re PR rtl-optimization/61325 (aarch64_be build fails)
[gcc.git] / gcc / lra-constraints.c
1 /* Code for RTL transformations to satisfy insn constraints.
2 Copyright (C) 2010-2014 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 "tm.h"
113 #include "hard-reg-set.h"
114 #include "rtl.h"
115 #include "tm_p.h"
116 #include "regs.h"
117 #include "insn-config.h"
118 #include "insn-codes.h"
119 #include "recog.h"
120 #include "output.h"
121 #include "addresses.h"
122 #include "target.h"
123 #include "function.h"
124 #include "expr.h"
125 #include "basic-block.h"
126 #include "except.h"
127 #include "optabs.h"
128 #include "df.h"
129 #include "ira.h"
130 #include "rtl-error.h"
131 #include "lra-int.h"
132
133 /* Value of LRA_CURR_RELOAD_NUM at the beginning of BB of the current
134 insn. Remember that LRA_CURR_RELOAD_NUM is the number of emitted
135 reload insns. */
136 static int bb_reload_num;
137
138 /* The current insn being processed and corresponding its single set
139 (NULL otherwise), its data (basic block, the insn data, the insn
140 static data, and the mode of each operand). */
141 static rtx curr_insn;
142 static rtx curr_insn_set;
143 static basic_block curr_bb;
144 static lra_insn_recog_data_t curr_id;
145 static struct lra_static_insn_data *curr_static_id;
146 static enum machine_mode curr_operand_mode[MAX_RECOG_OPERANDS];
147
148 \f
149
150 /* Start numbers for new registers and insns at the current constraints
151 pass start. */
152 static int new_regno_start;
153 static int new_insn_uid_start;
154
155 /* If LOC is nonnull, strip any outer subreg from it. */
156 static inline rtx *
157 strip_subreg (rtx *loc)
158 {
159 return loc && GET_CODE (*loc) == SUBREG ? &SUBREG_REG (*loc) : loc;
160 }
161
162 /* Return hard regno of REGNO or if it is was not assigned to a hard
163 register, use a hard register from its allocno class. */
164 static int
165 get_try_hard_regno (int regno)
166 {
167 int hard_regno;
168 enum reg_class rclass;
169
170 if ((hard_regno = regno) >= FIRST_PSEUDO_REGISTER)
171 hard_regno = lra_get_regno_hard_regno (regno);
172 if (hard_regno >= 0)
173 return hard_regno;
174 rclass = lra_get_allocno_class (regno);
175 if (rclass == NO_REGS)
176 return -1;
177 return ira_class_hard_regs[rclass][0];
178 }
179
180 /* Return final hard regno (plus offset) which will be after
181 elimination. We do this for matching constraints because the final
182 hard regno could have a different class. */
183 static int
184 get_final_hard_regno (int hard_regno, int offset)
185 {
186 if (hard_regno < 0)
187 return hard_regno;
188 hard_regno = lra_get_elimination_hard_regno (hard_regno);
189 return hard_regno + offset;
190 }
191
192 /* Return hard regno of X after removing subreg and making
193 elimination. If X is not a register or subreg of register, return
194 -1. For pseudo use its assignment. */
195 static int
196 get_hard_regno (rtx x)
197 {
198 rtx reg;
199 int offset, hard_regno;
200
201 reg = x;
202 if (GET_CODE (x) == SUBREG)
203 reg = SUBREG_REG (x);
204 if (! REG_P (reg))
205 return -1;
206 if ((hard_regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER)
207 hard_regno = lra_get_regno_hard_regno (hard_regno);
208 if (hard_regno < 0)
209 return -1;
210 offset = 0;
211 if (GET_CODE (x) == SUBREG)
212 offset += subreg_regno_offset (hard_regno, GET_MODE (reg),
213 SUBREG_BYTE (x), GET_MODE (x));
214 return get_final_hard_regno (hard_regno, offset);
215 }
216
217 /* If REGNO is a hard register or has been allocated a hard register,
218 return the class of that register. If REGNO is a reload pseudo
219 created by the current constraints pass, return its allocno class.
220 Return NO_REGS otherwise. */
221 static enum reg_class
222 get_reg_class (int regno)
223 {
224 int hard_regno;
225
226 if ((hard_regno = regno) >= FIRST_PSEUDO_REGISTER)
227 hard_regno = lra_get_regno_hard_regno (regno);
228 if (hard_regno >= 0)
229 {
230 hard_regno = get_final_hard_regno (hard_regno, 0);
231 return REGNO_REG_CLASS (hard_regno);
232 }
233 if (regno >= new_regno_start)
234 return lra_get_allocno_class (regno);
235 return NO_REGS;
236 }
237
238 /* Return true if REG satisfies (or will satisfy) reg class constraint
239 CL. Use elimination first if REG is a hard register. If REG is a
240 reload pseudo created by this constraints pass, assume that it will
241 be allocated a hard register from its allocno class, but allow that
242 class to be narrowed to CL if it is currently a superset of CL.
243
244 If NEW_CLASS is nonnull, set *NEW_CLASS to the new allocno class of
245 REGNO (reg), or NO_REGS if no change in its class was needed. */
246 static bool
247 in_class_p (rtx reg, enum reg_class cl, enum reg_class *new_class)
248 {
249 enum reg_class rclass, common_class;
250 enum machine_mode reg_mode;
251 int class_size, hard_regno, nregs, i, j;
252 int regno = REGNO (reg);
253
254 if (new_class != NULL)
255 *new_class = NO_REGS;
256 if (regno < FIRST_PSEUDO_REGISTER)
257 {
258 rtx final_reg = reg;
259 rtx *final_loc = &final_reg;
260
261 lra_eliminate_reg_if_possible (final_loc);
262 return TEST_HARD_REG_BIT (reg_class_contents[cl], REGNO (*final_loc));
263 }
264 reg_mode = GET_MODE (reg);
265 rclass = get_reg_class (regno);
266 if (regno < new_regno_start
267 /* Do not allow the constraints for reload instructions to
268 influence the classes of new pseudos. These reloads are
269 typically moves that have many alternatives, and restricting
270 reload pseudos for one alternative may lead to situations
271 where other reload pseudos are no longer allocatable. */
272 || (INSN_UID (curr_insn) >= new_insn_uid_start
273 && curr_insn_set != NULL
274 && ((OBJECT_P (SET_SRC (curr_insn_set))
275 && ! CONSTANT_P (SET_SRC (curr_insn_set)))
276 || (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
277 && OBJECT_P (SUBREG_REG (SET_SRC (curr_insn_set)))
278 && ! CONSTANT_P (SUBREG_REG (SET_SRC (curr_insn_set)))))))
279 /* When we don't know what class will be used finally for reload
280 pseudos, we use ALL_REGS. */
281 return ((regno >= new_regno_start && rclass == ALL_REGS)
282 || (rclass != NO_REGS && ira_class_subset_p[rclass][cl]
283 && ! hard_reg_set_subset_p (reg_class_contents[cl],
284 lra_no_alloc_regs)));
285 else
286 {
287 common_class = ira_reg_class_subset[rclass][cl];
288 if (new_class != NULL)
289 *new_class = common_class;
290 if (hard_reg_set_subset_p (reg_class_contents[common_class],
291 lra_no_alloc_regs))
292 return false;
293 /* Check that there are enough allocatable regs. */
294 class_size = ira_class_hard_regs_num[common_class];
295 for (i = 0; i < class_size; i++)
296 {
297 hard_regno = ira_class_hard_regs[common_class][i];
298 nregs = hard_regno_nregs[hard_regno][reg_mode];
299 if (nregs == 1)
300 return true;
301 for (j = 0; j < nregs; j++)
302 if (TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno + j)
303 || ! TEST_HARD_REG_BIT (reg_class_contents[common_class],
304 hard_regno + j))
305 break;
306 if (j >= nregs)
307 return true;
308 }
309 return false;
310 }
311 }
312
313 /* Return true if REGNO satisfies a memory constraint. */
314 static bool
315 in_mem_p (int regno)
316 {
317 return get_reg_class (regno) == NO_REGS;
318 }
319
320 /* Initiate equivalences for LRA. As we keep original equivalences
321 before any elimination, we need to make copies otherwise any change
322 in insns might change the equivalences. */
323 void
324 lra_init_equiv (void)
325 {
326 ira_expand_reg_equiv ();
327 for (int i = FIRST_PSEUDO_REGISTER; i < max_reg_num (); i++)
328 {
329 rtx res;
330
331 if ((res = ira_reg_equiv[i].memory) != NULL_RTX)
332 ira_reg_equiv[i].memory = copy_rtx (res);
333 if ((res = ira_reg_equiv[i].invariant) != NULL_RTX)
334 ira_reg_equiv[i].invariant = copy_rtx (res);
335 }
336 }
337
338 static rtx loc_equivalence_callback (rtx, const_rtx, void *);
339
340 /* Update equivalence for REGNO. We need to this as the equivalence
341 might contain other pseudos which are changed by their
342 equivalences. */
343 static void
344 update_equiv (int regno)
345 {
346 rtx x;
347
348 if ((x = ira_reg_equiv[regno].memory) != NULL_RTX)
349 ira_reg_equiv[regno].memory
350 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
351 NULL_RTX);
352 if ((x = ira_reg_equiv[regno].invariant) != NULL_RTX)
353 ira_reg_equiv[regno].invariant
354 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
355 NULL_RTX);
356 }
357
358 /* If we have decided to substitute X with another value, return that
359 value, otherwise return X. */
360 static rtx
361 get_equiv (rtx x)
362 {
363 int regno;
364 rtx res;
365
366 if (! REG_P (x) || (regno = REGNO (x)) < FIRST_PSEUDO_REGISTER
367 || ! ira_reg_equiv[regno].defined_p
368 || ! ira_reg_equiv[regno].profitable_p
369 || lra_get_regno_hard_regno (regno) >= 0)
370 return x;
371 if ((res = ira_reg_equiv[regno].memory) != NULL_RTX)
372 return res;
373 if ((res = ira_reg_equiv[regno].constant) != NULL_RTX)
374 return res;
375 if ((res = ira_reg_equiv[regno].invariant) != NULL_RTX)
376 return res;
377 gcc_unreachable ();
378 }
379
380 /* If we have decided to substitute X with the equivalent value,
381 return that value after elimination for INSN, otherwise return
382 X. */
383 static rtx
384 get_equiv_with_elimination (rtx x, rtx insn)
385 {
386 rtx res = get_equiv (x);
387
388 if (x == res || CONSTANT_P (res))
389 return res;
390 return lra_eliminate_regs_1 (insn, res, GET_MODE (res), false, false, true);
391 }
392
393 /* Set up curr_operand_mode. */
394 static void
395 init_curr_operand_mode (void)
396 {
397 int nop = curr_static_id->n_operands;
398 for (int i = 0; i < nop; i++)
399 {
400 enum machine_mode mode = GET_MODE (*curr_id->operand_loc[i]);
401 if (mode == VOIDmode)
402 {
403 /* The .md mode for address operands is the mode of the
404 addressed value rather than the mode of the address itself. */
405 if (curr_id->icode >= 0 && curr_static_id->operand[i].is_address)
406 mode = Pmode;
407 else
408 mode = curr_static_id->operand[i].mode;
409 }
410 curr_operand_mode[i] = mode;
411 }
412 }
413
414 \f
415
416 /* The page contains code to reuse input reloads. */
417
418 /* Structure describes input reload of the current insns. */
419 struct input_reload
420 {
421 /* Reloaded value. */
422 rtx input;
423 /* Reload pseudo used. */
424 rtx reg;
425 };
426
427 /* The number of elements in the following array. */
428 static int curr_insn_input_reloads_num;
429 /* Array containing info about input reloads. It is used to find the
430 same input reload and reuse the reload pseudo in this case. */
431 static struct input_reload curr_insn_input_reloads[LRA_MAX_INSN_RELOADS];
432
433 /* Initiate data concerning reuse of input reloads for the current
434 insn. */
435 static void
436 init_curr_insn_input_reloads (void)
437 {
438 curr_insn_input_reloads_num = 0;
439 }
440
441 /* Create a new pseudo using MODE, RCLASS, ORIGINAL or reuse already
442 created input reload pseudo (only if TYPE is not OP_OUT). Don't
443 reuse pseudo if IN_SUBREG_P is true and the reused pseudo should be
444 wrapped up in SUBREG. The result pseudo is returned through
445 RESULT_REG. Return TRUE if we created a new pseudo, FALSE if we
446 reused the already created input reload pseudo. Use TITLE to
447 describe new registers for debug purposes. */
448 static bool
449 get_reload_reg (enum op_type type, enum machine_mode mode, rtx original,
450 enum reg_class rclass, bool in_subreg_p,
451 const char *title, rtx *result_reg)
452 {
453 int i, regno;
454 enum reg_class new_class;
455
456 if (type == OP_OUT)
457 {
458 *result_reg
459 = lra_create_new_reg_with_unique_value (mode, original, rclass, title);
460 return true;
461 }
462 /* Prevent reuse value of expression with side effects,
463 e.g. volatile memory. */
464 if (! side_effects_p (original))
465 for (i = 0; i < curr_insn_input_reloads_num; i++)
466 if (rtx_equal_p (curr_insn_input_reloads[i].input, original)
467 && in_class_p (curr_insn_input_reloads[i].reg, rclass, &new_class))
468 {
469 rtx reg = curr_insn_input_reloads[i].reg;
470 regno = REGNO (reg);
471 /* If input is equal to original and both are VOIDmode,
472 GET_MODE (reg) might be still different from mode.
473 Ensure we don't return *result_reg with wrong mode. */
474 if (GET_MODE (reg) != mode)
475 {
476 if (in_subreg_p)
477 continue;
478 if (GET_MODE_SIZE (GET_MODE (reg)) < GET_MODE_SIZE (mode))
479 continue;
480 reg = lowpart_subreg (mode, reg, GET_MODE (reg));
481 if (reg == NULL_RTX || GET_CODE (reg) != SUBREG)
482 continue;
483 }
484 *result_reg = reg;
485 if (lra_dump_file != NULL)
486 {
487 fprintf (lra_dump_file, " Reuse r%d for reload ", regno);
488 dump_value_slim (lra_dump_file, original, 1);
489 }
490 if (new_class != lra_get_allocno_class (regno))
491 lra_change_class (regno, new_class, ", change to", false);
492 if (lra_dump_file != NULL)
493 fprintf (lra_dump_file, "\n");
494 return false;
495 }
496 *result_reg = lra_create_new_reg (mode, original, rclass, title);
497 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
498 curr_insn_input_reloads[curr_insn_input_reloads_num].input = original;
499 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = *result_reg;
500 return true;
501 }
502
503 \f
504
505 /* The page contains code to extract memory address parts. */
506
507 /* Wrapper around REGNO_OK_FOR_INDEX_P, to allow pseudos. */
508 static inline bool
509 ok_for_index_p_nonstrict (rtx reg)
510 {
511 unsigned regno = REGNO (reg);
512
513 return regno >= FIRST_PSEUDO_REGISTER || REGNO_OK_FOR_INDEX_P (regno);
514 }
515
516 /* A version of regno_ok_for_base_p for use here, when all pseudos
517 should count as OK. Arguments as for regno_ok_for_base_p. */
518 static inline bool
519 ok_for_base_p_nonstrict (rtx reg, enum machine_mode mode, addr_space_t as,
520 enum rtx_code outer_code, enum rtx_code index_code)
521 {
522 unsigned regno = REGNO (reg);
523
524 if (regno >= FIRST_PSEUDO_REGISTER)
525 return true;
526 return ok_for_base_p_1 (regno, mode, as, outer_code, index_code);
527 }
528
529 \f
530
531 /* The page contains major code to choose the current insn alternative
532 and generate reloads for it. */
533
534 /* Return the offset from REGNO of the least significant register
535 in (reg:MODE REGNO).
536
537 This function is used to tell whether two registers satisfy
538 a matching constraint. (reg:MODE1 REGNO1) matches (reg:MODE2 REGNO2) if:
539
540 REGNO1 + lra_constraint_offset (REGNO1, MODE1)
541 == REGNO2 + lra_constraint_offset (REGNO2, MODE2) */
542 int
543 lra_constraint_offset (int regno, enum machine_mode mode)
544 {
545 lra_assert (regno < FIRST_PSEUDO_REGISTER);
546 if (WORDS_BIG_ENDIAN && GET_MODE_SIZE (mode) > UNITS_PER_WORD
547 && SCALAR_INT_MODE_P (mode))
548 return hard_regno_nregs[regno][mode] - 1;
549 return 0;
550 }
551
552 /* Like rtx_equal_p except that it allows a REG and a SUBREG to match
553 if they are the same hard reg, and has special hacks for
554 auto-increment and auto-decrement. This is specifically intended for
555 process_alt_operands to use in determining whether two operands
556 match. X is the operand whose number is the lower of the two.
557
558 It is supposed that X is the output operand and Y is the input
559 operand. Y_HARD_REGNO is the final hard regno of register Y or
560 register in subreg Y as we know it now. Otherwise, it is a
561 negative value. */
562 static bool
563 operands_match_p (rtx x, rtx y, int y_hard_regno)
564 {
565 int i;
566 RTX_CODE code = GET_CODE (x);
567 const char *fmt;
568
569 if (x == y)
570 return true;
571 if ((code == REG || (code == SUBREG && REG_P (SUBREG_REG (x))))
572 && (REG_P (y) || (GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y)))))
573 {
574 int j;
575
576 i = get_hard_regno (x);
577 if (i < 0)
578 goto slow;
579
580 if ((j = y_hard_regno) < 0)
581 goto slow;
582
583 i += lra_constraint_offset (i, GET_MODE (x));
584 j += lra_constraint_offset (j, GET_MODE (y));
585
586 return i == j;
587 }
588
589 /* If two operands must match, because they are really a single
590 operand of an assembler insn, then two post-increments are invalid
591 because the assembler insn would increment only once. On the
592 other hand, a post-increment matches ordinary indexing if the
593 post-increment is the output operand. */
594 if (code == POST_DEC || code == POST_INC || code == POST_MODIFY)
595 return operands_match_p (XEXP (x, 0), y, y_hard_regno);
596
597 /* Two pre-increments are invalid because the assembler insn would
598 increment only once. On the other hand, a pre-increment matches
599 ordinary indexing if the pre-increment is the input operand. */
600 if (GET_CODE (y) == PRE_DEC || GET_CODE (y) == PRE_INC
601 || GET_CODE (y) == PRE_MODIFY)
602 return operands_match_p (x, XEXP (y, 0), -1);
603
604 slow:
605
606 if (code == REG && GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y))
607 && x == SUBREG_REG (y))
608 return true;
609 if (GET_CODE (y) == REG && code == SUBREG && REG_P (SUBREG_REG (x))
610 && SUBREG_REG (x) == y)
611 return true;
612
613 /* Now we have disposed of all the cases in which different rtx
614 codes can match. */
615 if (code != GET_CODE (y))
616 return false;
617
618 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
619 if (GET_MODE (x) != GET_MODE (y))
620 return false;
621
622 switch (code)
623 {
624 CASE_CONST_UNIQUE:
625 return false;
626
627 case LABEL_REF:
628 return XEXP (x, 0) == XEXP (y, 0);
629 case SYMBOL_REF:
630 return XSTR (x, 0) == XSTR (y, 0);
631
632 default:
633 break;
634 }
635
636 /* Compare the elements. If any pair of corresponding elements fail
637 to match, return false for the whole things. */
638
639 fmt = GET_RTX_FORMAT (code);
640 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
641 {
642 int val, j;
643 switch (fmt[i])
644 {
645 case 'w':
646 if (XWINT (x, i) != XWINT (y, i))
647 return false;
648 break;
649
650 case 'i':
651 if (XINT (x, i) != XINT (y, i))
652 return false;
653 break;
654
655 case 'e':
656 val = operands_match_p (XEXP (x, i), XEXP (y, i), -1);
657 if (val == 0)
658 return false;
659 break;
660
661 case '0':
662 break;
663
664 case 'E':
665 if (XVECLEN (x, i) != XVECLEN (y, i))
666 return false;
667 for (j = XVECLEN (x, i) - 1; j >= 0; --j)
668 {
669 val = operands_match_p (XVECEXP (x, i, j), XVECEXP (y, i, j), -1);
670 if (val == 0)
671 return false;
672 }
673 break;
674
675 /* It is believed that rtx's at this level will never
676 contain anything but integers and other rtx's, except for
677 within LABEL_REFs and SYMBOL_REFs. */
678 default:
679 gcc_unreachable ();
680 }
681 }
682 return true;
683 }
684
685 /* True if X is a constant that can be forced into the constant pool.
686 MODE is the mode of the operand, or VOIDmode if not known. */
687 #define CONST_POOL_OK_P(MODE, X) \
688 ((MODE) != VOIDmode \
689 && CONSTANT_P (X) \
690 && GET_CODE (X) != HIGH \
691 && !targetm.cannot_force_const_mem (MODE, X))
692
693 /* True if C is a non-empty register class that has too few registers
694 to be safely used as a reload target class. */
695 #define SMALL_REGISTER_CLASS_P(C) \
696 (ira_class_hard_regs_num [(C)] == 1 \
697 || (ira_class_hard_regs_num [(C)] >= 1 \
698 && targetm.class_likely_spilled_p (C)))
699
700 /* If REG is a reload pseudo, try to make its class satisfying CL. */
701 static void
702 narrow_reload_pseudo_class (rtx reg, enum reg_class cl)
703 {
704 enum reg_class rclass;
705
706 /* Do not make more accurate class from reloads generated. They are
707 mostly moves with a lot of constraints. Making more accurate
708 class may results in very narrow class and impossibility of find
709 registers for several reloads of one insn. */
710 if (INSN_UID (curr_insn) >= new_insn_uid_start)
711 return;
712 if (GET_CODE (reg) == SUBREG)
713 reg = SUBREG_REG (reg);
714 if (! REG_P (reg) || (int) REGNO (reg) < new_regno_start)
715 return;
716 if (in_class_p (reg, cl, &rclass) && rclass != cl)
717 lra_change_class (REGNO (reg), rclass, " Change to", true);
718 }
719
720 /* Generate reloads for matching OUT and INS (array of input operand
721 numbers with end marker -1) with reg class GOAL_CLASS. Add input
722 and output reloads correspondingly to the lists *BEFORE and *AFTER.
723 OUT might be negative. In this case we generate input reloads for
724 matched input operands INS. */
725 static void
726 match_reload (signed char out, signed char *ins, enum reg_class goal_class,
727 rtx *before, rtx *after)
728 {
729 int i, in;
730 rtx new_in_reg, new_out_reg, reg, clobber;
731 enum machine_mode inmode, outmode;
732 rtx in_rtx = *curr_id->operand_loc[ins[0]];
733 rtx out_rtx = out < 0 ? in_rtx : *curr_id->operand_loc[out];
734
735 inmode = curr_operand_mode[ins[0]];
736 outmode = out < 0 ? inmode : curr_operand_mode[out];
737 push_to_sequence (*before);
738 if (inmode != outmode)
739 {
740 if (GET_MODE_SIZE (inmode) > GET_MODE_SIZE (outmode))
741 {
742 reg = new_in_reg
743 = lra_create_new_reg_with_unique_value (inmode, in_rtx,
744 goal_class, "");
745 if (SCALAR_INT_MODE_P (inmode))
746 new_out_reg = gen_lowpart_SUBREG (outmode, reg);
747 else
748 new_out_reg = gen_rtx_SUBREG (outmode, reg, 0);
749 LRA_SUBREG_P (new_out_reg) = 1;
750 /* If the input reg is dying here, we can use the same hard
751 register for REG and IN_RTX. We do it only for original
752 pseudos as reload pseudos can die although original
753 pseudos still live where reload pseudos dies. */
754 if (REG_P (in_rtx) && (int) REGNO (in_rtx) < lra_new_regno_start
755 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx)))
756 lra_assign_reg_val (REGNO (in_rtx), REGNO (reg));
757 }
758 else
759 {
760 reg = new_out_reg
761 = lra_create_new_reg_with_unique_value (outmode, out_rtx,
762 goal_class, "");
763 if (SCALAR_INT_MODE_P (outmode))
764 new_in_reg = gen_lowpart_SUBREG (inmode, reg);
765 else
766 new_in_reg = gen_rtx_SUBREG (inmode, reg, 0);
767 /* NEW_IN_REG is non-paradoxical subreg. We don't want
768 NEW_OUT_REG living above. We add clobber clause for
769 this. This is just a temporary clobber. We can remove
770 it at the end of LRA work. */
771 clobber = emit_clobber (new_out_reg);
772 LRA_TEMP_CLOBBER_P (PATTERN (clobber)) = 1;
773 LRA_SUBREG_P (new_in_reg) = 1;
774 if (GET_CODE (in_rtx) == SUBREG)
775 {
776 rtx subreg_reg = SUBREG_REG (in_rtx);
777
778 /* If SUBREG_REG is dying here and sub-registers IN_RTX
779 and NEW_IN_REG are similar, we can use the same hard
780 register for REG and SUBREG_REG. */
781 if (REG_P (subreg_reg)
782 && (int) REGNO (subreg_reg) < lra_new_regno_start
783 && GET_MODE (subreg_reg) == outmode
784 && SUBREG_BYTE (in_rtx) == SUBREG_BYTE (new_in_reg)
785 && find_regno_note (curr_insn, REG_DEAD, REGNO (subreg_reg)))
786 lra_assign_reg_val (REGNO (subreg_reg), REGNO (reg));
787 }
788 }
789 }
790 else
791 {
792 /* Pseudos have values -- see comments for lra_reg_info.
793 Different pseudos with the same value do not conflict even if
794 they live in the same place. When we create a pseudo we
795 assign value of original pseudo (if any) from which we
796 created the new pseudo. If we create the pseudo from the
797 input pseudo, the new pseudo will no conflict with the input
798 pseudo which is wrong when the input pseudo lives after the
799 insn and as the new pseudo value is changed by the insn
800 output. Therefore we create the new pseudo from the output.
