PR middle-end/55401
[official-gcc.git] / gcc / lra-constraints.c
blobf6d5ac3467422a706b992a486898087bc7af189a
1 /* Code for RTL transformations to satisfy insn constraints.
2 Copyright (C) 2010, 2011, 2012
3 Free Software Foundation, Inc.
4 Contributed by Vladimir Makarov <vmakarov@redhat.com>.
6 This file is part of GCC.
8 GCC is free software; you can redistribute it and/or modify it under
9 the terms of the GNU General Public License as published by the Free
10 Software Foundation; either version 3, or (at your option) any later
11 version.
13 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
14 WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
16 for more details.
18 You should have received a copy of the GNU General Public License
19 along with GCC; see the file COPYING3. If not see
20 <http://www.gnu.org/licenses/>. */
23 /* This file contains code for 3 passes: constraint pass,
24 inheritance/split pass, and pass for undoing failed inheritance and
25 split.
27 The major goal of constraint pass is to transform RTL to satisfy
28 insn and address constraints by:
29 o choosing insn alternatives;
30 o generating *reload insns* (or reloads in brief) and *reload
31 pseudos* which will get necessary hard registers later;
32 o substituting pseudos with equivalent values and removing the
33 instructions that initialized those pseudos.
35 The constraint pass has biggest and most complicated code in LRA.
36 There are a lot of important details like:
37 o reuse of input reload pseudos to simplify reload pseudo
38 allocations;
39 o some heuristics to choose insn alternative to improve the
40 inheritance;
41 o early clobbers etc.
43 The pass is mimicking former reload pass in alternative choosing
44 because the reload pass is oriented to current machine description
45 model. It might be changed if the machine description model is
46 changed.
48 There is special code for preventing all LRA and this pass cycling
49 in case of bugs.
51 On the first iteration of the pass we process every instruction and
52 choose an alternative for each one. On subsequent iterations we try
53 to avoid reprocessing instructions if we can be sure that the old
54 choice is still valid.
56 The inheritance/spilt pass is to transform code to achieve
57 ineheritance and live range splitting. It is done on backward
58 traversal of EBBs.
60 The inheritance optimization goal is to reuse values in hard
61 registers. There is analogous optimization in old reload pass. The
62 inheritance is achieved by following transformation:
64 reload_p1 <- p reload_p1 <- p
65 ... new_p <- reload_p1
66 ... => ...
67 reload_p2 <- p reload_p2 <- new_p
69 where p is spilled and not changed between the insns. Reload_p1 is
70 also called *original pseudo* and new_p is called *inheritance
71 pseudo*.
73 The subsequent assignment pass will try to assign the same (or
74 another if it is not possible) hard register to new_p as to
75 reload_p1 or reload_p2.
77 If the assignment pass fails to assign a hard register to new_p,
78 this file will undo the inheritance and restore the original code.
79 This is because implementing the above sequence with a spilled
80 new_p would make the code much worse. The inheritance is done in
81 EBB scope. The above is just a simplified example to get an idea
82 of the inheritance as the inheritance is also done for non-reload
83 insns.
85 Splitting (transformation) is also done in EBB scope on the same
86 pass as the inheritance:
88 r <- ... or ... <- r r <- ... or ... <- r
89 ... s <- r (new insn -- save)
90 ... =>
91 ... r <- s (new insn -- restore)
92 ... <- r ... <- r
94 The *split pseudo* s is assigned to the hard register of the
95 original pseudo or hard register r.
97 Splitting is done:
98 o In EBBs with high register pressure for global pseudos (living
99 in at least 2 BBs) and assigned to hard registers when there
100 are more one reloads needing the hard registers;
101 o for pseudos needing save/restore code around calls.
