = *loc;
4530 bool result = false;
4531 enum rtx_code code = GET_CODE (x);
4532 const char *fmt;
4533 int i, j;
4534
4535 if (code == SUBREG)
4536 {
4537 reg = SUBREG_REG (x);
4538 if ((subst = get_equiv_with_elimination (reg, curr_insn)) != reg
4539 && GET_MODE (subst) == VOIDmode)
4540 {
4541 /* We cannot reload debug location. Simplify subreg here
4542 while we know the inner mode. */
4543 *loc = simplify_gen_subreg (GET_MODE (x), subst,
4544 GET_MODE (reg), SUBREG_BYTE (x));
4545 return true;
4546 }
4547 }
4548 if (code == REG && (subst = get_equiv_with_elimination (x, curr_insn)) != x)
4549 {
4550 *loc = subst;
4551 return true;
4552 }
4553
4554 /* Scan all the operand sub-expressions. */
4555 fmt = GET_RTX_FORMAT (code);
4556 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4557 {
4558 if (fmt[i] == 'e')
4559 result = loc_equivalence_change_p (&XEXP (x, i)) || result;
4560 else if (fmt[i] == 'E')
4561 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4562 result
4563 = loc_equivalence_change_p (&XVECEXP (x, i, j)) || result;
4564 }
4565 return result;
4566 }
4567
4568 /* Similar to loc_equivalence_change_p, but for use as
4569 simplify_replace_fn_rtx callback. DATA is insn for which the
4570 elimination is done. If it null we don't do the elimination. */
4571 static rtx
4572 loc_equivalence_callback (rtx loc, const_rtx, void *data)
4573 {
4574 if (!REG_P (loc))
4575 return NULL_RTX;
4576
4577 rtx subst = (data == NULL
4578 ? get_equiv (loc) : get_equiv_with_elimination (loc, (rtx_insn *) data));
4579 if (subst != loc)
4580 return subst;
4581
4582 return NULL_RTX;
4583 }
4584
4585 /* Maximum number of generated reload insns per an insn. It is for
4586 preventing this pass cycling in a bug case. */
4587 #define MAX_RELOAD_INSNS_NUMBER LRA_MAX_INSN_RELOADS
4588
4589 /* The current iteration number of this LRA pass. */
4590 int lra_constraint_iter;
4591
4592 /* True if we substituted equiv which needs checking register
4593 allocation correctness because the equivalent value contains
4594 allocatable hard registers or when we restore multi-register
4595 pseudo. */
4596 bool lra_risky_transformations_p;
4597
4598 /* Return true if REGNO is referenced in more than one block. */
4599 static bool
4600 multi_block_pseudo_p (int regno)
4601 {
4602 basic_block bb = NULL;
4603 unsigned int uid;
4604 bitmap_iterator bi;
4605
4606 if (regno < FIRST_PSEUDO_REGISTER)
4607 return false;
4608
4609 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi)
4610 if (bb == NULL)
4611 bb = BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn);
4612 else if (BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn) != bb)
4613 return true;
4614 return false;
4615 }
4616
4617 /* Return true if LIST contains a deleted insn. */
4618 static bool
4619 contains_deleted_insn_p (rtx_insn_list *list)
4620 {
4621 for (; list != NULL_RTX; list = list->next ())
4622 if (NOTE_P (list->insn ())
4623 && NOTE_KIND (list->insn ()) == NOTE_INSN_DELETED)
4624 return true;
4625 return false;
4626 }
4627
4628 /* Return true if X contains a pseudo dying in INSN. */
4629 static bool
4630 dead_pseudo_p (rtx x, rtx_insn *insn)
4631 {
4632 int i, j;
4633 const char *fmt;
4634 enum rtx_code code;
4635
4636 if (REG_P (x))
4637 return (insn != NULL_RTX
4638 && find_regno_note (insn, REG_DEAD, REGNO (x)) != NULL_RTX);
4639 code = GET_CODE (x);
4640 fmt = GET_RTX_FORMAT (code);
4641 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4642 {
4643 if (fmt[i] == 'e')
4644 {
4645 if (dead_pseudo_p (XEXP (x, i), insn))
4646 return true;
4647 }
4648 else if (fmt[i] == 'E')
4649 {
4650 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4651 if (dead_pseudo_p (XVECEXP (x, i, j), insn))
4652 return true;
4653 }
4654 }
4655 return false;
4656 }
4657
4658 /* Return true if INSN contains a dying pseudo in INSN right hand
4659 side. */
4660 static bool
4661 insn_rhs_dead_pseudo_p (rtx_insn *insn)
4662 {
4663 rtx set = single_set (insn);
4664
4665 gcc_assert (set != NULL);
4666 return dead_pseudo_p (SET_SRC (set), insn);
4667 }
4668
4669 /* Return true if any init insn of REGNO contains a dying pseudo in
4670 insn right hand side. */
4671 static bool
4672 init_insn_rhs_dead_pseudo_p (int regno)
4673 {
4674 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4675
4676 if (insns == NULL)
4677 return false;
4678 for (; insns != NULL_RTX; insns = insns->next ())
4679 if (insn_rhs_dead_pseudo_p (insns->insn ()))
4680 return true;
4681 return false;
4682 }
4683
4684 /* Return TRUE if REGNO has a reverse equivalence. The equivalence is
4685 reverse only if we have one init insn with given REGNO as a
4686 source. */
4687 static bool
4688 reverse_equiv_p (int regno)
4689 {
4690 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4691 rtx set;
4692
4693 if (insns == NULL)
4694 return false;
4695 if (! INSN_P (insns->insn ())
4696 || insns->next () != NULL)
4697 return false;
4698 if ((set = single_set (insns->insn ())) == NULL_RTX)
4699 return false;
4700 return REG_P (SET_SRC (set)) && (int) REGNO (SET_SRC (set)) == regno;
4701 }
4702
4703 /* Return TRUE if REGNO was reloaded in an equivalence init insn. We
4704 call this function only for non-reverse equivalence. */
4705 static bool
4706 contains_reloaded_insn_p (int regno)
4707 {
4708 rtx set;
4709 rtx_insn_list *list = ira_reg_equiv[regno].init_insns;
4710
4711 for (; list != NULL; list = list->next ())
4712 if ((set = single_set (list->insn ())) == NULL_RTX
4713 || ! REG_P (SET_DEST (set))
4714 || (int) REGNO (SET_DEST (set)) != regno)
4715 return true;
4716 return false;
4717 }
4718
4719 /* Entry function of LRA constraint pass. Return true if the
4720 constraint pass did change the code. */
4721 bool
4722 lra_constraints (bool first_p)
4723 {
4724 bool changed_p;
4725 int i, hard_regno, new_insns_num;
4726 unsigned int min_len, new_min_len, uid;
4727 rtx set, x, reg, dest_reg;
4728 basic_block last_bb;
4729 bitmap_iterator bi;
4730
4731 lra_constraint_iter++;
4732 if (lra_dump_file != NULL)
4733 fprintf (lra_dump_file, "\n********** Local #%d: **********\n\n",
4734 lra_constraint_iter);
4735 changed_p = false;
4736 if (pic_offset_table_rtx
4737 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
4738 lra_risky_transformations_p = true;
4739 else
4740 /* On the first iteration we should check IRA assignment
4741 correctness. In rare cases, the assignments can be wrong as
4742 early clobbers operands are ignored in IRA or usages of
4743 paradoxical sub-registers are not taken into account by
4744 IRA. */
4745 lra_risky_transformations_p = first_p;
4746 new_insn_uid_start = get_max_uid ();
4747 new_regno_start = first_p ? lra_constraint_new_regno_start : max_reg_num ();
4748 /* Mark used hard regs for target stack size calulations. */
4749 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4750 if (lra_reg_info[i].nrefs != 0
4751 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4752 {
4753 int j, nregs;
4754
4755 nregs = hard_regno_nregs (hard_regno, lra_reg_info[i].biggest_mode);
4756 for (j = 0; j < nregs; j++)
4757 df_set_regs_ever_live (hard_regno + j, true);
4758 }
4759 /* Do elimination before the equivalence processing as we can spill
4760 some pseudos during elimination. */
4761 lra_eliminate (false, first_p);
4762 auto_bitmap equiv_insn_bitmap (&reg_obstack);
4763 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4764 if (lra_reg_info[i].nrefs != 0)
4765 {
4766 ira_reg_equiv[i].profitable_p = true;
4767 reg = regno_reg_rtx[i];
4768 if (lra_get_regno_hard_regno (i) < 0 && (x = get_equiv (reg)) != reg)
4769 {
4770 bool pseudo_p = contains_reg_p (x, false, false);
4771
4772 /* After RTL transformation, we cannot guarantee that
4773 pseudo in the substitution was not reloaded which might
4774 make equivalence invalid. For example, in reverse
4775 equiv of p0
4776
4777 p0 <- ...
4778 ...
4779 equiv_mem <- p0
4780
4781 the memory address register was reloaded before the 2nd
4782 insn. */
4783 if ((! first_p && pseudo_p)
4784 /* We don't use DF for compilation speed sake. So it
4785 is problematic to update live info when we use an
4786 equivalence containing pseudos in more than one
4787 BB. */
4788 || (pseudo_p && multi_block_pseudo_p (i))
4789 /* If an init insn was deleted for some reason, cancel
4790 the equiv. We could update the equiv insns after
4791 transformations including an equiv insn deletion
4792 but it is not worthy as such cases are extremely
4793 rare. */
4794 || contains_deleted_insn_p (ira_reg_equiv[i].init_insns)
4795 /* If it is not a reverse equivalence, we check that a
4796 pseudo in rhs of the init insn is not dying in the
4797 insn. Otherwise, the live info at the beginning of
4798 the corresponding BB might be wrong after we
4799 removed the insn. When the equiv can be a
4800 constant, the right hand side of the init insn can
4801 be a pseudo. */
4802 || (! reverse_equiv_p (i)
4803 && (init_insn_rhs_dead_pseudo_p (i)
4804 /* If we reloaded the pseudo in an equivalence
4805 init insn, we cannot remove the equiv init
4806 insns and the init insns might write into
4807 const memory in this case. */
4808 || contains_reloaded_insn_p (i)))
4809 /* Prevent access beyond equivalent memory for
4810 paradoxical subregs. */
4811 || (MEM_P (x)
4812 && maybe_gt (GET_MODE_SIZE (lra_reg_info[i].biggest_mode),
4813 GET_MODE_SIZE (GET_MODE (x))))
4814 || (pic_offset_table_rtx
4815 && ((CONST_POOL_OK_P (PSEUDO_REGNO_MODE (i), x)
4816 && (targetm.preferred_reload_class
4817 (x, lra_get_allocno_class (i)) == NO_REGS))
4818 || contains_symbol_ref_p (x))))
4819 ira_reg_equiv[i].defined_p = false;
4820 if (contains_reg_p (x, false, true))
4821 ira_reg_equiv[i].profitable_p = false;
4822 if (get_equiv (reg) != reg)
4823 bitmap_ior_into (equiv_insn_bitmap, &lra_reg_info[i].insn_bitmap);
4824 }
4825 }
4826 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4827 update_equiv (i);
4828 /* We should add all insns containing pseudos which should be
4829 substituted by their equivalences. */
4830 EXECUTE_IF_SET_IN_BITMAP (equiv_insn_bitmap, 0, uid, bi)
4831 lra_push_insn_by_uid (uid);
4832 min_len = lra_insn_stack_length ();
4833 new_insns_num = 0;
4834 last_bb = NULL;
4835 changed_p = false;
4836 while ((new_min_len = lra_insn_stack_length ()) != 0)
4837 {
4838 curr_insn = lra_pop_insn ();
4839 --new_min_len;
4840 curr_bb = BLOCK_FOR_INSN (curr_insn);
4841 if (curr_bb != last_bb)
4842 {
4843 last_bb = curr_bb;
4844 bb_reload_num = lra_curr_reload_num;
4845 }
4846 if (min_len > new_min_len)
4847 {
4848 min_len = new_min_len;
4849 new_insns_num = 0;
4850 }
4851 if (new_insns_num > MAX_RELOAD_INSNS_NUMBER)
4852 internal_error
4853 ("Max. number of generated reload insns per insn is achieved (%d)\n",
4854 MAX_RELOAD_INSNS_NUMBER);
4855 new_insns_num++;
4856 if (DEBUG_INSN_P (curr_insn))
4857 {
4858 /* We need to check equivalence in debug insn and change
4859 pseudo to the equivalent value if necessary. */
4860 curr_id = lra_get_insn_recog_data (curr_insn);
4861 if (bitmap_bit_p (equiv_insn_bitmap, INSN_UID (curr_insn)))
4862 {
4863 rtx old = *curr_id->operand_loc[0];
4864 *curr_id->operand_loc[0]
4865 = simplify_replace_fn_rtx (old, NULL_RTX,
4866 loc_equivalence_callback, curr_insn);
4867 if (old != *curr_id->operand_loc[0])
4868 {
4869 lra_update_insn_regno_info (curr_insn);
4870 changed_p = true;
4871 }
4872 }
4873 }
4874 else if (INSN_P (curr_insn))
4875 {
4876 if ((set = single_set (curr_insn)) != NULL_RTX)
4877 {
4878 dest_reg = SET_DEST (set);
4879 /* The equivalence pseudo could be set up as SUBREG in a
4880 case when it is a call restore insn in a mode
4881 different from the pseudo mode. */
4882 if (GET_CODE (dest_reg) == SUBREG)
4883 dest_reg = SUBREG_REG (dest_reg);
4884 if ((REG_P (dest_reg)
4885 && (x = get_equiv (dest_reg)) != dest_reg
4886 /* Remove insns which set up a pseudo whose value
4887 cannot be changed. Such insns might be not in
4888 init_insns because we don't update equiv data
4889 during insn transformations.