801
802 We cannot reuse the current output register because we might
803 have a situation like "a <- a op b", where the constraints
804 force the second input operand ("b") to match the output
805 operand ("a"). "b" must then be copied into a new register
806 so that it doesn't clobber the current value of "a". */
807
808 new_in_reg = new_out_reg
809 = lra_create_new_reg_with_unique_value (outmode, out_rtx,
810 goal_class, "");
811 }
812 /* In operand can be got from transformations before processing insn
813 constraints. One example of such transformations is subreg
814 reloading (see function simplify_operand_subreg). The new
815 pseudos created by the transformations might have inaccurate
816 class (ALL_REGS) and we should make their classes more
817 accurate. */
818 narrow_reload_pseudo_class (in_rtx, goal_class);
819 lra_emit_move (copy_rtx (new_in_reg), in_rtx);
820 *before = get_insns ();
821 end_sequence ();
822 for (i = 0; (in = ins[i]) >= 0; i++)
823 {
824 lra_assert
825 (GET_MODE (*curr_id->operand_loc[in]) == VOIDmode
826 || GET_MODE (new_in_reg) == GET_MODE (*curr_id->operand_loc[in]));
827 *curr_id->operand_loc[in] = new_in_reg;
828 }
829 lra_update_dups (curr_id, ins);
830 if (out < 0)
831 return;
832 /* See a comment for the input operand above. */
833 narrow_reload_pseudo_class (out_rtx, goal_class);
834 if (find_reg_note (curr_insn, REG_UNUSED, out_rtx) == NULL_RTX)
835 {
836 start_sequence ();
837 lra_emit_move (out_rtx, copy_rtx (new_out_reg));
838 emit_insn (*after);
839 *after = get_insns ();
840 end_sequence ();
841 }
842 *curr_id->operand_loc[out] = new_out_reg;
843 lra_update_dup (curr_id, out);
844 }
845
846 /* Return register class which is union of all reg classes in insn
847 constraint alternative string starting with P. */
848 static enum reg_class
849 reg_class_from_constraints (const char *p)
850 {
851 int c, len;
852 enum reg_class op_class = NO_REGS;
853
854 do
855 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
856 {
857 case '#':
858 case ',':
859 return op_class;
860
861 case 'p':
862 op_class = (reg_class_subunion
863 [op_class][base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
864 ADDRESS, SCRATCH)]);
865 break;
866
867 case 'g':
868 case 'r':
869 op_class = reg_class_subunion[op_class][GENERAL_REGS];
870 break;
871
872 default:
873 if (REG_CLASS_FROM_CONSTRAINT (c, p) == NO_REGS)
874 {
875 #ifdef EXTRA_CONSTRAINT_STR
876 if (EXTRA_ADDRESS_CONSTRAINT (c, p))
877 op_class
878 = (reg_class_subunion
879 [op_class][base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
880 ADDRESS, SCRATCH)]);
881 #endif
882 break;
883 }
884
885 op_class
886 = reg_class_subunion[op_class][REG_CLASS_FROM_CONSTRAINT (c, p)];
887 break;
888 }
889 while ((p += len), c);
890 return op_class;
891 }
892
893 /* If OP is a register, return the class of the register as per
894 get_reg_class, otherwise return NO_REGS. */
895 static inline enum reg_class
896 get_op_class (rtx op)
897 {
898 return REG_P (op) ? get_reg_class (REGNO (op)) : NO_REGS;
899 }
900
901 /* Return generated insn mem_pseudo:=val if TO_P or val:=mem_pseudo
902 otherwise. If modes of MEM_PSEUDO and VAL are different, use
903 SUBREG for VAL to make them equal. */
904 static rtx
905 emit_spill_move (bool to_p, rtx mem_pseudo, rtx val)
906 {
907 if (GET_MODE (mem_pseudo) != GET_MODE (val))
908 {
909 /* Usually size of mem_pseudo is greater than val size but in
910 rare cases it can be less as it can be defined by target
911 dependent macro HARD_REGNO_CALLER_SAVE_MODE. */
912 if (! MEM_P (val))
913 {
914 val = gen_rtx_SUBREG (GET_MODE (mem_pseudo),
915 GET_CODE (val) == SUBREG ? SUBREG_REG (val) : val,
916 0);
917 LRA_SUBREG_P (val) = 1;
918 }
919 else
920 {
921 mem_pseudo = gen_lowpart_SUBREG (GET_MODE (val), mem_pseudo);
922 LRA_SUBREG_P (mem_pseudo) = 1;
923 }
924 }
925 return (to_p
926 ? gen_move_insn (mem_pseudo, val)
927 : gen_move_insn (val, mem_pseudo));
928 }
929
930 /* Process a special case insn (register move), return true if we
931 don't need to process it anymore. INSN should be a single set
932 insn. Set up that RTL was changed through CHANGE_P and macro
933 SECONDARY_MEMORY_NEEDED says to use secondary memory through
934 SEC_MEM_P. */
935 static bool
936 check_and_process_move (bool *change_p, bool *sec_mem_p ATTRIBUTE_UNUSED)
937 {
938 int sregno, dregno;
939 rtx dest, src, dreg, sreg, old_sreg, new_reg, before, scratch_reg;
940 enum reg_class dclass, sclass, secondary_class;
941 enum machine_mode sreg_mode;
942 secondary_reload_info sri;
943
944 lra_assert (curr_insn_set != NULL_RTX);
945 dreg = dest = SET_DEST (curr_insn_set);
946 sreg = src = SET_SRC (curr_insn_set);
947 if (GET_CODE (dest) == SUBREG)
948 dreg = SUBREG_REG (dest);
949 if (GET_CODE (src) == SUBREG)
950 sreg = SUBREG_REG (src);
951 if (! (REG_P (dreg) || MEM_P (dreg)) || ! (REG_P (sreg) || MEM_P (sreg)))
952 return false;
953 sclass = dclass = NO_REGS;
954 if (REG_P (dreg))
955 dclass = get_reg_class (REGNO (dreg));
956 if (dclass == ALL_REGS)
957 /* ALL_REGS is used for new pseudos created by transformations
958 like reload of SUBREG_REG (see function
959 simplify_operand_subreg). We don't know their class yet. We
960 should figure out the class from processing the insn
961 constraints not in this fast path function. Even if ALL_REGS
962 were a right class for the pseudo, secondary_... hooks usually
963 are not define for ALL_REGS. */
964 return false;
965 sreg_mode = GET_MODE (sreg);
966 old_sreg = sreg;
967 if (REG_P (sreg))
968 sclass = get_reg_class (REGNO (sreg));
969 if (sclass == ALL_REGS)
970 /* See comments above. */
971 return false;
972 if (sclass == NO_REGS && dclass == NO_REGS)
973 return false;
974 #ifdef SECONDARY_MEMORY_NEEDED
975 if (SECONDARY_MEMORY_NEEDED (sclass, dclass, GET_MODE (src))
976 #ifdef SECONDARY_MEMORY_NEEDED_MODE
977 && ((sclass != NO_REGS && dclass != NO_REGS)
978 || GET_MODE (src) != SECONDARY_MEMORY_NEEDED_MODE (GET_MODE (src)))
979 #endif
980 )
981 {
982 *sec_mem_p = true;
983 return false;
984 }
985 #endif
986 if (! REG_P (dreg) || ! REG_P (sreg))
987 return false;
988 sri.prev_sri = NULL;
989 sri.icode = CODE_FOR_nothing;
990 sri.extra_cost = 0;
991 secondary_class = NO_REGS;
992 /* Set up hard register for a reload pseudo for hook
993 secondary_reload because some targets just ignore unassigned
994 pseudos in the hook. */
995 if (dclass != NO_REGS && lra_get_regno_hard_regno (REGNO (dreg)) < 0)
996 {
997 dregno = REGNO (dreg);
998 reg_renumber[dregno] = ira_class_hard_regs[dclass][0];
999 }
1000 else
1001 dregno = -1;
1002 if (sclass != NO_REGS && lra_get_regno_hard_regno (REGNO (sreg)) < 0)
1003 {
1004 sregno = REGNO (sreg);
1005 reg_renumber[sregno] = ira_class_hard_regs[sclass][0];
1006 }
1007 else
1008 sregno = -1;
1009 if (sclass != NO_REGS)
1010 secondary_class
1011 = (enum reg_class) targetm.secondary_reload (false, dest,
1012 (reg_class_t) sclass,
1013 GET_MODE (src), &sri);
1014 if (sclass == NO_REGS
1015 || ((secondary_class != NO_REGS || sri.icode != CODE_FOR_nothing)
1016 && dclass != NO_REGS))
1017 {
1018 enum reg_class old_sclass = secondary_class;
1019 secondary_reload_info old_sri = sri;
1020
1021 sri.prev_sri = NULL;
1022 sri.icode = CODE_FOR_nothing;
1023 sri.extra_cost = 0;
1024 secondary_class
1025 = (enum reg_class) targetm.secondary_reload (true, sreg,
1026 (reg_class_t) dclass,
1027 sreg_mode, &sri);
1028 /* Check the target hook consistency. */
1029 lra_assert
1030 ((secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1031 || (old_sclass == NO_REGS && old_sri.icode == CODE_FOR_nothing)
1032 || (secondary_class == old_sclass && sri.icode == old_sri.icode));
1033 }
1034 if (sregno >= 0)
1035 reg_renumber [sregno] = -1;
1036 if (dregno >= 0)
1037 reg_renumber [dregno] = -1;
1038 if (secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1039 return false;
1040 *change_p = true;
1041 new_reg = NULL_RTX;
1042 if (secondary_class != NO_REGS)
1043 new_reg = lra_create_new_reg_with_unique_value (sreg_mode, NULL_RTX,
1044 secondary_class,
1045 "secondary");
1046 start_sequence ();
1047 if (old_sreg != sreg)
1048 sreg = copy_rtx (sreg);
1049 if (sri.icode == CODE_FOR_nothing)
1050 lra_emit_move (new_reg, sreg);
1051 else
1052 {
1053 enum reg_class scratch_class;
1054
1055 scratch_class = (reg_class_from_constraints
1056 (insn_data[sri.icode].operand[2].constraint));
1057 scratch_reg = (lra_create_new_reg_with_unique_value
1058 (insn_data[sri.icode].operand[2].mode, NULL_RTX,
1059 scratch_class, "scratch"));
1060 emit_insn (GEN_FCN (sri.icode) (new_reg != NULL_RTX ? new_reg : dest,
1061 sreg, scratch_reg));
1062 }
1063 before = get_insns ();
1064 end_sequence ();
1065 lra_process_new_insns (curr_insn, before, NULL_RTX, "Inserting the move");
1066 if (new_reg != NULL_RTX)
1067 {
1068 if (GET_CODE (src) == SUBREG)
1069 SUBREG_REG (src) = new_reg;
1070 else
1071 SET_SRC (curr_insn_set) = new_reg;
1072 }
1073 else
1074 {
1075 if (lra_dump_file != NULL)
1076 {
1077 fprintf (lra_dump_file, "Deleting move %u\n", INSN_UID (curr_insn));
1078 dump_insn_slim (lra_dump_file, curr_insn);
1079 }
1080 lra_set_insn_deleted (curr_insn);
1081 return true;
1082 }
1083 return false;
1084 }
1085
1086 /* The following data describe the result of process_alt_operands.
1087 The data are used in curr_insn_transform to generate reloads. */
1088
1089 /* The chosen reg classes which should be used for the corresponding
1090 operands. */
1091 static enum reg_class goal_alt[MAX_RECOG_OPERANDS];
1092 /* True if the operand should be the same as another operand and that
1093 other operand does not need a reload. */
1094 static bool goal_alt_match_win[MAX_RECOG_OPERANDS];
1095 /* True if the operand does not need a reload. */
1096 static bool goal_alt_win[MAX_RECOG_OPERANDS];
1097 /* True if the operand can be offsetable memory. */
1098 static bool goal_alt_offmemok[MAX_RECOG_OPERANDS];
1099 /* The number of an operand to which given operand can be matched to. */
1100 static int goal_alt_matches[MAX_RECOG_OPERANDS];
1101 /* The number of elements in the following array. */
1102 static int goal_alt_dont_inherit_ops_num;
1103 /* Numbers of operands whose reload pseudos should not be inherited. */
1104 static int goal_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1105 /* True if the insn commutative operands should be swapped. */
1106 static bool goal_alt_swapped;
1107 /* The chosen insn alternative. */
1108 static int goal_alt_number;
1109
1110 /* The following five variables are used to choose the best insn
1111 alternative. They reflect final characteristics of the best
1112 alternative. */
1113
1114 /* Number of necessary reloads and overall cost reflecting the
1115 previous value and other unpleasantness of the best alternative. */
1116 static int best_losers, best_overall;
1117 /* Overall number hard registers used for reloads. For example, on
1118 some targets we need 2 general registers to reload DFmode and only
1119 one floating point register. */
1120 static int best_reload_nregs;
1121 /* Overall number reflecting distances of previous reloading the same
1122 value. The distances are counted from the current BB start. It is
1123 used to improve inheritance chances. */
1124 static int best_reload_sum;
1125
1126 /* True if the current insn should have no correspondingly input or
1127 output reloads. */
1128 static bool no_input_reloads_p, no_output_reloads_p;
1129
1130 /* True if we swapped the commutative operands in the current
1131 insn. */
1132 static int curr_swapped;
1133
1134 /* Arrange for address element *LOC to be a register of class CL.
1135 Add any input reloads to list BEFORE. AFTER is nonnull if *LOC is an
1136 automodified value; handle that case by adding the required output
1137 reloads to list AFTER. Return true if the RTL was changed. */
1138 static bool
1139 process_addr_reg (rtx *loc, rtx *before, rtx *after, enum reg_class cl)
1140 {
1141 int regno;
1142 enum reg_class rclass, new_class;
1143 rtx reg;
1144 rtx new_reg;
1145 enum machine_mode mode;
1146 bool subreg_p, before_p = false;
1147
1148 subreg_p = GET_CODE (*loc) == SUBREG;
1149 if (subreg_p)
1150 loc = &SUBREG_REG (*loc);
1151 reg = *loc;
1152 mode = GET_MODE (reg);
1153 if (! REG_P (reg))
1154 {
1155 /* Always reload memory in an address even if the target supports
1156 such addresses. */
1157 new_reg = lra_create_new_reg_with_unique_value (mode, reg, cl, "address");
1158 before_p = true;
1159 }
1160 else
1161 {
1162 regno = REGNO (reg);
1163 rclass = get_reg_class (regno);
1164 if ((*loc = get_equiv_with_elimination (reg, curr_insn)) != reg)
1165 {
1166 if (lra_dump_file != NULL)
1167 {
1168 fprintf (lra_dump_file,
1169 "Changing pseudo %d in address of insn %u on equiv ",
1170 REGNO (reg), INSN_UID (curr_insn));
1171 dump_value_slim (lra_dump_file, *loc, 1);
1172 fprintf (lra_dump_file, "\n");
1173 }
1174 *loc = copy_rtx (*loc);
1175 }
1176 if (*loc != reg || ! in_class_p (reg, cl, &new_class))
1177 {
1178 reg = *loc;
1179 if (get_reload_reg (after == NULL ? OP_IN : OP_INOUT,
1180 mode, reg, cl, subreg_p, "address", &new_reg))
1181 before_p = true;
1182 }
1183 else if (new_class != NO_REGS && rclass != new_class)
1184 {
1185 lra_change_class (regno, new_class, " Change to", true);
1186 return false;
1187 }
1188 else
1189 return false;
1190 }
1191 if (before_p)
1192 {
1193 push_to_sequence (*before);
1194 lra_emit_move (new_reg, reg);
1195 *before = get_insns ();
1196 end_sequence ();
1197 }
1198 *loc = new_reg;
1199 if (after != NULL)
1200 {
1201 start_sequence ();
1202 lra_emit_move (reg, new_reg);
1203 emit_insn (*after);
1204 *after = get_insns ();
1205 end_sequence ();
1206 }
1207 return true;
1208 }
1209
1210 /* Insert move insn in simplify_operand_subreg. BEFORE returns
1211 the insn to be inserted before curr insn. AFTER returns the
1212 the insn to be inserted after curr insn. ORIGREG and NEWREG
1213 are the original reg and new reg for reload. */
1214 static void
1215 insert_move_for_subreg (rtx *before, rtx *after, rtx origreg, rtx newreg)
1216 {
1217 if (before)
1218 {
1219 push_to_sequence (*before);
1220 lra_emit_move (newreg, origreg);
1221 *before = get_insns ();
1222 end_sequence ();
1223 }
1224 if (after)
1225 {
1226 start_sequence ();
1227 lra_emit_move (origreg, newreg);
1228 emit_insn (*after);
1229 *after = get_insns ();
1230 end_sequence ();
1231 }
1232 }
1233
1234 /* Make reloads for subreg in operand NOP with internal subreg mode
1235 REG_MODE, add new reloads for further processing. Return true if
1236 any reload was generated. */
1237 static bool
1238 simplify_operand_subreg (int nop, enum machine_mode reg_mode)
1239 {
1240 int hard_regno;
1241 rtx before, after;
1242 enum machine_mode mode;
1243 rtx reg, new_reg;
1244 rtx operand = *curr_id->operand_loc[nop];
1245 enum reg_class regclass;
1246 enum op_type type;
1247
1248 before = after = NULL_RTX;
1249
1250 if (GET_CODE (operand) != SUBREG)
1251 return false;
1252
1253 mode = GET_MODE (operand);
1254 reg = SUBREG_REG (operand);
1255 type = curr_static_id->operand[nop].type;
1256 /* If we change address for paradoxical subreg of memory, the
1257 address might violate the necessary alignment or the access might
1258 be slow. So take this into consideration. We should not worry
1259 about access beyond allocated memory for paradoxical memory
1260 subregs as we don't substitute such equiv memory (see processing
1261 equivalences in function lra_constraints) and because for spilled
1262 pseudos we allocate stack memory enough for the biggest
1263 corresponding paradoxical subreg. */
1264 if ((MEM_P (reg)
1265 && (! SLOW_UNALIGNED_ACCESS (mode, MEM_ALIGN (reg))
1266 || MEM_ALIGN (reg) >= GET_MODE_ALIGNMENT (mode)))
1267 || (REG_P (reg) && REGNO (reg) < FIRST_PSEUDO_REGISTER))
1268 {
1269 alter_subreg (curr_id->operand_loc[nop], false);
1270 return true;
1271 }
1272 /* Put constant into memory when we have mixed modes. It generates
1273 a better code in most cases as it does not need a secondary
1274 reload memory. It also prevents LRA looping when LRA is using
1275 secondary reload memory again and again. */
1276 if (CONSTANT_P (reg) && CONST_POOL_OK_P (reg_mode, reg)
1277 && SCALAR_INT_MODE_P (reg_mode) != SCALAR_INT_MODE_P (mode))
1278 {
1279 SUBREG_REG (operand) = force_const_mem (reg_mode, reg);
1280 alter_subreg (curr_id->operand_loc[nop], false);
1281 return true;
1282 }
1283 /* Force a reload of the SUBREG_REG if this is a constant or PLUS or
1284 if there may be a problem accessing OPERAND in the outer
1285 mode. */
1286 if ((REG_P (reg)
1287 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1288 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1289 /* Don't reload paradoxical subregs because we could be looping
1290 having repeatedly final regno out of hard regs range. */
1291 && (hard_regno_nregs[hard_regno][GET_MODE (reg)]
1292 >= hard_regno_nregs[hard_regno][mode])
1293 && simplify_subreg_regno (hard_regno, GET_MODE (reg),
1294 SUBREG_BYTE (operand), mode) < 0
1295 /* Don't reload subreg for matching reload. It is actually
1296 valid subreg in LRA. */
1297 && ! LRA_SUBREG_P (operand))
1298 || CONSTANT_P (reg) || GET_CODE (reg) == PLUS || MEM_P (reg))
1299 {
1300 enum reg_class rclass;
1301
1302 if (REG_P (reg))
1303 /* There is a big probability that we will get the same class
1304 for the new pseudo and we will get the same insn which
1305 means infinite looping. So spill the new pseudo. */
1306 rclass = NO_REGS;
1307 else
1308 /* The class will be defined later in curr_insn_transform. */
1309 rclass
1310 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1311
1312 if (get_reload_reg (curr_static_id->operand[nop].type, reg_mode, reg,
1313 rclass, TRUE, "subreg reg", &new_reg))
1314 {
1315 bool insert_before, insert_after;
1316 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1317
1318 insert_before = (type != OP_OUT
1319 || GET_MODE_SIZE (GET_MODE (reg)) > GET_MODE_SIZE (mode));
1320 insert_after = (type != OP_IN);
1321 insert_move_for_subreg (insert_before ? &before : NULL,
1322 insert_after ? &after : NULL,
1323 reg, new_reg);
1324 }
1325 SUBREG_REG (operand) = new_reg;
1326 lra_process_new_insns (curr_insn, before, after,
1327 "Inserting subreg reload");
1328 return true;
1329 }
1330 /* Force a reload for a paradoxical subreg. For paradoxical subreg,
1331 IRA allocates hardreg to the inner pseudo reg according to its mode
1332 instead of the outermode, so the size of the hardreg may not be enough
1333 to contain the outermode operand, in that case we may need to insert
1334 reload for the reg. For the following two types of paradoxical subreg,
1335 we need to insert reload:
1336 1. If the op_type is OP_IN, and the hardreg could not be paired with
1337 other hardreg to contain the outermode operand
1338 (checked by in_hard_reg_set_p), we need to insert the reload.
1339 2. If the op_type is OP_OUT or OP_INOUT.
1340
1341 Here is a paradoxical subreg example showing how the reload is generated:
1342
1343 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1344 (subreg:TI (reg:DI 107 [ __comp ]) 0)) {*movti_internal_rex64}
1345
1346 In IRA, reg107 is allocated to a DImode hardreg. We use x86-64 as example
1347 here, if reg107 is assigned to hardreg R15, because R15 is the last
1348 hardreg, compiler cannot find another hardreg to pair with R15 to
1349 contain TImode data. So we insert a TImode reload reg180 for it.
1350 After reload is inserted:
1351
1352 (insn 283 0 0 (set (subreg:DI (reg:TI 180 [orig:107 __comp ] [107]) 0)
1353 (reg:DI 107 [ __comp ])) -1
1354 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1355 (subreg:TI (reg:TI 180 [orig:107 __comp ] [107]) 0)) {*movti_internal_rex64}
1356
1357 Two reload hard registers will be allocated to reg180 to save TImode data
1358 in LRA_assign. */
1359 else if (REG_P (reg)
1360 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1361 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1362 && (hard_regno_nregs[hard_regno][GET_MODE (reg)]
1363 < hard_regno_nregs[hard_regno][mode])
1364 && (regclass = lra_get_allocno_class (REGNO (reg)))
1365 && (type != OP_IN
1366 || !in_hard_reg_set_p (reg_class_contents[regclass],
1367 mode, hard_regno)))
1368 {
1369 /* The class will be defined later in curr_insn_transform. */
1370 enum reg_class rclass
1371 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1372
1373 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1374 rclass, TRUE, "paradoxical subreg", &new_reg))
1375 {
1376 rtx subreg;
1377 bool insert_before, insert_after;
1378
1379 PUT_MODE (new_reg, mode);
1380 subreg = simplify_gen_subreg (GET_MODE (reg), new_reg, mode, 0);
1381 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1382
1383 insert_before = (type != OP_OUT);
1384 insert_after = (type != OP_IN);
1385 insert_move_for_subreg (insert_before ? &before : NULL,
1386 insert_after ? &after : NULL,
1387 reg, subreg);
1388 }
1389 SUBREG_REG (operand) = new_reg;
1390 lra_process_new_insns (curr_insn, before, after,
1391 "Inserting paradoxical subreg reload");
1392 return true;
1393 }
1394 return false;
1395 }
1396
1397 /* Return TRUE if X refers for a hard register from SET. */
1398 static bool
1399 uses_hard_regs_p (rtx x, HARD_REG_SET set)
1400 {
1401 int i, j, x_hard_regno;
1402 enum machine_mode mode;
1403 const char *fmt;
1404 enum rtx_code code;
1405
1406 if (x == NULL_RTX)
1407 return false;
1408 code = GET_CODE (x);
1409 mode = GET_MODE (x);
1410 if (code == SUBREG)
1411 {
1412 x = SUBREG_REG (x);
1413 code = GET_CODE (x);
1414 if (GET_MODE_SIZE (GET_MODE (x)) > GET_MODE_SIZE (mode))
1415 mode = GET_MODE (x);
1416 }
1417
1418 if (REG_P (x))
1419 {
1420 x_hard_regno = get_hard_regno (x);
1421 return (x_hard_regno >= 0
1422 && overlaps_hard_reg_set_p (set, mode, x_hard_regno));
1423 }
1424 if (MEM_P (x))
1425 {
1426 struct address_info ad;
1427
1428 decompose_mem_address (&ad, x);
1429 if (ad.base_term != NULL && uses_hard_regs_p (*ad.base_term, set))
1430 return true;
1431 if (ad.index_term != NULL && uses_hard_regs_p (*ad.index_term, set))
1432 return true;
1433 }
1434 fmt = GET_RTX_FORMAT (code);
1435 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1436 {
1437 if (fmt[i] == 'e')
1438 {
1439 if (uses_hard_regs_p (XEXP (x, i), set))
1440 return true;
1441 }
1442 else if (fmt[i] == 'E')
1443 {
1444 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1445 if (uses_hard_regs_p (XVECEXP (x, i, j), set))
1446 return true;
1447 }
1448 }
1449 return false;
1450 }
1451
1452 /* Return true if OP is a spilled pseudo. */
1453 static inline bool
1454 spilled_pseudo_p (rtx op)
1455 {
1456 return (REG_P (op)
1457 && REGNO (op) >= FIRST_PSEUDO_REGISTER && in_mem_p (REGNO (op)));
1458 }
1459
1460 /* Return true if X is a general constant. */
1461 static inline bool
1462 general_constant_p (rtx x)
1463 {
1464 return CONSTANT_P (x) && (! flag_pic || LEGITIMATE_PIC_OPERAND_P (x));
1465 }
1466
1467 static bool
1468 reg_in_class_p (rtx reg, enum reg_class cl)
1469 {
1470 if (cl == NO_REGS)
1471 return get_reg_class (REGNO (reg)) == NO_REGS;
1472 return in_class_p (reg, cl, NULL);
1473 }
1474
1475 /* Major function to choose the current insn alternative and what
1476 operands should be reloaded and how. If ONLY_ALTERNATIVE is not
1477 negative we should consider only this alternative. Return false if
1478 we can not choose the alternative or find how to reload the
1479 operands. */
1480 static bool
1481 process_alt_operands (int only_alternative)
1482 {
1483 bool ok_p = false;
1484 int nop, overall, nalt;
1485 int n_alternatives = curr_static_id->n_alternatives;
1486 int n_operands = curr_static_id->n_operands;
1487 /* LOSERS counts the operands that don't fit this alternative and
1488 would require loading. */
1489 int losers;
1490 /* REJECT is a count of how undesirable this alternative says it is
1491 if any reloading is required. If the alternative matches exactly
1492 then REJECT is ignored, but otherwise it gets this much counted
1493 against it in addition to the reloading needed. */
1494 int reject;
1495 /* The number of elements in the following array. */
1496 int early_clobbered_regs_num;
1497 /* Numbers of operands which are early clobber registers. */
1498 int early_clobbered_nops[MAX_RECOG_OPERANDS];
1499 enum reg_class curr_alt[MAX_RECOG_OPERANDS];
1500 HARD_REG_SET curr_alt_set[MAX_RECOG_OPERANDS];
1501 bool curr_alt_match_win[MAX_RECOG_OPERANDS];
1502 bool curr_alt_win[MAX_RECOG_OPERANDS];
1503 bool curr_alt_offmemok[MAX_RECOG_OPERANDS];
1504 int curr_alt_matches[MAX_RECOG_OPERANDS];
1505 /* The number of elements in the following array. */
1506 int curr_alt_dont_inherit_ops_num;
1507 /* Numbers of operands whose reload pseudos should not be inherited. */
1508 int curr_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1509 rtx op;
1510 /* The register when the operand is a subreg of register, otherwise the
1511 operand itself. */
1512 rtx no_subreg_reg_operand[MAX_RECOG_OPERANDS];
1513 /* The register if the operand is a register or subreg of register,
1514 otherwise NULL. */
1515 rtx operand_reg[MAX_RECOG_OPERANDS];
1516 int hard_regno[MAX_RECOG_OPERANDS];
1517 enum machine_mode biggest_mode[MAX_RECOG_OPERANDS];
1518 int reload_nregs, reload_sum;
1519 bool costly_p;
1520 enum reg_class cl;
1521
1522 /* Calculate some data common for all alternatives to speed up the
1523 function. */
1524 for (nop = 0; nop < n_operands; nop++)
1525 {
1526 rtx reg;
1527
1528 op = no_subreg_reg_operand[nop] = *curr_id->operand_loc[nop];
1529 /* The real hard regno of the operand after the allocation. */
1530 hard_regno[nop] = get_hard_regno (op);
1531
1532 operand_reg[nop] = reg = op;
1533 biggest_mode[nop] = GET_MODE (op);
1534 if (GET_CODE (op) == SUBREG)
1535 {
1536 operand_reg[nop] = reg = SUBREG_REG (op);
1537 if (GET_MODE_SIZE (biggest_mode[nop])
1538 < GET_MODE_SIZE (GET_MODE (reg)))
1539 biggest_mode[nop] = GET_MODE (reg);
1540 }
1541 if (! REG_P (reg))
1542 operand_reg[nop] = NULL_RTX;
1543 else if (REGNO (reg) >= FIRST_PSEUDO_REGISTER
1544 || ((int) REGNO (reg)
1545 == lra_get_elimination_hard_regno (REGNO (reg))))
1546 no_subreg_reg_operand[nop] = reg;
1547 else
1548 operand_reg[nop] = no_subreg_reg_operand[nop]
1549 /* Just use natural mode for elimination result. It should
1550 be enough for extra constraints hooks. */
1551 = regno_reg_rtx[hard_regno[nop]];
1552 }
1553
1554 /* The constraints are made of several alternatives. Each operand's
1555 constraint looks like foo,bar,... with commas separating the
1556 alternatives. The first alternatives for all operands go
1557 together, the second alternatives go together, etc.