103 If the split pseudo still has the same hard register as the
104 original pseudo after the subsequent assignment pass or the
105 original pseudo was split, the opposite transformation is done on
106 the same pass for undoing inheritance. */
108 #undef REG_OK_STRICT
110 #include "config.h"
111 #include "system.h"
112 #include "coretypes.h"
113 #include "tm.h"
114 #include "hard-reg-set.h"
115 #include "rtl.h"
116 #include "tm_p.h"
117 #include "regs.h"
118 #include "insn-config.h"
119 #include "insn-codes.h"
120 #include "recog.h"
121 #include "output.h"
122 #include "addresses.h"
123 #include "target.h"
124 #include "function.h"
125 #include "expr.h"
126 #include "basic-block.h"
127 #include "except.h"
128 #include "optabs.h"
129 #include "df.h"
130 #include "ira.h"
131 #include "rtl-error.h"
132 #include "lra-int.h"
134 /* Value of LRA_CURR_RELOAD_NUM at the beginning of BB of the current
135 insn. Remember that LRA_CURR_RELOAD_NUM is the number of emitted
136 reload insns. */
137 static int bb_reload_num;
139 /* The current insn being processed and corresponding its data (basic
140 block, the insn data, the insn static data, and the mode of each
141 operand). */
142 static rtx curr_insn;
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];
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;
155 /* If LOC is nonnull, strip any outer subreg from it. */
156 static inline rtx *
157 strip_subreg (rtx *loc)
159 return loc && GET_CODE (*loc) == SUBREG ? &SUBREG_REG (*loc) : loc;
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)
167 int hard_regno;
168 enum reg_class rclass;
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];
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)
186 if (hard_regno < 0)
187 return hard_regno;
188 hard_regno = lra_get_elimination_hard_regno (hard_regno);
189 return hard_regno + offset;
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)
198 rtx reg;
199 int offset, hard_regno;
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);
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)
224 int hard_regno;
226 if ((hard_regno = regno) >= FIRST_PSEUDO_REGISTER)
227 hard_regno = lra_get_regno_hard_regno (regno);
228 if (hard_regno >= 0)
230 hard_regno = get_final_hard_regno (hard_regno, 0);
231 return REGNO_REG_CLASS (hard_regno);
233 if (regno >= new_regno_start)
234 return lra_get_allocno_class (regno);
235 return NO_REGS;
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.
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)
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);
254 if (new_class != NULL)
255 *new_class = NO_REGS;
256 if (regno < FIRST_PSEUDO_REGISTER)
258 rtx final_reg = reg;
259 rtx *final_loc = &final_reg;
261 lra_eliminate_reg_if_possible (final_loc);
262 return TEST_HARD_REG_BIT (reg_class_contents[cl], REGNO (*final_loc));
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 /* When we don't know what class will be used finally for reload
274 pseudos, we use ALL_REGS. */
275 return ((regno >= new_regno_start && rclass == ALL_REGS)
276 || (rclass != NO_REGS && ira_class_subset_p[rclass][cl]
277 && ! hard_reg_set_subset_p (reg_class_contents[cl],
278 lra_no_alloc_regs)));
279 else
281 common_class = ira_reg_class_subset[rclass][cl];
282 if (new_class != NULL)
283 *new_class = common_class;
284 if (hard_reg_set_subset_p (reg_class_contents[common_class],
285 lra_no_alloc_regs))
286 return false;
287 /* Check that there are enough allocatable regs. */
288 class_size = ira_class_hard_regs_num[common_class];
289 for (i = 0; i < class_size; i++)
291 hard_regno = ira_class_hard_regs[common_class][i];
292 nregs = hard_regno_nregs[hard_regno][reg_mode];
293 if (nregs == 1)
294 return true;
295 for (j = 0; j < nregs; j++)
296 if (TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno + j)
297 || ! TEST_HARD_REG_BIT (reg_class_contents[common_class],
298 hard_regno + j))
299 break;
300 if (j >= nregs)
301 return true;
303 return false;
307 /* Return true if REGNO satisfies a memory constraint. */
308 static bool
309 in_mem_p (int regno)
311 return get_reg_class (regno) == NO_REGS;
314 /* If we have decided to substitute X with another value, return that
315 value, otherwise return X. */
316 static rtx
317 get_equiv_substitution (rtx x)
319 int regno;
320 rtx res;
322 if (! REG_P (x) || (regno = REGNO (x)) < FIRST_PSEUDO_REGISTER
323 || ! ira_reg_equiv[regno].defined_p
324 || ! ira_reg_equiv[regno].profitable_p
325 || lra_get_regno_hard_regno (regno) >= 0)
326 return x;
327 if ((res = ira_reg_equiv[regno].memory) != NULL_RTX)
328 return res;