4890
4891 As an example, let suppose that a pseudo got
4892 hard register and on the 1st pass was not
4893 changed to equivalent constant. We generate an
4894 additional insn setting up the pseudo because of
4895 secondary memory movement. Then the pseudo is
4896 spilled and we use the equiv constant. In this
4897 case we should remove the additional insn and
4898 this insn is not init_insns list. */
4899 && (! MEM_P (x) || MEM_READONLY_P (x)
4900 /* Check that this is actually an insn setting
4901 up the equivalence. */
4902 || in_list_p (curr_insn,
4903 ira_reg_equiv
4904 [REGNO (dest_reg)].init_insns)))
4905 || (((x = get_equiv (SET_SRC (set))) != SET_SRC (set))
4906 && in_list_p (curr_insn,
4907 ira_reg_equiv
4908 [REGNO (SET_SRC (set))].init_insns)))
4909 {
4910 /* This is equiv init insn of pseudo which did not get a
4911 hard register -- remove the insn. */
4912 if (lra_dump_file != NULL)
4913 {
4914 fprintf (lra_dump_file,
4915 " Removing equiv init insn %i (freq=%d)\n",
4916 INSN_UID (curr_insn),
4917 REG_FREQ_FROM_BB (BLOCK_FOR_INSN (curr_insn)));
4918 dump_insn_slim (lra_dump_file, curr_insn);
4919 }
4920 if (contains_reg_p (x, true, false))
4921 lra_risky_transformations_p = true;
4922 lra_set_insn_deleted (curr_insn);
4923 continue;
4924 }
4925 }
4926 curr_id = lra_get_insn_recog_data (curr_insn);
4927 curr_static_id = curr_id->insn_static_data;
4928 init_curr_insn_input_reloads ();
4929 init_curr_operand_mode ();
4930 if (curr_insn_transform (false))
4931 changed_p = true;
4932 /* Check non-transformed insns too for equiv change as USE
4933 or CLOBBER don't need reloads but can contain pseudos
4934 being changed on their equivalences. */
4935 else if (bitmap_bit_p (equiv_insn_bitmap, INSN_UID (curr_insn))
4936 && loc_equivalence_change_p (&PATTERN (curr_insn)))
4937 {
4938 lra_update_insn_regno_info (curr_insn);
4939 changed_p = true;
4940 }
4941 }
4942 }
4943
4944 /* If we used a new hard regno, changed_p should be true because the
4945 hard reg is assigned to a new pseudo. */
4946 if (flag_checking && !changed_p)
4947 {
4948 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4949 if (lra_reg_info[i].nrefs != 0
4950 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4951 {
4952 int j, nregs = hard_regno_nregs (hard_regno,
4953 PSEUDO_REGNO_MODE (i));
4954
4955 for (j = 0; j < nregs; j++)
4956 lra_assert (df_regs_ever_live_p (hard_regno + j));
4957 }
4958 }
4959 return changed_p;
4960 }
4961
4962 static void initiate_invariants (void);
4963 static void finish_invariants (void);
4964
4965 /* Initiate the LRA constraint pass. It is done once per
4966 function. */
4967 void
4968 lra_constraints_init (void)
4969 {
4970 initiate_invariants ();
4971 }
4972
4973 /* Finalize the LRA constraint pass. It is done once per
4974 function. */
4975 void
4976 lra_constraints_finish (void)
4977 {
4978 finish_invariants ();
4979 }
4980
4981 \f
4982
4983 /* Structure describes invariants for ineheritance. */
4984 struct lra_invariant
4985 {
4986 /* The order number of the invariant. */
4987 int num;
4988 /* The invariant RTX. */
4989 rtx invariant_rtx;
4990 /* The origin insn of the invariant. */
4991 rtx_insn *insn;
4992 };
4993
4994 typedef lra_invariant invariant_t;
4995 typedef invariant_t *invariant_ptr_t;
4996 typedef const invariant_t *const_invariant_ptr_t;
4997
4998 /* Pointer to the inheritance invariants. */
4999 static vec<invariant_ptr_t> invariants;
5000
5001 /* Allocation pool for the invariants. */
5002 static object_allocator<lra_invariant> *invariants_pool;
5003
5004 /* Hash table for the invariants. */
5005 static htab_t invariant_table;
5006
5007 /* Hash function for INVARIANT. */
5008 static hashval_t
5009 invariant_hash (const void *invariant)
5010 {
5011 rtx inv = ((const_invariant_ptr_t) invariant)->invariant_rtx;
5012 return lra_rtx_hash (inv);
5013 }
5014
5015 /* Equal function for invariants INVARIANT1 and INVARIANT2. */
5016 static int
5017 invariant_eq_p (const void *invariant1, const void *invariant2)
5018 {
5019 rtx inv1 = ((const_invariant_ptr_t) invariant1)->invariant_rtx;
5020 rtx inv2 = ((const_invariant_ptr_t) invariant2)->invariant_rtx;
5021
5022 return rtx_equal_p (inv1, inv2);
5023 }
5024
5025 /* Insert INVARIANT_RTX into the table if it is not there yet. Return
5026 invariant which is in the table. */
5027 static invariant_ptr_t
5028 insert_invariant (rtx invariant_rtx)
5029 {
5030 void **entry_ptr;
5031 invariant_t invariant;
5032 invariant_ptr_t invariant_ptr;
5033
5034 invariant.invariant_rtx = invariant_rtx;
5035 entry_ptr = htab_find_slot (invariant_table, &invariant, INSERT);
5036 if (*entry_ptr == NULL)
5037 {
5038 invariant_ptr = invariants_pool->allocate ();
5039 invariant_ptr->invariant_rtx = invariant_rtx;
5040 invariant_ptr->insn = NULL;
5041 invariants.safe_push (invariant_ptr);
5042 *entry_ptr = (void *) invariant_ptr;
5043 }
5044 return (invariant_ptr_t) *entry_ptr;
5045 }
5046
5047 /* Initiate the invariant table. */
5048 static void
5049 initiate_invariants (void)
5050 {
5051 invariants.create (100);
5052 invariants_pool
5053 = new object_allocator<lra_invariant> ("Inheritance invariants");
5054 invariant_table = htab_create (100, invariant_hash, invariant_eq_p, NULL);
5055 }
5056
5057 /* Finish the invariant table. */
5058 static void
5059 finish_invariants (void)
5060 {
5061 htab_delete (invariant_table);
5062 delete invariants_pool;
5063 invariants.release ();
5064 }
5065
5066 /* Make the invariant table empty. */
5067 static void
5068 clear_invariants (void)
5069 {
5070 htab_empty (invariant_table);
5071 invariants_pool->release ();
5072 invariants.truncate (0);
5073 }
5074
5075 \f
5076
5077 /* This page contains code to do inheritance/split
5078 transformations. */
5079
5080 /* Number of reloads passed so far in current EBB. */
5081 static int reloads_num;
5082
5083 /* Number of calls passed so far in current EBB. */
5084 static int calls_num;
5085
5086 /* Current reload pseudo check for validity of elements in
5087 USAGE_INSNS. */
5088 static int curr_usage_insns_check;
5089
5090 /* Info about last usage of registers in EBB to do inheritance/split
5091 transformation. Inheritance transformation is done from a spilled
5092 pseudo and split transformations from a hard register or a pseudo
5093 assigned to a hard register. */
5094 struct usage_insns
5095 {
5096 /* If the value is equal to CURR_USAGE_INSNS_CHECK, then the member
5097 value INSNS is valid. The insns is chain of optional debug insns
5098 and a finishing non-debug insn using the corresponding reg. The
5099 value is also used to mark the registers which are set up in the
5100 current insn. The negated insn uid is used for this. */
5101 int check;
5102 /* Value of global reloads_num at the last insn in INSNS. */
5103 int reloads_num;
5104 /* Value of global reloads_nums at the last insn in INSNS. */
5105 int calls_num;
5106 /* It can be true only for splitting. And it means that the restore
5107 insn should be put after insn given by the following member. */
5108 bool after_p;
5109 /* Next insns in the current EBB which use the original reg and the
5110 original reg value is not changed between the current insn and
5111 the next insns. In order words, e.g. for inheritance, if we need
5112 to use the original reg value again in the next insns we can try
5113 to use the value in a hard register from a reload insn of the
5114 current insn. */
5115 rtx insns;
5116 };
5117
5118 /* Map: regno -> corresponding pseudo usage insns. */
5119 static struct usage_insns *usage_insns;
5120
5121 static void
5122 setup_next_usage_insn (int regno, rtx insn, int reloads_num, bool after_p)
5123 {
5124 usage_insns[regno].check = curr_usage_insns_check;
5125 usage_insns[regno].insns = insn;
5126 usage_insns[regno].reloads_num = reloads_num;
5127 usage_insns[regno].calls_num = calls_num;
5128 usage_insns[regno].after_p = after_p;
5129 }
5130
5131 /* The function is used to form list REGNO usages which consists of
5132 optional debug insns finished by a non-debug insn using REGNO.
5133 RELOADS_NUM is current number of reload insns processed so far. */
5134 static void
5135 add_next_usage_insn (int regno, rtx_insn *insn, int reloads_num)
5136 {
5137 rtx next_usage_insns;
5138
5139 if (usage_insns[regno].check == curr_usage_insns_check
5140 && (next_usage_insns = usage_insns[regno].insns) != NULL_RTX
5141 && DEBUG_INSN_P (insn))
5142 {
5143 /* Check that we did not add the debug insn yet. */
5144 if (next_usage_insns != insn
5145 && (GET_CODE (next_usage_insns) != INSN_LIST
5146 || XEXP (next_usage_insns, 0) != insn))
5147 usage_insns[regno].insns = gen_rtx_INSN_LIST (VOIDmode, insn,
5148 next_usage_insns);
5149 }
5150 else if (NONDEBUG_INSN_P (insn))
5151 setup_next_usage_insn (regno, insn, reloads_num, false);
5152 else
5153 usage_insns[regno].check = 0;
5154 }
5155
5156 /* Return first non-debug insn in list USAGE_INSNS. */
5157 static rtx_insn *
5158 skip_usage_debug_insns (rtx usage_insns)
5159 {
5160 rtx insn;
5161
5162 /* Skip debug insns. */
5163 for (insn = usage_insns;
5164 insn != NULL_RTX && GET_CODE (insn) == INSN_LIST;
5165 insn = XEXP (insn, 1))
5166 ;
5167 return safe_as_a <rtx_insn *> (insn);
5168 }
5169
5170 /* Return true if we need secondary memory moves for insn in
5171 USAGE_INSNS after inserting inherited pseudo of class INHER_CL
5172 into the insn. */
5173 static bool
5174 check_secondary_memory_needed_p (enum reg_class inher_cl ATTRIBUTE_UNUSED,
5175 rtx usage_insns ATTRIBUTE_UNUSED)
5176 {
5177 rtx_insn *insn;
5178 rtx set, dest;
5179 enum reg_class cl;
5180
5181 if (inher_cl == ALL_REGS
5182 || (insn = skip_usage_debug_insns (usage_insns)) == NULL_RTX)
5183 return false;
5184 lra_assert (INSN_P (insn));
5185 if ((set = single_set (insn)) == NULL_RTX || ! REG_P (SET_DEST (set)))
5186 return false;
5187 dest = SET_DEST (set);
5188 if (! REG_P (dest))
5189 return false;
5190 lra_assert (inher_cl != NO_REGS);
5191 cl = get_reg_class (REGNO (dest));
5192 return (cl != NO_REGS && cl != ALL_REGS
5193 && targetm.secondary_memory_needed (GET_MODE (dest), inher_cl, cl));
5194 }
5195
5196 /* Registers involved in inheritance/split in the current EBB
5197 (inheritance/split pseudos and original registers). */
5198 static bitmap_head check_only_regs;
5199
5200 /* Reload pseudos cannot be involded in invariant inheritance in the
5201 current EBB. */
5202 static bitmap_head invalid_invariant_regs;
5203
5204 /* Do inheritance transformations for insn INSN, which defines (if
5205 DEF_P) or uses ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which
5206 instruction in the EBB next uses ORIGINAL_REGNO; it has the same
5207 form as the "insns" field of usage_insns. Return true if we
5208 succeed in such transformation.
5209
5210 The transformations look like:
5211
5212 p <- ... i <- ...
5213 ... p <- i (new insn)
5214 ... =>
5215 <- ... p ... <- ... i ...
5216 or
5217 ... i <- p (new insn)
5218 <- ... p ... <- ... i ...
5219 ... =>
5220 <- ... p ... <- ... i ...
5221 where p is a spilled original pseudo and i is a new inheritance pseudo.