1558
1559 First loop over alternatives. */
1560 alternative_mask enabled = curr_id->enabled_alternatives;
1561 if (only_alternative >= 0)
1562 enabled &= ALTERNATIVE_BIT (only_alternative);
1563
1564 for (nalt = 0; nalt < n_alternatives; nalt++)
1565 {
1566 /* Loop over operands for one constraint alternative. */
1567 if (!TEST_BIT (enabled, nalt))
1568 continue;
1569
1570 overall = losers = reject = reload_nregs = reload_sum = 0;
1571 for (nop = 0; nop < n_operands; nop++)
1572 {
1573 int inc = (curr_static_id
1574 ->operand_alternative[nalt * n_operands + nop].reject);
1575 if (lra_dump_file != NULL && inc != 0)
1576 fprintf (lra_dump_file,
1577 " Staticly defined alt reject+=%d\n", inc);
1578 reject += inc;
1579 }
1580 early_clobbered_regs_num = 0;
1581
1582 for (nop = 0; nop < n_operands; nop++)
1583 {
1584 const char *p;
1585 char *end;
1586 int len, c, m, i, opalt_num, this_alternative_matches;
1587 bool win, did_match, offmemok, early_clobber_p;
1588 /* false => this operand can be reloaded somehow for this
1589 alternative. */
1590 bool badop;
1591 /* true => this operand can be reloaded if the alternative
1592 allows regs. */
1593 bool winreg;
1594 /* True if a constant forced into memory would be OK for
1595 this operand. */
1596 bool constmemok;
1597 enum reg_class this_alternative, this_costly_alternative;
1598 HARD_REG_SET this_alternative_set, this_costly_alternative_set;
1599 bool this_alternative_match_win, this_alternative_win;
1600 bool this_alternative_offmemok;
1601 bool scratch_p;
1602 enum machine_mode mode;
1603
1604 opalt_num = nalt * n_operands + nop;
1605 if (curr_static_id->operand_alternative[opalt_num].anything_ok)
1606 {
1607 /* Fast track for no constraints at all. */
1608 curr_alt[nop] = NO_REGS;
1609 CLEAR_HARD_REG_SET (curr_alt_set[nop]);
1610 curr_alt_win[nop] = true;
1611 curr_alt_match_win[nop] = false;
1612 curr_alt_offmemok[nop] = false;
1613 curr_alt_matches[nop] = -1;
1614 continue;
1615 }
1616
1617 op = no_subreg_reg_operand[nop];
1618 mode = curr_operand_mode[nop];
1619
1620 win = did_match = winreg = offmemok = constmemok = false;
1621 badop = true;
1622
1623 early_clobber_p = false;
1624 p = curr_static_id->operand_alternative[opalt_num].constraint;
1625
1626 this_costly_alternative = this_alternative = NO_REGS;
1627 /* We update set of possible hard regs besides its class
1628 because reg class might be inaccurate. For example,
1629 union of LO_REGS (l), HI_REGS(h), and STACK_REG(k) in ARM
1630 is translated in HI_REGS because classes are merged by
1631 pairs and there is no accurate intermediate class. */
1632 CLEAR_HARD_REG_SET (this_alternative_set);
1633 CLEAR_HARD_REG_SET (this_costly_alternative_set);
1634 this_alternative_win = false;
1635 this_alternative_match_win = false;
1636 this_alternative_offmemok = false;
1637 this_alternative_matches = -1;
1638
1639 /* An empty constraint should be excluded by the fast
1640 track. */
1641 lra_assert (*p != 0 && *p != ',');
1642
1643 /* Scan this alternative's specs for this operand; set WIN
1644 if the operand fits any letter in this alternative.
1645 Otherwise, clear BADOP if this operand could fit some
1646 letter after reloads, or set WINREG if this operand could
1647 fit after reloads provided the constraint allows some
1648 registers. */
1649 costly_p = false;
1650 do
1651 {
1652 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
1653 {
1654 case '\0':
1655 len = 0;
1656 break;
1657 case ',':
1658 c = '\0';
1659 break;
1660
1661 case '=': case '+': case '?': case '*': case '!':
1662 case ' ': case '\t':
1663 break;
1664
1665 case '%':
1666 /* We only support one commutative marker, the first
1667 one. We already set commutative above. */
1668 break;
1669
1670 case '&':
1671 early_clobber_p = true;
1672 break;
1673
1674 case '#':
1675 /* Ignore rest of this alternative. */
1676 c = '\0';
1677 break;
1678
1679 case '0': case '1': case '2': case '3': case '4':
1680 case '5': case '6': case '7': case '8': case '9':
1681 {
1682 int m_hregno;
1683 bool match_p;
1684
1685 m = strtoul (p, &end, 10);
1686 p = end;
1687 len = 0;
1688 lra_assert (nop > m);
1689
1690 this_alternative_matches = m;
1691 m_hregno = get_hard_regno (*curr_id->operand_loc[m]);
1692 /* We are supposed to match a previous operand.
1693 If we do, we win if that one did. If we do
1694 not, count both of the operands as losers.
1695 (This is too conservative, since most of the
1696 time only a single reload insn will be needed
1697 to make the two operands win. As a result,
1698 this alternative may be rejected when it is
1699 actually desirable.) */
1700 match_p = false;
1701 if (operands_match_p (*curr_id->operand_loc[nop],
1702 *curr_id->operand_loc[m], m_hregno))
1703 {
1704 /* We should reject matching of an early
1705 clobber operand if the matching operand is
1706 not dying in the insn. */
1707 if (! curr_static_id->operand[m].early_clobber
1708 || operand_reg[nop] == NULL_RTX
1709 || (find_regno_note (curr_insn, REG_DEAD,
1710 REGNO (op))
1711 || REGNO (op) == REGNO (operand_reg[m])))
1712 match_p = true;
1713 }
1714 if (match_p)
1715 {
1716 /* If we are matching a non-offsettable
1717 address where an offsettable address was
1718 expected, then we must reject this
1719 combination, because we can't reload
1720 it. */
1721 if (curr_alt_offmemok[m]
1722 && MEM_P (*curr_id->operand_loc[m])
1723 && curr_alt[m] == NO_REGS && ! curr_alt_win[m])
1724 continue;
1725 }
1726 else
1727 {
1728 /* Operands don't match. Both operands must
1729 allow a reload register, otherwise we
1730 cannot make them match. */
1731 if (curr_alt[m] == NO_REGS)
1732 break;
1733 /* Retroactively mark the operand we had to
1734 match as a loser, if it wasn't already and
1735 it wasn't matched to a register constraint
1736 (e.g it might be matched by memory). */
1737 if (curr_alt_win[m]
1738 && (operand_reg[m] == NULL_RTX
1739 || hard_regno[m] < 0))
1740 {
1741 losers++;
1742 reload_nregs
1743 += (ira_reg_class_max_nregs[curr_alt[m]]
1744 [GET_MODE (*curr_id->operand_loc[m])]);
1745 }
1746
1747 /* Prefer matching earlyclobber alternative as
1748 it results in less hard regs required for
1749 the insn than a non-matching earlyclobber
1750 alternative. */
1751 if (curr_static_id->operand[m].early_clobber)
1752 {
1753 if (lra_dump_file != NULL)
1754 fprintf
1755 (lra_dump_file,
1756 " %d Matching earlyclobber alt:"
1757 " reject--\n",
1758 nop);
1759 reject--;
1760 }
1761 /* Otherwise we prefer no matching
1762 alternatives because it gives more freedom
1763 in RA. */
1764 else if (operand_reg[nop] == NULL_RTX
1765 || (find_regno_note (curr_insn, REG_DEAD,
1766 REGNO (operand_reg[nop]))
1767 == NULL_RTX))
1768 {
1769 if (lra_dump_file != NULL)
1770 fprintf
1771 (lra_dump_file,
1772 " %d Matching alt: reject+=2\n",
1773 nop);
1774 reject += 2;
1775 }
1776 }
1777 /* If we have to reload this operand and some
1778 previous operand also had to match the same
1779 thing as this operand, we don't know how to do
1780 that. */
1781 if (!match_p || !curr_alt_win[m])
1782 {
1783 for (i = 0; i < nop; i++)
1784 if (curr_alt_matches[i] == m)
1785 break;
1786 if (i < nop)
1787 break;
1788 }
1789 else
1790 did_match = true;
1791
1792 /* This can be fixed with reloads if the operand
1793 we are supposed to match can be fixed with
1794 reloads. */
1795 badop = false;
1796 this_alternative = curr_alt[m];
1797 COPY_HARD_REG_SET (this_alternative_set, curr_alt_set[m]);
1798 winreg = this_alternative != NO_REGS;
1799 break;
1800 }
1801
1802 case 'p':
1803 cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
1804 ADDRESS, SCRATCH);
1805 this_alternative = reg_class_subunion[this_alternative][cl];
1806 IOR_HARD_REG_SET (this_alternative_set,
1807 reg_class_contents[cl]);
1808 if (costly_p)
1809 {
1810 this_costly_alternative
1811 = reg_class_subunion[this_costly_alternative][cl];
1812 IOR_HARD_REG_SET (this_costly_alternative_set,
1813 reg_class_contents[cl]);
1814 }
1815 win = true;
1816 badop = false;
1817 break;
1818
1819 case TARGET_MEM_CONSTRAINT:
1820 if (MEM_P (op) || spilled_pseudo_p (op))
1821 win = true;
1822 /* We can put constant or pseudo value into memory
1823 to satisfy the constraint. */
1824 if (CONST_POOL_OK_P (mode, op) || REG_P (op))
1825 badop = false;
1826 constmemok = true;
1827 break;
1828
1829 case '<':
1830 if (MEM_P (op)
1831 && (GET_CODE (XEXP (op, 0)) == PRE_DEC
1832 || GET_CODE (XEXP (op, 0)) == POST_DEC))
1833 win = true;
1834 break;
1835
1836 case '>':
1837 if (MEM_P (op)
1838 && (GET_CODE (XEXP (op, 0)) == PRE_INC
1839 || GET_CODE (XEXP (op, 0)) == POST_INC))
1840 win = true;
1841 break;
1842
1843 /* Memory op whose address is not offsettable. */
1844 case 'V':
1845 if (MEM_P (op)
1846 && ! offsettable_nonstrict_memref_p (op))
1847 win = true;
1848 break;
1849
1850 /* Memory operand whose address is offsettable. */
1851 case 'o':
1852 if ((MEM_P (op)
1853 && offsettable_nonstrict_memref_p (op))
1854 || spilled_pseudo_p (op))
1855 win = true;
1856 /* We can put constant or pseudo value into memory
1857 or make memory address offsetable to satisfy the
1858 constraint. */
1859 if (CONST_POOL_OK_P (mode, op) || MEM_P (op) || REG_P (op))
1860 badop = false;
1861 constmemok = true;
1862 offmemok = true;
1863 break;
1864
1865 case 'E':
1866 case 'F':
1867 if (GET_CODE (op) == CONST_DOUBLE
1868 || (GET_CODE (op) == CONST_VECTOR
1869 && (GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT)))
1870 win = true;
1871 break;
1872
1873 case 'G':
1874 case 'H':
1875 if (CONST_DOUBLE_AS_FLOAT_P (op)
1876 && CONST_DOUBLE_OK_FOR_CONSTRAINT_P (op, c, p))
1877 win = true;
1878 break;
1879
1880 case 's':
1881 if (CONST_SCALAR_INT_P (op))
1882 break;
1883
1884 case 'i':
1885 if (general_constant_p (op))
1886 win = true;
1887 break;
1888
1889 case 'n':
1890 if (CONST_SCALAR_INT_P (op))
1891 win = true;
1892 break;
1893
1894 case 'I':
1895 case 'J':
1896 case 'K':
1897 case 'L':
1898 case 'M':
1899 case 'N':
1900 case 'O':
1901 case 'P':
1902 if (CONST_INT_P (op)
1903 && CONST_OK_FOR_CONSTRAINT_P (INTVAL (op), c, p))
1904 win = true;
1905 break;
1906
1907 case 'X':
1908 /* This constraint should be excluded by the fast
1909 track. */
1910 gcc_unreachable ();
1911 break;
1912
1913 case 'g':
1914 if (MEM_P (op)
1915 || general_constant_p (op)
1916 || spilled_pseudo_p (op))
1917 win = true;
1918 /* Drop through into 'r' case. */
1919
1920 case 'r':
1921 this_alternative
1922 = reg_class_subunion[this_alternative][GENERAL_REGS];
1923 IOR_HARD_REG_SET (this_alternative_set,
1924 reg_class_contents[GENERAL_REGS]);
1925 if (costly_p)
1926 {
1927 this_costly_alternative
1928 = (reg_class_subunion
1929 [this_costly_alternative][GENERAL_REGS]);
1930 IOR_HARD_REG_SET (this_costly_alternative_set,
1931 reg_class_contents[GENERAL_REGS]);
1932 }
1933 goto reg;
1934
1935 default:
1936 if (REG_CLASS_FROM_CONSTRAINT (c, p) == NO_REGS)
1937 {
1938 #ifdef EXTRA_CONSTRAINT_STR
1939 if (EXTRA_MEMORY_CONSTRAINT (c, p))
1940 {
1941 if (EXTRA_CONSTRAINT_STR (op, c, p))
1942 win = true;
1943 else if (spilled_pseudo_p (op))
1944 win = true;
1945
1946 /* If we didn't already win, we can reload
1947 constants via force_const_mem or put the
1948 pseudo value into memory, or make other
1949 memory by reloading the address like for
1950 'o'. */
1951 if (CONST_POOL_OK_P (mode, op)
1952 || MEM_P (op) || REG_P (op))
1953 badop = false;
1954 constmemok = true;
1955 offmemok = true;
1956 break;
1957 }
1958 if (EXTRA_ADDRESS_CONSTRAINT (c, p))
1959 {
1960 if (EXTRA_CONSTRAINT_STR (op, c, p))
1961 win = true;
1962
1963 /* If we didn't already win, we can reload
1964 the address into a base register. */
1965 cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
1966 ADDRESS, SCRATCH);
1967 this_alternative
1968 = reg_class_subunion[this_alternative][cl];
1969 IOR_HARD_REG_SET (this_alternative_set,
1970 reg_class_contents[cl]);
1971 if (costly_p)
1972 {
1973 this_costly_alternative
1974 = (reg_class_subunion
1975 [this_costly_alternative][cl]);
1976 IOR_HARD_REG_SET (this_costly_alternative_set,
1977 reg_class_contents[cl]);
1978 }
1979 badop = false;
1980 break;
1981 }
1982
1983 if (EXTRA_CONSTRAINT_STR (op, c, p))
1984 win = true;
1985 #endif
1986 break;
1987 }
1988
1989 cl = REG_CLASS_FROM_CONSTRAINT (c, p);
1990 this_alternative = reg_class_subunion[this_alternative][cl];
1991 IOR_HARD_REG_SET (this_alternative_set,
1992 reg_class_contents[cl]);
1993 if (costly_p)
1994 {
1995 this_costly_alternative
1996 = reg_class_subunion[this_costly_alternative][cl];
1997 IOR_HARD_REG_SET (this_costly_alternative_set,
1998 reg_class_contents[cl]);
1999 }
2000 reg:
2001 if (mode == BLKmode)
2002 break;
2003 winreg = true;
2004 if (REG_P (op))
2005 {
2006 if (hard_regno[nop] >= 0
2007 && in_hard_reg_set_p (this_alternative_set,
2008 mode, hard_regno[nop]))
2009 win = true;
2010 else if (hard_regno[nop] < 0
2011 && in_class_p (op, this_alternative, NULL))
2012 win = true;
2013 }
2014 break;
2015 }
2016 if (c != ' ' && c != '\t')
2017 costly_p = c == '*';
2018 }
2019 while ((p += len), c);
2020
2021 scratch_p = (operand_reg[nop] != NULL_RTX
2022 && lra_former_scratch_p (REGNO (operand_reg[nop])));
2023 /* Record which operands fit this alternative. */
2024 if (win)
2025 {
2026 this_alternative_win = true;
2027 if (operand_reg[nop] != NULL_RTX)
2028 {
2029 if (hard_regno[nop] >= 0)
2030 {
2031 if (in_hard_reg_set_p (this_costly_alternative_set,
2032 mode, hard_regno[nop]))
2033 {
2034 if (lra_dump_file != NULL)
2035 fprintf (lra_dump_file,
2036 " %d Costly set: reject++\n",
2037 nop);
2038 reject++;
2039 }
2040 }
2041 else
2042 {
2043 /* Prefer won reg to spilled pseudo under other
2044 equal conditions for possibe inheritance. */
2045 if (! scratch_p)
2046 {
2047 if (lra_dump_file != NULL)
2048 fprintf
2049 (lra_dump_file,
2050 " %d Non pseudo reload: reject++\n",
2051 nop);
2052 reject++;
2053 }
2054 if (in_class_p (operand_reg[nop],
2055 this_costly_alternative, NULL))
2056 {
2057 if (lra_dump_file != NULL)
2058 fprintf
2059 (lra_dump_file,
2060 " %d Non pseudo costly reload:"
2061 " reject++\n",
2062 nop);
2063 reject++;
2064 }
2065 }
2066 /* We simulate the behaviour of old reload here.
2067 Although scratches need hard registers and it
2068 might result in spilling other pseudos, no reload
2069 insns are generated for the scratches. So it
2070 might cost something but probably less than old
2071 reload pass believes. */
2072 if (scratch_p)
2073 {
2074 if (lra_dump_file != NULL)
2075 fprintf (lra_dump_file,
2076 " %d Scratch win: reject+=2\n",
2077 nop);
2078 reject += 2;
2079 }
2080 }
2081 }
2082 else if (did_match)
2083 this_alternative_match_win = true;
2084 else
2085 {
2086 int const_to_mem = 0;
2087 bool no_regs_p;
2088
2089 /* Never do output reload of stack pointer. It makes
2090 impossible to do elimination when SP is changed in
2091 RTL. */
2092 if (op == stack_pointer_rtx && ! frame_pointer_needed
2093 && curr_static_id->operand[nop].type != OP_IN)
2094 goto fail;
2095
2096 /* If this alternative asks for a specific reg class, see if there
2097 is at least one allocatable register in that class. */
2098 no_regs_p
2099 = (this_alternative == NO_REGS
2100 || (hard_reg_set_subset_p
2101 (reg_class_contents[this_alternative],
2102 lra_no_alloc_regs)));
2103
2104 /* For asms, verify that the class for this alternative is possible
2105 for the mode that is specified. */
2106 if (!no_regs_p && INSN_CODE (curr_insn) < 0)
2107 {
2108 int i;
2109 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2110 if (HARD_REGNO_MODE_OK (i, mode)
2111 && in_hard_reg_set_p (reg_class_contents[this_alternative],
2112 mode, i))
2113 break;
2114 if (i == FIRST_PSEUDO_REGISTER)
2115 winreg = false;
2116 }
2117
2118 /* If this operand accepts a register, and if the
2119 register class has at least one allocatable register,
2120 then this operand can be reloaded. */
2121 if (winreg && !no_regs_p)
2122 badop = false;
2123
2124 if (badop)
2125 {
2126 if (lra_dump_file != NULL)
2127 fprintf (lra_dump_file,
2128 " alt=%d: Bad operand -- refuse\n",
2129 nalt);
2130 goto fail;
2131 }
2132
2133 /* If not assigned pseudo has a class which a subset of
2134 required reg class, it is a less costly alternative
2135 as the pseudo still can get a hard reg of necessary
2136 class. */
2137 if (! no_regs_p && REG_P (op) && hard_regno[nop] < 0
2138 && (cl = get_reg_class (REGNO (op))) != NO_REGS
2139 && ira_class_subset_p[this_alternative][cl])
2140 {
2141 if (lra_dump_file != NULL)
2142 fprintf
2143 (lra_dump_file,
2144 " %d Super set class reg: reject-=3\n", nop);
2145 reject -= 3;
2146 }
2147
2148 this_alternative_offmemok = offmemok;
2149 if (this_costly_alternative != NO_REGS)
2150 {
2151 if (lra_dump_file != NULL)
2152 fprintf (lra_dump_file,
2153 " %d Costly loser: reject++\n", nop);
2154 reject++;
2155 }
2156 /* If the operand is dying, has a matching constraint,
2157 and satisfies constraints of the matched operand
2158 which failed to satisfy the own constraints, most probably
2159 the reload for this operand will be gone. */
2160 if (this_alternative_matches >= 0
2161 && !curr_alt_win[this_alternative_matches]
2162 && REG_P (op)
2163 && find_regno_note (curr_insn, REG_DEAD, REGNO (op))
2164 && (hard_regno[nop] >= 0
2165 ? in_hard_reg_set_p (this_alternative_set,
2166 mode, hard_regno[nop])
2167 : in_class_p (op, this_alternative, NULL)))
2168 {
2169 if (lra_dump_file != NULL)
2170 fprintf
2171 (lra_dump_file,
2172 " %d Dying matched operand reload: reject++\n",
2173 nop);
2174 reject++;
2175 }
2176 else
2177 {
2178 /* Strict_low_part requires to reload the register
2179 not the sub-register. In this case we should
2180 check that a final reload hard reg can hold the
2181 value mode. */
2182 if (curr_static_id->operand[nop].strict_low
2183 && REG_P (op)
2184 && hard_regno[nop] < 0
2185 && GET_CODE (*curr_id->operand_loc[nop]) == SUBREG
2186 && ira_class_hard_regs_num[this_alternative] > 0
2187 && ! HARD_REGNO_MODE_OK (ira_class_hard_regs
2188 [this_alternative][0],
2189 GET_MODE
2190 (*curr_id->operand_loc[nop])))
2191 {
2192 if (lra_dump_file != NULL)
2193 fprintf
2194 (lra_dump_file,
2195 " alt=%d: Strict low subreg reload -- refuse\n",
2196 nalt);
2197 goto fail;
2198 }
2199 losers++;
2200 }
2201 if (operand_reg[nop] != NULL_RTX
2202 /* Output operands and matched input operands are
2203 not inherited. The following conditions do not
2204 exactly describe the previous statement but they
2205 are pretty close. */
2206 && curr_static_id->operand[nop].type != OP_OUT
2207 && (this_alternative_matches < 0
2208 || curr_static_id->operand[nop].type != OP_IN))
2209 {
2210 int last_reload = (lra_reg_info[ORIGINAL_REGNO
2211 (operand_reg[nop])]
2212 .last_reload);
2213
2214 /* The value of reload_sum has sense only if we
2215 process insns in their order. It happens only on
2216 the first constraints sub-pass when we do most of
2217 reload work. */
2218 if (lra_constraint_iter == 1 && last_reload > bb_reload_num)
2219 reload_sum += last_reload - bb_reload_num;
2220 }
2221 /* If this is a constant that is reloaded into the
2222 desired class by copying it to memory first, count
2223 that as another reload. This is consistent with
2224 other code and is required to avoid choosing another
2225 alternative when the constant is moved into memory.
2226 Note that the test here is precisely the same as in
2227 the code below that calls force_const_mem. */
2228 if (CONST_POOL_OK_P (mode, op)
2229 && ((targetm.preferred_reload_class
2230 (op, this_alternative) == NO_REGS)
2231 || no_input_reloads_p))
2232 {
2233 const_to_mem = 1;
2234 if (! no_regs_p)
2235 losers++;
2236 }
2237
2238 /* Alternative loses if it requires a type of reload not
2239 permitted for this insn. We can always reload
2240 objects with a REG_UNUSED note. */
2241 if ((curr_static_id->operand[nop].type != OP_IN
2242 && no_output_reloads_p
2243 && ! find_reg_note (curr_insn, REG_UNUSED, op))
2244 || (curr_static_id->operand[nop].type != OP_OUT
2245 && no_input_reloads_p && ! const_to_mem)
2246 || (this_alternative_matches >= 0
2247 && (no_input_reloads_p
2248 || (no_output_reloads_p
2249 && (curr_static_id->operand
2250 [this_alternative_matches].type != OP_IN)
2251 && ! find_reg_note (curr_insn, REG_UNUSED,
2252 no_subreg_reg_operand
2253 [this_alternative_matches])))))
2254 {
2255 if (lra_dump_file != NULL)
2256 fprintf
2257 (lra_dump_file,
2258 " alt=%d: No input/otput reload -- refuse\n",
2259 nalt);
2260 goto fail;
2261 }
2262
2263 /* Check strong discouragement of reload of non-constant
2264 into class THIS_ALTERNATIVE. */
2265 if (! CONSTANT_P (op) && ! no_regs_p
2266 && (targetm.preferred_reload_class
2267 (op, this_alternative) == NO_REGS
2268 || (curr_static_id->operand[nop].type == OP_OUT
2269 && (targetm.preferred_output_reload_class
2270 (op, this_alternative) == NO_REGS))))
2271 {
2272 if (lra_dump_file != NULL)
2273 fprintf (lra_dump_file,
2274 " %d Non-prefered reload: reject+=%d\n",
2275 nop, LRA_MAX_REJECT);
2276 reject += LRA_MAX_REJECT;
2277 }
2278
2279 if (! (MEM_P (op) && offmemok)
2280 && ! (const_to_mem && constmemok))
2281 {
2282 /* We prefer to reload pseudos over reloading other
2283 things, since such reloads may be able to be
2284 eliminated later. So bump REJECT in other cases.
2285 Don't do this in the case where we are forcing a
2286 constant into memory and it will then win since
2287 we don't want to have a different alternative
2288 match then. */
2289 if (! (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER))
2290 {
2291 if (lra_dump_file != NULL)
2292 fprintf
2293 (lra_dump_file,
2294 " %d Non-pseudo reload: reject+=2\n",
2295 nop);
2296 reject += 2;
2297 }
2298
2299 if (! no_regs_p)
2300 reload_nregs
2301 += ira_reg_class_max_nregs[this_alternative][mode];
2302
2303 if (SMALL_REGISTER_CLASS_P (this_alternative))
2304 {
2305 if (lra_dump_file != NULL)
2306 fprintf
2307 (lra_dump_file,
2308 " %d Small class reload: reject+=%d\n",
2309 nop, LRA_LOSER_COST_FACTOR / 2);
2310 reject += LRA_LOSER_COST_FACTOR / 2;
2311 }
2312 }
2313
2314 /* We are trying to spill pseudo into memory. It is
2315 usually more costly than moving to a hard register
2316 although it might takes the same number of
2317 reloads. */
2318 if (no_regs_p && REG_P (op) && hard_regno[nop] >= 0)
2319 {
2320 if (lra_dump_file != NULL)
2321 fprintf
2322 (lra_dump_file,
2323 " %d Spill pseudo into memory: reject+=3\n",
2324 nop);
2325 reject += 3;
2326 if (VECTOR_MODE_P (mode))
2327 {
2328 /* Spilling vectors into memory is usually more
2329 costly as they contain big values. */
2330 if (lra_dump_file != NULL)
2331 fprintf
2332 (lra_dump_file,
2333 " %d Spill vector pseudo: reject+=2\n",
2334 nop);
2335 reject += 2;
2336 }
2337 }
2338
2339 #ifdef SECONDARY_MEMORY_NEEDED
2340 /* If reload requires moving value through secondary
2341 memory, it will need one more insn at least. */
2342 if (this_alternative != NO_REGS
2343 && REG_P (op) && (cl = get_reg_class (REGNO (op))) != NO_REGS
2344 && ((curr_static_id->operand[nop].type != OP_OUT
2345 && SECONDARY_MEMORY_NEEDED (cl, this_alternative,
2346 GET_MODE (op)))
2347 || (curr_static_id->operand[nop].type != OP_IN
2348 && SECONDARY_MEMORY_NEEDED (this_alternative, cl,
2349 GET_MODE (op)))))
2350 losers++;
2351 #endif
2352 /* Input reloads can be inherited more often than output
2353 reloads can be removed, so penalize output
2354 reloads. */
2355 if (!REG_P (op) || curr_static_id->operand[nop].type != OP_IN)
2356 {
2357 if (lra_dump_file != NULL)
2358 fprintf
2359 (lra_dump_file,
2360 " %d Non input pseudo reload: reject++\n",
2361 nop);
2362 reject++;
2363 }
2364 }
2365
2366 if (early_clobber_p && ! scratch_p)
2367 {
2368 if (lra_dump_file != NULL)
2369 fprintf (lra_dump_file,
2370 " %d Early clobber: reject++\n", nop);
2371 reject++;
2372 }
2373 /* ??? We check early clobbers after processing all operands
2374 (see loop below) and there we update the costs more.