329 if ((res = ira_reg_equiv[regno].constant) != NULL_RTX)
330 return res;
331 if ((res = ira_reg_equiv[regno].invariant) != NULL_RTX)
332 return res;
333 gcc_unreachable ();
336 /* Set up curr_operand_mode. */
337 static void
338 init_curr_operand_mode (void)
340 int nop = curr_static_id->n_operands;
341 for (int i = 0; i < nop; i++)
343 enum machine_mode mode = GET_MODE (*curr_id->operand_loc[i]);
344 if (mode == VOIDmode)
346 /* The .md mode for address operands is the mode of the
347 addressed value rather than the mode of the address itself. */
348 if (curr_id->icode >= 0 && curr_static_id->operand[i].is_address)
349 mode = Pmode;
350 else
351 mode = curr_static_id->operand[i].mode;
353 curr_operand_mode[i] = mode;
359 /* The page contains code to reuse input reloads. */
361 /* Structure describes input reload of the current insns. */
362 struct input_reload
364 /* Reloaded value. */
365 rtx input;
366 /* Reload pseudo used. */
367 rtx reg;
370 /* The number of elements in the following array. */
371 static int curr_insn_input_reloads_num;
372 /* Array containing info about input reloads. It is used to find the
373 same input reload and reuse the reload pseudo in this case. */
374 static struct input_reload curr_insn_input_reloads[LRA_MAX_INSN_RELOADS];
376 /* Initiate data concerning reuse of input reloads for the current
377 insn. */
378 static void
379 init_curr_insn_input_reloads (void)
381 curr_insn_input_reloads_num = 0;
384 /* Change class of pseudo REGNO to NEW_CLASS. Print info about it
385 using TITLE. Output a new line if NL_P. */
386 static void
387 change_class (int regno, enum reg_class new_class,
388 const char *title, bool nl_p)
390 lra_assert (regno >= FIRST_PSEUDO_REGISTER);
391 if (lra_dump_file != NULL)
392 fprintf (lra_dump_file, "%s to class %s for r%d",
393 title, reg_class_names[new_class], regno);
394 setup_reg_classes (regno, new_class, NO_REGS, new_class);
395 if (lra_dump_file != NULL && nl_p)
396 fprintf (lra_dump_file, "\n");
399 /* Create a new pseudo using MODE, RCLASS, ORIGINAL or reuse already
400 created input reload pseudo (only if TYPE is not OP_OUT). The
401 result pseudo is returned through RESULT_REG. Return TRUE if we
402 created a new pseudo, FALSE if we reused the already created input
403 reload pseudo. Use TITLE to describe new registers for debug
404 purposes. */
405 static bool
406 get_reload_reg (enum op_type type, enum machine_mode mode, rtx original,
407 enum reg_class rclass, const char *title, rtx *result_reg)
409 int i, regno;
410 enum reg_class new_class;
412 if (type == OP_OUT)
414 *result_reg
415 = lra_create_new_reg_with_unique_value (mode, original, rclass, title);
416 return true;
418 for (i = 0; i < curr_insn_input_reloads_num; i++)
419 if (rtx_equal_p (curr_insn_input_reloads[i].input, original)
420 && in_class_p (curr_insn_input_reloads[i].reg, rclass, &new_class))
422 lra_assert (! side_effects_p (original));
423 *result_reg = curr_insn_input_reloads[i].reg;
424 regno = REGNO (*result_reg);
425 if (lra_dump_file != NULL)
427 fprintf (lra_dump_file, " Reuse r%d for reload ", regno);
428 dump_value_slim (lra_dump_file, original, 1);
430 if (new_class != lra_get_allocno_class (regno))
431 change_class (regno, new_class, ", change", false);
432 if (lra_dump_file != NULL)
433 fprintf (lra_dump_file, "\n");
434 return false;
436 *result_reg = lra_create_new_reg (mode, original, rclass, title);
437 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
438 curr_insn_input_reloads[curr_insn_input_reloads_num].input = original;
439 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = *result_reg;
440 return true;
445 /* The page contains code to extract memory address parts. */
447 /* Wrapper around REGNO_OK_FOR_INDEX_P, to allow pseudos. */
448 static inline bool
449 ok_for_index_p_nonstrict (rtx reg)
451 unsigned regno = REGNO (reg);
453 return regno >= FIRST_PSEUDO_REGISTER || REGNO_OK_FOR_INDEX_P (regno);
456 /* A version of regno_ok_for_base_p for use here, when all pseudos
457 should count as OK. Arguments as for regno_ok_for_base_p. */
458 static inline bool
459 ok_for_base_p_nonstrict (rtx reg, enum machine_mode mode, addr_space_t as,
460 enum rtx_code outer_code, enum rtx_code index_code)
462 unsigned regno = REGNO (reg);
464 if (regno >= FIRST_PSEUDO_REGISTER)
465 return true;
466 return ok_for_base_p_1 (regno, mode, as, outer_code, index_code);
471 /* The page contains major code to choose the current insn alternative
472 and generate reloads for it. */
474 /* Return the offset from REGNO of the least significant register
475 in (reg:MODE REGNO).