5222
5223
5224 The inheritance pseudo has the smallest class of two classes CL and
5225 class of ORIGINAL REGNO. */
5226 static bool
5227 inherit_reload_reg (bool def_p, int original_regno,
5228 enum reg_class cl, rtx_insn *insn, rtx next_usage_insns)
5229 {
5230 if (optimize_function_for_size_p (cfun))
5231 return false;
5232
5233 enum reg_class rclass = lra_get_allocno_class (original_regno);
5234 rtx original_reg = regno_reg_rtx[original_regno];
5235 rtx new_reg, usage_insn;
5236 rtx_insn *new_insns;
5237
5238 lra_assert (! usage_insns[original_regno].after_p);
5239 if (lra_dump_file != NULL)
5240 fprintf (lra_dump_file,
5241 " <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n");
5242 if (! ira_reg_classes_intersect_p[cl][rclass])
5243 {
5244 if (lra_dump_file != NULL)
5245 {
5246 fprintf (lra_dump_file,
5247 " Rejecting inheritance for %d "
5248 "because of disjoint classes %s and %s\n",
5249 original_regno, reg_class_names[cl],
5250 reg_class_names[rclass]);
5251 fprintf (lra_dump_file,
5252 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5253 }
5254 return false;
5255 }
5256 if ((ira_class_subset_p[cl][rclass] && cl != rclass)
5257 /* We don't use a subset of two classes because it can be
5258 NO_REGS. This transformation is still profitable in most
5259 cases even if the classes are not intersected as register
5260 move is probably cheaper than a memory load. */
5261 || ira_class_hard_regs_num[cl] < ira_class_hard_regs_num[rclass])
5262 {
5263 if (lra_dump_file != NULL)
5264 fprintf (lra_dump_file, " Use smallest class of %s and %s\n",
5265 reg_class_names[cl], reg_class_names[rclass]);
5266
5267 rclass = cl;
5268 }
5269 if (check_secondary_memory_needed_p (rclass, next_usage_insns))
5270 {
5271 /* Reject inheritance resulting in secondary memory moves.
5272 Otherwise, there is a danger in LRA cycling. Also such
5273 transformation will be unprofitable. */
5274 if (lra_dump_file != NULL)
5275 {
5276 rtx_insn *insn = skip_usage_debug_insns (next_usage_insns);
5277 rtx set = single_set (insn);
5278
5279 lra_assert (set != NULL_RTX);
5280
5281 rtx dest = SET_DEST (set);
5282
5283 lra_assert (REG_P (dest));
5284 fprintf (lra_dump_file,
5285 " Rejecting inheritance for insn %d(%s)<-%d(%s) "
5286 "as secondary mem is needed\n",
5287 REGNO (dest), reg_class_names[get_reg_class (REGNO (dest))],
5288 original_regno, reg_class_names[rclass]);
5289 fprintf (lra_dump_file,
5290 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5291 }
5292 return false;
5293 }
5294 new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg,
5295 rclass, "inheritance");
5296 start_sequence ();
5297 if (def_p)
5298 lra_emit_move (original_reg, new_reg);
5299 else
5300 lra_emit_move (new_reg, original_reg);
5301 new_insns = get_insns ();
5302 end_sequence ();
5303 if (NEXT_INSN (new_insns) != NULL_RTX)
5304 {
5305 if (lra_dump_file != NULL)
5306 {
5307 fprintf (lra_dump_file,
5308 " Rejecting inheritance %d->%d "
5309 "as it results in 2 or more insns:\n",
5310 original_regno, REGNO (new_reg));
5311 dump_rtl_slim (lra_dump_file, new_insns, NULL, -1, 0);
5312 fprintf (lra_dump_file,
5313 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5314 }
5315 return false;
5316 }
5317 lra_substitute_pseudo_within_insn (insn, original_regno, new_reg, false);
5318 lra_update_insn_regno_info (insn);
5319 if (! def_p)
5320 /* We now have a new usage insn for original regno. */
5321 setup_next_usage_insn (original_regno, new_insns, reloads_num, false);
5322 if (lra_dump_file != NULL)
5323 fprintf (lra_dump_file, " Original reg change %d->%d (bb%d):\n",
5324 original_regno, REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5325 lra_reg_info[REGNO (new_reg)].restore_rtx = regno_reg_rtx[original_regno];
5326 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5327 bitmap_set_bit (&check_only_regs, original_regno);
5328 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5329 if (def_p)
5330 lra_process_new_insns (insn, NULL, new_insns,
5331 "Add original<-inheritance");
5332 else
5333 lra_process_new_insns (insn, new_insns, NULL,
5334 "Add inheritance<-original");
5335 while (next_usage_insns != NULL_RTX)
5336 {
5337 if (GET_CODE (next_usage_insns) != INSN_LIST)
5338 {
5339 usage_insn = next_usage_insns;
5340 lra_assert (NONDEBUG_INSN_P (usage_insn));
5341 next_usage_insns = NULL;
5342 }
5343 else
5344 {
5345 usage_insn = XEXP (next_usage_insns, 0);
5346 lra_assert (DEBUG_INSN_P (usage_insn));
5347 next_usage_insns = XEXP (next_usage_insns, 1);
5348 }
5349 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false,
5350 DEBUG_INSN_P (usage_insn));
5351 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5352 if (lra_dump_file != NULL)
5353 {
5354 basic_block bb = BLOCK_FOR_INSN (usage_insn);
5355 fprintf (lra_dump_file,
5356 " Inheritance reuse change %d->%d (bb%d):\n",
5357 original_regno, REGNO (new_reg),
5358 bb ? bb->index : -1);
5359 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5360 }
5361 }
5362 if (lra_dump_file != NULL)
5363 fprintf (lra_dump_file,
5364 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
5365 return true;
5366 }
5367
5368 /* Return true if we need a caller save/restore for pseudo REGNO which
5369 was assigned to a hard register. */
5370 static inline bool
5371 need_for_call_save_p (int regno)
5372 {
5373 lra_assert (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] >= 0);
5374 return (usage_insns[regno].calls_num < calls_num
5375 && (overlaps_hard_reg_set_p
5376 ((flag_ipa_ra &&
5377 ! hard_reg_set_empty_p (lra_reg_info[regno].actual_call_used_reg_set))
5378 ? lra_reg_info[regno].actual_call_used_reg_set
5379 : call_used_reg_set,
5380 PSEUDO_REGNO_MODE (regno), reg_renumber[regno])
5381 || (targetm.hard_regno_call_part_clobbered
5382 (reg_renumber[regno], PSEUDO_REGNO_MODE (regno)))));
5383 }
5384
5385 /* Global registers occurring in the current EBB. */
5386 static bitmap_head ebb_global_regs;
5387
5388 /* Return true if we need a split for hard register REGNO or pseudo
5389 REGNO which was assigned to a hard register.
5390 POTENTIAL_RELOAD_HARD_REGS contains hard registers which might be
5391 used for reloads since the EBB end. It is an approximation of the
5392 used hard registers in the split range. The exact value would
5393 require expensive calculations. If we were aggressive with
5394 splitting because of the approximation, the split pseudo will save
5395 the same hard register assignment and will be removed in the undo
5396 pass. We still need the approximation because too aggressive
5397 splitting would result in too inaccurate cost calculation in the
5398 assignment pass because of too many generated moves which will be
5399 probably removed in the undo pass. */
5400 static inline bool
5401 need_for_split_p (HARD_REG_SET potential_reload_hard_regs, int regno)
5402 {
5403 int hard_regno = regno < FIRST_PSEUDO_REGISTER ? regno : reg_renumber[regno];
5404
5405 lra_assert (hard_regno >= 0);
5406 return ((TEST_HARD_REG_BIT (potential_reload_hard_regs, hard_regno)
5407 /* Don't split eliminable hard registers, otherwise we can
5408 split hard registers like hard frame pointer, which
5409 lives on BB start/end according to DF-infrastructure,
5410 when there is a pseudo assigned to the register and
5411 living in the same BB. */
5412 && (regno >= FIRST_PSEUDO_REGISTER
5413 || ! TEST_HARD_REG_BIT (eliminable_regset, hard_regno))
5414 && ! TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno)
5415 /* Don't split call clobbered hard regs living through
5416 calls, otherwise we might have a check problem in the
5417 assign sub-pass as in the most cases (exception is a
5418 situation when lra_risky_transformations_p value is
5419 true) the assign pass assumes that all pseudos living
5420 through calls are assigned to call saved hard regs. */
5421 && (regno >= FIRST_PSEUDO_REGISTER
5422 || ! TEST_HARD_REG_BIT (call_used_reg_set, regno)
5423 || usage_insns[regno].calls_num == calls_num)
5424 /* We need at least 2 reloads to make pseudo splitting
5425 profitable. We should provide hard regno splitting in
5426 any case to solve 1st insn scheduling problem when
5427 moving hard register definition up might result in
5428 impossibility to find hard register for reload pseudo of
5429 small register class. */
5430 && (usage_insns[regno].reloads_num
5431 + (regno < FIRST_PSEUDO_REGISTER ? 0 : 3) < reloads_num)
5432 && (regno < FIRST_PSEUDO_REGISTER
5433 /* For short living pseudos, spilling + inheritance can
5434 be considered a substitution for splitting.
5435 Therefore we do not splitting for local pseudos. It
5436 decreases also aggressiveness of splitting. The
5437 minimal number of references is chosen taking into
5438 account that for 2 references splitting has no sense
5439 as we can just spill the pseudo. */
5440 || (regno >= FIRST_PSEUDO_REGISTER
5441 && lra_reg_info[regno].nrefs > 3
5442 && bitmap_bit_p (&ebb_global_regs, regno))))
5443 || (regno >= FIRST_PSEUDO_REGISTER && need_for_call_save_p (regno)));
5444 }
5445
5446 /* Return class for the split pseudo created from original pseudo with
5447 ALLOCNO_CLASS and MODE which got a hard register HARD_REGNO. We
5448 choose subclass of ALLOCNO_CLASS which contains HARD_REGNO and
5449 results in no secondary memory movements. */
5450 static enum reg_class
5451 choose_split_class (enum reg_class allocno_class,
5452 int hard_regno ATTRIBUTE_UNUSED,
5453 machine_mode mode ATTRIBUTE_UNUSED)
5454 {
5455 int i;
5456 enum reg_class cl, best_cl = NO_REGS;
5457 enum reg_class hard_reg_class ATTRIBUTE_UNUSED
5458 = REGNO_REG_CLASS (hard_regno);
5459
5460 if (! targetm.secondary_memory_needed (mode, allocno_class, allocno_class)
5461 && TEST_HARD_REG_BIT (reg_class_contents[allocno_class], hard_regno))
5462 return allocno_class;
5463 for (i = 0;
5464 (cl = reg_class_subclasses[allocno_class][i]) != LIM_REG_CLASSES;
5465 i++)
5466 if (! targetm.secondary_memory_needed (mode, cl, hard_reg_class)
5467 && ! targetm.secondary_memory_needed (mode, hard_reg_class, cl)
5468 && TEST_HARD_REG_BIT (reg_class_contents[cl], hard_regno)
5469 && (best_cl == NO_REGS
5470 || ira_class_hard_regs_num[best_cl] < ira_class_hard_regs_num[cl]))
5471 best_cl = cl;
5472 return best_cl;
5473 }
5474
5475 /* Copy any equivalence information from ORIGINAL_REGNO to NEW_REGNO.
5476 It only makes sense to call this function if NEW_REGNO is always
5477 equal to ORIGINAL_REGNO. */
5478
5479 static void
5480 lra_copy_reg_equiv (unsigned int new_regno, unsigned int original_regno)
5481 {
5482 if (!ira_reg_equiv[original_regno].defined_p)
5483 return;
5484
5485 ira_expand_reg_equiv ();
5486 ira_reg_equiv[new_regno].defined_p = true;
5487 if (ira_reg_equiv[original_regno].memory)
5488 ira_reg_equiv[new_regno].memory
5489 = copy_rtx (ira_reg_equiv[original_regno].memory);
5490 if (ira_reg_equiv[original_regno].constant)
5491 ira_reg_equiv[new_regno].constant
5492 = copy_rtx (ira_reg_equiv[original_regno].constant);
5493 if (ira_reg_equiv[original_regno].invariant)
5494 ira_reg_equiv[new_regno].invariant
5495 = copy_rtx (ira_reg_equiv[original_regno].invariant);
5496 }
5497
5498 /* Do split transformations for insn INSN, which defines or uses
5499 ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which instruction in
5500 the EBB next uses ORIGINAL_REGNO; it has the same form as the
5501 "insns" field of usage_insns. If TO is not NULL, we don't use
5502 usage_insns, we put restore insns after TO insn. It is a case when
5503 we call it from lra_split_hard_reg_for, outside the inheritance
5504 pass.
5505
5506 The transformations look like:
5507
5508 p <- ... p <- ...
5509 ... s <- p (new insn -- save)
5510 ... =>
5511 ... p <- s (new insn -- restore)
5512 <- ... p ... <- ... p ...
5513 or
5514 <- ... p ... <- ... p ...
5515 ... s <- p (new insn -- save)
5516 ... =>
5517 ... p <- s (new insn -- restore)
5518 <- ... p ... <- ... p ...