2375 Should we update the cost (may be approximately) here
2376 because of early clobber register reloads or it is a rare
2377 or non-important thing to be worth to do it. */
2378 overall = losers * LRA_LOSER_COST_FACTOR + reject;
2379 if ((best_losers == 0 || losers != 0) && best_overall < overall)
2380 {
2381 if (lra_dump_file != NULL)
2382 fprintf (lra_dump_file,
2383 " alt=%d,overall=%d,losers=%d -- refuse\n",
2384 nalt, overall, losers);
2385 goto fail;
2386 }
2387
2388 curr_alt[nop] = this_alternative;
2389 COPY_HARD_REG_SET (curr_alt_set[nop], this_alternative_set);
2390 curr_alt_win[nop] = this_alternative_win;
2391 curr_alt_match_win[nop] = this_alternative_match_win;
2392 curr_alt_offmemok[nop] = this_alternative_offmemok;
2393 curr_alt_matches[nop] = this_alternative_matches;
2394
2395 if (this_alternative_matches >= 0
2396 && !did_match && !this_alternative_win)
2397 curr_alt_win[this_alternative_matches] = false;
2398
2399 if (early_clobber_p && operand_reg[nop] != NULL_RTX)
2400 early_clobbered_nops[early_clobbered_regs_num++] = nop;
2401 }
2402 if (curr_insn_set != NULL_RTX && n_operands == 2
2403 /* Prevent processing non-move insns. */
2404 && (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
2405 || SET_SRC (curr_insn_set) == no_subreg_reg_operand[1])
2406 && ((! curr_alt_win[0] && ! curr_alt_win[1]
2407 && REG_P (no_subreg_reg_operand[0])
2408 && REG_P (no_subreg_reg_operand[1])
2409 && (reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2410 || reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0])))
2411 || (! curr_alt_win[0] && curr_alt_win[1]
2412 && REG_P (no_subreg_reg_operand[1])
2413 && reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0]))
2414 || (curr_alt_win[0] && ! curr_alt_win[1]
2415 && REG_P (no_subreg_reg_operand[0])
2416 && reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2417 && (! CONST_POOL_OK_P (curr_operand_mode[1],
2418 no_subreg_reg_operand[1])
2419 || (targetm.preferred_reload_class
2420 (no_subreg_reg_operand[1],
2421 (enum reg_class) curr_alt[1]) != NO_REGS))
2422 /* If it is a result of recent elimination in move
2423 insn we can transform it into an add still by
2424 using this alternative. */
2425 && GET_CODE (no_subreg_reg_operand[1]) != PLUS)))
2426 {
2427 /* We have a move insn and a new reload insn will be similar
2428 to the current insn. We should avoid such situation as it
2429 results in LRA cycling. */
2430 overall += LRA_MAX_REJECT;
2431 }
2432 ok_p = true;
2433 curr_alt_dont_inherit_ops_num = 0;
2434 for (nop = 0; nop < early_clobbered_regs_num; nop++)
2435 {
2436 int i, j, clobbered_hard_regno, first_conflict_j, last_conflict_j;
2437 HARD_REG_SET temp_set;
2438
2439 i = early_clobbered_nops[nop];
2440 if ((! curr_alt_win[i] && ! curr_alt_match_win[i])
2441 || hard_regno[i] < 0)
2442 continue;
2443 lra_assert (operand_reg[i] != NULL_RTX);
2444 clobbered_hard_regno = hard_regno[i];
2445 CLEAR_HARD_REG_SET (temp_set);
2446 add_to_hard_reg_set (&temp_set, biggest_mode[i], clobbered_hard_regno);
2447 first_conflict_j = last_conflict_j = -1;
2448 for (j = 0; j < n_operands; j++)
2449 if (j == i
2450 /* We don't want process insides of match_operator and
2451 match_parallel because otherwise we would process
2452 their operands once again generating a wrong
2453 code. */
2454 || curr_static_id->operand[j].is_operator)
2455 continue;
2456 else if ((curr_alt_matches[j] == i && curr_alt_match_win[j])
2457 || (curr_alt_matches[i] == j && curr_alt_match_win[i]))
2458 continue;
2459 /* If we don't reload j-th operand, check conflicts. */
2460 else if ((curr_alt_win[j] || curr_alt_match_win[j])
2461 && uses_hard_regs_p (*curr_id->operand_loc[j], temp_set))
2462 {
2463 if (first_conflict_j < 0)
2464 first_conflict_j = j;
2465 last_conflict_j = j;
2466 }
2467 if (last_conflict_j < 0)
2468 continue;
2469 /* If earlyclobber operand conflicts with another
2470 non-matching operand which is actually the same register
2471 as the earlyclobber operand, it is better to reload the
2472 another operand as an operand matching the earlyclobber
2473 operand can be also the same. */
2474 if (first_conflict_j == last_conflict_j
2475 && operand_reg[last_conflict_j]
2476 != NULL_RTX && ! curr_alt_match_win[last_conflict_j]
2477 && REGNO (operand_reg[i]) == REGNO (operand_reg[last_conflict_j]))
2478 {
2479 curr_alt_win[last_conflict_j] = false;
2480 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++]
2481 = last_conflict_j;
2482 losers++;
2483 /* Early clobber was already reflected in REJECT. */
2484 lra_assert (reject > 0);
2485 if (lra_dump_file != NULL)
2486 fprintf
2487 (lra_dump_file,
2488 " %d Conflict early clobber reload: reject--\n",
2489 i);
2490 reject--;
2491 overall += LRA_LOSER_COST_FACTOR - 1;
2492 }
2493 else
2494 {
2495 /* We need to reload early clobbered register and the
2496 matched registers. */
2497 for (j = 0; j < n_operands; j++)
2498 if (curr_alt_matches[j] == i)
2499 {
2500 curr_alt_match_win[j] = false;
2501 losers++;
2502 overall += LRA_LOSER_COST_FACTOR;
2503 }
2504 if (! curr_alt_match_win[i])
2505 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++] = i;
2506 else
2507 {
2508 /* Remember pseudos used for match reloads are never
2509 inherited. */
2510 lra_assert (curr_alt_matches[i] >= 0);
2511 curr_alt_win[curr_alt_matches[i]] = false;
2512 }
2513 curr_alt_win[i] = curr_alt_match_win[i] = false;
2514 losers++;
2515 /* Early clobber was already reflected in REJECT. */
2516 lra_assert (reject > 0);
2517 if (lra_dump_file != NULL)
2518 fprintf
2519 (lra_dump_file,
2520 " %d Matched conflict early clobber reloads:"
2521 "reject--\n",
2522 i);
2523 reject--;
2524 overall += LRA_LOSER_COST_FACTOR - 1;
2525 }
2526 }
2527 if (lra_dump_file != NULL)
2528 fprintf (lra_dump_file, " alt=%d,overall=%d,losers=%d,rld_nregs=%d\n",
2529 nalt, overall, losers, reload_nregs);
2530
2531 /* If this alternative can be made to work by reloading, and it
2532 needs less reloading than the others checked so far, record
2533 it as the chosen goal for reloading. */
2534 if ((best_losers != 0 && losers == 0)
2535 || (((best_losers == 0 && losers == 0)
2536 || (best_losers != 0 && losers != 0))
2537 && (best_overall > overall
2538 || (best_overall == overall
2539 /* If the cost of the reloads is the same,
2540 prefer alternative which requires minimal
2541 number of reload regs. */
2542 && (reload_nregs < best_reload_nregs
2543 || (reload_nregs == best_reload_nregs
2544 && (best_reload_sum < reload_sum
2545 || (best_reload_sum == reload_sum
2546 && nalt < goal_alt_number))))))))
2547 {
2548 for (nop = 0; nop < n_operands; nop++)
2549 {
2550 goal_alt_win[nop] = curr_alt_win[nop];
2551 goal_alt_match_win[nop] = curr_alt_match_win[nop];
2552 goal_alt_matches[nop] = curr_alt_matches[nop];
2553 goal_alt[nop] = curr_alt[nop];
2554 goal_alt_offmemok[nop] = curr_alt_offmemok[nop];
2555 }
2556 goal_alt_dont_inherit_ops_num = curr_alt_dont_inherit_ops_num;
2557 for (nop = 0; nop < curr_alt_dont_inherit_ops_num; nop++)
2558 goal_alt_dont_inherit_ops[nop] = curr_alt_dont_inherit_ops[nop];
2559 goal_alt_swapped = curr_swapped;
2560 best_overall = overall;
2561 best_losers = losers;
2562 best_reload_nregs = reload_nregs;
2563 best_reload_sum = reload_sum;
2564 goal_alt_number = nalt;
2565 }
2566 if (losers == 0)
2567 /* Everything is satisfied. Do not process alternatives
2568 anymore. */
2569 break;
2570 fail:
2571 ;
2572 }
2573 return ok_p;
2574 }
2575
2576 /* Return 1 if ADDR is a valid memory address for mode MODE in address
2577 space AS, and check that each pseudo has the proper kind of hard
2578 reg. */
2579 static int
2580 valid_address_p (enum machine_mode mode ATTRIBUTE_UNUSED,
2581 rtx addr, addr_space_t as)
2582 {
2583 #ifdef GO_IF_LEGITIMATE_ADDRESS
2584 lra_assert (ADDR_SPACE_GENERIC_P (as));
2585 GO_IF_LEGITIMATE_ADDRESS (mode, addr, win);
2586 return 0;
2587
2588 win:
2589 return 1;
2590 #else
2591 return targetm.addr_space.legitimate_address_p (mode, addr, 0, as);
2592 #endif
2593 }
2594
2595 /* Return whether address AD is valid. */
2596
2597 static bool
2598 valid_address_p (struct address_info *ad)
2599 {
2600 /* Some ports do not check displacements for eliminable registers,
2601 so we replace them temporarily with the elimination target. */
2602 rtx saved_base_reg = NULL_RTX;
2603 rtx saved_index_reg = NULL_RTX;
2604 rtx *base_term = strip_subreg (ad->base_term);
2605 rtx *index_term = strip_subreg (ad->index_term);
2606 if (base_term != NULL)
2607 {
2608 saved_base_reg = *base_term;
2609 lra_eliminate_reg_if_possible (base_term);
2610 if (ad->base_term2 != NULL)
2611 *ad->base_term2 = *ad->base_term;
2612 }
2613 if (index_term != NULL)
2614 {
2615 saved_index_reg = *index_term;
2616 lra_eliminate_reg_if_possible (index_term);
2617 }
2618 bool ok_p = valid_address_p (ad->mode, *ad->outer, ad->as);
2619 if (saved_base_reg != NULL_RTX)
2620 {
2621 *base_term = saved_base_reg;
2622 if (ad->base_term2 != NULL)
2623 *ad->base_term2 = *ad->base_term;
2624 }
2625 if (saved_index_reg != NULL_RTX)
2626 *index_term = saved_index_reg;
2627 return ok_p;
2628 }
2629
2630 /* Make reload base reg + disp from address AD. Return the new pseudo. */
2631 static rtx
2632 base_plus_disp_to_reg (struct address_info *ad)
2633 {
2634 enum reg_class cl;
2635 rtx new_reg;
2636
2637 lra_assert (ad->base == ad->base_term && ad->disp == ad->disp_term);
2638 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
2639 get_index_code (ad));
2640 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
2641 cl, "base + disp");
2642 lra_emit_add (new_reg, *ad->base_term, *ad->disp_term);
2643 return new_reg;
2644 }
2645
2646 /* Make reload of index part of address AD. Return the new
2647 pseudo. */
2648 static rtx
2649 index_part_to_reg (struct address_info *ad)
2650 {
2651 rtx new_reg;
2652
2653 new_reg = lra_create_new_reg (GET_MODE (*ad->index), NULL_RTX,
2654 INDEX_REG_CLASS, "index term");
2655 expand_mult (GET_MODE (*ad->index), *ad->index_term,
2656 GEN_INT (get_index_scale (ad)), new_reg, 1);
2657 return new_reg;
2658 }
2659
2660 /* Return true if we can add a displacement to address AD, even if that
2661 makes the address invalid. The fix-up code requires any new address
2662 to be the sum of the BASE_TERM, INDEX and DISP_TERM fields. */
2663 static bool
2664 can_add_disp_p (struct address_info *ad)
2665 {
2666 return (!ad->autoinc_p
2667 && ad->segment == NULL
2668 && ad->base == ad->base_term
2669 && ad->disp == ad->disp_term);
2670 }
2671
2672 /* Make equiv substitution in address AD. Return true if a substitution
2673 was made. */
2674 static bool
2675 equiv_address_substitution (struct address_info *ad)
2676 {
2677 rtx base_reg, new_base_reg, index_reg, new_index_reg, *base_term, *index_term;
2678 HOST_WIDE_INT disp, scale;
2679 bool change_p;
2680
2681 base_term = strip_subreg (ad->base_term);
2682 if (base_term == NULL)
2683 base_reg = new_base_reg = NULL_RTX;
2684 else
2685 {
2686 base_reg = *base_term;
2687 new_base_reg = get_equiv_with_elimination (base_reg, curr_insn);
2688 }
2689 index_term = strip_subreg (ad->index_term);
2690 if (index_term == NULL)
2691 index_reg = new_index_reg = NULL_RTX;
2692 else
2693 {
2694 index_reg = *index_term;
2695 new_index_reg = get_equiv_with_elimination (index_reg, curr_insn);
2696 }
2697 if (base_reg == new_base_reg && index_reg == new_index_reg)
2698 return false;
2699 disp = 0;
2700 change_p = false;
2701 if (lra_dump_file != NULL)
2702 {
2703 fprintf (lra_dump_file, "Changing address in insn %d ",
2704 INSN_UID (curr_insn));
2705 dump_value_slim (lra_dump_file, *ad->outer, 1);
2706 }
2707 if (base_reg != new_base_reg)
2708 {
2709 if (REG_P (new_base_reg))
2710 {
2711 *base_term = new_base_reg;
2712 change_p = true;
2713 }
2714 else if (GET_CODE (new_base_reg) == PLUS
2715 && REG_P (XEXP (new_base_reg, 0))
2716 && CONST_INT_P (XEXP (new_base_reg, 1))
2717 && can_add_disp_p (ad))
2718 {
2719 disp += INTVAL (XEXP (new_base_reg, 1));
2720 *base_term = XEXP (new_base_reg, 0);
2721 change_p = true;
2722 }
2723 if (ad->base_term2 != NULL)
2724 *ad->base_term2 = *ad->base_term;
2725 }
2726 if (index_reg != new_index_reg)
2727 {
2728 if (REG_P (new_index_reg))
2729 {
2730 *index_term = new_index_reg;
2731 change_p = true;
2732 }
2733 else if (GET_CODE (new_index_reg) == PLUS
2734 && REG_P (XEXP (new_index_reg, 0))
2735 && CONST_INT_P (XEXP (new_index_reg, 1))
2736 && can_add_disp_p (ad)
2737 && (scale = get_index_scale (ad)))
2738 {
2739 disp += INTVAL (XEXP (new_index_reg, 1)) * scale;
2740 *index_term = XEXP (new_index_reg, 0);
2741 change_p = true;
2742 }
2743 }
2744 if (disp != 0)
2745 {
2746 if (ad->disp != NULL)
2747 *ad->disp = plus_constant (GET_MODE (*ad->inner), *ad->disp, disp);
2748 else
2749 {
2750 *ad->inner = plus_constant (GET_MODE (*ad->inner), *ad->inner, disp);
2751 update_address (ad);
2752 }
2753 change_p = true;
2754 }
2755 if (lra_dump_file != NULL)
2756 {
2757 if (! change_p)
2758 fprintf (lra_dump_file, " -- no change\n");
2759 else
2760 {
2761 fprintf (lra_dump_file, " on equiv ");
2762 dump_value_slim (lra_dump_file, *ad->outer, 1);
2763 fprintf (lra_dump_file, "\n");
2764 }
2765 }
2766 return change_p;
2767 }
2768
2769 /* Major function to make reloads for an address in operand NOP.
2770 The supported cases are:
2771
2772 1) an address that existed before LRA started, at which point it
2773 must have been valid. These addresses are subject to elimination
2774 and may have become invalid due to the elimination offset being out
2775 of range.
2776
2777 2) an address created by forcing a constant to memory
2778 (force_const_to_mem). The initial form of these addresses might
2779 not be valid, and it is this function's job to make them valid.
2780
2781 3) a frame address formed from a register and a (possibly zero)
2782 constant offset. As above, these addresses might not be valid and
2783 this function must make them so.
2784
2785 Add reloads to the lists *BEFORE and *AFTER. We might need to add
2786 reloads to *AFTER because of inc/dec, {pre, post} modify in the
2787 address. Return true for any RTL change.
2788
2789 The function is a helper function which does not produce all
2790 transformations which can be necessary. It does just basic steps.
2791 To do all necessary transformations use function
2792 process_address. */
2793 static bool
2794 process_address_1 (int nop, rtx *before, rtx *after)
2795 {
2796 struct address_info ad;
2797 rtx new_reg;
2798 rtx op = *curr_id->operand_loc[nop];
2799 const char *constraint = curr_static_id->operand[nop].constraint;
2800 bool change_p;
2801
2802 if (constraint[0] == 'p'
2803 || EXTRA_ADDRESS_CONSTRAINT (constraint[0], constraint))
2804 decompose_lea_address (&ad, curr_id->operand_loc[nop]);
2805 else if (MEM_P (op))
2806 decompose_mem_address (&ad, op);
2807 else if (GET_CODE (op) == SUBREG
2808 && MEM_P (SUBREG_REG (op)))
2809 decompose_mem_address (&ad, SUBREG_REG (op));
2810 else
2811 return false;
2812 change_p = equiv_address_substitution (&ad);
2813 if (ad.base_term != NULL
2814 && (process_addr_reg
2815 (ad.base_term, before,
2816 (ad.autoinc_p
2817 && !(REG_P (*ad.base_term)
2818 && find_regno_note (curr_insn, REG_DEAD,
2819 REGNO (*ad.base_term)) != NULL_RTX)
2820 ? after : NULL),
2821 base_reg_class (ad.mode, ad.as, ad.base_outer_code,
2822 get_index_code (&ad)))))
2823 {
2824 change_p = true;
2825 if (ad.base_term2 != NULL)
2826 *ad.base_term2 = *ad.base_term;
2827 }
2828 if (ad.index_term != NULL
2829 && process_addr_reg (ad.index_term, before, NULL, INDEX_REG_CLASS))
2830 change_p = true;
2831
2832 #ifdef EXTRA_CONSTRAINT_STR
2833 /* Target hooks sometimes reject extra constraint addresses -- use
2834 EXTRA_CONSTRAINT_STR for the validation. */
2835 if (constraint[0] != 'p'
2836 && EXTRA_ADDRESS_CONSTRAINT (constraint[0], constraint)
2837 && EXTRA_CONSTRAINT_STR (op, constraint[0], constraint))
2838 return change_p;
2839 #endif
2840
2841 /* There are three cases where the shape of *AD.INNER may now be invalid:
2842
2843 1) the original address was valid, but either elimination or
2844 equiv_address_substitution was applied and that made
2845 the address invalid.
2846
2847 2) the address is an invalid symbolic address created by
2848 force_const_to_mem.
2849
2850 3) the address is a frame address with an invalid offset.
2851
2852 All these cases involve a non-autoinc address, so there is no
2853 point revalidating other types. */
2854 if (ad.autoinc_p || valid_address_p (&ad))
2855 return change_p;
2856
2857 /* Any index existed before LRA started, so we can assume that the
2858 presence and shape of the index is valid. */
2859 push_to_sequence (*before);
2860 lra_assert (ad.disp == ad.disp_term);
2861 if (ad.base == NULL)
2862 {
2863 if (ad.index == NULL)
2864 {
2865 int code = -1;
2866 enum reg_class cl = base_reg_class (ad.mode, ad.as,
2867 SCRATCH, SCRATCH);
2868 rtx addr = *ad.inner;
2869
2870 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
2871 #ifdef HAVE_lo_sum
2872 {
2873 rtx insn;
2874 rtx last = get_last_insn ();
2875
2876 /* addr => lo_sum (new_base, addr), case (2) above. */
2877 insn = emit_insn (gen_rtx_SET
2878 (VOIDmode, new_reg,
2879 gen_rtx_HIGH (Pmode, copy_rtx (addr))));
2880 code = recog_memoized (insn);
2881 if (code >= 0)
2882 {
2883 *ad.inner = gen_rtx_LO_SUM (Pmode, new_reg, addr);
2884 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
2885 {
2886 /* Try to put lo_sum into register. */
2887 insn = emit_insn (gen_rtx_SET
2888 (VOIDmode, new_reg,
2889 gen_rtx_LO_SUM (Pmode, new_reg, addr)));
2890 code = recog_memoized (insn);
2891 if (code >= 0)
2892 {
2893 *ad.inner = new_reg;
2894 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
2895 {
2896 *ad.inner = addr;
2897 code = -1;
2898 }
2899 }
2900
2901 }
2902 }
2903 if (code < 0)
2904 delete_insns_since (last);
2905 }
2906 #endif
2907 if (code < 0)
2908 {
2909 /* addr => new_base, case (2) above. */
2910 lra_emit_move (new_reg, addr);
2911 *ad.inner = new_reg;
2912 }
2913 }
2914 else
2915 {
2916 /* index * scale + disp => new base + index * scale,
2917 case (1) above. */
2918 enum reg_class cl = base_reg_class (ad.mode, ad.as, PLUS,
2919 GET_CODE (*ad.index));
2920
2921 lra_assert (INDEX_REG_CLASS != NO_REGS);
2922 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp");
2923 lra_emit_move (new_reg, *ad.disp);
2924 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
2925 new_reg, *ad.index);
2926 }
2927 }
2928 else if (ad.index == NULL)
2929 {
2930 int regno;
2931 enum reg_class cl;
2932 rtx set, insns, last_insn;
2933 /* base + disp => new base, cases (1) and (3) above. */
2934 /* Another option would be to reload the displacement into an
2935 index register. However, postreload has code to optimize
2936 address reloads that have the same base and different
2937 displacements, so reloading into an index register would
2938 not necessarily be a win. */
2939 start_sequence ();
2940 new_reg = base_plus_disp_to_reg (&ad);
2941 insns = get_insns ();
2942 last_insn = get_last_insn ();
2943 /* If we generated at least two insns, try last insn source as
2944 an address. If we succeed, we generate one less insn. */
2945 if (last_insn != insns && (set = single_set (last_insn)) != NULL_RTX
2946 && GET_CODE (SET_SRC (set)) == PLUS
2947 && REG_P (XEXP (SET_SRC (set), 0))
2948 && CONSTANT_P (XEXP (SET_SRC (set), 1)))
2949 {
2950 *ad.inner = SET_SRC (set);
2951 if (valid_address_p (ad.mode, *ad.outer, ad.as))
2952 {
2953 *ad.base_term = XEXP (SET_SRC (set), 0);
2954 *ad.disp_term = XEXP (SET_SRC (set), 1);
2955 cl = base_reg_class (ad.mode, ad.as, ad.base_outer_code,
2956 get_index_code (&ad));
2957 regno = REGNO (*ad.base_term);
2958 if (regno >= FIRST_PSEUDO_REGISTER
2959 && cl != lra_get_allocno_class (regno))
2960 lra_change_class (regno, cl, " Change to", true);
2961 new_reg = SET_SRC (set);
2962 delete_insns_since (PREV_INSN (last_insn));
2963 }
2964 }
2965 end_sequence ();
2966 emit_insn (insns);
2967 *ad.inner = new_reg;
2968 }
2969 else if (ad.disp_term != NULL)
2970 {
2971 /* base + scale * index + disp => new base + scale * index,
2972 case (1) above. */
2973 new_reg = base_plus_disp_to_reg (&ad);
2974 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
2975 new_reg, *ad.index);
2976 }
2977 else
2978 {
2979 /* base + scale * index => base + new_reg,
2980 case (1) above.
2981 Index part of address may become invalid. For example, we
2982 changed pseudo on the equivalent memory and a subreg of the
2983 pseudo onto the memory of different mode for which the scale is
2984 prohibitted. */
2985 new_reg = index_part_to_reg (&ad);
2986 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
2987 *ad.base_term, new_reg);
2988 }
2989 *before = get_insns ();
2990 end_sequence ();
2991 return true;
2992 }
2993
2994 /* Do address reloads until it is necessary. Use process_address_1 as
2995 a helper function. Return true for any RTL changes. */
2996 static bool
2997 process_address (int nop, rtx *before, rtx *after)
2998 {
2999 bool res = false;
3000
3001 while (process_address_1 (nop, before, after))
3002 res = true;
3003 return res;
3004 }
3005
3006 /* Emit insns to reload VALUE into a new register. VALUE is an
3007 auto-increment or auto-decrement RTX whose operand is a register or
3008 memory location; so reloading involves incrementing that location.
3009 IN is either identical to VALUE, or some cheaper place to reload
3010 value being incremented/decremented from.
3011
3012 INC_AMOUNT is the number to increment or decrement by (always
3013 positive and ignored for POST_MODIFY/PRE_MODIFY).
3014
3015 Return pseudo containing the result. */
3016 static rtx
3017 emit_inc (enum reg_class new_rclass, rtx in, rtx value, int inc_amount)
3018 {
3019 /* REG or MEM to be copied and incremented. */
3020 rtx incloc = XEXP (value, 0);
3021 /* Nonzero if increment after copying. */
3022 int post = (GET_CODE (value) == POST_DEC || GET_CODE (value) == POST_INC
3023 || GET_CODE (value) == POST_MODIFY);
3024 rtx last;
3025 rtx inc;
3026 rtx add_insn;
3027 int code;
3028 rtx real_in = in == value ? incloc : in;
3029 rtx result;
3030 bool plus_p = true;
3031
3032 if (GET_CODE (value) == PRE_MODIFY || GET_CODE (value) == POST_MODIFY)
3033 {
3034 lra_assert (GET_CODE (XEXP (value, 1)) == PLUS
3035 || GET_CODE (XEXP (value, 1)) == MINUS);
3036 lra_assert (rtx_equal_p (XEXP (XEXP (value, 1), 0), XEXP (value, 0)));
3037 plus_p = GET_CODE (XEXP (value, 1)) == PLUS;
3038 inc = XEXP (XEXP (value, 1), 1);
3039 }
3040 else
3041 {
3042 if (GET_CODE (value) == PRE_DEC || GET_CODE (value) == POST_DEC)
3043 inc_amount = -inc_amount;
3044
3045 inc = GEN_INT (inc_amount);
3046 }
3047
3048 if (! post && REG_P (incloc))
3049 result = incloc;
3050 else
3051 result = lra_create_new_reg (GET_MODE (value), value, new_rclass,
3052 "INC/DEC result");
3053
3054 if (real_in != result)
3055 {
3056 /* First copy the location to the result register. */
3057 lra_assert (REG_P (result));
3058 emit_insn (gen_move_insn (result, real_in));
3059 }
3060
3061 /* We suppose that there are insns to add/sub with the constant
3062 increment permitted in {PRE/POST)_{DEC/INC/MODIFY}. At least the
3063 old reload worked with this assumption. If the assumption
3064 becomes wrong, we should use approach in function
3065 base_plus_disp_to_reg. */
3066 if (in == value)
3067 {
3068 /* See if we can directly increment INCLOC. */
3069 last = get_last_insn ();
3070 add_insn = emit_insn (plus_p
3071 ? gen_add2_insn (incloc, inc)
3072 : gen_sub2_insn (incloc, inc));
3073
3074 code = recog_memoized (add_insn);
3075 if (code >= 0)
3076 {
3077 if (! post && result != incloc)
3078 emit_insn (gen_move_insn (result, incloc));
3079 return result;
3080 }
3081 delete_insns_since (last);
3082 }
3083
3084 /* If couldn't do the increment directly, must increment in RESULT.
3085 The way we do this depends on whether this is pre- or
3086 post-increment. For pre-increment, copy INCLOC to the reload
3087 register, increment it there, then save back. */
3088 if (! post)
3089 {
3090 if (real_in != result)
3091 emit_insn (gen_move_insn (result, real_in));
3092 if (plus_p)
3093 emit_insn (gen_add2_insn (result, inc));
3094 else
3095 emit_insn (gen_sub2_insn (result, inc));
3096 if (result != incloc)
3097 emit_insn (gen_move_insn (incloc, result));
3098 }
3099 else
3100 {
3101 /* Post-increment.
3102
3103 Because this might be a jump insn or a compare, and because
3104 RESULT may not be available after the insn in an input
3105 reload, we must do the incrementing before the insn being
3106 reloaded for.