477 This function is used to tell whether two registers satisfy
478 a matching constraint. (reg:MODE1 REGNO1) matches (reg:MODE2 REGNO2) if:
480 REGNO1 + lra_constraint_offset (REGNO1, MODE1)
481 == REGNO2 + lra_constraint_offset (REGNO2, MODE2) */
483 lra_constraint_offset (int regno, enum machine_mode mode)
485 lra_assert (regno < FIRST_PSEUDO_REGISTER);
486 if (WORDS_BIG_ENDIAN && GET_MODE_SIZE (mode) > UNITS_PER_WORD
487 && SCALAR_INT_MODE_P (mode))
488 return hard_regno_nregs[regno][mode] - 1;
489 return 0;
492 /* Like rtx_equal_p except that it allows a REG and a SUBREG to match
493 if they are the same hard reg, and has special hacks for
494 auto-increment and auto-decrement. This is specifically intended for
495 process_alt_operands to use in determining whether two operands
496 match. X is the operand whose number is the lower of the two.
498 It is supposed that X is the output operand and Y is the input
499 operand. Y_HARD_REGNO is the final hard regno of register Y or
500 register in subreg Y as we know it now. Otherwise, it is a
501 negative value. */
502 static bool
503 operands_match_p (rtx x, rtx y, int y_hard_regno)
505 int i;
506 RTX_CODE code = GET_CODE (x);
507 const char *fmt;
509 if (x == y)
510 return true;
511 if ((code == REG || (code == SUBREG && REG_P (SUBREG_REG (x))))
512 && (REG_P (y) || (GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y)))))
514 int j;
516 i = get_hard_regno (x);
517 if (i < 0)
518 goto slow;
520 if ((j = y_hard_regno) < 0)
521 goto slow;
523 i += lra_constraint_offset (i, GET_MODE (x));
524 j += lra_constraint_offset (j, GET_MODE (y));
526 return i == j;
529 /* If two operands must match, because they are really a single
530 operand of an assembler insn, then two post-increments are invalid
531 because the assembler insn would increment only once. On the
532 other hand, a post-increment matches ordinary indexing if the
533 post-increment is the output operand. */
534 if (code == POST_DEC || code == POST_INC || code == POST_MODIFY)
535 return operands_match_p (XEXP (x, 0), y, y_hard_regno);
537 /* Two pre-increments are invalid because the assembler insn would
538 increment only once. On the other hand, a pre-increment matches
539 ordinary indexing if the pre-increment is the input operand. */
540 if (GET_CODE (y) == PRE_DEC || GET_CODE (y) == PRE_INC
541 || GET_CODE (y) == PRE_MODIFY)
542 return operands_match_p (x, XEXP (y, 0), -1);
544 slow:
546 if (code == REG && GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y))
547 && x == SUBREG_REG (y))
548 return true;
549 if (GET_CODE (y) == REG && code == SUBREG && REG_P (SUBREG_REG (x))
550 && SUBREG_REG (x) == y)
551 return true;
553 /* Now we have disposed of all the cases in which different rtx
554 codes can match. */
555 if (code != GET_CODE (y))
556 return false;
558 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
559 if (GET_MODE (x) != GET_MODE (y))
560 return false;
562 switch (code)
564 CASE_CONST_UNIQUE:
565 return false;
567 case LABEL_REF:
568 return XEXP (x, 0) == XEXP (y, 0);
569 case SYMBOL_REF:
570 return XSTR (x, 0) == XSTR (y, 0);
572 default:
573 break;
576 /* Compare the elements. If any pair of corresponding elements fail
577 to match, return false for the whole things. */
579 fmt = GET_RTX_FORMAT (code);