5519
5520 where p is an original pseudo got a hard register or a hard
5521 register and s is a new split pseudo. The save is put before INSN
5522 if BEFORE_P is true. Return true if we succeed in such
5523 transformation. */
5524 static bool
5525 split_reg (bool before_p, int original_regno, rtx_insn *insn,
5526 rtx next_usage_insns, rtx_insn *to)
5527 {
5528 enum reg_class rclass;
5529 rtx original_reg;
5530 int hard_regno, nregs;
5531 rtx new_reg, usage_insn;
5532 rtx_insn *restore, *save;
5533 bool after_p;
5534 bool call_save_p;
5535 machine_mode mode;
5536
5537 if (original_regno < FIRST_PSEUDO_REGISTER)
5538 {
5539 rclass = ira_allocno_class_translate[REGNO_REG_CLASS (original_regno)];
5540 hard_regno = original_regno;
5541 call_save_p = false;
5542 nregs = 1;
5543 mode = lra_reg_info[hard_regno].biggest_mode;
5544 machine_mode reg_rtx_mode = GET_MODE (regno_reg_rtx[hard_regno]);
5545 /* A reg can have a biggest_mode of VOIDmode if it was only ever seen
5546 as part of a multi-word register. In that case, or if the biggest
5547 mode was larger than a register, just use the reg_rtx. Otherwise,
5548 limit the size to that of the biggest access in the function. */
5549 if (mode == VOIDmode
5550 || paradoxical_subreg_p (mode, reg_rtx_mode))
5551 {
5552 original_reg = regno_reg_rtx[hard_regno];
5553 mode = reg_rtx_mode;
5554 }
5555 else
5556 original_reg = gen_rtx_REG (mode, hard_regno);
5557 }
5558 else
5559 {
5560 mode = PSEUDO_REGNO_MODE (original_regno);
5561 hard_regno = reg_renumber[original_regno];
5562 nregs = hard_regno_nregs (hard_regno, mode);
5563 rclass = lra_get_allocno_class (original_regno);
5564 original_reg = regno_reg_rtx[original_regno];
5565 call_save_p = need_for_call_save_p (original_regno);
5566 }
5567 lra_assert (hard_regno >= 0);
5568 if (lra_dump_file != NULL)
5569 fprintf (lra_dump_file,
5570 " ((((((((((((((((((((((((((((((((((((((((((((((((\n");
5571
5572 if (call_save_p)
5573 {
5574 mode = HARD_REGNO_CALLER_SAVE_MODE (hard_regno,
5575 hard_regno_nregs (hard_regno, mode),
5576 mode);
5577 new_reg = lra_create_new_reg (mode, NULL_RTX, NO_REGS, "save");
5578 }
5579 else
5580 {
5581 rclass = choose_split_class (rclass, hard_regno, mode);
5582 if (rclass == NO_REGS)
5583 {
5584 if (lra_dump_file != NULL)
5585 {
5586 fprintf (lra_dump_file,
5587 " Rejecting split of %d(%s): "
5588 "no good reg class for %d(%s)\n",
5589 original_regno,
5590 reg_class_names[lra_get_allocno_class (original_regno)],
5591 hard_regno,
5592 reg_class_names[REGNO_REG_CLASS (hard_regno)]);
5593 fprintf
5594 (lra_dump_file,
5595 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5596 }
5597 return false;
5598 }
5599 /* Split_if_necessary can split hard registers used as part of a
5600 multi-register mode but splits each register individually. The
5601 mode used for each independent register may not be supported
5602 so reject the split. Splitting the wider mode should theoretically
5603 be possible but is not implemented. */
5604 if (!targetm.hard_regno_mode_ok (hard_regno, mode))
5605 {
5606 if (lra_dump_file != NULL)
5607 {
5608 fprintf (lra_dump_file,
5609 " Rejecting split of %d(%s): unsuitable mode %s\n",
5610 original_regno,
5611 reg_class_names[lra_get_allocno_class (original_regno)],
5612 GET_MODE_NAME (mode));
5613 fprintf
5614 (lra_dump_file,
5615 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5616 }
5617 return false;
5618 }
5619 new_reg = lra_create_new_reg (mode, original_reg, rclass, "split");
5620 reg_renumber[REGNO (new_reg)] = hard_regno;
5621 }
5622 int new_regno = REGNO (new_reg);
5623 save = emit_spill_move (true, new_reg, original_reg);
5624 if (NEXT_INSN (save) != NULL_RTX && !call_save_p)
5625 {
5626 if (lra_dump_file != NULL)
5627 {
5628 fprintf
5629 (lra_dump_file,
5630 " Rejecting split %d->%d resulting in > 2 save insns:\n",
5631 original_regno, new_regno);
5632 dump_rtl_slim (lra_dump_file, save, NULL, -1, 0);
5633 fprintf (lra_dump_file,
5634 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5635 }
5636 return false;
5637 }
5638 restore = emit_spill_move (false, new_reg, original_reg);
5639 if (NEXT_INSN (restore) != NULL_RTX && !call_save_p)
5640 {
5641 if (lra_dump_file != NULL)
5642 {
5643 fprintf (lra_dump_file,
5644 " Rejecting split %d->%d "
5645 "resulting in > 2 restore insns:\n",
5646 original_regno, new_regno);
5647 dump_rtl_slim (lra_dump_file, restore, NULL, -1, 0);
5648 fprintf (lra_dump_file,
5649 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5650 }
5651 return false;
5652 }
5653 /* Transfer equivalence information to the spill register, so that
5654 if we fail to allocate the spill register, we have the option of
5655 rematerializing the original value instead of spilling to the stack. */
5656 if (!HARD_REGISTER_NUM_P (original_regno)
5657 && mode == PSEUDO_REGNO_MODE (original_regno))
5658 lra_copy_reg_equiv (new_regno, original_regno);
5659 lra_reg_info[new_regno].restore_rtx = regno_reg_rtx[original_regno];
5660 bitmap_set_bit (&lra_split_regs, new_regno);
5661 if (to != NULL)
5662 {
5663 lra_assert (next_usage_insns == NULL);
5664 usage_insn = to;
5665 after_p = TRUE;
5666 }
5667 else
5668 {
5669 /* We need check_only_regs only inside the inheritance pass. */
5670 bitmap_set_bit (&check_only_regs, new_regno);
5671 bitmap_set_bit (&check_only_regs, original_regno);
5672 after_p = usage_insns[original_regno].after_p;
5673 for (;;)
5674 {
5675 if (GET_CODE (next_usage_insns) != INSN_LIST)
5676 {
5677 usage_insn = next_usage_insns;
5678 break;
5679 }
5680 usage_insn = XEXP (next_usage_insns, 0);
5681 lra_assert (DEBUG_INSN_P (usage_insn));
5682 next_usage_insns = XEXP (next_usage_insns, 1);
5683 lra_substitute_pseudo (&usage_insn, original_regno, new_reg, false,
5684 true);
5685 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
5686 if (lra_dump_file != NULL)
5687 {
5688 fprintf (lra_dump_file, " Split reuse change %d->%d:\n",
5689 original_regno, new_regno);
5690 dump_insn_slim (lra_dump_file, as_a <rtx_insn *> (usage_insn));
5691 }
5692 }
5693 }
5694 lra_assert (NOTE_P (usage_insn) || NONDEBUG_INSN_P (usage_insn));
5695 lra_assert (usage_insn != insn || (after_p && before_p));
5696 lra_process_new_insns (as_a <rtx_insn *> (usage_insn),
5697 after_p ? NULL : restore,
5698 after_p ? restore : NULL,
5699 call_save_p
5700 ? "Add reg<-save" : "Add reg<-split");
5701 lra_process_new_insns (insn, before_p ? save : NULL,
5702 before_p ? NULL : save,
5703 call_save_p
5704 ? "Add save<-reg" : "Add split<-reg");
5705 if (nregs > 1)
5706 /* If we are trying to split multi-register. We should check
5707 conflicts on the next assignment sub-pass. IRA can allocate on
5708 sub-register levels, LRA do this on pseudos level right now and
5709 this discrepancy may create allocation conflicts after
5710 splitting. */
5711 lra_risky_transformations_p = true;
5712 if (lra_dump_file != NULL)
5713 fprintf (lra_dump_file,
5714 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
5715 return true;
5716 }
5717
5718 /* Split a hard reg for reload pseudo REGNO having RCLASS and living
5719 in the range [FROM, TO]. Return true if did a split. Otherwise,
5720 return false. */
5721 bool
5722 spill_hard_reg_in_range (int regno, enum reg_class rclass, rtx_insn *from, rtx_insn *to)
5723 {
5724 int i, hard_regno;
5725 int rclass_size;
5726 rtx_insn *insn;
5727 unsigned int uid;
5728 bitmap_iterator bi;
5729 HARD_REG_SET ignore;
5730
5731 lra_assert (from != NULL && to != NULL);
5732 CLEAR_HARD_REG_SET (ignore);
5733 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi)
5734 {
5735 lra_insn_recog_data_t id = lra_insn_recog_data[uid];
5736 struct lra_static_insn_data *static_id = id->insn_static_data;
5737 struct lra_insn_reg *reg;
5738
5739 for (reg = id->regs; reg != NULL; reg = reg->next)
5740 if (reg->regno < FIRST_PSEUDO_REGISTER)
5741 SET_HARD_REG_BIT (ignore, reg->regno);
5742 for (reg = static_id->hard_regs; reg != NULL; reg = reg->next)
5743 SET_HARD_REG_BIT (ignore, reg->regno);
5744 }
5745 rclass_size = ira_class_hard_regs_num[rclass];
5746 for (i = 0; i < rclass_size; i++)
5747 {
5748 hard_regno = ira_class_hard_regs[rclass][i];
5749 if (! TEST_HARD_REG_BIT (lra_reg_info[regno].conflict_hard_regs, hard_regno)
5750 || TEST_HARD_REG_BIT (ignore, hard_regno))
5751 continue;
5752 for (insn = from; insn != NEXT_INSN (to); insn = NEXT_INSN (insn))
5753 {
5754 struct lra_static_insn_data *static_id;
5755 struct lra_insn_reg *reg;
5756
5757 if (!INSN_P (insn))
5758 continue;
5759 if (bitmap_bit_p (&lra_reg_info[hard_regno].insn_bitmap,
5760 INSN_UID (insn)))
5761 break;
5762 static_id = lra_get_insn_recog_data (insn)->insn_static_data;
5763 for (reg = static_id->hard_regs; reg != NULL; reg = reg->next)
5764 if (reg->regno == hard_regno)
5765 break;
5766 if (reg != NULL)
5767 break;
5768 }
5769 if (insn != NEXT_INSN (to))
5770 continue;
5771 if (split_reg (TRUE, hard_regno, from, NULL, to))
5772 return true;
5773 }
5774 return false;
5775 }
5776
5777 /* Recognize that we need a split transformation for insn INSN, which
5778 defines or uses REGNO in its insn biggest MODE (we use it only if
5779 REGNO is a hard register). POTENTIAL_RELOAD_HARD_REGS contains
5780 hard registers which might be used for reloads since the EBB end.