3107
3108 We have already copied IN to RESULT. Increment the copy in
3109 RESULT, save that back, then decrement RESULT so it has
3110 the original value. */
3111 if (plus_p)
3112 emit_insn (gen_add2_insn (result, inc));
3113 else
3114 emit_insn (gen_sub2_insn (result, inc));
3115 emit_insn (gen_move_insn (incloc, result));
3116 /* Restore non-modified value for the result. We prefer this
3117 way because it does not require an additional hard
3118 register. */
3119 if (plus_p)
3120 {
3121 if (CONST_INT_P (inc))
3122 emit_insn (gen_add2_insn (result,
3123 gen_int_mode (-INTVAL (inc),
3124 GET_MODE (result))));
3125 else
3126 emit_insn (gen_sub2_insn (result, inc));
3127 }
3128 else
3129 emit_insn (gen_add2_insn (result, inc));
3130 }
3131 return result;
3132 }
3133
3134 /* Return true if the current move insn does not need processing as we
3135 already know that it satisfies its constraints. */
3136 static bool
3137 simple_move_p (void)
3138 {
3139 rtx dest, src;
3140 enum reg_class dclass, sclass;
3141
3142 lra_assert (curr_insn_set != NULL_RTX);
3143 dest = SET_DEST (curr_insn_set);
3144 src = SET_SRC (curr_insn_set);
3145 return ((dclass = get_op_class (dest)) != NO_REGS
3146 && (sclass = get_op_class (src)) != NO_REGS
3147 /* The backend guarantees that register moves of cost 2
3148 never need reloads. */
3149 && targetm.register_move_cost (GET_MODE (src), dclass, sclass) == 2);
3150 }
3151
3152 /* Swap operands NOP and NOP + 1. */
3153 static inline void
3154 swap_operands (int nop)
3155 {
3156 enum machine_mode mode = curr_operand_mode[nop];
3157 curr_operand_mode[nop] = curr_operand_mode[nop + 1];
3158 curr_operand_mode[nop + 1] = mode;
3159 rtx x = *curr_id->operand_loc[nop];
3160 *curr_id->operand_loc[nop] = *curr_id->operand_loc[nop + 1];
3161 *curr_id->operand_loc[nop + 1] = x;
3162 /* Swap the duplicates too. */
3163 lra_update_dup (curr_id, nop);
3164 lra_update_dup (curr_id, nop + 1);
3165 }
3166
3167 /* Main entry point of the constraint code: search the body of the
3168 current insn to choose the best alternative. It is mimicking insn
3169 alternative cost calculation model of former reload pass. That is
3170 because machine descriptions were written to use this model. This
3171 model can be changed in future. Make commutative operand exchange
3172 if it is chosen.
3173
3174 Return true if some RTL changes happened during function call. */
3175 static bool
3176 curr_insn_transform (void)
3177 {
3178 int i, j, k;
3179 int n_operands;
3180 int n_alternatives;
3181 int commutative;
3182 signed char goal_alt_matched[MAX_RECOG_OPERANDS][MAX_RECOG_OPERANDS];
3183 signed char match_inputs[MAX_RECOG_OPERANDS + 1];
3184 rtx before, after;
3185 bool alt_p = false;
3186 /* Flag that the insn has been changed through a transformation. */
3187 bool change_p;
3188 bool sec_mem_p;
3189 #ifdef SECONDARY_MEMORY_NEEDED
3190 bool use_sec_mem_p;
3191 #endif
3192 int max_regno_before;
3193 int reused_alternative_num;
3194
3195 curr_insn_set = single_set (curr_insn);
3196 if (curr_insn_set != NULL_RTX && simple_move_p ())
3197 return false;
3198
3199 no_input_reloads_p = no_output_reloads_p = false;
3200 goal_alt_number = -1;
3201 change_p = sec_mem_p = false;
3202 /* JUMP_INSNs and CALL_INSNs are not allowed to have any output
3203 reloads; neither are insns that SET cc0. Insns that use CC0 are
3204 not allowed to have any input reloads. */
3205 if (JUMP_P (curr_insn) || CALL_P (curr_insn))
3206 no_output_reloads_p = true;
3207
3208 #ifdef HAVE_cc0
3209 if (reg_referenced_p (cc0_rtx, PATTERN (curr_insn)))
3210 no_input_reloads_p = true;
3211 if (reg_set_p (cc0_rtx, PATTERN (curr_insn)))
3212 no_output_reloads_p = true;
3213 #endif
3214
3215 n_operands = curr_static_id->n_operands;
3216 n_alternatives = curr_static_id->n_alternatives;
3217
3218 /* Just return "no reloads" if insn has no operands with
3219 constraints. */
3220 if (n_operands == 0 || n_alternatives == 0)
3221 return false;
3222
3223 max_regno_before = max_reg_num ();
3224
3225 for (i = 0; i < n_operands; i++)
3226 {
3227 goal_alt_matched[i][0] = -1;
3228 goal_alt_matches[i] = -1;
3229 }
3230
3231 commutative = curr_static_id->commutative;
3232
3233 /* Now see what we need for pseudos that didn't get hard regs or got
3234 the wrong kind of hard reg. For this, we must consider all the
3235 operands together against the register constraints. */
3236
3237 best_losers = best_overall = INT_MAX;
3238 best_reload_sum = 0;
3239
3240 curr_swapped = false;
3241 goal_alt_swapped = false;
3242
3243 /* Make equivalence substitution and memory subreg elimination
3244 before address processing because an address legitimacy can
3245 depend on memory mode. */
3246 for (i = 0; i < n_operands; i++)
3247 {
3248 rtx op = *curr_id->operand_loc[i];
3249 rtx subst, old = op;
3250 bool op_change_p = false;
3251
3252 if (GET_CODE (old) == SUBREG)
3253 old = SUBREG_REG (old);
3254 subst = get_equiv_with_elimination (old, curr_insn);
3255 if (subst != old)
3256 {
3257 subst = copy_rtx (subst);
3258 lra_assert (REG_P (old));
3259 if (GET_CODE (op) == SUBREG)
3260 SUBREG_REG (op) = subst;
3261 else
3262 *curr_id->operand_loc[i] = subst;
3263 if (lra_dump_file != NULL)
3264 {
3265 fprintf (lra_dump_file,
3266 "Changing pseudo %d in operand %i of insn %u on equiv ",
3267 REGNO (old), i, INSN_UID (curr_insn));
3268 dump_value_slim (lra_dump_file, subst, 1);
3269 fprintf (lra_dump_file, "\n");
3270 }
3271 op_change_p = change_p = true;
3272 }
3273 if (simplify_operand_subreg (i, GET_MODE (old)) || op_change_p)
3274 {
3275 change_p = true;
3276 lra_update_dup (curr_id, i);
3277 }
3278 }
3279
3280 /* Reload address registers and displacements. We do it before
3281 finding an alternative because of memory constraints. */
3282 before = after = NULL_RTX;
3283 for (i = 0; i < n_operands; i++)
3284 if (! curr_static_id->operand[i].is_operator
3285 && process_address (i, &before, &after))
3286 {
3287 change_p = true;
3288 lra_update_dup (curr_id, i);
3289 }
3290
3291 if (change_p)
3292 /* If we've changed the instruction then any alternative that
3293 we chose previously may no longer be valid. */
3294 lra_set_used_insn_alternative (curr_insn, -1);
3295
3296 if (curr_insn_set != NULL_RTX
3297 && check_and_process_move (&change_p, &sec_mem_p))
3298 return change_p;
3299
3300 try_swapped:
3301
3302 reused_alternative_num = curr_id->used_insn_alternative;
3303 if (lra_dump_file != NULL && reused_alternative_num >= 0)
3304 fprintf (lra_dump_file, "Reusing alternative %d for insn #%u\n",
3305 reused_alternative_num, INSN_UID (curr_insn));
3306
3307 if (process_alt_operands (reused_alternative_num))
3308 alt_p = true;
3309
3310 /* If insn is commutative (it's safe to exchange a certain pair of
3311 operands) then we need to try each alternative twice, the second
3312 time matching those two operands as if we had exchanged them. To
3313 do this, really exchange them in operands.
3314
3315 If we have just tried the alternatives the second time, return
3316 operands to normal and drop through. */
3317
3318 if (reused_alternative_num < 0 && commutative >= 0)
3319 {
3320 curr_swapped = !curr_swapped;
3321 if (curr_swapped)
3322 {
3323 swap_operands (commutative);
3324 goto try_swapped;
3325 }
3326 else
3327 swap_operands (commutative);
3328 }
3329
3330 if (! alt_p && ! sec_mem_p)
3331 {
3332 /* No alternative works with reloads?? */
3333 if (INSN_CODE (curr_insn) >= 0)
3334 fatal_insn ("unable to generate reloads for:", curr_insn);
3335 error_for_asm (curr_insn,
3336 "inconsistent operand constraints in an %<asm%>");
3337 /* Avoid further trouble with this insn. */
3338 PATTERN (curr_insn) = gen_rtx_USE (VOIDmode, const0_rtx);
3339 lra_invalidate_insn_data (curr_insn);
3340 return true;
3341 }
3342
3343 /* If the best alternative is with operands 1 and 2 swapped, swap
3344 them. Update the operand numbers of any reloads already
3345 pushed. */
3346
3347 if (goal_alt_swapped)
3348 {
3349 if (lra_dump_file != NULL)
3350 fprintf (lra_dump_file, " Commutative operand exchange in insn %u\n",
3351 INSN_UID (curr_insn));
3352
3353 /* Swap the duplicates too. */
3354 swap_operands (commutative);
3355 change_p = true;
3356 }
3357
3358 #ifdef SECONDARY_MEMORY_NEEDED
3359 /* Some target macros SECONDARY_MEMORY_NEEDED (e.g. x86) are defined
3360 too conservatively. So we use the secondary memory only if there
3361 is no any alternative without reloads. */
3362 use_sec_mem_p = false;
3363 if (! alt_p)
3364 use_sec_mem_p = true;
3365 else if (sec_mem_p)
3366 {
3367 for (i = 0; i < n_operands; i++)
3368 if (! goal_alt_win[i] && ! goal_alt_match_win[i])
3369 break;
3370 use_sec_mem_p = i < n_operands;
3371 }
3372
3373 if (use_sec_mem_p)
3374 {
3375 rtx new_reg, src, dest, rld;
3376 enum machine_mode sec_mode, rld_mode;
3377
3378 lra_assert (sec_mem_p);
3379 lra_assert (curr_static_id->operand[0].type == OP_OUT
3380 && curr_static_id->operand[1].type == OP_IN);
3381 dest = *curr_id->operand_loc[0];
3382 src = *curr_id->operand_loc[1];
3383 rld = (GET_MODE_SIZE (GET_MODE (dest)) <= GET_MODE_SIZE (GET_MODE (src))
3384 ? dest : src);
3385 rld_mode = GET_MODE (rld);
3386 #ifdef SECONDARY_MEMORY_NEEDED_MODE
3387 sec_mode = SECONDARY_MEMORY_NEEDED_MODE (rld_mode);
3388 #else
3389 sec_mode = rld_mode;
3390 #endif
3391 new_reg = lra_create_new_reg (sec_mode, NULL_RTX,
3392 NO_REGS, "secondary");
3393 /* If the mode is changed, it should be wider. */
3394 lra_assert (GET_MODE_SIZE (sec_mode) >= GET_MODE_SIZE (rld_mode));
3395 if (sec_mode != rld_mode)
3396 {
3397 /* If the target says specifically to use another mode for
3398 secondary memory moves we can not reuse the original
3399 insn. */
3400 after = emit_spill_move (false, new_reg, dest);
3401 lra_process_new_insns (curr_insn, NULL_RTX, after,
3402 "Inserting the sec. move");
3403 /* We may have non null BEFORE here (e.g. after address
3404 processing. */
3405 push_to_sequence (before);
3406 before = emit_spill_move (true, new_reg, src);
3407 emit_insn (before);
3408 before = get_insns ();
3409 end_sequence ();
3410 lra_process_new_insns (curr_insn, before, NULL_RTX, "Changing on");
3411 lra_set_insn_deleted (curr_insn);
3412 }
3413 else if (dest == rld)
3414 {
3415 *curr_id->operand_loc[0] = new_reg;
3416 after = emit_spill_move (false, new_reg, dest);
3417 lra_process_new_insns (curr_insn, NULL_RTX, after,
3418 "Inserting the sec. move");
3419 }
3420 else
3421 {
3422 *curr_id->operand_loc[1] = new_reg;
3423 /* See comments above. */
3424 push_to_sequence (before);
3425 before = emit_spill_move (true, new_reg, src);
3426 emit_insn (before);
3427 before = get_insns ();
3428 end_sequence ();
3429 lra_process_new_insns (curr_insn, before, NULL_RTX,
3430 "Inserting the sec. move");
3431 }
3432 lra_update_insn_regno_info (curr_insn);
3433 return true;
3434 }
3435 #endif
3436
3437 lra_assert (goal_alt_number >= 0);
3438 lra_set_used_insn_alternative (curr_insn, goal_alt_number);
3439
3440 if (lra_dump_file != NULL)
3441 {
3442 const char *p;
3443
3444 fprintf (lra_dump_file, " Choosing alt %d in insn %u:",
3445 goal_alt_number, INSN_UID (curr_insn));
3446 for (i = 0; i < n_operands; i++)
3447 {
3448 p = (curr_static_id->operand_alternative
3449 [goal_alt_number * n_operands + i].constraint);
3450 if (*p == '\0')
3451 continue;
3452 fprintf (lra_dump_file, " (%d) ", i);
3453 for (; *p != '\0' && *p != ',' && *p != '#'; p++)
3454 fputc (*p, lra_dump_file);
3455 }
3456 if (INSN_CODE (curr_insn) >= 0
3457 && (p = get_insn_name (INSN_CODE (curr_insn))) != NULL)
3458 fprintf (lra_dump_file, " {%s}", p);
3459 if (curr_id->sp_offset != 0)
3460 fprintf (lra_dump_file, " (sp_off=%" HOST_WIDE_INT_PRINT "d)",
3461 curr_id->sp_offset);
3462 fprintf (lra_dump_file, "\n");
3463 }
3464
3465 /* Right now, for any pair of operands I and J that are required to
3466 match, with J < I, goal_alt_matches[I] is J. Add I to
3467 goal_alt_matched[J]. */
3468
3469 for (i = 0; i < n_operands; i++)
3470 if ((j = goal_alt_matches[i]) >= 0)
3471 {
3472 for (k = 0; goal_alt_matched[j][k] >= 0; k++)
3473 ;
3474 /* We allow matching one output operand and several input
3475 operands. */
3476 lra_assert (k == 0
3477 || (curr_static_id->operand[j].type == OP_OUT
3478 && curr_static_id->operand[i].type == OP_IN
3479 && (curr_static_id->operand
3480 [goal_alt_matched[j][0]].type == OP_IN)));
3481 goal_alt_matched[j][k] = i;
3482 goal_alt_matched[j][k + 1] = -1;
3483 }
3484
3485 for (i = 0; i < n_operands; i++)
3486 goal_alt_win[i] |= goal_alt_match_win[i];
3487
3488 /* Any constants that aren't allowed and can't be reloaded into
3489 registers are here changed into memory references. */
3490 for (i = 0; i < n_operands; i++)
3491 if (goal_alt_win[i])
3492 {
3493 int regno;
3494 enum reg_class new_class;
3495 rtx reg = *curr_id->operand_loc[i];
3496
3497 if (GET_CODE (reg) == SUBREG)
3498 reg = SUBREG_REG (reg);
3499
3500 if (REG_P (reg) && (regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER)
3501 {
3502 bool ok_p = in_class_p (reg, goal_alt[i], &new_class);
3503
3504 if (new_class != NO_REGS && get_reg_class (regno) != new_class)
3505 {
3506 lra_assert (ok_p);
3507 lra_change_class (regno, new_class, " Change to", true);
3508 }
3509 }
3510 }
3511 else
3512 {
3513 const char *constraint;
3514 char c;
3515 rtx op = *curr_id->operand_loc[i];
3516 rtx subreg = NULL_RTX;
3517 enum machine_mode mode = curr_operand_mode[i];
3518
3519 if (GET_CODE (op) == SUBREG)
3520 {
3521 subreg = op;
3522 op = SUBREG_REG (op);
3523 mode = GET_MODE (op);
3524 }
3525
3526 if (CONST_POOL_OK_P (mode, op)
3527 && ((targetm.preferred_reload_class
3528 (op, (enum reg_class) goal_alt[i]) == NO_REGS)
3529 || no_input_reloads_p))
3530 {
3531 rtx tem = force_const_mem (mode, op);
3532
3533 change_p = true;
3534 if (subreg != NULL_RTX)
3535 tem = gen_rtx_SUBREG (mode, tem, SUBREG_BYTE (subreg));
3536
3537 *curr_id->operand_loc[i] = tem;
3538 lra_update_dup (curr_id, i);
3539 process_address (i, &before, &after);
3540
3541 /* If the alternative accepts constant pool refs directly
3542 there will be no reload needed at all. */
3543 if (subreg != NULL_RTX)
3544 continue;
3545 /* Skip alternatives before the one requested. */
3546 constraint = (curr_static_id->operand_alternative
3547 [goal_alt_number * n_operands + i].constraint);
3548 for (;
3549 (c = *constraint) && c != ',' && c != '#';
3550 constraint += CONSTRAINT_LEN (c, constraint))
3551 {
3552 if (c == TARGET_MEM_CONSTRAINT || c == 'o')
3553 break;
3554 #ifdef EXTRA_CONSTRAINT_STR
3555 if (EXTRA_MEMORY_CONSTRAINT (c, constraint)
3556 && EXTRA_CONSTRAINT_STR (tem, c, constraint))
3557 break;
3558 #endif
3559 }
3560 if (c == '\0' || c == ',' || c == '#')
3561 continue;
3562
3563 goal_alt_win[i] = true;
3564 }
3565 }
3566
3567 for (i = 0; i < n_operands; i++)
3568 {
3569 int regno;
3570 bool optional_p = false;
3571 rtx old, new_reg;
3572 rtx op = *curr_id->operand_loc[i];
3573
3574 if (goal_alt_win[i])
3575 {
3576 if (goal_alt[i] == NO_REGS
3577 && REG_P (op)
3578 /* When we assign NO_REGS it means that we will not
3579 assign a hard register to the scratch pseudo by
3580 assigment pass and the scratch pseudo will be
3581 spilled. Spilled scratch pseudos are transformed
3582 back to scratches at the LRA end. */
3583 && lra_former_scratch_operand_p (curr_insn, i))
3584 {
3585 int regno = REGNO (op);
3586 lra_change_class (regno, NO_REGS, " Change to", true);
3587 if (lra_get_regno_hard_regno (regno) >= 0)
3588 /* We don't have to mark all insn affected by the
3589 spilled pseudo as there is only one such insn, the
3590 current one. */
3591 reg_renumber[regno] = -1;
3592 }
3593 /* We can do an optional reload. If the pseudo got a hard
3594 reg, we might improve the code through inheritance. If
3595 it does not get a hard register we coalesce memory/memory
3596 moves later. Ignore move insns to avoid cycling. */
3597 if (! lra_simple_p
3598 && lra_undo_inheritance_iter < LRA_MAX_INHERITANCE_PASSES
3599 && goal_alt[i] != NO_REGS && REG_P (op)
3600 && (regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER
3601 && regno < new_regno_start
3602 && ! lra_former_scratch_p (regno)
3603 && reg_renumber[regno] < 0
3604 && (curr_insn_set == NULL_RTX
3605 || !((REG_P (SET_SRC (curr_insn_set))
3606 || MEM_P (SET_SRC (curr_insn_set))
3607 || GET_CODE (SET_SRC (curr_insn_set)) == SUBREG)
3608 && (REG_P (SET_DEST (curr_insn_set))
3609 || MEM_P (SET_DEST (curr_insn_set))
3610 || GET_CODE (SET_DEST (curr_insn_set)) == SUBREG))))
3611 optional_p = true;
3612 else
3613 continue;
3614 }
3615
3616 /* Operands that match previous ones have already been handled. */
3617 if (goal_alt_matches[i] >= 0)
3618 continue;
3619
3620 /* We should not have an operand with a non-offsettable address
3621 appearing where an offsettable address will do. It also may
3622 be a case when the address should be special in other words
3623 not a general one (e.g. it needs no index reg). */
3624 if (goal_alt_matched[i][0] == -1 && goal_alt_offmemok[i] && MEM_P (op))
3625 {
3626 enum reg_class rclass;
3627 rtx *loc = &XEXP (op, 0);
3628 enum rtx_code code = GET_CODE (*loc);
3629
3630 push_to_sequence (before);
3631 rclass = base_reg_class (GET_MODE (op), MEM_ADDR_SPACE (op),
3632 MEM, SCRATCH);
3633 if (GET_RTX_CLASS (code) == RTX_AUTOINC)
3634 new_reg = emit_inc (rclass, *loc, *loc,
3635 /* This value does not matter for MODIFY. */
3636 GET_MODE_SIZE (GET_MODE (op)));
3637 else if (get_reload_reg (OP_IN, Pmode, *loc, rclass, FALSE,
3638 "offsetable address", &new_reg))
3639 lra_emit_move (new_reg, *loc);
3640 before = get_insns ();
3641 end_sequence ();
3642 *loc = new_reg;
3643 lra_update_dup (curr_id, i);
3644 }
3645 else if (goal_alt_matched[i][0] == -1)
3646 {
3647 enum machine_mode mode;
3648 rtx reg, *loc;
3649 int hard_regno, byte;
3650 enum op_type type = curr_static_id->operand[i].type;
3651
3652 loc = curr_id->operand_loc[i];
3653 mode = curr_operand_mode[i];
3654 if (GET_CODE (*loc) == SUBREG)
3655 {
3656 reg = SUBREG_REG (*loc);
3657 byte = SUBREG_BYTE (*loc);
3658 if (REG_P (reg)
3659 /* Strict_low_part requires reload the register not
3660 the sub-register. */
3661 && (curr_static_id->operand[i].strict_low
3662 || (GET_MODE_SIZE (mode)
3663 <= GET_MODE_SIZE (GET_MODE (reg))
3664 && (hard_regno
3665 = get_try_hard_regno (REGNO (reg))) >= 0
3666 && (simplify_subreg_regno
3667 (hard_regno,
3668 GET_MODE (reg), byte, mode) < 0)
3669 && (goal_alt[i] == NO_REGS
3670 || (simplify_subreg_regno
3671 (ira_class_hard_regs[goal_alt[i]][0],
3672 GET_MODE (reg), byte, mode) >= 0)))))
3673 {
3674 loc = &SUBREG_REG (*loc);
3675 mode = GET_MODE (*loc);
3676 }
3677 }
3678 old = *loc;
3679 if (get_reload_reg (type, mode, old, goal_alt[i],
3680 loc != curr_id->operand_loc[i], "", &new_reg)
3681 && type != OP_OUT)
3682 {
3683 push_to_sequence (before);
3684 lra_emit_move (new_reg, old);
3685 before = get_insns ();
3686 end_sequence ();
3687 }
3688 *loc = new_reg;
3689 if (type != OP_IN
3690 && find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX)
3691 {
3692 start_sequence ();
3693 lra_emit_move (type == OP_INOUT ? copy_rtx (old) : old, new_reg);
3694 emit_insn (after);
3695 after = get_insns ();
3696 end_sequence ();
3697 *loc = new_reg;
3698 }
3699 for (j = 0; j < goal_alt_dont_inherit_ops_num; j++)
3700 if (goal_alt_dont_inherit_ops[j] == i)
3701 {
3702 lra_set_regno_unique_value (REGNO (new_reg));
3703 break;
3704 }
3705 lra_update_dup (curr_id, i);
3706 }
3707 else if (curr_static_id->operand[i].type == OP_IN
3708 && (curr_static_id->operand[goal_alt_matched[i][0]].type
3709 == OP_OUT))
3710 {
3711 /* generate reloads for input and matched outputs. */
3712 match_inputs[0] = i;
3713 match_inputs[1] = -1;
3714 match_reload (goal_alt_matched[i][0], match_inputs,
3715 goal_alt[i], &before, &after);
3716 }
3717 else if (curr_static_id->operand[i].type == OP_OUT
3718 && (curr_static_id->operand[goal_alt_matched[i][0]].type
3719 == OP_IN))
3720 /* Generate reloads for output and matched inputs. */
3721 match_reload (i, goal_alt_matched[i], goal_alt[i], &before, &after);
3722 else if (curr_static_id->operand[i].type == OP_IN
3723 && (curr_static_id->operand[goal_alt_matched[i][0]].type
3724 == OP_IN))
3725 {
3726 /* Generate reloads for matched inputs. */
3727 match_inputs[0] = i;
3728 for (j = 0; (k = goal_alt_matched[i][j]) >= 0; j++)
3729 match_inputs[j + 1] = k;
3730 match_inputs[j + 1] = -1;
3731 match_reload (-1, match_inputs, goal_alt[i], &before, &after);
3732 }
3733 else
3734 /* We must generate code in any case when function
3735 process_alt_operands decides that it is possible. */
3736 gcc_unreachable ();
3737 if (optional_p)
3738 {
3739 lra_assert (REG_P (op));
3740 regno = REGNO (op);
3741 op = *curr_id->operand_loc[i]; /* Substitution. */
3742 if (GET_CODE (op) == SUBREG)
3743 op = SUBREG_REG (op);
3744 gcc_assert (REG_P (op) && (int) REGNO (op) >= new_regno_start);
3745 bitmap_set_bit (&lra_optional_reload_pseudos, REGNO (op));
3746 lra_reg_info[REGNO (op)].restore_regno = regno;
3747 if (lra_dump_file != NULL)
3748 fprintf (lra_dump_file,
3749 " Making reload reg %d for reg %d optional\n",
3750 REGNO (op), regno);
3751 }
3752 }
3753 if (before != NULL_RTX || after != NULL_RTX
3754 || max_regno_before != max_reg_num ())
3755 change_p = true;
3756 if (change_p)
3757 {
3758 lra_update_operator_dups (curr_id);
3759 /* Something changes -- process the insn. */
3760 lra_update_insn_regno_info (curr_insn);
3761 }
3762 lra_process_new_insns (curr_insn, before, after, "Inserting insn reload");
3763 return change_p;
3764 }
3765
3766 /* Return true if X is in LIST. */
3767 static bool
3768 in_list_p (rtx x, rtx list)
3769 {
3770 for (; list != NULL_RTX; list = XEXP (list, 1))
3771 if (XEXP (list, 0) == x)
3772 return true;
3773 return false;
3774 }
3775
3776 /* Return true if X contains an allocatable hard register (if
3777 HARD_REG_P) or a (spilled if SPILLED_P) pseudo. */
3778 static bool
3779 contains_reg_p (rtx x, bool hard_reg_p, bool spilled_p)
3780 {
3781 int i, j;
3782 const char *fmt;
3783 enum rtx_code code;
3784
3785 code = GET_CODE (x);
3786 if (REG_P (x))
3787 {
3788 int regno = REGNO (x);
3789 HARD_REG_SET alloc_regs;
3790
3791 if (hard_reg_p)
3792 {
3793 if (regno >= FIRST_PSEUDO_REGISTER)
3794 regno = lra_get_regno_hard_regno (regno);
3795 if (regno < 0)
3796 return false;
3797 COMPL_HARD_REG_SET (alloc_regs, lra_no_alloc_regs);
3798 return overlaps_hard_reg_set_p (alloc_regs, GET_MODE (x), regno);
3799 }
3800 else
3801 {
3802 if (regno < FIRST_PSEUDO_REGISTER)
3803 return false;
3804 if (! spilled_p)
3805 return true;
3806 return lra_get_regno_hard_regno (regno) < 0;
3807 }
3808 }
3809 fmt = GET_RTX_FORMAT (code);
3810 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3811 {
3812 if (fmt[i] == 'e')
3813 {
3814 if (contains_reg_p (XEXP (x, i), hard_reg_p, spilled_p))
3815 return true;
3816 }
3817 else if (fmt[i] == 'E')
3818 {
3819 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3820 if (contains_reg_p (XVECEXP (x, i, j), hard_reg_p, spilled_p))
3821 return true;
3822 }
3823 }
3824 return false;
3825 }
3826
3827 /* Process all regs in location *LOC and change them on equivalent
3828 substitution. Return true if any change was done. */
3829 static bool
3830 loc_equivalence_change_p (rtx *loc)
3831 {
3832 rtx subst, reg, x = *loc;
3833 bool result = false;
3834 enum rtx_code code = GET_CODE (x);
3835 const char *fmt;
3836 int i, j;
3837
3838 if (code == SUBREG)
3839 {
3840 reg = SUBREG_REG (x);
3841 if ((subst = get_equiv_with_elimination (reg, curr_insn)) != reg
3842 && GET_MODE (subst) == VOIDmode)
3843 {
3844 /* We cannot reload debug location. Simplify subreg here
3845 while we know the inner mode. */
3846 *loc = simplify_gen_subreg (GET_MODE (x), subst,
3847 GET_MODE (reg), SUBREG_BYTE (x));
3848 return true;
3849 }
3850 }
3851 if (code == REG && (subst = get_equiv_with_elimination (x, curr_insn)) != x)
3852 {
3853 *loc = subst;
3854 return true;
3855 }
3856
3857 /* Scan all the operand sub-expressions. */
3858 fmt = GET_RTX_FORMAT (code);
3859 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3860 {
3861 if (fmt[i] == 'e')
3862 result = loc_equivalence_change_p (&XEXP (x, i)) || result;
3863 else if (fmt[i] == 'E')
3864 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3865 result
3866 = loc_equivalence_change_p (&XVECEXP (x, i, j)) || result;
3867 }
3868 return result;
3869 }
3870
3871 /* Similar to loc_equivalence_change_p, but for use as
3872 simplify_replace_fn_rtx callback. DATA is insn for which the
3873 elimination is done. If it null we don't do the elimination. */
3874 static rtx
3875 loc_equivalence_callback (rtx loc, const_rtx, void *data)
3876 {
3877 if (!REG_P (loc))
3878 return NULL_RTX;
3879
3880 rtx subst = (data == NULL
3881 ? get_equiv (loc) : get_equiv_with_elimination (loc, (rtx) data));
3882 if (subst != loc)
3883 return subst;
3884
3885 return NULL_RTX;
3886 }
3887
3888 /* Maximum number of generated reload insns per an insn. It is for
3889 preventing this pass cycling in a bug case. */
3890 #define MAX_RELOAD_INSNS_NUMBER LRA_MAX_INSN_RELOADS
3891
3892 /* The current iteration number of this LRA pass. */
3893 int lra_constraint_iter;
3894
3895 /* The current iteration number of this LRA pass after the last spill
3896 pass. */
3897 int lra_constraint_iter_after_spill;
3898
3899 /* True if we substituted equiv which needs checking register
3900 allocation correctness because the equivalent value contains
3901 allocatable hard registers or when we restore multi-register
3902 pseudo. */
3903 bool lra_risky_transformations_p;
3904
3905 /* Return true if REGNO is referenced in more than one block. */
3906 static bool
3907 multi_block_pseudo_p (int regno)
3908 {
3909 basic_block bb = NULL;
3910 unsigned int uid;
3911 bitmap_iterator bi;
3912
3913 if (regno < FIRST_PSEUDO_REGISTER)
3914 return false;
3915
3916 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi)
3917 if (bb == NULL)
3918 bb = BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn);
3919 else if (BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn) != bb)
3920 return true;
3921 return false;
3922 }
3923
3924 /* Return true if LIST contains a deleted insn. */
3925 static bool
3926 contains_deleted_insn_p (rtx list)
3927 {
3928 for (; list != NULL_RTX; list = XEXP (list, 1))
3929 if (NOTE_P (XEXP (list, 0))
3930 && NOTE_KIND (XEXP (list, 0)) == NOTE_INSN_DELETED)
3931 return true;
3932 return false;
3933 }
3934
3935 /* Return true if X contains a pseudo dying in INSN. */
3936 static bool
3937 dead_pseudo_p (rtx x, rtx insn)
3938 {
3939 int i, j;
3940 const char *fmt;
3941 enum rtx_code code;
3942
3943 if (REG_P (x))
3944 return (insn != NULL_RTX
3945 && find_regno_note (insn, REG_DEAD, REGNO (x)) != NULL_RTX);
3946 code = GET_CODE (x);
3947 fmt = GET_RTX_FORMAT (code);
3948 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3949 {
3950 if (fmt[i] == 'e')
3951 {
3952 if (dead_pseudo_p (XEXP (x, i), insn))
3953 return true;
3954 }
3955 else if (fmt[i] == 'E')
3956 {
3957 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3958 if (dead_pseudo_p (XVECEXP (x, i, j), insn))
3959 return true;
3960 }
3961 }
3962 return false;
3963 }
3964
3965 /* Return true if INSN contains a dying pseudo in INSN right hand
3966 side. */
3967 static bool
3968 insn_rhs_dead_pseudo_p (rtx insn)
3969 {
3970 rtx set = single_set (insn);
3971
3972 gcc_assert (set != NULL);
3973 return dead_pseudo_p (SET_SRC (set), insn);
3974 }
3975
3976 /* Return true if any init insn of REGNO contains a dying pseudo in
3977 insn right hand side. */
3978 static bool
3979 init_insn_rhs_dead_pseudo_p (int regno)
3980 {
3981 rtx insns = ira_reg_equiv[regno].init_insns;
3982
3983 if (insns == NULL)
3984 return false;
3985 if (INSN_P (insns))
3986 return insn_rhs_dead_pseudo_p (insns);
3987 for (; insns != NULL_RTX; insns = XEXP (insns, 1))
3988 if (insn_rhs_dead_pseudo_p (XEXP (insns, 0)))
3989 return true;
3990 return false;
3991 }
3992
3993 /* Return TRUE if REGNO has a reverse equivalence. The equivalence is
3994 reverse only if we have one init insn with given REGNO as a
3995 source. */
3996 static bool
3997 reverse_equiv_p (int regno)
3998 {
3999 rtx insns, set;
4000
4001 if ((insns = ira_reg_equiv[regno].init_insns) == NULL_RTX)
4002 return false;
4003 if (! INSN_P (XEXP (insns, 0))
4004 || XEXP (insns, 1) != NULL_RTX)
4005 return false;
4006 if ((set = single_set (XEXP (insns, 0))) == NULL_RTX)
4007 return false;
4008 return REG_P (SET_SRC (set)) && (int) REGNO (SET_SRC (set)) == regno;
4009 }
4010
4011 /* Return TRUE if REGNO was reloaded in an equivalence init insn. We
4012 call this function only for non-reverse equivalence. */
4013 static bool
4014 contains_reloaded_insn_p (int regno)
4015 {
4016 rtx set;
4017 rtx list = ira_reg_equiv[regno].init_insns;
4018
4019 for (; list != NULL_RTX; list = XEXP (list, 1))
4020 if ((set = single_set (XEXP (list, 0))) == NULL_RTX
4021 || ! REG_P (SET_DEST (set))
4022 || (int) REGNO (SET_DEST (set)) != regno)
4023 return true;
4024 return false;
4025 }
4026
4027 /* Entry function of LRA constraint pass. Return true if the
4028 constraint pass did change the code. */
4029 bool
4030 lra_constraints (bool first_p)
4031 {
4032 bool changed_p;
4033 int i, hard_regno, new_insns_num;
4034 unsigned int min_len, new_min_len, uid;
4035 rtx set, x, reg, dest_reg;
4036 basic_block last_bb;
4037 bitmap_head equiv_insn_bitmap;
4038 bitmap_iterator bi;
4039
4040 lra_constraint_iter++;
4041 if (lra_dump_file != NULL)
4042 fprintf (lra_dump_file, "\n********** Local #%d: **********\n\n",
4043 lra_constraint_iter);
4044 lra_constraint_iter_after_spill++;
4045 if (lra_constraint_iter_after_spill > LRA_MAX_CONSTRAINT_ITERATION_NUMBER)
4046 internal_error
4047 ("Maximum number of LRA constraint passes is achieved (%d)\n",
4048 LRA_MAX_CONSTRAINT_ITERATION_NUMBER);
4049 changed_p = false;
4050 lra_risky_transformations_p = false;
4051 new_insn_uid_start = get_max_uid ();
4052 new_regno_start = first_p ? lra_constraint_new_regno_start : max_reg_num ();
4053 /* Mark used hard regs for target stack size calulations. */
4054 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4055 if (lra_reg_info[i].nrefs != 0
4056 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4057 {
4058 int j, nregs;
4059
4060 nregs = hard_regno_nregs[hard_regno][lra_reg_info[i].biggest_mode];
4061 for (j = 0; j < nregs; j++)
4062 df_set_regs_ever_live (hard_regno + j, true);
4063 }
4064 /* Do elimination before the equivalence processing as we can spill
4065 some pseudos during elimination. */
4066 lra_eliminate (false, first_p);
4067 bitmap_initialize (&equiv_insn_bitmap, &reg_obstack);
4068 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4069 if (lra_reg_info[i].nrefs != 0)
4070 {
4071 ira_reg_equiv[i].profitable_p = true;
4072 reg = regno_reg_rtx[i];
4073 if (lra_get_regno_hard_regno (i) < 0 && (x = get_equiv (reg)) != reg)
4074 {
4075 bool pseudo_p = contains_reg_p (x, false, false);
4076
4077 /* After RTL transformation, we can not guarantee that
4078 pseudo in the substitution was not reloaded which might
4079 make equivalence invalid. For example, in reverse
4080 equiv of p0
4081
4082 p0 <- ...