5781 Put the save before INSN if BEFORE_P is true. MAX_UID is maximla
5782 uid before starting INSN processing. Return true if we succeed in
5783 such transformation. */
5784 static bool
5785 split_if_necessary (int regno, machine_mode mode,
5786 HARD_REG_SET potential_reload_hard_regs,
5787 bool before_p, rtx_insn *insn, int max_uid)
5788 {
5789 bool res = false;
5790 int i, nregs = 1;
5791 rtx next_usage_insns;
5792
5793 if (regno < FIRST_PSEUDO_REGISTER)
5794 nregs = hard_regno_nregs (regno, mode);
5795 for (i = 0; i < nregs; i++)
5796 if (usage_insns[regno + i].check == curr_usage_insns_check
5797 && (next_usage_insns = usage_insns[regno + i].insns) != NULL_RTX
5798 /* To avoid processing the register twice or more. */
5799 && ((GET_CODE (next_usage_insns) != INSN_LIST
5800 && INSN_UID (next_usage_insns) < max_uid)
5801 || (GET_CODE (next_usage_insns) == INSN_LIST
5802 && (INSN_UID (XEXP (next_usage_insns, 0)) < max_uid)))
5803 && need_for_split_p (potential_reload_hard_regs, regno + i)
5804 && split_reg (before_p, regno + i, insn, next_usage_insns, NULL))
5805 res = true;
5806 return res;
5807 }
5808
5809 /* Return TRUE if rtx X is considered as an invariant for
5810 inheritance. */
5811 static bool
5812 invariant_p (const_rtx x)
5813 {
5814 machine_mode mode;
5815 const char *fmt;
5816 enum rtx_code code;
5817 int i, j;
5818
5819 code = GET_CODE (x);
5820 mode = GET_MODE (x);
5821 if (code == SUBREG)
5822 {
5823 x = SUBREG_REG (x);
5824 code = GET_CODE (x);
5825 mode = wider_subreg_mode (mode, GET_MODE (x));
5826 }
5827
5828 if (MEM_P (x))
5829 return false;
5830
5831 if (REG_P (x))
5832 {
5833 int i, nregs, regno = REGNO (x);
5834
5835 if (regno >= FIRST_PSEUDO_REGISTER || regno == STACK_POINTER_REGNUM
5836 || TEST_HARD_REG_BIT (eliminable_regset, regno)
5837 || GET_MODE_CLASS (GET_MODE (x)) == MODE_CC)
5838 return false;
5839 nregs = hard_regno_nregs (regno, mode);
5840 for (i = 0; i < nregs; i++)
5841 if (! fixed_regs[regno + i]
5842 /* A hard register may be clobbered in the current insn
5843 but we can ignore this case because if the hard
5844 register is used it should be set somewhere after the
5845 clobber. */
5846 || bitmap_bit_p (&invalid_invariant_regs, regno + i))
5847 return false;
5848 }
5849 fmt = GET_RTX_FORMAT (code);
5850 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
5851 {
5852 if (fmt[i] == 'e')
5853 {
5854 if (! invariant_p (XEXP (x, i)))
5855 return false;
5856 }
5857 else if (fmt[i] == 'E')
5858 {
5859 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
5860 if (! invariant_p (XVECEXP (x, i, j)))
5861 return false;
5862 }
5863 }
5864 return true;
5865 }
5866
5867 /* We have 'dest_reg <- invariant'. Let us try to make an invariant
5868 inheritance transformation (using dest_reg instead invariant in a
5869 subsequent insn). */
5870 static bool
5871 process_invariant_for_inheritance (rtx dst_reg, rtx invariant_rtx)
5872 {
5873 invariant_ptr_t invariant_ptr;
5874 rtx_insn *insn, *new_insns;
5875 rtx insn_set, insn_reg, new_reg;
5876 int insn_regno;
5877 bool succ_p = false;
5878 int dst_regno = REGNO (dst_reg);
5879 machine_mode dst_mode = GET_MODE (dst_reg);
5880 enum reg_class cl = lra_get_allocno_class (dst_regno), insn_reg_cl;
5881
5882 invariant_ptr = insert_invariant (invariant_rtx);
5883 if ((insn = invariant_ptr->insn) != NULL_RTX)
5884 {
5885 /* We have a subsequent insn using the invariant. */
5886 insn_set = single_set (insn);
5887 lra_assert (insn_set != NULL);
5888 insn_reg = SET_DEST (insn_set);
5889 lra_assert (REG_P (insn_reg));
5890 insn_regno = REGNO (insn_reg);
5891 insn_reg_cl = lra_get_allocno_class (insn_regno);
5892
5893 if (dst_mode == GET_MODE (insn_reg)
5894 /* We should consider only result move reg insns which are
5895 cheap. */
5896 && targetm.register_move_cost (dst_mode, cl, insn_reg_cl) == 2
5897 && targetm.register_move_cost (dst_mode, cl, cl) == 2)
5898 {
5899 if (lra_dump_file != NULL)
5900 fprintf (lra_dump_file,
5901 " [[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[[\n");
5902 new_reg = lra_create_new_reg (dst_mode, dst_reg,
5903 cl, "invariant inheritance");
5904 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
5905 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
5906 lra_reg_info[REGNO (new_reg)].restore_rtx = PATTERN (insn);
5907 start_sequence ();
5908 lra_emit_move (new_reg, dst_reg);
5909 new_insns = get_insns ();
5910 end_sequence ();
5911 lra_process_new_insns (curr_insn, NULL, new_insns,
5912 "Add invariant inheritance<-original");
5913 start_sequence ();
5914 lra_emit_move (SET_DEST (insn_set), new_reg);
5915 new_insns = get_insns ();
5916 end_sequence ();
5917 lra_process_new_insns (insn, NULL, new_insns,
5918 "Changing reload<-inheritance");
5919 lra_set_insn_deleted (insn);
5920 succ_p = true;
5921 if (lra_dump_file != NULL)
5922 {
5923 fprintf (lra_dump_file,
5924 " Invariant inheritance reuse change %d (bb%d):\n",
5925 REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
5926 dump_insn_slim (lra_dump_file, insn);
5927 fprintf (lra_dump_file,
5928 " ]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]]\n");
5929 }
5930 }
5931 }
5932 invariant_ptr->insn = curr_insn;
5933 return succ_p;
5934 }
5935
5936 /* Check only registers living at the current program point in the
5937 current EBB. */
5938 static bitmap_head live_regs;
5939
5940 /* Update live info in EBB given by its HEAD and TAIL insns after
5941 inheritance/split transformation. The function removes dead moves
5942 too. */
5943 static void
5944 update_ebb_live_info (rtx_insn *head, rtx_insn *tail)
5945 {
5946 unsigned int j;
5947 int i, regno;
5948 bool live_p;
5949 rtx_insn *prev_insn;
5950 rtx set;
5951 bool remove_p;
5952 basic_block last_bb, prev_bb, curr_bb;
5953 bitmap_iterator bi;
5954 struct lra_insn_reg *reg;
5955 edge e;
5956 edge_iterator ei;
5957
5958 last_bb = BLOCK_FOR_INSN (tail);
5959 prev_bb = NULL;
5960 for (curr_insn = tail;
5961 curr_insn != PREV_INSN (head);
5962 curr_insn = prev_insn)
5963 {
5964 prev_insn = PREV_INSN (curr_insn);
5965 /* We need to process empty blocks too. They contain
5966 NOTE_INSN_BASIC_BLOCK referring for the basic block. */
5967 if (NOTE_P (curr_insn) && NOTE_KIND (curr_insn) != NOTE_INSN_BASIC_BLOCK)
5968 continue;
5969 curr_bb = BLOCK_FOR_INSN (curr_insn);
5970 if (curr_bb != prev_bb)
5971 {
5972 if (prev_bb != NULL)
5973 {
5974 /* Update df_get_live_in (prev_bb): */
5975 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5976 if (bitmap_bit_p (&live_regs, j))
5977 bitmap_set_bit (df_get_live_in (prev_bb), j);
5978 else
5979 bitmap_clear_bit (df_get_live_in (prev_bb), j);
5980 }
5981 if (curr_bb != last_bb)
5982 {
5983 /* Update df_get_live_out (curr_bb): */
5984 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5985 {
5986 live_p = bitmap_bit_p (&live_regs, j);
5987 if (! live_p)
5988 FOR_EACH_EDGE (e, ei, curr_bb->succs)
5989 if (bitmap_bit_p (df_get_live_in (e->dest), j))
5990 {
5991 live_p = true;
5992 break;
5993 }
5994 if (live_p)
5995 bitmap_set_bit (df_get_live_out (curr_bb), j);
5996 else
5997 bitmap_clear_bit (df_get_live_out (curr_bb), j);
5998 }
5999 }
6000 prev_bb = curr_bb;
6001 bitmap_and (&live_regs, &check_only_regs, df_get_live_out (curr_bb));
6002 }
6003 if (! NONDEBUG_INSN_P (curr_insn))
6004 continue;
6005 curr_id = lra_get_insn_recog_data (curr_insn);
6006 curr_static_id = curr_id->insn_static_data;
6007 remove_p = false;
6008 if ((set = single_set (curr_insn)) != NULL_RTX
6009 && REG_P (SET_DEST (set))
6010 && (regno = REGNO (SET_DEST (set))) >= FIRST_PSEUDO_REGISTER
6011 && SET_DEST (set) != pic_offset_table_rtx
6012 && bitmap_bit_p (&check_only_regs, regno)
6013 && ! bitmap_bit_p (&live_regs, regno))
6014 remove_p = true;
6015 /* See which defined values die here. */
6016 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6017 if (reg->type == OP_OUT && ! reg->subreg_p)
6018 bitmap_clear_bit (&live_regs, reg->regno);
6019 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
6020 if (reg->type == OP_OUT && ! reg->subreg_p)
6021 bitmap_clear_bit (&live_regs, reg->regno);
6022 if (curr_id->arg_hard_regs != NULL)
6023 /* Make clobbered argument hard registers die. */
6024 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6025 if (regno >= FIRST_PSEUDO_REGISTER)
6026 bitmap_clear_bit (&live_regs, regno - FIRST_PSEUDO_REGISTER);
6027 /* Mark each used value as live. */
6028 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6029 if (reg->type != OP_OUT
6030 && bitmap_bit_p (&check_only_regs, reg->regno))
6031 bitmap_set_bit (&live_regs, reg->regno);
6032 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
6033 if (reg->type != OP_OUT
6034 && bitmap_bit_p (&check_only_regs, reg->regno))
6035 bitmap_set_bit (&live_regs, reg->regno);
6036 if (curr_id->arg_hard_regs != NULL)
6037 /* Make used argument hard registers live. */
6038 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6039 if (regno < FIRST_PSEUDO_REGISTER
6040 && bitmap_bit_p (&check_only_regs, regno))
6041 bitmap_set_bit (&live_regs, regno);
6042 /* It is quite important to remove dead move insns because it
6043 means removing dead store. We don't need to process them for
6044 constraints. */
6045 if (remove_p)
6046 {
6047 if (lra_dump_file != NULL)
6048 {
6049 fprintf (lra_dump_file, " Removing dead insn:\n ");
6050 dump_insn_slim (lra_dump_file, curr_insn);
6051 }
6052 lra_set_insn_deleted (curr_insn);
6053 }
6054 }
6055 }
6056
6057 /* The structure describes info to do an inheritance for the current
6058 insn. We need to collect such info first before doing the
6059 transformations because the transformations change the insn
6060 internal representation. */
6061 struct to_inherit
6062 {
6063 /* Original regno. */
6064 int regno;
6065 /* Subsequent insns which can inherit original reg value. */
6066 rtx insns;
6067 };
6068
6069 /* Array containing all info for doing inheritance from the current
6070 insn. */
6071 static struct to_inherit to_inherit[LRA_MAX_INSN_RELOADS];
6072
6073 /* Number elements in the previous array. */
6074 static int to_inherit_num;
6075
6076 /* Add inheritance info REGNO and INSNS. Their meaning is described in
6077 structure to_inherit. */
6078 static void
6079 add_to_inherit (int regno, rtx insns)
6080 {
6081 int i;
6082
6083 for (i = 0; i < to_inherit_num; i++)
6084 if (to_inherit[i].regno == regno)
6085 return;
6086 lra_assert (to_inherit_num < LRA_MAX_INSN_RELOADS);
6087 to_inherit[to_inherit_num].regno = regno;
6088 to_inherit[to_inherit_num++].insns = insns;
6089 }
6090
6091 /* Return the last non-debug insn in basic block BB, or the block begin
6092 note if none. */
6093 static rtx_insn *
6094 get_last_insertion_point (basic_block bb)
6095 {
6096 rtx_insn *insn;
6097
6098 FOR_BB_INSNS_REVERSE (bb, insn)
6099 if (NONDEBUG_INSN_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn))
6100 return insn;
6101 gcc_unreachable ();
6102 }
6103
6104 /* Set up RES by registers living on edges FROM except the edge (FROM,
6105 TO) or by registers set up in a jump insn in BB FROM. */
6106 static void
6107 get_live_on_other_edges (basic_block from, basic_block to, bitmap res)
6108 {
6109 rtx_insn *last;
6110 struct lra_insn_reg *reg;
6111 edge e;
6112 edge_iterator ei;
6113
6114 lra_assert (to != NULL);
6115 bitmap_clear (res);
6116 FOR_EACH_EDGE (e, ei, from->succs)
6117 if (e->dest != to)
6118 bitmap_ior_into (res, df_get_live_in (e->dest));
6119 last = get_last_insertion_point (from);
6120 if (! JUMP_P (last))
6121 return;
6122 curr_id = lra_get_insn_recog_data (last);
6123 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6124 if (reg->type != OP_IN)
6125 bitmap_set_bit (res, reg->regno);
6126 }
6127
6128 /* Used as a temporary results of some bitmap calculations. */
6129 static bitmap_head temp_bitmap;
6130
6131 /* We split for reloads of small class of hard regs. The following
6132 defines how many hard regs the class should have to be qualified as
6133 small. The code is mostly oriented to x86/x86-64 architecture
6134 where some insns need to use only specific register or pair of
6135 registers and these register can live in RTL explicitly, e.g. for
6136 parameter passing. */
6137 static const int max_small_class_regs_num = 2;
6138
6139 /* Do inheritance/split transformations in EBB starting with HEAD and
6140 finishing on TAIL. We process EBB insns in the reverse order.
6141 Return true if we did any inheritance/split transformation in the
6142 EBB.
6143
6144 We should avoid excessive splitting which results in worse code
6145 because of inaccurate cost calculations for spilling new split
6146 pseudos in such case. To achieve this we do splitting only if
6147 register pressure is high in given basic block and there are reload
6148 pseudos requiring hard registers. We could do more register
6149 pressure calculations at any given program point to avoid necessary
6150 splitting even more but it is to expensive and the current approach
6151 works well enough. */
6152 static bool
6153 inherit_in_ebb (rtx_insn *head, rtx_insn *tail)
6154 {
6155 int i, src_regno, dst_regno, nregs;
6156 bool change_p, succ_p, update_reloads_num_p;
6157 rtx_insn *prev_insn, *last_insn;
6158 rtx next_usage_insns, curr_set;
6159 enum reg_class cl;
6160 struct lra_insn_reg *reg;
6161 basic_block last_processed_bb, curr_bb = NULL;
6162 HARD_REG_SET potential_reload_hard_regs, live_hard_regs;
6163 bitmap to_process;
6164 unsigned int j;
6165 bitmap_iterator bi;
6166 bool head_p, after_p;
6167
6168 change_p = false;
6169 curr_usage_insns_check++;
6170 clear_invariants ();
6171 reloads_num = calls_num = 0;
6172 bitmap_clear (&check_only_regs);
6173 bitmap_clear (&invalid_invariant_regs);
6174 last_processed_bb = NULL;
6175 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
6176 COPY_HARD_REG_SET (live_hard_regs, eliminable_regset);
6177 IOR_HARD_REG_SET (live_hard_regs, lra_no_alloc_regs);
6178 /* We don't process new insns generated in the loop. */
6179 for (curr_insn = tail; curr_insn != PREV_INSN (head); curr_insn = prev_insn)
6180 {
6181 prev_insn = PREV_INSN (curr_insn);
6182 if (BLOCK_FOR_INSN (curr_insn) != NULL)
6183 curr_bb = BLOCK_FOR_INSN (curr_insn);
6184 if (last_processed_bb != curr_bb)
6185 {
6186 /* We are at the end of BB. Add qualified living
6187 pseudos for potential splitting. */
6188 to_process = df_get_live_out (curr_bb);
6189 if (last_processed_bb != NULL)
6190 {
6191 /* We are somewhere in the middle of EBB. */
6192 get_live_on_other_edges (curr_bb, last_processed_bb,
6193 &temp_bitmap);
6194 to_process = &temp_bitmap;
6195 }
6196 last_processed_bb = curr_bb;
6197 last_insn = get_last_insertion_point (curr_bb);
6198 after_p = (! JUMP_P (last_insn)
6199 && (! CALL_P (last_insn)
6200 || (find_reg_note (last_insn,
6201 REG_NORETURN, NULL_RTX) == NULL_RTX
6202 && ! SIBLING_CALL_P (last_insn))));
6203 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
6204 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6205 {
6206 if ((int) j >= lra_constraint_new_regno_start)
6207 break;
6208 if (j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6209 {
6210 if (j < FIRST_PSEUDO_REGISTER)
6211 SET_HARD_REG_BIT (live_hard_regs, j);
6212 else
6213 add_to_hard_reg_set (&live_hard_regs,
6214 PSEUDO_REGNO_MODE (j),
6215 reg_renumber[j]);
6216 setup_next_usage_insn (j, last_insn, reloads_num, after_p);
6217 }
6218 }
6219 }
6220 src_regno = dst_regno = -1;
6221 curr_set = single_set (curr_insn);
6222 if (curr_set != NULL_RTX && REG_P (SET_DEST (curr_set)))
6223 dst_regno = REGNO (SET_DEST (curr_set));
6224 if (curr_set != NULL_RTX && REG_P (SET_SRC (curr_set)))
6225 src_regno = REGNO (SET_SRC (curr_set));
6226 update_reloads_num_p = true;
6227 if (src_regno < lra_constraint_new_regno_start
6228 && src_regno >= FIRST_PSEUDO_REGISTER
6229 && reg_renumber[src_regno] < 0
6230 && dst_regno >= lra_constraint_new_regno_start
6231 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS)
6232 {
6233 /* 'reload_pseudo <- original_pseudo'. */
6234 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6235 reloads_num++;
6236 update_reloads_num_p = false;
6237 succ_p = false;
6238 if (usage_insns[src_regno].check == curr_usage_insns_check
6239 && (next_usage_insns = usage_insns[src_regno].insns) != NULL_RTX)
6240 succ_p = inherit_reload_reg (false, src_regno, cl,
6241 curr_insn, next_usage_insns);
6242 if (succ_p)
6243 change_p = true;
6244 else
6245 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6246 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6247 IOR_HARD_REG_SET (potential_reload_hard_regs,
6248 reg_class_contents[cl]);
6249 }
6250 else if (src_regno < 0
6251 && dst_regno >= lra_constraint_new_regno_start
6252 && invariant_p (SET_SRC (curr_set))
6253 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS
6254 && ! bitmap_bit_p (&invalid_invariant_regs, dst_regno)
6255 && ! bitmap_bit_p (&invalid_invariant_regs,
6256 ORIGINAL_REGNO(regno_reg_rtx[dst_regno])))
6257 {
6258 /* 'reload_pseudo <- invariant'. */
6259 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6260 reloads_num++;
6261 update_reloads_num_p = false;
6262 if (process_invariant_for_inheritance (SET_DEST (curr_set), SET_SRC (curr_set)))
6263 change_p = true;
6264 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6265 IOR_HARD_REG_SET (potential_reload_hard_regs,
6266 reg_class_contents[cl]);
6267 }
6268 else if (src_regno >= lra_constraint_new_regno_start
6269 && dst_regno < lra_constraint_new_regno_start
6270 && dst_regno >= FIRST_PSEUDO_REGISTER
6271 && reg_renumber[dst_regno] < 0
6272 && (cl = lra_get_allocno_class (src_regno)) != NO_REGS
6273 && usage_insns[dst_regno].check == curr_usage_insns_check
6274 && (next_usage_insns
6275 = usage_insns[dst_regno].insns) != NULL_RTX)
6276 {
6277 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6278 reloads_num++;
6279 update_reloads_num_p = false;
6280 /* 'original_pseudo <- reload_pseudo'. */
6281 if (! JUMP_P (curr_insn)
6282 && inherit_reload_reg (true, dst_regno, cl,
6283 curr_insn, next_usage_insns))
6284 change_p = true;
6285 /* Invalidate. */
6286 usage_insns[dst_regno].check = 0;
6287 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6288 IOR_HARD_REG_SET (potential_reload_hard_regs,
6289 reg_class_contents[cl]);
6290 }
6291 else if (INSN_P (curr_insn))
6292 {
6293 int iter;
6294 int max_uid = get_max_uid ();
6295
6296 curr_id = lra_get_insn_recog_data (curr_insn);
6297 curr_static_id = curr_id->insn_static_data;
6298 to_inherit_num = 0;
6299 /* Process insn definitions. */
6300 for (iter = 0; iter < 2; iter++)
6301 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6302 reg != NULL;
6303 reg = reg->next)
6304 if (reg->type != OP_IN
6305 && (dst_regno = reg->regno) < lra_constraint_new_regno_start)
6306 {
6307 if (dst_regno >= FIRST_PSEUDO_REGISTER && reg->type == OP_OUT
6308 && reg_renumber[dst_regno] < 0 && ! reg->subreg_p
6309 && usage_insns[dst_regno].check == curr_usage_insns_check
6310 && (next_usage_insns
6311 = usage_insns[dst_regno].insns) != NULL_RTX)
6312 {
6313 struct lra_insn_reg *r;
6314
6315 for (r = curr_id->regs; r != NULL; r = r->next)
6316 if (r->type != OP_OUT && r->regno == dst_regno)
6317 break;
6318 /* Don't do inheritance if the pseudo is also
6319 used in the insn. */
6320 if (r == NULL)
6321 /* We cannot do inheritance right now
6322 because the current insn reg info (chain
6323 regs) can change after that. */
6324 add_to_inherit (dst_regno, next_usage_insns);
6325 }
6326 /* We cannot process one reg twice here because of
6327 usage_insns invalidation. */
6328 if ((dst_regno < FIRST_PSEUDO_REGISTER
6329 || reg_renumber[dst_regno] >= 0)
6330 && ! reg->subreg_p && reg->type != OP_IN)
6331 {
6332 HARD_REG_SET s;
6333
6334 if (split_if_necessary (dst_regno, reg->biggest_mode,
6335 potential_reload_hard_regs,
6336 false, curr_insn, max_uid))
6337 change_p = true;
6338 CLEAR_HARD_REG_SET (s);
6339 if (dst_regno < FIRST_PSEUDO_REGISTER)
6340 add_to_hard_reg_set (&s, reg->biggest_mode, dst_regno);
6341 else
6342 add_to_hard_reg_set (&s, PSEUDO_REGNO_MODE (dst_regno),
6343 reg_renumber[dst_regno]);
6344 AND_COMPL_HARD_REG_SET (live_hard_regs, s);
6345 }
6346 /* We should invalidate potential inheritance or
6347 splitting for the current insn usages to the next
6348 usage insns (see code below) as the output pseudo
6349 prevents this. */
6350 if ((dst_regno >= FIRST_PSEUDO_REGISTER
6351 && reg_renumber[dst_regno] < 0)
6352 || (reg->type == OP_OUT && ! reg->subreg_p
6353 && (dst_regno < FIRST_PSEUDO_REGISTER
6354 || reg_renumber[dst_regno] >= 0)))
6355 {
6356 /* Invalidate and mark definitions. */
6357 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6358 usage_insns[dst_regno].check = -(int) INSN_UID (curr_insn);
6359 else
6360 {
6361 nregs = hard_regno_nregs (dst_regno,
6362 reg->biggest_mode);
6363 for (i = 0; i < nregs; i++)
6364 usage_insns[dst_regno + i].check
6365 = -(int) INSN_UID (curr_insn);
6366 }
6367 }
6368 }
6369 /* Process clobbered call regs. */
6370 if (curr_id->arg_hard_regs != NULL)
6371 for (i = 0; (dst_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6372 if (dst_regno >= FIRST_PSEUDO_REGISTER)
6373 usage_insns[dst_regno - FIRST_PSEUDO_REGISTER].check
6374 = -(int) INSN_UID (curr_insn);
6375 if (! JUMP_P (curr_insn))
6376 for (i = 0; i < to_inherit_num; i++)
6377 if (inherit_reload_reg (true, to_inherit[i].regno,
6378 ALL_REGS, curr_insn,
6379 to_inherit[i].insns))
6380 change_p = true;
6381 if (CALL_P (curr_insn))
6382 {
6383 rtx cheap, pat, dest;
6384 rtx_insn *restore;
6385 int regno, hard_regno;
6386
6387 calls_num++;
6388 if ((cheap = find_reg_note (curr_insn,
6389 REG_RETURNED, NULL_RTX)) != NULL_RTX
6390 && ((cheap = XEXP (cheap, 0)), true)
6391 && (regno = REGNO (cheap)) >= FIRST_PSEUDO_REGISTER
6392 && (hard_regno = reg_renumber[regno]) >= 0
6393 && usage_insns[regno].check == curr_usage_insns_check
6394 /* If there are pending saves/restores, the
6395 optimization is not worth. */
6396 && usage_insns[regno].calls_num == calls_num - 1
6397 && TEST_HARD_REG_BIT (call_used_reg_set, hard_regno))
6398 {
6399 /* Restore the pseudo from the call result as
6400 REG_RETURNED note says that the pseudo value is
6401 in the call result and the pseudo is an argument
6402 of the call. */
6403 pat = PATTERN (curr_insn);
6404 if (GET_CODE (pat) == PARALLEL)
6405 pat = XVECEXP (pat, 0, 0);
6406 dest = SET_DEST (pat);
6407 /* For multiple return values dest is PARALLEL.