4083 ...
4084 equiv_mem <- p0
4085
4086 the memory address register was reloaded before the 2nd
4087 insn. */
4088 if ((! first_p && pseudo_p)
4089 /* We don't use DF for compilation speed sake. So it
4090 is problematic to update live info when we use an
4091 equivalence containing pseudos in more than one
4092 BB. */
4093 || (pseudo_p && multi_block_pseudo_p (i))
4094 /* If an init insn was deleted for some reason, cancel
4095 the equiv. We could update the equiv insns after
4096 transformations including an equiv insn deletion
4097 but it is not worthy as such cases are extremely
4098 rare. */
4099 || contains_deleted_insn_p (ira_reg_equiv[i].init_insns)
4100 /* If it is not a reverse equivalence, we check that a
4101 pseudo in rhs of the init insn is not dying in the
4102 insn. Otherwise, the live info at the beginning of
4103 the corresponding BB might be wrong after we
4104 removed the insn. When the equiv can be a
4105 constant, the right hand side of the init insn can
4106 be a pseudo. */
4107 || (! reverse_equiv_p (i)
4108 && (init_insn_rhs_dead_pseudo_p (i)
4109 /* If we reloaded the pseudo in an equivalence
4110 init insn, we can not remove the equiv init
4111 insns and the init insns might write into
4112 const memory in this case. */
4113 || contains_reloaded_insn_p (i)))
4114 /* Prevent access beyond equivalent memory for
4115 paradoxical subregs. */
4116 || (MEM_P (x)
4117 && (GET_MODE_SIZE (lra_reg_info[i].biggest_mode)
4118 > GET_MODE_SIZE (GET_MODE (x)))))
4119 ira_reg_equiv[i].defined_p = false;
4120 if (contains_reg_p (x, false, true))
4121 ira_reg_equiv[i].profitable_p = false;
4122 if (get_equiv (reg) != reg)
4123 bitmap_ior_into (&equiv_insn_bitmap, &lra_reg_info[i].insn_bitmap);
4124 }
4125 }
4126 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4127 update_equiv (i);
4128 /* We should add all insns containing pseudos which should be
4129 substituted by their equivalences. */
4130 EXECUTE_IF_SET_IN_BITMAP (&equiv_insn_bitmap, 0, uid, bi)
4131 lra_push_insn_by_uid (uid);
4132 min_len = lra_insn_stack_length ();
4133 new_insns_num = 0;
4134 last_bb = NULL;
4135 changed_p = false;
4136 while ((new_min_len = lra_insn_stack_length ()) != 0)
4137 {
4138 curr_insn = lra_pop_insn ();
4139 --new_min_len;
4140 curr_bb = BLOCK_FOR_INSN (curr_insn);
4141 if (curr_bb != last_bb)
4142 {
4143 last_bb = curr_bb;
4144 bb_reload_num = lra_curr_reload_num;
4145 }
4146 if (min_len > new_min_len)
4147 {
4148 min_len = new_min_len;
4149 new_insns_num = 0;
4150 }
4151 if (new_insns_num > MAX_RELOAD_INSNS_NUMBER)
4152 internal_error
4153 ("Max. number of generated reload insns per insn is achieved (%d)\n",
4154 MAX_RELOAD_INSNS_NUMBER);
4155 new_insns_num++;
4156 if (DEBUG_INSN_P (curr_insn))
4157 {
4158 /* We need to check equivalence in debug insn and change
4159 pseudo to the equivalent value if necessary. */
4160 curr_id = lra_get_insn_recog_data (curr_insn);
4161 if (bitmap_bit_p (&equiv_insn_bitmap, INSN_UID (curr_insn)))
4162 {
4163 rtx old = *curr_id->operand_loc[0];
4164 *curr_id->operand_loc[0]
4165 = simplify_replace_fn_rtx (old, NULL_RTX,
4166 loc_equivalence_callback, curr_insn);
4167 if (old != *curr_id->operand_loc[0])
4168 {
4169 lra_update_insn_regno_info (curr_insn);
4170 changed_p = true;
4171 }
4172 }
4173 }
4174 else if (INSN_P (curr_insn))
4175 {
4176 if ((set = single_set (curr_insn)) != NULL_RTX)
4177 {
4178 dest_reg = SET_DEST (set);
4179 /* The equivalence pseudo could be set up as SUBREG in a
4180 case when it is a call restore insn in a mode
4181 different from the pseudo mode. */
4182 if (GET_CODE (dest_reg) == SUBREG)
4183 dest_reg = SUBREG_REG (dest_reg);
4184 if ((REG_P (dest_reg)
4185 && (x = get_equiv (dest_reg)) != dest_reg
4186 /* Remove insns which set up a pseudo whose value
4187 can not be changed. Such insns might be not in
4188 init_insns because we don't update equiv data
4189 during insn transformations.
4190
4191 As an example, let suppose that a pseudo got
4192 hard register and on the 1st pass was not
4193 changed to equivalent constant. We generate an
4194 additional insn setting up the pseudo because of
4195 secondary memory movement. Then the pseudo is
4196 spilled and we use the equiv constant. In this
4197 case we should remove the additional insn and
4198 this insn is not init_insns list. */
4199 && (! MEM_P (x) || MEM_READONLY_P (x)
4200 /* Check that this is actually an insn setting
4201 up the equivalence. */
4202 || in_list_p (curr_insn,
4203 ira_reg_equiv
4204 [REGNO (dest_reg)].init_insns)))
4205 || (((x = get_equiv (SET_SRC (set))) != SET_SRC (set))
4206 && in_list_p (curr_insn,
4207 ira_reg_equiv
4208 [REGNO (SET_SRC (set))].init_insns)))
4209 {
4210 /* This is equiv init insn of pseudo which did not get a
4211 hard register -- remove the insn. */
4212 if (lra_dump_file != NULL)
4213 {
4214 fprintf (lra_dump_file,
4215 " Removing equiv init insn %i (freq=%d)\n",
4216 INSN_UID (curr_insn),
4217 REG_FREQ_FROM_BB (BLOCK_FOR_INSN (curr_insn)));
4218 dump_insn_slim (lra_dump_file, curr_insn);
4219 }
4220 if (contains_reg_p (x, true, false))
4221 lra_risky_transformations_p = true;
4222 lra_set_insn_deleted (curr_insn);
4223 continue;
4224 }
4225 }
4226 curr_id = lra_get_insn_recog_data (curr_insn);
4227 curr_static_id = curr_id->insn_static_data;
4228 init_curr_insn_input_reloads ();
4229 init_curr_operand_mode ();
4230 if (curr_insn_transform ())
4231 changed_p = true;
4232 /* Check non-transformed insns too for equiv change as USE
4233 or CLOBBER don't need reloads but can contain pseudos
4234 being changed on their equivalences. */
4235 else if (bitmap_bit_p (&equiv_insn_bitmap, INSN_UID (curr_insn))
4236 && loc_equivalence_change_p (&PATTERN (curr_insn)))
4237 {
4238 lra_update_insn_regno_info (curr_insn);
4239 changed_p = true;
4240 }
4241 }
4242 }
4243 bitmap_clear (&equiv_insn_bitmap);
4244 /* If we used a new hard regno, changed_p should be true because the
4245 hard reg is assigned to a new pseudo. */
4246 #ifdef ENABLE_CHECKING
4247 if (! changed_p)
4248 {
4249 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4250 if (lra_reg_info[i].nrefs != 0
4251 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4252 {
4253 int j, nregs = hard_regno_nregs[hard_regno][PSEUDO_REGNO_MODE (i)];
4254
4255 for (j = 0; j < nregs; j++)
4256 lra_assert (df_regs_ever_live_p (hard_regno + j));
4257 }
4258 }
4259 #endif
4260 return changed_p;
4261 }
4262
4263 /* Initiate the LRA constraint pass. It is done once per
4264 function. */
4265 void
4266 lra_constraints_init (void)
4267 {
4268 }
4269
4270 /* Finalize the LRA constraint pass. It is done once per
4271 function. */
4272 void
4273 lra_constraints_finish (void)
4274 {
4275 }
4276
4277 \f
4278
4279 /* This page contains code to do inheritance/split
4280 transformations. */
4281
4282 /* Number of reloads passed so far in current EBB. */
4283 static int reloads_num;
4284
4285 /* Number of calls passed so far in current EBB. */
4286 static int calls_num;
4287
4288 /* Current reload pseudo check for validity of elements in
4289 USAGE_INSNS. */
4290 static int curr_usage_insns_check;
4291
4292 /* Info about last usage of registers in EBB to do inheritance/split
4293 transformation. Inheritance transformation is done from a spilled
4294 pseudo and split transformations from a hard register or a pseudo
4295 assigned to a hard register. */
4296 struct usage_insns
4297 {
4298 /* If the value is equal to CURR_USAGE_INSNS_CHECK, then the member
4299 value INSNS is valid. The insns is chain of optional debug insns
4300 and a finishing non-debug insn using the corresponding reg. The
4301 value is also used to mark the registers which are set up in the
4302 current insn. The negated insn uid is used for this. */
4303 int check;
4304 /* Value of global reloads_num at the last insn in INSNS. */
4305 int reloads_num;
4306 /* Value of global reloads_nums at the last insn in INSNS. */
4307 int calls_num;
4308 /* It can be true only for splitting. And it means that the restore
4309 insn should be put after insn given by the following member. */
4310 bool after_p;
4311 /* Next insns in the current EBB which use the original reg and the
4312 original reg value is not changed between the current insn and
4313 the next insns. In order words, e.g. for inheritance, if we need
4314 to use the original reg value again in the next insns we can try
4315 to use the value in a hard register from a reload insn of the
4316 current insn. */
4317 rtx insns;
4318 };
4319
4320 /* Map: regno -> corresponding pseudo usage insns. */
4321 static struct usage_insns *usage_insns;
4322
4323 static void
4324 setup_next_usage_insn (int regno, rtx insn, int reloads_num, bool after_p)
4325 {
4326 usage_insns[regno].check = curr_usage_insns_check;
4327 usage_insns[regno].insns = insn;
4328 usage_insns[regno].reloads_num = reloads_num;
4329 usage_insns[regno].calls_num = calls_num;
4330 usage_insns[regno].after_p = after_p;
4331 }
4332
4333 /* The function is used to form list REGNO usages which consists of
4334 optional debug insns finished by a non-debug insn using REGNO.
4335 RELOADS_NUM is current number of reload insns processed so far. */
4336 static void
4337 add_next_usage_insn (int regno, rtx insn, int reloads_num)
4338 {
4339 rtx next_usage_insns;
4340
4341 if (usage_insns[regno].check == curr_usage_insns_check
4342 && (next_usage_insns = usage_insns[regno].insns) != NULL_RTX
4343 && DEBUG_INSN_P (insn))
4344 {
4345 /* Check that we did not add the debug insn yet. */
4346 if (next_usage_insns != insn
4347 && (GET_CODE (next_usage_insns) != INSN_LIST
4348 || XEXP (next_usage_insns, 0) != insn))
4349 usage_insns[regno].insns = gen_rtx_INSN_LIST (VOIDmode, insn,
4350 next_usage_insns);
4351 }
4352 else if (NONDEBUG_INSN_P (insn))
4353 setup_next_usage_insn (regno, insn, reloads_num, false);
4354 else
4355 usage_insns[regno].check = 0;
4356 }
4357
4358 /* Replace all references to register OLD_REGNO in *LOC with pseudo
4359 register NEW_REG. Return true if any change was made. */
4360 static bool
4361 substitute_pseudo (rtx *loc, int old_regno, rtx new_reg)
4362 {
4363 rtx x = *loc;
4364 bool result = false;
4365 enum rtx_code code;
4366 const char *fmt;
4367 int i, j;
4368
4369 if (x == NULL_RTX)
4370 return false;
4371
4372 code = GET_CODE (x);
4373 if (code == REG && (int) REGNO (x) == old_regno)
4374 {
4375 enum machine_mode mode = GET_MODE (*loc);
4376 enum machine_mode inner_mode = GET_MODE (new_reg);
4377
4378 if (mode != inner_mode)
4379 {
4380 if (GET_MODE_SIZE (mode) >= GET_MODE_SIZE (inner_mode)
4381 || ! SCALAR_INT_MODE_P (inner_mode))
4382 new_reg = gen_rtx_SUBREG (mode, new_reg, 0);
4383 else
4384 new_reg = gen_lowpart_SUBREG (mode, new_reg);
4385 }
4386 *loc = new_reg;
4387 return true;
4388 }
4389
4390 /* Scan all the operand sub-expressions. */
4391 fmt = GET_RTX_FORMAT (code);
4392 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4393 {
4394 if (fmt[i] == 'e')
4395 {
4396 if (substitute_pseudo (&XEXP (x, i), old_regno, new_reg))
4397 result = true;
4398 }
4399 else if (fmt[i] == 'E')
4400 {
4401 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4402 if (substitute_pseudo (&XVECEXP (x, i, j), old_regno, new_reg))
4403 result = true;
4404 }
4405 }
4406 return result;
4407 }
4408
4409 /* Return first non-debug insn in list USAGE_INSNS. */
4410 static rtx
4411 skip_usage_debug_insns (rtx usage_insns)
4412 {
4413 rtx insn;
4414
4415 /* Skip debug insns. */
4416 for (insn = usage_insns;
4417 insn != NULL_RTX && GET_CODE (insn) == INSN_LIST;
4418 insn = XEXP (insn, 1))
4419 ;
4420 return insn;
4421 }
4422
4423 /* Return true if we need secondary memory moves for insn in
4424 USAGE_INSNS after inserting inherited pseudo of class INHER_CL
4425 into the insn. */
4426 static bool
4427 check_secondary_memory_needed_p (enum reg_class inher_cl ATTRIBUTE_UNUSED,
4428 rtx usage_insns ATTRIBUTE_UNUSED)
4429 {
4430 #ifndef SECONDARY_MEMORY_NEEDED
4431 return false;
4432 #else
4433 rtx insn, set, dest;
4434 enum reg_class cl;
4435
4436 if (inher_cl == ALL_REGS
4437 || (insn = skip_usage_debug_insns (usage_insns)) == NULL_RTX)
4438 return false;
4439 lra_assert (INSN_P (insn));
4440 if ((set = single_set (insn)) == NULL_RTX || ! REG_P (SET_DEST (set)))
4441 return false;
4442 dest = SET_DEST (set);
4443 if (! REG_P (dest))
4444 return false;
4445 lra_assert (inher_cl != NO_REGS);
4446 cl = get_reg_class (REGNO (dest));
4447 return (cl != NO_REGS && cl != ALL_REGS
4448 && SECONDARY_MEMORY_NEEDED (inher_cl, cl, GET_MODE (dest)));
4449 #endif
4450 }
4451
4452 /* Registers involved in inheritance/split in the current EBB
4453 (inheritance/split pseudos and original registers). */
4454 static bitmap_head check_only_regs;
4455
4456 /* Do inheritance transformations for insn INSN, which defines (if
4457 DEF_P) or uses ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which
4458 instruction in the EBB next uses ORIGINAL_REGNO; it has the same
4459 form as the "insns" field of usage_insns. Return true if we
4460 succeed in such transformation.
4461
4462 The transformations look like:
4463
4464 p <- ... i <- ...
4465 ... p <- i (new insn)
4466 ... =>
4467 <- ... p ... <- ... i ...
4468 or
4469 ... i <- p (new insn)
4470 <- ... p ... <- ... i ...
4471 ... =>
4472 <- ... p ... <- ... i ...
4473 where p is a spilled original pseudo and i is a new inheritance pseudo.
4474
4475
4476 The inheritance pseudo has the smallest class of two classes CL and
4477 class of ORIGINAL REGNO. */
4478 static bool
4479 inherit_reload_reg (bool def_p, int original_regno,
4480 enum reg_class cl, rtx insn, rtx next_usage_insns)
4481 {
4482 if (optimize_function_for_size_p (cfun))
4483 return false;
4484
4485 enum reg_class rclass = lra_get_allocno_class (original_regno);
4486 rtx original_reg = regno_reg_rtx[original_regno];
4487 rtx new_reg, new_insns, usage_insn;
4488
4489 lra_assert (! usage_insns[original_regno].after_p);
4490 if (lra_dump_file != NULL)
4491 fprintf (lra_dump_file,
4492 " <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n");
4493 if (! ira_reg_classes_intersect_p[cl][rclass])
4494 {
4495 if (lra_dump_file != NULL)
4496 {
4497 fprintf (lra_dump_file,
4498 " Rejecting inheritance for %d "
4499 "because of disjoint classes %s and %s\n",
4500 original_regno, reg_class_names[cl],
4501 reg_class_names[rclass]);
4502 fprintf (lra_dump_file,
4503 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
4504 }
4505 return false;
4506 }
4507 if ((ira_class_subset_p[cl][rclass] && cl != rclass)
4508 /* We don't use a subset of two classes because it can be
4509 NO_REGS. This transformation is still profitable in most
4510 cases even if the classes are not intersected as register
4511 move is probably cheaper than a memory load. */
4512 || ira_class_hard_regs_num[cl] < ira_class_hard_regs_num[rclass])
4513 {
4514 if (lra_dump_file != NULL)
4515 fprintf (lra_dump_file, " Use smallest class of %s and %s\n",
4516 reg_class_names[cl], reg_class_names[rclass]);
4517
4518 rclass = cl;
4519 }
4520 if (check_secondary_memory_needed_p (rclass, next_usage_insns))
4521 {
4522 /* Reject inheritance resulting in secondary memory moves.
4523 Otherwise, there is a danger in LRA cycling. Also such
4524 transformation will be unprofitable. */
4525 if (lra_dump_file != NULL)
4526 {
4527 rtx insn = skip_usage_debug_insns (next_usage_insns);
4528 rtx set = single_set (insn);
4529
4530 lra_assert (set != NULL_RTX);
4531
4532 rtx dest = SET_DEST (set);
4533
4534 lra_assert (REG_P (dest));
4535 fprintf (lra_dump_file,
4536 " Rejecting inheritance for insn %d(%s)<-%d(%s) "
4537 "as secondary mem is needed\n",
4538 REGNO (dest), reg_class_names[get_reg_class (REGNO (dest))],
4539 original_regno, reg_class_names[rclass]);
4540 fprintf (lra_dump_file,
4541 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
4542 }
4543 return false;
4544 }
4545 new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg,
4546 rclass, "inheritance");
4547 start_sequence ();
4548 if (def_p)
4549 lra_emit_move (original_reg, new_reg);
4550 else
4551 lra_emit_move (new_reg, original_reg);
4552 new_insns = get_insns ();
4553 end_sequence ();
4554 if (NEXT_INSN (new_insns) != NULL_RTX)
4555 {
4556 if (lra_dump_file != NULL)
4557 {
4558 fprintf (lra_dump_file,
4559 " Rejecting inheritance %d->%d "
4560 "as it results in 2 or more insns:\n",
4561 original_regno, REGNO (new_reg));
4562 dump_rtl_slim (lra_dump_file, new_insns, NULL_RTX, -1, 0);
4563 fprintf (lra_dump_file,
4564 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
4565 }
4566 return false;
4567 }
4568 substitute_pseudo (&insn, original_regno, new_reg);
4569 lra_update_insn_regno_info (insn);
4570 if (! def_p)
4571 /* We now have a new usage insn for original regno. */
4572 setup_next_usage_insn (original_regno, new_insns, reloads_num, false);
4573 if (lra_dump_file != NULL)
4574 fprintf (lra_dump_file, " Original reg change %d->%d (bb%d):\n",
4575 original_regno, REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
4576 lra_reg_info[REGNO (new_reg)].restore_regno = original_regno;
4577 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
4578 bitmap_set_bit (&check_only_regs, original_regno);
4579 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
4580 if (def_p)
4581 lra_process_new_insns (insn, NULL_RTX, new_insns,
4582 "Add original<-inheritance");
4583 else
4584 lra_process_new_insns (insn, new_insns, NULL_RTX,
4585 "Add inheritance<-original");
4586 while (next_usage_insns != NULL_RTX)
4587 {
4588 if (GET_CODE (next_usage_insns) != INSN_LIST)
4589 {
4590 usage_insn = next_usage_insns;
4591 lra_assert (NONDEBUG_INSN_P (usage_insn));
4592 next_usage_insns = NULL;
4593 }
4594 else
4595 {
4596 usage_insn = XEXP (next_usage_insns, 0);
4597 lra_assert (DEBUG_INSN_P (usage_insn));
4598 next_usage_insns = XEXP (next_usage_insns, 1);
4599 }
4600 substitute_pseudo (&usage_insn, original_regno, new_reg);
4601 lra_update_insn_regno_info (usage_insn);
4602 if (lra_dump_file != NULL)
4603 {
4604 fprintf (lra_dump_file,
4605 " Inheritance reuse change %d->%d (bb%d):\n",
4606 original_regno, REGNO (new_reg),
4607 BLOCK_FOR_INSN (usage_insn)->index);
4608 dump_insn_slim (lra_dump_file, usage_insn);
4609 }
4610 }
4611 if (lra_dump_file != NULL)
4612 fprintf (lra_dump_file,
4613 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
4614 return true;
4615 }
4616
4617 /* Return true if we need a caller save/restore for pseudo REGNO which
4618 was assigned to a hard register. */
4619 static inline bool
4620 need_for_call_save_p (int regno)
4621 {
4622 lra_assert (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] >= 0);
4623 return (usage_insns[regno].calls_num < calls_num
4624 && (overlaps_hard_reg_set_p
4625 (call_used_reg_set,
4626 PSEUDO_REGNO_MODE (regno), reg_renumber[regno])
4627 || HARD_REGNO_CALL_PART_CLOBBERED (reg_renumber[regno],
4628 PSEUDO_REGNO_MODE (regno))));
4629 }
4630
4631 /* Global registers occurring in the current EBB. */
4632 static bitmap_head ebb_global_regs;
4633
4634 /* Return true if we need a split for hard register REGNO or pseudo
4635 REGNO which was assigned to a hard register.
4636 POTENTIAL_RELOAD_HARD_REGS contains hard registers which might be
4637 used for reloads since the EBB end. It is an approximation of the
4638 used hard registers in the split range. The exact value would
4639 require expensive calculations. If we were aggressive with
4640 splitting because of the approximation, the split pseudo will save
4641 the same hard register assignment and will be removed in the undo
4642 pass. We still need the approximation because too aggressive
4643 splitting would result in too inaccurate cost calculation in the
4644 assignment pass because of too many generated moves which will be
4645 probably removed in the undo pass. */
4646 static inline bool
4647 need_for_split_p (HARD_REG_SET potential_reload_hard_regs, int regno)
4648 {
4649 int hard_regno = regno < FIRST_PSEUDO_REGISTER ? regno : reg_renumber[regno];
4650
4651 lra_assert (hard_regno >= 0);
4652 return ((TEST_HARD_REG_BIT (potential_reload_hard_regs, hard_regno)
4653 /* Don't split eliminable hard registers, otherwise we can
4654 split hard registers like hard frame pointer, which
4655 lives on BB start/end according to DF-infrastructure,
4656 when there is a pseudo assigned to the register and
4657 living in the same BB. */
4658 && (regno >= FIRST_PSEUDO_REGISTER
4659 || ! TEST_HARD_REG_BIT (eliminable_regset, hard_regno))
4660 && ! TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno)
4661 /* Don't split call clobbered hard regs living through
4662 calls, otherwise we might have a check problem in the
4663 assign sub-pass as in the most cases (exception is a
4664 situation when lra_risky_transformations_p value is
4665 true) the assign pass assumes that all pseudos living
4666 through calls are assigned to call saved hard regs. */
4667 && (regno >= FIRST_PSEUDO_REGISTER
4668 || ! TEST_HARD_REG_BIT (call_used_reg_set, regno)
4669 || usage_insns[regno].calls_num == calls_num)
4670 /* We need at least 2 reloads to make pseudo splitting
4671 profitable. We should provide hard regno splitting in
4672 any case to solve 1st insn scheduling problem when
4673 moving hard register definition up might result in
4674 impossibility to find hard register for reload pseudo of
4675 small register class. */
4676 && (usage_insns[regno].reloads_num
4677 + (regno < FIRST_PSEUDO_REGISTER ? 0 : 3) < reloads_num)
4678 && (regno < FIRST_PSEUDO_REGISTER
4679 /* For short living pseudos, spilling + inheritance can
4680 be considered a substitution for splitting.