6408 Currently we handle only single return value case. */
6409 if (REG_P (dest))
6410 {
6411 start_sequence ();
6412 emit_move_insn (cheap, copy_rtx (dest));
6413 restore = get_insns ();
6414 end_sequence ();
6415 lra_process_new_insns (curr_insn, NULL, restore,
6416 "Inserting call parameter restore");
6417 /* We don't need to save/restore of the pseudo from
6418 this call. */
6419 usage_insns[regno].calls_num = calls_num;
6420 bitmap_set_bit (&check_only_regs, regno);
6421 }
6422 }
6423 }
6424 to_inherit_num = 0;
6425 /* Process insn usages. */
6426 for (iter = 0; iter < 2; iter++)
6427 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
6428 reg != NULL;
6429 reg = reg->next)
6430 if ((reg->type != OP_OUT
6431 || (reg->type == OP_OUT && reg->subreg_p))
6432 && (src_regno = reg->regno) < lra_constraint_new_regno_start)
6433 {
6434 if (src_regno >= FIRST_PSEUDO_REGISTER
6435 && reg_renumber[src_regno] < 0 && reg->type == OP_IN)
6436 {
6437 if (usage_insns[src_regno].check == curr_usage_insns_check
6438 && (next_usage_insns
6439 = usage_insns[src_regno].insns) != NULL_RTX
6440 && NONDEBUG_INSN_P (curr_insn))
6441 add_to_inherit (src_regno, next_usage_insns);
6442 else if (usage_insns[src_regno].check
6443 != -(int) INSN_UID (curr_insn))
6444 /* Add usages but only if the reg is not set up
6445 in the same insn. */
6446 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6447 }
6448 else if (src_regno < FIRST_PSEUDO_REGISTER
6449 || reg_renumber[src_regno] >= 0)
6450 {
6451 bool before_p;
6452 rtx_insn *use_insn = curr_insn;
6453
6454 before_p = (JUMP_P (curr_insn)
6455 || (CALL_P (curr_insn) && reg->type == OP_IN));
6456 if (NONDEBUG_INSN_P (curr_insn)
6457 && (! JUMP_P (curr_insn) || reg->type == OP_IN)
6458 && split_if_necessary (src_regno, reg->biggest_mode,
6459 potential_reload_hard_regs,
6460 before_p, curr_insn, max_uid))
6461 {
6462 if (reg->subreg_p)
6463 lra_risky_transformations_p = true;
6464 change_p = true;
6465 /* Invalidate. */
6466 usage_insns[src_regno].check = 0;
6467 if (before_p)
6468 use_insn = PREV_INSN (curr_insn);
6469 }
6470 if (NONDEBUG_INSN_P (curr_insn))
6471 {
6472 if (src_regno < FIRST_PSEUDO_REGISTER)
6473 add_to_hard_reg_set (&live_hard_regs,
6474 reg->biggest_mode, src_regno);
6475 else
6476 add_to_hard_reg_set (&live_hard_regs,
6477 PSEUDO_REGNO_MODE (src_regno),
6478 reg_renumber[src_regno]);
6479 }
6480 if (src_regno >= FIRST_PSEUDO_REGISTER)
6481 add_next_usage_insn (src_regno, use_insn, reloads_num);
6482 else
6483 {
6484 for (i = 0; i < hard_regno_nregs (src_regno, reg->biggest_mode); i++)
6485 add_next_usage_insn (src_regno + i, use_insn, reloads_num);
6486 }
6487 }
6488 }
6489 /* Process used call regs. */
6490 if (curr_id->arg_hard_regs != NULL)
6491 for (i = 0; (src_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6492 if (src_regno < FIRST_PSEUDO_REGISTER)
6493 {
6494 SET_HARD_REG_BIT (live_hard_regs, src_regno);
6495 add_next_usage_insn (src_regno, curr_insn, reloads_num);
6496 }
6497 for (i = 0; i < to_inherit_num; i++)
6498 {
6499 src_regno = to_inherit[i].regno;
6500 if (inherit_reload_reg (false, src_regno, ALL_REGS,
6501 curr_insn, to_inherit[i].insns))
6502 change_p = true;
6503 else
6504 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
6505 }
6506 }
6507 if (update_reloads_num_p
6508 && NONDEBUG_INSN_P (curr_insn) && curr_set != NULL_RTX)
6509 {
6510 int regno = -1;
6511 if ((REG_P (SET_DEST (curr_set))
6512 && (regno = REGNO (SET_DEST (curr_set))) >= lra_constraint_new_regno_start
6513 && reg_renumber[regno] < 0
6514 && (cl = lra_get_allocno_class (regno)) != NO_REGS)
6515 || (REG_P (SET_SRC (curr_set))
6516 && (regno = REGNO (SET_SRC (curr_set))) >= lra_constraint_new_regno_start
6517 && reg_renumber[regno] < 0
6518 && (cl = lra_get_allocno_class (regno)) != NO_REGS))
6519 {
6520 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
6521 reloads_num++;
6522 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
6523 IOR_HARD_REG_SET (potential_reload_hard_regs,
6524 reg_class_contents[cl]);
6525 }
6526 }
6527 if (NONDEBUG_INSN_P (curr_insn))
6528 {
6529 int regno;
6530
6531 /* Invalidate invariants with changed regs. */
6532 curr_id = lra_get_insn_recog_data (curr_insn);
6533 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6534 if (reg->type != OP_IN)
6535 {
6536 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6537 bitmap_set_bit (&invalid_invariant_regs,
6538 ORIGINAL_REGNO (regno_reg_rtx[reg->regno]));
6539 }
6540 curr_static_id = curr_id->insn_static_data;
6541 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
6542 if (reg->type != OP_IN)
6543 bitmap_set_bit (&invalid_invariant_regs, reg->regno);
6544 if (curr_id->arg_hard_regs != NULL)
6545 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
6546 if (regno >= FIRST_PSEUDO_REGISTER)
6547 bitmap_set_bit (&invalid_invariant_regs,
6548 regno - FIRST_PSEUDO_REGISTER);
6549 }
6550 /* We reached the start of the current basic block. */
6551 if (prev_insn == NULL_RTX || prev_insn == PREV_INSN (head)
6552 || BLOCK_FOR_INSN (prev_insn) != curr_bb)
6553 {
6554 /* We reached the beginning of the current block -- do
6555 rest of spliting in the current BB. */
6556 to_process = df_get_live_in (curr_bb);
6557 if (BLOCK_FOR_INSN (head) != curr_bb)
6558 {
6559 /* We are somewhere in the middle of EBB. */
6560 get_live_on_other_edges (EDGE_PRED (curr_bb, 0)->src,
6561 curr_bb, &temp_bitmap);
6562 to_process = &temp_bitmap;
6563 }
6564 head_p = true;
6565 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
6566 {
6567 if ((int) j >= lra_constraint_new_regno_start)
6568 break;
6569 if (((int) j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
6570 && usage_insns[j].check == curr_usage_insns_check
6571 && (next_usage_insns = usage_insns[j].insns) != NULL_RTX)
6572 {
6573 if (need_for_split_p (potential_reload_hard_regs, j))
6574 {
6575 if (lra_dump_file != NULL && head_p)
6576 {
6577 fprintf (lra_dump_file,
6578 " ----------------------------------\n");
6579 head_p = false;
6580 }
6581 if (split_reg (false, j, bb_note (curr_bb),
6582 next_usage_insns, NULL))
6583 change_p = true;
6584 }
6585 usage_insns[j].check = 0;
6586 }
6587 }
6588 }
6589 }
6590 return change_p;
6591 }
6592
6593 /* This value affects EBB forming. If probability of edge from EBB to
6594 a BB is not greater than the following value, we don't add the BB
6595 to EBB. */
6596 #define EBB_PROBABILITY_CUTOFF \
6597 ((REG_BR_PROB_BASE * LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF) / 100)
6598
6599 /* Current number of inheritance/split iteration. */
6600 int lra_inheritance_iter;
6601
6602 /* Entry function for inheritance/split pass. */
6603 void
6604 lra_inheritance (void)
6605 {
6606 int i;
6607 basic_block bb, start_bb;
6608 edge e;
6609
6610 lra_inheritance_iter++;
6611 if (lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6612 return;
6613 timevar_push (TV_LRA_INHERITANCE);
6614 if (lra_dump_file != NULL)
6615 fprintf (lra_dump_file, "\n********** Inheritance #%d: **********\n\n",
6616 lra_inheritance_iter);
6617 curr_usage_insns_check = 0;
6618 usage_insns = XNEWVEC (struct usage_insns, lra_constraint_new_regno_start);
6619 for (i = 0; i < lra_constraint_new_regno_start; i++)
6620 usage_insns[i].check = 0;
6621 bitmap_initialize (&check_only_regs, &reg_obstack);
6622 bitmap_initialize (&invalid_invariant_regs, &reg_obstack);
6623 bitmap_initialize (&live_regs, &reg_obstack);
6624 bitmap_initialize (&temp_bitmap, &reg_obstack);
6625 bitmap_initialize (&ebb_global_regs, &reg_obstack);
6626 FOR_EACH_BB_FN (bb, cfun)
6627 {
6628 start_bb = bb;
6629 if (lra_dump_file != NULL)
6630 fprintf (lra_dump_file, "EBB");
6631 /* Form a EBB starting with BB. */
6632 bitmap_clear (&ebb_global_regs);
6633 bitmap_ior_into (&ebb_global_regs, df_get_live_in (bb));
6634 for (;;)
6635 {
6636 if (lra_dump_file != NULL)
6637 fprintf (lra_dump_file, " %d", bb->index);
6638 if (bb->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
6639 || LABEL_P (BB_HEAD (bb->next_bb)))
6640 break;
6641 e = find_fallthru_edge (bb->succs);
6642 if (! e)
6643 break;
6644 if (e->probability.initialized_p ()
6645 && e->probability.to_reg_br_prob_base () < EBB_PROBABILITY_CUTOFF)
6646 break;
6647 bb = bb->next_bb;
6648 }
6649 bitmap_ior_into (&ebb_global_regs, df_get_live_out (bb));
6650 if (lra_dump_file != NULL)
6651 fprintf (lra_dump_file, "\n");
6652 if (inherit_in_ebb (BB_HEAD (start_bb), BB_END (bb)))
6653 /* Remember that the EBB head and tail can change in
6654 inherit_in_ebb. */
6655 update_ebb_live_info (BB_HEAD (start_bb), BB_END (bb));
6656 }
6657 bitmap_release (&ebb_global_regs);
6658 bitmap_release (&temp_bitmap);
6659 bitmap_release (&live_regs);
6660 bitmap_release (&invalid_invariant_regs);
6661 bitmap_release (&check_only_regs);
6662 free (usage_insns);
6663
6664 timevar_pop (TV_LRA_INHERITANCE);
6665 }
6666
6667 \f
6668
6669 /* This page contains code to undo failed inheritance/split
6670 transformations. */
6671
6672 /* Current number of iteration undoing inheritance/split. */
6673 int lra_undo_inheritance_iter;
6674
6675 /* Fix BB live info LIVE after removing pseudos created on pass doing
6676 inheritance/split which are REMOVED_PSEUDOS. */
6677 static void
6678 fix_bb_live_info (bitmap live, bitmap removed_pseudos)
6679 {
6680 unsigned int regno;
6681 bitmap_iterator bi;
6682
6683 EXECUTE_IF_SET_IN_BITMAP (removed_pseudos, 0, regno, bi)
6684 if (bitmap_clear_bit (live, regno)
6685 && REG_P (lra_reg_info[regno].restore_rtx))
6686 bitmap_set_bit (live, REGNO (lra_reg_info[regno].restore_rtx));
6687 }
6688
6689 /* Return regno of the (subreg of) REG. Otherwise, return a negative
6690 number. */
6691 static int
6692 get_regno (rtx reg)
6693 {
6694 if (GET_CODE (reg) == SUBREG)
6695 reg = SUBREG_REG (reg);
6696 if (REG_P (reg))
6697 return REGNO (reg);
6698 return -1;
6699 }
6700
6701 /* Delete a move INSN with destination reg DREGNO and a previous
6702 clobber insn with the same regno. The inheritance/split code can
6703 generate moves with preceding clobber and when we delete such moves
6704 we should delete the clobber insn too to keep the correct life
6705 info. */
6706 static void
6707 delete_move_and_clobber (rtx_insn *insn, int dregno)
6708 {
6709 rtx_insn *prev_insn = PREV_INSN (insn);
6710
6711 lra_set_insn_deleted (insn);
6712 lra_assert (dregno >= 0);
6713 if (prev_insn != NULL && NONDEBUG_INSN_P (prev_insn)
6714 && GET_CODE (PATTERN (prev_insn)) == CLOBBER
6715 && dregno == get_regno (XEXP (PATTERN (prev_insn), 0)))
6716 lra_set_insn_deleted (prev_insn);
6717 }
6718
6719 /* Remove inheritance/split pseudos which are in REMOVE_PSEUDOS and
6720 return true if we did any change. The undo transformations for
6721 inheritance looks like
6722 i <- i2
6723 p <- i => p <- i2
6724 or removing
6725 p <- i, i <- p, and i <- i3
6726 where p is original pseudo from which inheritance pseudo i was
6727 created, i and i3 are removed inheritance pseudos, i2 is another
6728 not removed inheritance pseudo. All split pseudos or other
6729 occurrences of removed inheritance pseudos are changed on the
6730 corresponding original pseudos.
6731
6732 The function also schedules insns changed and created during
6733 inheritance/split pass for processing by the subsequent constraint
6734 pass. */
6735 static bool
6736 remove_inheritance_pseudos (bitmap remove_pseudos)
6737 {
6738 basic_block bb;
6739 int regno, sregno, prev_sregno, dregno;
6740 rtx restore_rtx;
6741 rtx set, prev_set;
6742 rtx_insn *prev_insn;
6743 bool change_p, done_p;
6744
6745 change_p = ! bitmap_empty_p (remove_pseudos);
6746 /* We cannot finish the function right away if CHANGE_P is true
6747 because we need to marks insns affected by previous
6748 inheritance/split pass for processing by the subsequent
6749 constraint pass. */
6750 FOR_EACH_BB_FN (bb, cfun)
6751 {
6752 fix_bb_live_info (df_get_live_in (bb), remove_pseudos);
6753 fix_bb_live_info (df_get_live_out (bb), remove_pseudos);
6754 FOR_BB_INSNS_REVERSE (bb, curr_insn)
6755 {
6756 if (! INSN_P (curr_insn))
6757 continue;
6758 done_p = false;
6759 sregno = dregno = -1;
6760 if (change_p && NONDEBUG_INSN_P (curr_insn)
6761 && (set = single_set (curr_insn)) != NULL_RTX)
6762 {
6763 dregno = get_regno (SET_DEST (set));
6764 sregno = get_regno (SET_SRC (set));
6765 }
6766
6767 if (sregno >= 0 && dregno >= 0)
6768 {
6769 if (bitmap_bit_p (remove_pseudos, dregno)
6770 && ! REG_P (lra_reg_info[dregno].restore_rtx))
6771 {
6772 /* invariant inheritance pseudo <- original pseudo */
6773 if (lra_dump_file != NULL)
6774 {
6775 fprintf (lra_dump_file, " Removing invariant inheritance:\n");
6776 dump_insn_slim (lra_dump_file, curr_insn);
6777 fprintf (lra_dump_file, "\n");
6778 }
6779 delete_move_and_clobber (curr_insn, dregno);
6780 done_p = true;
6781 }
6782 else if (bitmap_bit_p (remove_pseudos, sregno)
6783 && ! REG_P (lra_reg_info[sregno].restore_rtx))
6784 {
6785 /* reload pseudo <- invariant inheritance pseudo */
6786 start_sequence ();
6787 /* We cannot just change the source. It might be
6788 an insn different from the move. */
6789 emit_insn (lra_reg_info[sregno].restore_rtx);
6790 rtx_insn *new_insns = get_insns ();
6791 end_sequence ();
6792 lra_assert (single_set (new_insns) != NULL
6793 && SET_DEST (set) == SET_DEST (single_set (new_insns)));
6794 lra_process_new_insns (curr_insn, NULL, new_insns,
6795 "Changing reload<-invariant inheritance");
6796 delete_move_and_clobber (curr_insn, dregno);
6797 done_p = true;
6798 }
6799 else if ((bitmap_bit_p (remove_pseudos, sregno)
6800 && (get_regno (lra_reg_info[sregno].restore_rtx) == dregno
6801 || (bitmap_bit_p (remove_pseudos, dregno)
6802 && get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6803 && (get_regno (lra_reg_info[sregno].restore_rtx)
6804 == get_regno (lra_reg_info[dregno].restore_rtx)))))
6805 || (bitmap_bit_p (remove_pseudos, dregno)
6806 && get_regno (lra_reg_info[dregno].restore_rtx) == sregno))
6807 /* One of the following cases:
6808 original <- removed inheritance pseudo
6809 removed inherit pseudo <- another removed inherit pseudo
6810 removed inherit pseudo <- original pseudo
6811 Or
6812 removed_split_pseudo <- original_reg
6813 original_reg <- removed_split_pseudo */
6814 {
6815 if (lra_dump_file != NULL)
6816 {
6817 fprintf (lra_dump_file, " Removing %s:\n",
6818 bitmap_bit_p (&lra_split_regs, sregno)
6819 || bitmap_bit_p (&lra_split_regs, dregno)
6820 ? "split" : "inheritance");
6821 dump_insn_slim (lra_dump_file, curr_insn);
6822 }
6823 delete_move_and_clobber (curr_insn, dregno);
6824 done_p = true;
6825 }
6826 else if (bitmap_bit_p (remove_pseudos, sregno)
6827 && bitmap_bit_p (&lra_inheritance_pseudos, sregno))
6828 {
6829 /* Search the following pattern:
6830 inherit_or_split_pseudo1 <- inherit_or_split_pseudo2
6831 original_pseudo <- inherit_or_split_pseudo1
6832 where the 2nd insn is the current insn and
6833 inherit_or_split_pseudo2 is not removed. If it is found,
6834 change the current insn onto:
6835 original_pseudo <- inherit_or_split_pseudo2. */
6836 for (prev_insn = PREV_INSN (curr_insn);
6837 prev_insn != NULL_RTX && ! NONDEBUG_INSN_P (prev_insn);
6838 prev_insn = PREV_INSN (prev_insn))
6839 ;
6840 if (prev_insn != NULL_RTX && BLOCK_FOR_INSN (prev_insn) == bb
6841 && (prev_set = single_set (prev_insn)) != NULL_RTX
6842 /* There should be no subregs in insn we are
6843 searching because only the original reg might
6844 be in subreg when we changed the mode of
6845 load/store for splitting. */
6846 && REG_P (SET_DEST (prev_set))
6847 && REG_P (SET_SRC (prev_set))
6848 && (int) REGNO (SET_DEST (prev_set)) == sregno
6849 && ((prev_sregno = REGNO (SET_SRC (prev_set)))
6850 >= FIRST_PSEUDO_REGISTER)
6851 && (lra_reg_info[prev_sregno].restore_rtx == NULL_RTX
6852 ||
6853 /* As we consider chain of inheritance or
6854 splitting described in above comment we should
6855 check that sregno and prev_sregno were
6856 inheritance/split pseudos created from the
6857 same original regno. */
6858 (get_regno (lra_reg_info[sregno].restore_rtx) >= 0
6859 && (get_regno (lra_reg_info[sregno].restore_rtx)
6860 == get_regno (lra_reg_info[prev_sregno].restore_rtx))))
6861 && ! bitmap_bit_p (remove_pseudos, prev_sregno))
6862 {
6863 lra_assert (GET_MODE (SET_SRC (prev_set))
6864 == GET_MODE (regno_reg_rtx[sregno]));
6865 /* Although we have a single set, the insn can
6866 contain more one sregno register occurrence
6867 as a source. Change all occurrences. */
6868 lra_substitute_pseudo_within_insn (curr_insn, sregno,
6869 SET_SRC (prev_set),
6870 false);
6871 /* As we are finishing with processing the insn
6872 here, check the destination too as it might
6873 inheritance pseudo for another pseudo. */
6874 if (bitmap_bit_p (remove_pseudos, dregno)
6875 && bitmap_bit_p (&lra_inheritance_pseudos, dregno)
6876 && (restore_rtx
6877 = lra_reg_info[dregno].restore_rtx) != NULL_RTX)
6878 {
6879 if (GET_CODE (SET_DEST (set)) == SUBREG)
6880 SUBREG_REG (SET_DEST (set)) = restore_rtx;
6881 else
6882 SET_DEST (set) = restore_rtx;
6883 }
6884 lra_push_insn_and_update_insn_regno_info (curr_insn);
6885 lra_set_used_insn_alternative_by_uid
6886 (INSN_UID (curr_insn), LRA_UNKNOWN_ALT);
6887 done_p = true;
6888 if (lra_dump_file != NULL)
6889 {
6890 fprintf (lra_dump_file, " Change reload insn:\n");
6891 dump_insn_slim (lra_dump_file, curr_insn);
6892 }
6893 }
6894 }
6895 }
6896 if (! done_p)
6897 {
6898 struct lra_insn_reg *reg;
6899 bool restored_regs_p = false;
6900 bool kept_regs_p = false;
6901
6902 curr_id = lra_get_insn_recog_data (curr_insn);
6903 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
6904 {
6905 regno = reg->regno;
6906 restore_rtx = lra_reg_info[regno].restore_rtx;
6907 if (restore_rtx != NULL_RTX)
6908 {
6909 if (change_p && bitmap_bit_p (remove_pseudos, regno))
6910 {
6911 lra_substitute_pseudo_within_insn
6912 (curr_insn, regno, restore_rtx, false);
6913 restored_regs_p = true;
6914 }
6915 else
6916 kept_regs_p = true;
6917 }
6918 }
6919 if (NONDEBUG_INSN_P (curr_insn) && kept_regs_p)
6920 {
6921 /* The instruction has changed since the previous
6922 constraints pass. */
6923 lra_push_insn_and_update_insn_regno_info (curr_insn);
6924 lra_set_used_insn_alternative_by_uid
6925 (INSN_UID (curr_insn), LRA_UNKNOWN_ALT);
6926 }
6927 else if (restored_regs_p)
6928 /* The instruction has been restored to the form that
6929 it had during the previous constraints pass. */
6930 lra_update_insn_regno_info (curr_insn);
6931 if (restored_regs_p && lra_dump_file != NULL)
6932 {
6933 fprintf (lra_dump_file, " Insn after restoring regs:\n");
6934 dump_insn_slim (lra_dump_file, curr_insn);
6935 }
6936 }
6937 }
6938 }
6939 return change_p;
6940 }
6941
6942 /* If optional reload pseudos failed to get a hard register or was not
6943 inherited, it is better to remove optional reloads. We do this
6944 transformation after undoing inheritance to figure out necessity to
6945 remove optional reloads easier. Return true if we do any
6946 change. */
6947 static bool
6948 undo_optional_reloads (void)
6949 {
6950 bool change_p, keep_p;
6951 unsigned int regno, uid;
6952 bitmap_iterator bi, bi2;
6953 rtx_insn *insn;
6954 rtx set, src, dest;
6955 auto_bitmap removed_optional_reload_pseudos (&reg_obstack);
6956
6957 bitmap_copy (removed_optional_reload_pseudos, &lra_optional_reload_pseudos);
6958 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
6959 {
6960 keep_p = false;
6961 /* Keep optional reloads from previous subpasses. */
6962 if (lra_reg_info[regno].restore_rtx == NULL_RTX
6963 /* If the original pseudo changed its allocation, just
6964 removing the optional pseudo is dangerous as the original
6965 pseudo will have longer live range. */
6966 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] >= 0)
6967 keep_p = true;
6968 else if (reg_renumber[regno] >= 0)
6969 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi2)
6970 {
6971 insn = lra_insn_recog_data[uid]->insn;
6972 if ((set = single_set (insn)) == NULL_RTX)
6973 continue;
6974 src = SET_SRC (set);
6975 dest = SET_DEST (set);
6976 if (! REG_P (src) || ! REG_P (dest))
6977 continue;
6978 if (REGNO (dest) == regno
6979 /* Ignore insn for optional reloads itself. */
6980 && REGNO (lra_reg_info[regno].restore_rtx) != REGNO (src)
6981 /* Check only inheritance on last inheritance pass. */
6982 && (int) REGNO (src) >= new_regno_start
6983 /* Check that the optional reload was inherited. */
6984 && bitmap_bit_p (&lra_inheritance_pseudos, REGNO (src)))
6985 {
6986 keep_p = true;
6987 break;
6988 }
6989 }
6990 if (keep_p)
6991 {
6992 bitmap_clear_bit (removed_optional_reload_pseudos, regno);
6993 if (lra_dump_file != NULL)
6994 fprintf (lra_dump_file, "Keep optional reload reg %d\n", regno);
6995 }
6996 }
6997 change_p = ! bitmap_empty_p (removed_optional_reload_pseudos);
6998 auto_bitmap insn_bitmap (&reg_obstack);
6999 EXECUTE_IF_SET_IN_BITMAP (removed_optional_reload_pseudos, 0, regno, bi)
7000 {
7001 if (lra_dump_file != NULL)
7002 fprintf (lra_dump_file, "Remove optional reload reg %d\n", regno);
7003 bitmap_copy (insn_bitmap, &lra_reg_info[regno].insn_bitmap);
7004 EXECUTE_IF_SET_IN_BITMAP (insn_bitmap, 0, uid, bi2)
7005 {
7006 insn = lra_insn_recog_data[uid]->insn;
7007 if ((set = single_set (insn)) != NULL_RTX)
7008 {
7009 src = SET_SRC (set);
7010 dest = SET_DEST (set);
7011 if (REG_P (src) && REG_P (dest)
7012 && ((REGNO (src) == regno
7013 && (REGNO (lra_reg_info[regno].restore_rtx)
7014 == REGNO (dest)))
7015 || (REGNO (dest) == regno
7016 && (REGNO (lra_reg_info[regno].restore_rtx)
7017 == REGNO (src)))))
7018 {
7019 if (lra_dump_file != NULL)
7020 {
7021 fprintf (lra_dump_file, " Deleting move %u\n",
7022 INSN_UID (insn));
7023 dump_insn_slim (lra_dump_file, insn);
7024 }
7025 delete_move_and_clobber (insn, REGNO (dest));
7026 continue;
7027 }
7028 /* We should not worry about generation memory-memory
7029 moves here as if the corresponding inheritance did
7030 not work (inheritance pseudo did not get a hard reg),
7031 we remove the inheritance pseudo and the optional
7032 reload. */
7033 }
7034 lra_substitute_pseudo_within_insn
7035 (insn, regno, lra_reg_info[regno].restore_rtx, false);
7036 lra_update_insn_regno_info (insn);
7037 if (lra_dump_file != NULL)
7038 {
7039 fprintf (lra_dump_file,
7040 " Restoring original insn:\n");
7041 dump_insn_slim (lra_dump_file, insn);
7042 }
7043 }
7044 }
7045 /* Clear restore_regnos. */
7046 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
7047 lra_reg_info[regno].restore_rtx = NULL_RTX;
7048 return change_p;
7049 }
7050
7051 /* Entry function for undoing inheritance/split transformation. Return true
7052 if we did any RTL change in this pass. */
7053 bool
7054 lra_undo_inheritance (void)
7055 {
7056 unsigned int regno;
7057 int hard_regno;
7058 int n_all_inherit, n_inherit, n_all_split, n_split;
7059 rtx restore_rtx;
7060 bitmap_iterator bi;
7061 bool change_p;
7062
7063 lra_undo_inheritance_iter++;
7064 if (lra_undo_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
7065 return false;
7066 if (lra_dump_file != NULL)
7067 fprintf (lra_dump_file,
7068 "\n********** Undoing inheritance #%d: **********\n\n",
7069 lra_undo_inheritance_iter);
7070 auto_bitmap remove_pseudos (&reg_obstack);
7071 n_inherit = n_all_inherit = 0;
7072 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
7073 if (lra_reg_info[regno].restore_rtx != NULL_RTX)
7074 {
7075 n_all_inherit++;
7076 if (reg_renumber[regno] < 0
7077 /* If the original pseudo changed its allocation, just
7078 removing inheritance is dangerous as for changing
7079 allocation we used shorter live-ranges. */
7080 && (! REG_P (lra_reg_info[regno].restore_rtx)
7081 || reg_renumber[REGNO (lra_reg_info[regno].restore_rtx)] < 0))
7082 bitmap_set_bit (remove_pseudos, regno);
7083 else
7084 n_inherit++;
7085 }
7086 if (lra_dump_file != NULL && n_all_inherit != 0)
7087 fprintf (lra_dump_file, "Inherit %d out of %d (%.2f%%)\n",
7088 n_inherit, n_all_inherit,
7089 (double) n_inherit / n_all_inherit * 100);
7090 n_split = n_all_split = 0;
7091 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
7092 if ((restore_rtx = lra_reg_info[regno].restore_rtx) != NULL_RTX)
7093 {
7094 int restore_regno = REGNO (restore_rtx);
7095
7096 n_all_split++;
7097 hard_regno = (restore_regno >= FIRST_PSEUDO_REGISTER
7098 ? reg_renumber[restore_regno] : restore_regno);
7099 if (hard_regno < 0 || reg_renumber[regno] == hard_regno)
7100 bitmap_set_bit (remove_pseudos, regno);
7101 else
7102 {
7103 n_split++;
7104 if (lra_dump_file != NULL)
7105 fprintf (lra_dump_file, " Keep split r%d (orig=r%d)\n",
7106 regno, restore_regno);
7107 }
7108 }
7109 if (lra_dump_file != NULL && n_all_split != 0)
7110 fprintf (lra_dump_file, "Split %d out of %d (%.2f%%)\n",
7111 n_split, n_all_split,
7112 (double) n_split / n_all_split * 100);
7113 change_p = remove_inheritance_pseudos (remove_pseudos);
7114 /* Clear restore_regnos. */
7115 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
7116 lra_reg_info[regno].restore_rtx = NULL_RTX;
7117 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
7118 lra_reg_info[regno].restore_rtx = NULL_RTX;
7119 change_p = undo_optional_reloads () || change_p;
7120 return change_p;
7121 }