4681 Therefore we do not splitting for local pseudos. It
4682 decreases also aggressiveness of splitting. The
4683 minimal number of references is chosen taking into
4684 account that for 2 references splitting has no sense
4685 as we can just spill the pseudo. */
4686 || (regno >= FIRST_PSEUDO_REGISTER
4687 && lra_reg_info[regno].nrefs > 3
4688 && bitmap_bit_p (&ebb_global_regs, regno))))
4689 || (regno >= FIRST_PSEUDO_REGISTER && need_for_call_save_p (regno)));
4690 }
4691
4692 /* Return class for the split pseudo created from original pseudo with
4693 ALLOCNO_CLASS and MODE which got a hard register HARD_REGNO. We
4694 choose subclass of ALLOCNO_CLASS which contains HARD_REGNO and
4695 results in no secondary memory movements. */
4696 static enum reg_class
4697 choose_split_class (enum reg_class allocno_class,
4698 int hard_regno ATTRIBUTE_UNUSED,
4699 enum machine_mode mode ATTRIBUTE_UNUSED)
4700 {
4701 #ifndef SECONDARY_MEMORY_NEEDED
4702 return allocno_class;
4703 #else
4704 int i;
4705 enum reg_class cl, best_cl = NO_REGS;
4706 enum reg_class hard_reg_class ATTRIBUTE_UNUSED
4707 = REGNO_REG_CLASS (hard_regno);
4708
4709 if (! SECONDARY_MEMORY_NEEDED (allocno_class, allocno_class, mode)
4710 && TEST_HARD_REG_BIT (reg_class_contents[allocno_class], hard_regno))
4711 return allocno_class;
4712 for (i = 0;
4713 (cl = reg_class_subclasses[allocno_class][i]) != LIM_REG_CLASSES;
4714 i++)
4715 if (! SECONDARY_MEMORY_NEEDED (cl, hard_reg_class, mode)
4716 && ! SECONDARY_MEMORY_NEEDED (hard_reg_class, cl, mode)
4717 && TEST_HARD_REG_BIT (reg_class_contents[cl], hard_regno)
4718 && (best_cl == NO_REGS
4719 || ira_class_hard_regs_num[best_cl] < ira_class_hard_regs_num[cl]))
4720 best_cl = cl;
4721 return best_cl;
4722 #endif
4723 }
4724
4725 /* Do split transformations for insn INSN, which defines or uses
4726 ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which instruction in
4727 the EBB next uses ORIGINAL_REGNO; it has the same form as the
4728 "insns" field of usage_insns.
4729
4730 The transformations look like:
4731
4732 p <- ... p <- ...
4733 ... s <- p (new insn -- save)
4734 ... =>
4735 ... p <- s (new insn -- restore)
4736 <- ... p ... <- ... p ...
4737 or
4738 <- ... p ... <- ... p ...
4739 ... s <- p (new insn -- save)
4740 ... =>
4741 ... p <- s (new insn -- restore)
4742 <- ... p ... <- ... p ...
4743
4744 where p is an original pseudo got a hard register or a hard
4745 register and s is a new split pseudo. The save is put before INSN
4746 if BEFORE_P is true. Return true if we succeed in such
4747 transformation. */
4748 static bool
4749 split_reg (bool before_p, int original_regno, rtx insn, rtx next_usage_insns)
4750 {
4751 enum reg_class rclass;
4752 rtx original_reg;
4753 int hard_regno, nregs;
4754 rtx new_reg, save, restore, usage_insn;
4755 bool after_p;
4756 bool call_save_p;
4757
4758 if (original_regno < FIRST_PSEUDO_REGISTER)
4759 {
4760 rclass = ira_allocno_class_translate[REGNO_REG_CLASS (original_regno)];
4761 hard_regno = original_regno;
4762 call_save_p = false;
4763 nregs = 1;
4764 }
4765 else
4766 {
4767 hard_regno = reg_renumber[original_regno];
4768 nregs = hard_regno_nregs[hard_regno][PSEUDO_REGNO_MODE (original_regno)];
4769 rclass = lra_get_allocno_class (original_regno);
4770 original_reg = regno_reg_rtx[original_regno];
4771 call_save_p = need_for_call_save_p (original_regno);
4772 }
4773 original_reg = regno_reg_rtx[original_regno];
4774 lra_assert (hard_regno >= 0);
4775 if (lra_dump_file != NULL)
4776 fprintf (lra_dump_file,
4777 " ((((((((((((((((((((((((((((((((((((((((((((((((\n");
4778 if (call_save_p)
4779 {
4780 enum machine_mode mode = GET_MODE (original_reg);
4781
4782 mode = HARD_REGNO_CALLER_SAVE_MODE (hard_regno,
4783 hard_regno_nregs[hard_regno][mode],
4784 mode);
4785 new_reg = lra_create_new_reg (mode, NULL_RTX, NO_REGS, "save");
4786 }
4787 else
4788 {
4789 rclass = choose_split_class (rclass, hard_regno,
4790 GET_MODE (original_reg));
4791 if (rclass == NO_REGS)
4792 {
4793 if (lra_dump_file != NULL)
4794 {
4795 fprintf (lra_dump_file,
4796 " Rejecting split of %d(%s): "
4797 "no good reg class for %d(%s)\n",
4798 original_regno,
4799 reg_class_names[lra_get_allocno_class (original_regno)],
4800 hard_regno,
4801 reg_class_names[REGNO_REG_CLASS (hard_regno)]);
4802 fprintf
4803 (lra_dump_file,
4804 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
4805 }
4806 return false;
4807 }
4808 new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg,
4809 rclass, "split");
4810 reg_renumber[REGNO (new_reg)] = hard_regno;
4811 }
4812 save = emit_spill_move (true, new_reg, original_reg);
4813 if (NEXT_INSN (save) != NULL_RTX)
4814 {
4815 lra_assert (! call_save_p);
4816 if (lra_dump_file != NULL)
4817 {
4818 fprintf
4819 (lra_dump_file,
4820 " Rejecting split %d->%d resulting in > 2 %s save insns:\n",
4821 original_regno, REGNO (new_reg), call_save_p ? "call" : "");
4822 dump_rtl_slim (lra_dump_file, save, NULL_RTX, -1, 0);
4823 fprintf (lra_dump_file,
4824 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
4825 }
4826 return false;
4827 }
4828 restore = emit_spill_move (false, new_reg, original_reg);
4829 if (NEXT_INSN (restore) != NULL_RTX)
4830 {
4831 lra_assert (! call_save_p);
4832 if (lra_dump_file != NULL)
4833 {
4834 fprintf (lra_dump_file,
4835 " Rejecting split %d->%d "
4836 "resulting in > 2 %s restore insns:\n",
4837 original_regno, REGNO (new_reg), call_save_p ? "call" : "");
4838 dump_rtl_slim (lra_dump_file, restore, NULL_RTX, -1, 0);
4839 fprintf (lra_dump_file,
4840 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
4841 }
4842 return false;
4843 }
4844 after_p = usage_insns[original_regno].after_p;
4845 lra_reg_info[REGNO (new_reg)].restore_regno = original_regno;
4846 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
4847 bitmap_set_bit (&check_only_regs, original_regno);
4848 bitmap_set_bit (&lra_split_regs, REGNO (new_reg));
4849 for (;;)
4850 {
4851 if (GET_CODE (next_usage_insns) != INSN_LIST)
4852 {
4853 usage_insn = next_usage_insns;
4854 break;
4855 }
4856 usage_insn = XEXP (next_usage_insns, 0);
4857 lra_assert (DEBUG_INSN_P (usage_insn));
4858 next_usage_insns = XEXP (next_usage_insns, 1);
4859 substitute_pseudo (&usage_insn, original_regno, new_reg);
4860 lra_update_insn_regno_info (usage_insn);
4861 if (lra_dump_file != NULL)
4862 {
4863 fprintf (lra_dump_file, " Split reuse change %d->%d:\n",
4864 original_regno, REGNO (new_reg));
4865 dump_insn_slim (lra_dump_file, usage_insn);
4866 }
4867 }
4868 lra_assert (NOTE_P (usage_insn) || NONDEBUG_INSN_P (usage_insn));
4869 lra_assert (usage_insn != insn || (after_p && before_p));
4870 lra_process_new_insns (usage_insn, after_p ? NULL_RTX : restore,
4871 after_p ? restore : NULL_RTX,
4872 call_save_p
4873 ? "Add reg<-save" : "Add reg<-split");
4874 lra_process_new_insns (insn, before_p ? save : NULL_RTX,
4875 before_p ? NULL_RTX : save,
4876 call_save_p
4877 ? "Add save<-reg" : "Add split<-reg");
4878 if (nregs > 1)
4879 /* If we are trying to split multi-register. We should check
4880 conflicts on the next assignment sub-pass. IRA can allocate on
4881 sub-register levels, LRA do this on pseudos level right now and
4882 this discrepancy may create allocation conflicts after
4883 splitting. */
4884 lra_risky_transformations_p = true;
4885 if (lra_dump_file != NULL)
4886 fprintf (lra_dump_file,
4887 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
4888 return true;
4889 }
4890
4891 /* Recognize that we need a split transformation for insn INSN, which
4892 defines or uses REGNO in its insn biggest MODE (we use it only if
4893 REGNO is a hard register). POTENTIAL_RELOAD_HARD_REGS contains
4894 hard registers which might be used for reloads since the EBB end.
4895 Put the save before INSN if BEFORE_P is true. MAX_UID is maximla
4896 uid before starting INSN processing. Return true if we succeed in
4897 such transformation. */
4898 static bool
4899 split_if_necessary (int regno, enum machine_mode mode,
4900 HARD_REG_SET potential_reload_hard_regs,
4901 bool before_p, rtx insn, int max_uid)
4902 {
4903 bool res = false;
4904 int i, nregs = 1;
4905 rtx next_usage_insns;
4906
4907 if (regno < FIRST_PSEUDO_REGISTER)
4908 nregs = hard_regno_nregs[regno][mode];
4909 for (i = 0; i < nregs; i++)
4910 if (usage_insns[regno + i].check == curr_usage_insns_check
4911 && (next_usage_insns = usage_insns[regno + i].insns) != NULL_RTX
4912 /* To avoid processing the register twice or more. */
4913 && ((GET_CODE (next_usage_insns) != INSN_LIST
4914 && INSN_UID (next_usage_insns) < max_uid)
4915 || (GET_CODE (next_usage_insns) == INSN_LIST
4916 && (INSN_UID (XEXP (next_usage_insns, 0)) < max_uid)))
4917 && need_for_split_p (potential_reload_hard_regs, regno + i)
4918 && split_reg (before_p, regno + i, insn, next_usage_insns))
4919 res = true;
4920 return res;
4921 }
4922
4923 /* Check only registers living at the current program point in the
4924 current EBB. */
4925 static bitmap_head live_regs;
4926
4927 /* Update live info in EBB given by its HEAD and TAIL insns after
4928 inheritance/split transformation. The function removes dead moves
4929 too. */
4930 static void
4931 update_ebb_live_info (rtx head, rtx tail)
4932 {
4933 unsigned int j;
4934 int i, regno;
4935 bool live_p;
4936 rtx prev_insn, set;
4937 bool remove_p;
4938 basic_block last_bb, prev_bb, curr_bb;
4939 bitmap_iterator bi;
4940 struct lra_insn_reg *reg;
4941 edge e;
4942 edge_iterator ei;
4943
4944 last_bb = BLOCK_FOR_INSN (tail);
4945 prev_bb = NULL;
4946 for (curr_insn = tail;
4947 curr_insn != PREV_INSN (head);
4948 curr_insn = prev_insn)
4949 {
4950 prev_insn = PREV_INSN (curr_insn);
4951 /* We need to process empty blocks too. They contain
4952 NOTE_INSN_BASIC_BLOCK referring for the basic block. */
4953 if (NOTE_P (curr_insn) && NOTE_KIND (curr_insn) != NOTE_INSN_BASIC_BLOCK)
4954 continue;
4955 curr_bb = BLOCK_FOR_INSN (curr_insn);
4956 if (curr_bb != prev_bb)
4957 {
4958 if (prev_bb != NULL)
4959 {
4960 /* Update df_get_live_in (prev_bb): */
4961 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
4962 if (bitmap_bit_p (&live_regs, j))
4963 bitmap_set_bit (df_get_live_in (prev_bb), j);
4964 else
4965 bitmap_clear_bit (df_get_live_in (prev_bb), j);
4966 }
4967 if (curr_bb != last_bb)
4968 {
4969 /* Update df_get_live_out (curr_bb): */
4970 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
4971 {
4972 live_p = bitmap_bit_p (&live_regs, j);
4973 if (! live_p)
4974 FOR_EACH_EDGE (e, ei, curr_bb->succs)
4975 if (bitmap_bit_p (df_get_live_in (e->dest), j))
4976 {
4977 live_p = true;
4978 break;
4979 }
4980 if (live_p)
4981 bitmap_set_bit (df_get_live_out (curr_bb), j);
4982 else
4983 bitmap_clear_bit (df_get_live_out (curr_bb), j);
4984 }
4985 }
4986 prev_bb = curr_bb;
4987 bitmap_and (&live_regs, &check_only_regs, df_get_live_out (curr_bb));
4988 }
4989 if (! NONDEBUG_INSN_P (curr_insn))
4990 continue;
4991 curr_id = lra_get_insn_recog_data (curr_insn);
4992 curr_static_id = curr_id->insn_static_data;
4993 remove_p = false;
4994 if ((set = single_set (curr_insn)) != NULL_RTX && REG_P (SET_DEST (set))
4995 && (regno = REGNO (SET_DEST (set))) >= FIRST_PSEUDO_REGISTER
4996 && bitmap_bit_p (&check_only_regs, regno)
4997 && ! bitmap_bit_p (&live_regs, regno))
4998 remove_p = true;
4999 /* See which defined values die here. */
5000 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5001 if (reg->type == OP_OUT && ! reg->subreg_p)
5002 bitmap_clear_bit (&live_regs, reg->regno);
5003 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5004 if (reg->type == OP_OUT && ! reg->subreg_p)
5005 bitmap_clear_bit (&live_regs, reg->regno);
5006 /* Mark each used value as live. */
5007 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5008 if (reg->type != OP_OUT
5009 && bitmap_bit_p (&check_only_regs, reg->regno))
5010 bitmap_set_bit (&live_regs, reg->regno);
5011 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5012 if (reg->type != OP_OUT
5013 && bitmap_bit_p (&check_only_regs, reg->regno))
5014 bitmap_set_bit (&live_regs, reg->regno);
5015 if (curr_id->arg_hard_regs != NULL)
5016 /* Make argument hard registers live. */
5017 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5018 if (bitmap_bit_p (&check_only_regs, regno))
5019 bitmap_set_bit (&live_regs, regno);
5020 /* It is quite important to remove dead move insns because it
5021 means removing dead store. We don't need to process them for
5022 constraints. */
5023 if (remove_p)
5024 {
5025 if (lra_dump_file != NULL)
5026 {
5027 fprintf (lra_dump_file, " Removing dead insn:\n ");
5028 dump_insn_slim (lra_dump_file, curr_insn);
5029 }
5030 lra_set_insn_deleted (curr_insn);
5031 }
5032 }
5033 }
5034
5035 /* The structure describes info to do an inheritance for the current
5036 insn. We need to collect such info first before doing the
5037 transformations because the transformations change the insn
5038 internal representation. */
5039 struct to_inherit
5040 {
5041 /* Original regno. */
5042 int regno;
5043 /* Subsequent insns which can inherit original reg value. */
5044 rtx insns;
5045 };
5046
5047 /* Array containing all info for doing inheritance from the current
5048 insn. */
5049 static struct to_inherit to_inherit[LRA_MAX_INSN_RELOADS];
5050
5051 /* Number elements in the previous array. */
5052 static int to_inherit_num;
5053
5054 /* Add inheritance info REGNO and INSNS. Their meaning is described in
5055 structure to_inherit. */
5056 static void
5057 add_to_inherit (int regno, rtx insns)
5058 {
5059 int i;
5060
5061 for (i = 0; i < to_inherit_num; i++)
5062 if (to_inherit[i].regno == regno)
5063 return;
5064 lra_assert (to_inherit_num < LRA_MAX_INSN_RELOADS);
5065 to_inherit[to_inherit_num].regno = regno;
5066 to_inherit[to_inherit_num++].insns = insns;
5067 }
5068
5069 /* Return the last non-debug insn in basic block BB, or the block begin
5070 note if none. */
5071 static rtx
5072 get_last_insertion_point (basic_block bb)
5073 {
5074 rtx insn;
5075
5076 FOR_BB_INSNS_REVERSE (bb, insn)
5077 if (NONDEBUG_INSN_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn))
5078 return insn;
5079 gcc_unreachable ();
5080 }
5081
5082 /* Set up RES by registers living on edges FROM except the edge (FROM,
5083 TO) or by registers set up in a jump insn in BB FROM. */
5084 static void
5085 get_live_on_other_edges (basic_block from, basic_block to, bitmap res)
5086 {
5087 rtx last;
5088 struct lra_insn_reg *reg;
5089 edge e;
5090 edge_iterator ei;
5091
5092 lra_assert (to != NULL);
5093 bitmap_clear (res);
5094 FOR_EACH_EDGE (e, ei, from->succs)
5095 if (e->dest != to)
5096 bitmap_ior_into (res, df_get_live_in (e->dest));
5097 last = get_last_insertion_point (from);
5098 if (! JUMP_P (last))
5099 return;
5100 curr_id = lra_get_insn_recog_data (last);
5101 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5102 if (reg->type != OP_IN)
5103 bitmap_set_bit (res, reg->regno);
5104 }
5105
5106 /* Used as a temporary results of some bitmap calculations. */
5107 static bitmap_head temp_bitmap;
5108
5109 /* We split for reloads of small class of hard regs. The following
5110 defines how many hard regs the class should have to be qualified as
5111 small. The code is mostly oriented to x86/x86-64 architecture
5112 where some insns need to use only specific register or pair of
5113 registers and these register can live in RTL explicitly, e.g. for
5114 parameter passing. */
5115 static const int max_small_class_regs_num = 2;
5116
5117 /* Do inheritance/split transformations in EBB starting with HEAD and
5118 finishing on TAIL. We process EBB insns in the reverse order.
5119 Return true if we did any inheritance/split transformation in the
5120 EBB.
5121
5122 We should avoid excessive splitting which results in worse code
5123 because of inaccurate cost calculations for spilling new split
5124 pseudos in such case. To achieve this we do splitting only if
5125 register pressure is high in given basic block and there are reload
5126 pseudos requiring hard registers. We could do more register
5127 pressure calculations at any given program point to avoid necessary
5128 splitting even more but it is to expensive and the current approach
5129 works well enough. */
5130 static bool
5131 inherit_in_ebb (rtx head, rtx tail)
5132 {
5133 int i, src_regno, dst_regno, nregs;
5134 bool change_p, succ_p, update_reloads_num_p;
5135 rtx prev_insn, next_usage_insns, set, last_insn;
5136 enum reg_class cl;
5137 struct lra_insn_reg *reg;
5138 basic_block last_processed_bb, curr_bb = NULL;
5139 HARD_REG_SET potential_reload_hard_regs, live_hard_regs;
5140 bitmap to_process;
5141 unsigned int j;
5142 bitmap_iterator bi;
5143 bool head_p, after_p;
5144
5145 change_p = false;
5146 curr_usage_insns_check++;
5147 reloads_num = calls_num = 0;
5148 bitmap_clear (&check_only_regs);
5149 last_processed_bb = NULL;
5150 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
5151 COPY_HARD_REG_SET (live_hard_regs, eliminable_regset);
5152 IOR_HARD_REG_SET (live_hard_regs, lra_no_alloc_regs);
5153 /* We don't process new insns generated in the loop. */
5154 for (curr_insn = tail; curr_insn != PREV_INSN (head); curr_insn = prev_insn)
5155 {
5156 prev_insn = PREV_INSN (curr_insn);
5157 if (BLOCK_FOR_INSN (curr_insn) != NULL)
5158 curr_bb = BLOCK_FOR_INSN (curr_insn);
5159 if (last_processed_bb != curr_bb)
5160 {
5161 /* We are at the end of BB. Add qualified living
5162 pseudos for potential splitting. */
5163 to_process = df_get_live_out (curr_bb);
5164 if (last_processed_bb != NULL)
5165 {
5166 /* We are somewhere in the middle of EBB. */
5167 get_live_on_other_edges (curr_bb, last_processed_bb,
5168 &temp_bitmap);
5169 to_process = &temp_bitmap;
5170 }
5171 last_processed_bb = curr_bb;
5172 last_insn = get_last_insertion_point (curr_bb);
5173 after_p = (! JUMP_P (last_insn)
5174 && (! CALL_P (last_insn)
5175 || (find_reg_note (last_insn,
5176 REG_NORETURN, NULL_RTX) == NULL_RTX
5177 && ! SIBLING_CALL_P (last_insn))));
5178 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
5179 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
5180 {
5181 if ((int) j >= lra_constraint_new_regno_start)
5182 break;
5183 if (j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
5184 {
5185 if (j < FIRST_PSEUDO_REGISTER)
5186 SET_HARD_REG_BIT (live_hard_regs, j);
5187 else
5188 add_to_hard_reg_set (&live_hard_regs,
5189 PSEUDO_REGNO_MODE (j),
5190 reg_renumber[j]);
5191 setup_next_usage_insn (j, last_insn, reloads_num, after_p);
5192 }
5193 }
5194 }
5195 src_regno = dst_regno = -1;
5196 if (NONDEBUG_INSN_P (curr_insn)
5197 && (set = single_set (curr_insn)) != NULL_RTX
5198 && REG_P (SET_DEST (set)) && REG_P (SET_SRC (set)))
5199 {
5200 src_regno = REGNO (SET_SRC (set));
5201 dst_regno = REGNO (SET_DEST (set));
5202 }
5203 update_reloads_num_p = true;
5204 if (src_regno < lra_constraint_new_regno_start
5205 && src_regno >= FIRST_PSEUDO_REGISTER
5206 && reg_renumber[src_regno] < 0
5207 && dst_regno >= lra_constraint_new_regno_start
5208 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS)
5209 {
5210 /* 'reload_pseudo <- original_pseudo'. */
5211 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
5212 reloads_num++;
5213 update_reloads_num_p = false;
5214 succ_p = false;
5215 if (usage_insns[src_regno].check == curr_usage_insns_check
5216 && (next_usage_insns = usage_insns[src_regno].insns) != NULL_RTX)
5217 succ_p = inherit_reload_reg (false, src_regno, cl,
5218 curr_insn, next_usage_insns);
5219 if (succ_p)
5220 change_p = true;
5221 else
5222 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
5223 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
5224 IOR_HARD_REG_SET (potential_reload_hard_regs,
5225 reg_class_contents[cl]);
5226 }
5227 else if (src_regno >= lra_constraint_new_regno_start
5228 && dst_regno < lra_constraint_new_regno_start
5229 && dst_regno >= FIRST_PSEUDO_REGISTER
5230 && reg_renumber[dst_regno] < 0
5231 && (cl = lra_get_allocno_class (src_regno)) != NO_REGS
5232 && usage_insns[dst_regno].check == curr_usage_insns_check
5233 && (next_usage_insns
5234 = usage_insns[dst_regno].insns) != NULL_RTX)
5235 {
5236 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
5237 reloads_num++;
5238 update_reloads_num_p = false;
5239 /* 'original_pseudo <- reload_pseudo'. */
5240 if (! JUMP_P (curr_insn)
5241 && inherit_reload_reg (true, dst_regno, cl,
5242 curr_insn, next_usage_insns))
5243 change_p = true;
5244 /* Invalidate. */
5245 usage_insns[dst_regno].check = 0;
5246 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
5247 IOR_HARD_REG_SET (potential_reload_hard_regs,
5248 reg_class_contents[cl]);
5249 }
5250 else if (INSN_P (curr_insn))
5251 {
5252 int iter;
5253 int max_uid = get_max_uid ();
5254
5255 curr_id = lra_get_insn_recog_data (curr_insn);
5256 curr_static_id = curr_id->insn_static_data;
5257 to_inherit_num = 0;
5258 /* Process insn definitions. */
5259 for (iter = 0; iter < 2; iter++)
5260 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
5261 reg != NULL;
5262 reg = reg->next)
5263 if (reg->type != OP_IN
5264 && (dst_regno = reg->regno) < lra_constraint_new_regno_start)
5265 {
5266 if (dst_regno >= FIRST_PSEUDO_REGISTER && reg->type == OP_OUT
5267 && reg_renumber[dst_regno] < 0 && ! reg->subreg_p
5268 && usage_insns[dst_regno].check == curr_usage_insns_check
5269 && (next_usage_insns
5270 = usage_insns[dst_regno].insns) != NULL_RTX)
5271 {
5272 struct lra_insn_reg *r;
5273
5274 for (r = curr_id->regs; r != NULL; r = r->next)
5275 if (r->type != OP_OUT && r->regno == dst_regno)
5276 break;
5277 /* Don't do inheritance if the pseudo is also
5278 used in the insn. */
5279 if (r == NULL)
5280 /* We can not do inheritance right now
5281 because the current insn reg info (chain
5282 regs) can change after that. */
5283 add_to_inherit (dst_regno, next_usage_insns);
5284 }
5285 /* We can not process one reg twice here because of
5286 usage_insns invalidation. */
5287 if ((dst_regno < FIRST_PSEUDO_REGISTER
5288 || reg_renumber[dst_regno] >= 0)
5289 && ! reg->subreg_p && reg->type != OP_IN)
5290 {
5291 HARD_REG_SET s;
5292
5293 if (split_if_necessary (dst_regno, reg->biggest_mode,
5294 potential_reload_hard_regs,
5295 false, curr_insn, max_uid))
5296 change_p = true;
5297 CLEAR_HARD_REG_SET (s);
5298 if (dst_regno < FIRST_PSEUDO_REGISTER)
5299 add_to_hard_reg_set (&s, reg->biggest_mode, dst_regno);
5300 else
5301 add_to_hard_reg_set (&s, PSEUDO_REGNO_MODE (dst_regno),
5302 reg_renumber[dst_regno]);
5303 AND_COMPL_HARD_REG_SET (live_hard_regs, s);
5304 }
5305 /* We should invalidate potential inheritance or
5306 splitting for the current insn usages to the next
5307 usage insns (see code below) as the output pseudo
5308 prevents this. */
5309 if ((dst_regno >= FIRST_PSEUDO_REGISTER
5310 && reg_renumber[dst_regno] < 0)
5311 || (reg->type == OP_OUT && ! reg->subreg_p
5312 && (dst_regno < FIRST_PSEUDO_REGISTER
5313 || reg_renumber[dst_regno] >= 0)))
5314 {
5315 /* Invalidate and mark definitions. */
5316 if (dst_regno >= FIRST_PSEUDO_REGISTER)
5317 usage_insns[dst_regno].check = -(int) INSN_UID (curr_insn);
5318 else
5319 {
5320 nregs = hard_regno_nregs[dst_regno][reg->biggest_mode];
5321 for (i = 0; i < nregs; i++)
5322 usage_insns[dst_regno + i].check
5323 = -(int) INSN_UID (curr_insn);
5324 }
5325 }
5326 }
5327 if (! JUMP_P (curr_insn))
5328 for (i = 0; i < to_inherit_num; i++)
5329 if (inherit_reload_reg (true, to_inherit[i].regno,
5330 ALL_REGS, curr_insn,
5331 to_inherit[i].insns))
5332 change_p = true;
5333 if (CALL_P (curr_insn))
5334 {
5335 rtx cheap, pat, dest, restore;
5336 int regno, hard_regno;
5337
5338 calls_num++;
5339 if ((cheap = find_reg_note (curr_insn,
5340 REG_RETURNED, NULL_RTX)) != NULL_RTX
5341 && ((cheap = XEXP (cheap, 0)), true)
5342 && (regno = REGNO (cheap)) >= FIRST_PSEUDO_REGISTER
5343 && (hard_regno = reg_renumber[regno]) >= 0
5344 /* If there are pending saves/restores, the
5345 optimization is not worth. */
5346 && usage_insns[regno].calls_num == calls_num - 1
5347 && TEST_HARD_REG_BIT (call_used_reg_set, hard_regno))
5348 {
5349 /* Restore the pseudo from the call result as
5350 REG_RETURNED note says that the pseudo value is
5351 in the call result and the pseudo is an argument
5352 of the call. */
5353 pat = PATTERN (curr_insn);
5354 if (GET_CODE (pat) == PARALLEL)
5355 pat = XVECEXP (pat, 0, 0);
5356 dest = SET_DEST (pat);
5357 start_sequence ();
5358 emit_move_insn (cheap, copy_rtx (dest));
5359 restore = get_insns ();
5360 end_sequence ();
5361 lra_process_new_insns (curr_insn, NULL, restore,
5362 "Inserting call parameter restore");
5363 /* We don't need to save/restore of the pseudo from
5364 this call. */
5365 usage_insns[regno].calls_num = calls_num;
5366 bitmap_set_bit (&check_only_regs, regno);
5367 }
5368 }
5369 to_inherit_num = 0;
5370 /* Process insn usages. */
5371 for (iter = 0; iter < 2; iter++)
5372 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
5373 reg != NULL;
5374 reg = reg->next)
5375 if ((reg->type != OP_OUT
5376 || (reg->type == OP_OUT && reg->subreg_p))
5377 && (src_regno = reg->regno) < lra_constraint_new_regno_start)
5378 {
5379 if (src_regno >= FIRST_PSEUDO_REGISTER
5380 && reg_renumber[src_regno] < 0 && reg->type == OP_IN)
5381 {
5382 if (usage_insns[src_regno].check == curr_usage_insns_check
5383 && (next_usage_insns
5384 = usage_insns[src_regno].insns) != NULL_RTX
5385 && NONDEBUG_INSN_P (curr_insn))
5386 add_to_inherit (src_regno, next_usage_insns);
5387 else if (usage_insns[src_regno].check
5388 != -(int) INSN_UID (curr_insn))
5389 /* Add usages but only if the reg is not set up
5390 in the same insn. */
5391 add_next_usage_insn (src_regno, curr_insn, reloads_num);
5392 }
5393 else if (src_regno < FIRST_PSEUDO_REGISTER
5394 || reg_renumber[src_regno] >= 0)
5395 {
5396 bool before_p;
5397 rtx use_insn = curr_insn;
5398
5399 before_p = (JUMP_P (curr_insn)
5400 || (CALL_P (curr_insn) && reg->type == OP_IN));
5401 if (NONDEBUG_INSN_P (curr_insn)
5402 && (! JUMP_P (curr_insn) || reg->type == OP_IN)
5403 && split_if_necessary (src_regno, reg->biggest_mode,
5404 potential_reload_hard_regs,
5405 before_p, curr_insn, max_uid))
5406 {
5407 if (reg->subreg_p)
5408 lra_risky_transformations_p = true;
5409 change_p = true;
5410 /* Invalidate. */
5411 usage_insns[src_regno].check = 0;
5412 if (before_p)
5413 use_insn = PREV_INSN (curr_insn);
5414 }
5415 if (NONDEBUG_INSN_P (curr_insn))
5416 {
5417 if (src_regno < FIRST_PSEUDO_REGISTER)
5418 add_to_hard_reg_set (&live_hard_regs,
5419 reg->biggest_mode, src_regno);
5420 else
5421 add_to_hard_reg_set (&live_hard_regs,
5422 PSEUDO_REGNO_MODE (src_regno),
5423 reg_renumber[src_regno]);
5424 }
5425 add_next_usage_insn (src_regno, use_insn, reloads_num);
5426 }
5427 }
5428 /* Process call args. */
5429 if (curr_id->arg_hard_regs != NULL)
5430 for (i = 0; (src_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5431 if (src_regno < FIRST_PSEUDO_REGISTER)
5432 {
5433 SET_HARD_REG_BIT (live_hard_regs, src_regno);
5434 add_next_usage_insn (src_regno, curr_insn, reloads_num);
5435 }
5436 for (i = 0; i < to_inherit_num; i++)
5437 {
5438 src_regno = to_inherit[i].regno;
5439 if (inherit_reload_reg (false, src_regno, ALL_REGS,
5440 curr_insn, to_inherit[i].insns))
5441 change_p = true;
5442 else
5443 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
5444 }
5445 }
5446 if (update_reloads_num_p
5447 && NONDEBUG_INSN_P (curr_insn)
5448 && (set = single_set (curr_insn)) != NULL_RTX)
5449 {
5450 int regno = -1;
5451 if ((REG_P (SET_DEST (set))
5452 && (regno = REGNO (SET_DEST (set))) >= lra_constraint_new_regno_start
5453 && reg_renumber[regno] < 0
5454 && (cl = lra_get_allocno_class (regno)) != NO_REGS)
5455 || (REG_P (SET_SRC (set))
5456 && (regno = REGNO (SET_SRC (set))) >= lra_constraint_new_regno_start
5457 && reg_renumber[regno] < 0
5458 && (cl = lra_get_allocno_class (regno)) != NO_REGS))
5459 {
5460 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
5461 reloads_num++;
5462 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
5463 IOR_HARD_REG_SET (potential_reload_hard_regs,
5464 reg_class_contents[cl]);
5465 }
5466 }
5467 /* We reached the start of the current basic block. */
5468 if (prev_insn == NULL_RTX || prev_insn == PREV_INSN (head)
5469 || BLOCK_FOR_INSN (prev_insn) != curr_bb)
5470 {
5471 /* We reached the beginning of the current block -- do
5472 rest of spliting in the current BB. */
5473 to_process = df_get_live_in (curr_bb);
5474 if (BLOCK_FOR_INSN (head) != curr_bb)
5475 {
5476 /* We are somewhere in the middle of EBB. */
5477 get_live_on_other_edges (EDGE_PRED (curr_bb, 0)->src,
5478 curr_bb, &temp_bitmap);
5479 to_process = &temp_bitmap;
5480 }
5481 head_p = true;
5482 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
5483 {
5484 if ((int) j >= lra_constraint_new_regno_start)
5485 break;
5486 if (((int) j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
5487 && usage_insns[j].check == curr_usage_insns_check
5488 && (next_usage_insns = usage_insns[j].insns) != NULL_RTX)
5489 {
5490 if (need_for_split_p (potential_reload_hard_regs, j))
5491 {
5492 if (lra_dump_file != NULL && head_p)
5493 {
5494 fprintf (lra_dump_file,
5495 " ----------------------------------\n");
5496 head_p = false;
5497 }
5498 if (split_reg (false, j, bb_note (curr_bb),
5499 next_usage_insns))
5500 change_p = true;
5501 }
5502 usage_insns[j].check = 0;
5503 }
5504 }
5505 }
5506 }
5507 return change_p;
5508 }
5509
5510 /* This value affects EBB forming. If probability of edge from EBB to
5511 a BB is not greater than the following value, we don't add the BB
5512 to EBB. */
5513 #define EBB_PROBABILITY_CUTOFF ((REG_BR_PROB_BASE * 50) / 100)
5514
5515 /* Current number of inheritance/split iteration. */
5516 int lra_inheritance_iter;
5517
5518 /* Entry function for inheritance/split pass. */
5519 void
5520 lra_inheritance (void)
5521 {
5522 int i;
5523 basic_block bb, start_bb;
5524 edge e;
5525
5526 lra_inheritance_iter++;
5527 if (lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
5528 return;
5529 timevar_push (TV_LRA_INHERITANCE);
5530 if (lra_dump_file != NULL)
5531 fprintf (lra_dump_file, "\n********** Inheritance #%d: **********\n\n",
5532 lra_inheritance_iter);
5533 curr_usage_insns_check = 0;
5534 usage_insns = XNEWVEC (struct usage_insns, lra_constraint_new_regno_start);
5535 for (i = 0; i < lra_constraint_new_regno_start; i++)
5536 usage_insns[i].check = 0;
5537 bitmap_initialize (&check_only_regs, &reg_obstack);
5538 bitmap_initialize (&live_regs, &reg_obstack);
5539 bitmap_initialize (&temp_bitmap, &reg_obstack);
5540 bitmap_initialize (&ebb_global_regs, &reg_obstack);
5541 FOR_EACH_BB_FN (bb, cfun)
5542 {
5543 start_bb = bb;
5544 if (lra_dump_file != NULL)
5545 fprintf (lra_dump_file, "EBB");
5546 /* Form a EBB starting with BB. */
5547 bitmap_clear (&ebb_global_regs);
5548 bitmap_ior_into (&ebb_global_regs, df_get_live_in (bb));
5549 for (;;)
5550 {
5551 if (lra_dump_file != NULL)
5552 fprintf (lra_dump_file, " %d", bb->index);
5553 if (bb->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
5554 || LABEL_P (BB_HEAD (bb->next_bb)))
5555 break;
5556 e = find_fallthru_edge (bb->succs);
5557 if (! e)
5558 break;
5559 if (e->probability <= EBB_PROBABILITY_CUTOFF)
5560 break;
5561 bb = bb->next_bb;
5562 }
5563 bitmap_ior_into (&ebb_global_regs, df_get_live_out (bb));
5564 if (lra_dump_file != NULL)
5565 fprintf (lra_dump_file, "\n");
5566 if (inherit_in_ebb (BB_HEAD (start_bb), BB_END (bb)))
5567 /* Remember that the EBB head and tail can change in
5568 inherit_in_ebb. */
5569 update_ebb_live_info (BB_HEAD (start_bb), BB_END (bb));
5570 }
5571 bitmap_clear (&ebb_global_regs);
5572 bitmap_clear (&temp_bitmap);
5573 bitmap_clear (&live_regs);
5574 bitmap_clear (&check_only_regs);
5575 free (usage_insns);
5576
5577 timevar_pop (TV_LRA_INHERITANCE);
5578 }
5579
5580 \f
5581
5582 /* This page contains code to undo failed inheritance/split
5583 transformations. */
5584
5585 /* Current number of iteration undoing inheritance/split. */
5586 int lra_undo_inheritance_iter;
5587
5588 /* Fix BB live info LIVE after removing pseudos created on pass doing
5589 inheritance/split which are REMOVED_PSEUDOS. */
5590 static void
5591 fix_bb_live_info (bitmap live, bitmap removed_pseudos)
5592 {
5593 unsigned int regno;
5594 bitmap_iterator bi;
5595
5596 EXECUTE_IF_SET_IN_BITMAP (removed_pseudos, 0, regno, bi)
5597 if (bitmap_clear_bit (live, regno))
5598 bitmap_set_bit (live, lra_reg_info[regno].restore_regno);
5599 }
5600
5601 /* Return regno of the (subreg of) REG. Otherwise, return a negative
5602 number. */
5603 static int
5604 get_regno (rtx reg)
5605 {
5606 if (GET_CODE (reg) == SUBREG)
5607 reg = SUBREG_REG (reg);
5608 if (REG_P (reg))
5609 return REGNO (reg);
5610 return -1;
5611 }
5612
5613 /* Remove inheritance/split pseudos which are in REMOVE_PSEUDOS and
5614 return true if we did any change. The undo transformations for
5615 inheritance looks like
5616 i <- i2
5617 p <- i => p <- i2
5618 or removing
5619 p <- i, i <- p, and i <- i3
5620 where p is original pseudo from which inheritance pseudo i was
5621 created, i and i3 are removed inheritance pseudos, i2 is another
5622 not removed inheritance pseudo. All split pseudos or other
5623 occurrences of removed inheritance pseudos are changed on the
5624 corresponding original pseudos.
5625
5626 The function also schedules insns changed and created during
5627 inheritance/split pass for processing by the subsequent constraint
5628 pass. */
5629 static bool
5630 remove_inheritance_pseudos (bitmap remove_pseudos)
5631 {
5632 basic_block bb;
5633 int regno, sregno, prev_sregno, dregno, restore_regno;
5634 rtx set, prev_set, prev_insn;
5635 bool change_p, done_p;
5636
5637 change_p = ! bitmap_empty_p (remove_pseudos);
5638 /* We can not finish the function right away if CHANGE_P is true
5639 because we need to marks insns affected by previous
5640 inheritance/split pass for processing by the subsequent
5641 constraint pass. */
5642 FOR_EACH_BB_FN (bb, cfun)
5643 {
5644 fix_bb_live_info (df_get_live_in (bb), remove_pseudos);
5645 fix_bb_live_info (df_get_live_out (bb), remove_pseudos);
5646 FOR_BB_INSNS_REVERSE (bb, curr_insn)
5647 {
5648 if (! INSN_P (curr_insn))
5649 continue;
5650 done_p = false;
5651 sregno = dregno = -1;
5652 if (change_p && NONDEBUG_INSN_P (curr_insn)
5653 && (set = single_set (curr_insn)) != NULL_RTX)
5654 {
5655 dregno = get_regno (SET_DEST (set));
5656 sregno = get_regno (SET_SRC (set));
5657 }
5658
5659 if (sregno >= 0 && dregno >= 0)
5660 {
5661 if ((bitmap_bit_p (remove_pseudos, sregno)
5662 && (lra_reg_info[sregno].restore_regno == dregno
5663 || (bitmap_bit_p (remove_pseudos, dregno)
5664 && (lra_reg_info[sregno].restore_regno
5665 == lra_reg_info[dregno].restore_regno))))
5666 || (bitmap_bit_p (remove_pseudos, dregno)
5667 && lra_reg_info[dregno].restore_regno == sregno))
5668 /* One of the following cases:
5669 original <- removed inheritance pseudo
5670 removed inherit pseudo <- another removed inherit pseudo
5671 removed inherit pseudo <- original pseudo
5672 Or
5673 removed_split_pseudo <- original_reg
5674 original_reg <- removed_split_pseudo */
5675 {
5676 if (lra_dump_file != NULL)
5677 {
5678 fprintf (lra_dump_file, " Removing %s:\n",
5679 bitmap_bit_p (&lra_split_regs, sregno)
5680 || bitmap_bit_p (&lra_split_regs, dregno)
5681 ? "split" : "inheritance");
5682 dump_insn_slim (lra_dump_file, curr_insn);
5683 }
5684 lra_set_insn_deleted (curr_insn);
5685 done_p = true;
5686 }
5687 else if (bitmap_bit_p (remove_pseudos, sregno)
5688 && bitmap_bit_p (&lra_inheritance_pseudos, sregno))
5689 {
5690 /* Search the following pattern:
5691 inherit_or_split_pseudo1 <- inherit_or_split_pseudo2
5692 original_pseudo <- inherit_or_split_pseudo1
5693 where the 2nd insn is the current insn and
5694 inherit_or_split_pseudo2 is not removed. If it is found,
5695 change the current insn onto:
5696 original_pseudo <- inherit_or_split_pseudo2. */
5697 for (prev_insn = PREV_INSN (curr_insn);
5698 prev_insn != NULL_RTX && ! NONDEBUG_INSN_P (prev_insn);
5699 prev_insn = PREV_INSN (prev_insn))
5700 ;
5701 if (prev_insn != NULL_RTX && BLOCK_FOR_INSN (prev_insn) == bb
5702 && (prev_set = single_set (prev_insn)) != NULL_RTX
5703 /* There should be no subregs in insn we are
5704 searching because only the original reg might
5705 be in subreg when we changed the mode of
5706 load/store for splitting. */
5707 && REG_P (SET_DEST (prev_set))
5708 && REG_P (SET_SRC (prev_set))
5709 && (int) REGNO (SET_DEST (prev_set)) == sregno
5710 && ((prev_sregno = REGNO (SET_SRC (prev_set)))
5711 >= FIRST_PSEUDO_REGISTER)
5712 /* As we consider chain of inheritance or
5713 splitting described in above comment we should
5714 check that sregno and prev_sregno were
5715 inheritance/split pseudos created from the
5716 same original regno. */
5717 && (lra_reg_info[sregno].restore_regno
5718 == lra_reg_info[prev_sregno].restore_regno)
5719 && ! bitmap_bit_p (remove_pseudos, prev_sregno))
5720 {
5721 lra_assert (GET_MODE (SET_SRC (prev_set))
5722 == GET_MODE (regno_reg_rtx[sregno]));
5723 if (GET_CODE (SET_SRC (set)) == SUBREG)
5724 SUBREG_REG (SET_SRC (set)) = SET_SRC (prev_set);
5725 else
5726 SET_SRC (set) = SET_SRC (prev_set);
5727 lra_push_insn_and_update_insn_regno_info (curr_insn);
5728 lra_set_used_insn_alternative_by_uid
5729 (INSN_UID (curr_insn), -1);
5730 done_p = true;
5731 if (lra_dump_file != NULL)
5732 {
5733 fprintf (lra_dump_file, " Change reload insn:\n");
5734 dump_insn_slim (lra_dump_file, curr_insn);
5735 }
5736 }
5737 }
5738 }
5739 if (! done_p)
5740 {
5741 struct lra_insn_reg *reg;
5742 bool restored_regs_p = false;
5743 bool kept_regs_p = false;
5744
5745 curr_id = lra_get_insn_recog_data (curr_insn);
5746 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5747 {
5748 regno = reg->regno;
5749 restore_regno = lra_reg_info[regno].restore_regno;
5750 if (restore_regno >= 0)
5751 {
5752 if (change_p && bitmap_bit_p (remove_pseudos, regno))
5753 {
5754 substitute_pseudo (&curr_insn, regno,
5755 regno_reg_rtx[restore_regno]);
5756 restored_regs_p = true;
5757 }
5758 else
5759 kept_regs_p = true;
5760 }
5761 }
5762 if (NONDEBUG_INSN_P (curr_insn) && kept_regs_p)
5763 {
5764 /* The instruction has changed since the previous
5765 constraints pass. */
5766 lra_push_insn_and_update_insn_regno_info (curr_insn);
5767 lra_set_used_insn_alternative_by_uid
5768 (INSN_UID (curr_insn), -1);
5769 }
5770 else if (restored_regs_p)
5771 /* The instruction has been restored to the form that
5772 it had during the previous constraints pass. */
5773 lra_update_insn_regno_info (curr_insn);
5774 if (restored_regs_p && lra_dump_file != NULL)
5775 {
5776 fprintf (lra_dump_file, " Insn after restoring regs:\n");
5777 dump_insn_slim (lra_dump_file, curr_insn);
5778 }
5779 }
5780 }
5781 }
5782 return change_p;
5783 }
5784
5785 /* If optional reload pseudos failed to get a hard register or was not
5786 inherited, it is better to remove optional reloads. We do this
5787 transformation after undoing inheritance to figure out necessity to
5788 remove optional reloads easier. Return true if we do any
5789 change. */
5790 static bool
5791 undo_optional_reloads (void)
5792 {
5793 bool change_p, keep_p;
5794 unsigned int regno, uid;
5795 bitmap_iterator bi, bi2;
5796 rtx insn, set, src, dest;
5797 bitmap_head removed_optional_reload_pseudos, insn_bitmap;
5798
5799 bitmap_initialize (&removed_optional_reload_pseudos, &reg_obstack);
5800 bitmap_copy (&removed_optional_reload_pseudos, &lra_optional_reload_pseudos);
5801 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
5802 {
5803 keep_p = false;
5804 /* Keep optional reloads from previous subpasses. */
5805 if (lra_reg_info[regno].restore_regno < 0
5806 /* If the original pseudo changed its allocation, just
5807 removing the optional pseudo is dangerous as the original
5808 pseudo will have longer live range. */
5809 || reg_renumber[lra_reg_info[regno].restore_regno] >= 0)
5810 keep_p = true;
5811 else if (reg_renumber[regno] >= 0)
5812 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi2)
5813 {
5814 insn = lra_insn_recog_data[uid]->insn;
5815 if ((set = single_set (insn)) == NULL_RTX)
5816 continue;
5817 src = SET_SRC (set);
5818 dest = SET_DEST (set);
5819 if (! REG_P (src) || ! REG_P (dest))
5820 continue;
5821 if (REGNO (dest) == regno
5822 /* Ignore insn for optional reloads itself. */
5823 && lra_reg_info[regno].restore_regno != (int) REGNO (src)
5824 /* Check only inheritance on last inheritance pass. */
5825 && (int) REGNO (src) >= new_regno_start
5826 /* Check that the optional reload was inherited. */
5827 && bitmap_bit_p (&lra_inheritance_pseudos, REGNO (src)))
5828 {
5829 keep_p = true;
5830 break;
5831 }
5832 }
5833 if (keep_p)
5834 {
5835 bitmap_clear_bit (&removed_optional_reload_pseudos, regno);
5836 if (lra_dump_file != NULL)
5837 fprintf (lra_dump_file, "Keep optional reload reg %d\n", regno);
5838 }
5839 }
5840 change_p = ! bitmap_empty_p (&removed_optional_reload_pseudos);
5841 bitmap_initialize (&insn_bitmap, &reg_obstack);
5842 EXECUTE_IF_SET_IN_BITMAP (&removed_optional_reload_pseudos, 0, regno, bi)
5843 {
5844 if (lra_dump_file != NULL)
5845 fprintf (lra_dump_file, "Remove optional reload reg %d\n", regno);
5846 bitmap_copy (&insn_bitmap, &lra_reg_info[regno].insn_bitmap);
5847 EXECUTE_IF_SET_IN_BITMAP (&insn_bitmap, 0, uid, bi2)
5848 {
5849 insn = lra_insn_recog_data[uid]->insn;
5850 if ((set = single_set (insn)) != NULL_RTX)
5851 {
5852 src = SET_SRC (set);
5853 dest = SET_DEST (set);
5854 if (REG_P (src) && REG_P (dest)
5855 && ((REGNO (src) == regno
5856 && (lra_reg_info[regno].restore_regno
5857 == (int) REGNO (dest)))
5858 || (REGNO (dest) == regno
5859 && (lra_reg_info[regno].restore_regno
5860 == (int) REGNO (src)))))
5861 {
5862 if (lra_dump_file != NULL)
5863 {
5864 fprintf (lra_dump_file, " Deleting move %u\n",
5865 INSN_UID (insn));
5866 dump_insn_slim (lra_dump_file, insn);
5867 }
5868 lra_set_insn_deleted (insn);
5869 continue;
5870 }
5871 /* We should not worry about generation memory-memory
5872 moves here as if the corresponding inheritance did
5873 not work (inheritance pseudo did not get a hard reg),
5874 we remove the inheritance pseudo and the optional
5875 reload. */
5876 }
5877 substitute_pseudo (&insn, regno,
5878 regno_reg_rtx[lra_reg_info[regno].restore_regno]);
5879 lra_update_insn_regno_info (insn);
5880 if (lra_dump_file != NULL)
5881 {
5882 fprintf (lra_dump_file,
5883 " Restoring original insn:\n");
5884 dump_insn_slim (lra_dump_file, insn);
5885 }
5886 }
5887 }
5888 /* Clear restore_regnos. */
5889 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
5890 lra_reg_info[regno].restore_regno = -1;
5891 bitmap_clear (&insn_bitmap);
5892 bitmap_clear (&removed_optional_reload_pseudos);
5893 return change_p;
5894 }
5895
5896 /* Entry function for undoing inheritance/split transformation. Return true
5897 if we did any RTL change in this pass. */
5898 bool
5899 lra_undo_inheritance (void)
5900 {
5901 unsigned int regno;
5902 int restore_regno, hard_regno;
5903 int n_all_inherit, n_inherit, n_all_split, n_split;
5904 bitmap_head remove_pseudos;
5905 bitmap_iterator bi;
5906 bool change_p;
5907
5908 lra_undo_inheritance_iter++;
5909 if (lra_undo_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
5910 return false;
5911 if (lra_dump_file != NULL)
5912 fprintf (lra_dump_file,
5913 "\n********** Undoing inheritance #%d: **********\n\n",
5914 lra_undo_inheritance_iter);
5915 bitmap_initialize (&remove_pseudos, &reg_obstack);
5916 n_inherit = n_all_inherit = 0;
5917 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
5918 if (lra_reg_info[regno].restore_regno >= 0)
5919 {
5920 n_all_inherit++;
5921 if (reg_renumber[regno] < 0
5922 /* If the original pseudo changed its allocation, just
5923 removing inheritance is dangerous as for changing
5924 allocation we used shorter live-ranges. */
5925 && reg_renumber[lra_reg_info[regno].restore_regno] < 0)
5926 bitmap_set_bit (&remove_pseudos, regno);
5927 else
5928 n_inherit++;
5929 }
5930 if (lra_dump_file != NULL && n_all_inherit != 0)
5931 fprintf (lra_dump_file, "Inherit %d out of %d (%.2f%%)\n",
5932 n_inherit, n_all_inherit,
5933 (double) n_inherit / n_all_inherit * 100);
5934 n_split = n_all_split = 0;
5935 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
5936 if ((restore_regno = lra_reg_info[regno].restore_regno) >= 0)
5937 {
5938 n_all_split++;
5939 hard_regno = (restore_regno >= FIRST_PSEUDO_REGISTER
5940 ? reg_renumber[restore_regno] : restore_regno);
5941 if (hard_regno < 0 || reg_renumber[regno] == hard_regno)
5942 bitmap_set_bit (&remove_pseudos, regno);
5943 else
5944 {
5945 n_split++;
5946 if (lra_dump_file != NULL)
5947 fprintf (lra_dump_file, " Keep split r%d (orig=r%d)\n",
5948 regno, restore_regno);
5949 }
5950 }
5951 if (lra_dump_file != NULL && n_all_split != 0)
5952 fprintf (lra_dump_file, "Split %d out of %d (%.2f%%)\n",
5953 n_split, n_all_split,
5954 (double) n_split / n_all_split * 100);
5955 change_p = remove_inheritance_pseudos (&remove_pseudos);
5956 bitmap_clear (&remove_pseudos);
5957 /* Clear restore_regnos. */
5958 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
5959 lra_reg_info[regno].restore_regno = -1;
5960 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
5961 lra_reg_info[regno].restore_regno = -1;
5962 change_p = undo_optional_reloads () || change_p;
5963 return change_p;
5964 }