580 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
582 int val, j;
583 switch (fmt[i])
585 case 'w':
586 if (XWINT (x, i) != XWINT (y, i))
587 return false;
588 break;
590 case 'i':
591 if (XINT (x, i) != XINT (y, i))
592 return false;
593 break;
595 case 'e':
596 val = operands_match_p (XEXP (x, i), XEXP (y, i), -1);
597 if (val == 0)
598 return false;
599 break;
601 case '0':
602 break;
604 case 'E':
605 if (XVECLEN (x, i) != XVECLEN (y, i))
606 return false;
607 for (j = XVECLEN (x, i) - 1; j >= 0; --j)
609 val = operands_match_p (XVECEXP (x, i, j), XVECEXP (y, i, j), -1);
610 if (val == 0)
611 return false;
613 break;
615 /* It is believed that rtx's at this level will never
616 contain anything but integers and other rtx's, except for
617 within LABEL_REFs and SYMBOL_REFs. */
618 default:
619 gcc_unreachable ();
622 return true;
625 /* True if X is a constant that can be forced into the constant pool.
626 MODE is the mode of the operand, or VOIDmode if not known. */
627 #define CONST_POOL_OK_P(MODE, X) \
628 ((MODE) != VOIDmode \
629 && CONSTANT_P (X) \
630 && GET_CODE (X) != HIGH \
631 && !targetm.cannot_force_const_mem (MODE, X))
633 /* True if C is a non-empty register class that has too few registers
634 to be safely used as a reload target class. */
635 #define SMALL_REGISTER_CLASS_P(C) \
636 (reg_class_size [(C)] == 1 \
637 || (reg_class_size [(C)] >= 1 && targetm.class_likely_spilled_p (C)))
639 /* If REG is a reload pseudo, try to make its class satisfying CL. */
640 static void
641 narrow_reload_pseudo_class (rtx reg, enum reg_class cl)
643 enum reg_class rclass;
645 /* Do not make more accurate class from reloads generated. They are
646 mostly moves with a lot of constraints. Making more accurate
647 class may results in very narrow class and impossibility of find
648 registers for several reloads of one insn. */
649 if (INSN_UID (curr_insn) >= new_insn_uid_start)
650 return;
651 if (GET_CODE (reg) == SUBREG)
652 reg = SUBREG_REG (reg);
653 if (! REG_P (reg) || (int) REGNO (reg) < new_regno_start)
654 return;
655 if (in_class_p (reg, cl, &rclass) && rclass != cl)
656 change_class (REGNO (reg), rclass, " Change", true);
659 /* Generate reloads for matching OUT and INS (array of input operand
660 numbers with end marker -1) with reg class GOAL_CLASS. Add input
661 and output reloads correspondingly to the lists *BEFORE and
662 *AFTER. */
663 static void
664 match_reload (signed char out, signed char *ins, enum reg_class goal_class,
665 rtx *before, rtx *after)
667 int i, in;
668 rtx new_in_reg, new_out_reg, reg, clobber;
669 enum machine_mode inmode, outmode;
670 rtx in_rtx = *curr_id->operand_loc[ins[0]];
671 rtx out_rtx = *curr_id->operand_loc[out];
673 outmode = curr_operand_mode[out];
674 inmode = curr_operand_mode[ins[0]];
675 push_to_sequence (*before);
676 if (inmode != outmode)
678 if (GET_MODE_SIZE (inmode) > GET_MODE_SIZE (outmode))
680 reg = new_in_reg
681 = lra_create_new_reg_with_unique_value (inmode, in_rtx,
682 goal_class, "");
683 if (SCALAR_INT_MODE_P (inmode))
684 new_out_reg = gen_lowpart_SUBREG (outmode, reg);
685 else
686 new_out_reg = gen_rtx_SUBREG (outmode, reg, 0);
687 /* If the input reg is dying here, we can use the same hard
688 register for REG and IN_RTX. We do it only for original
689 pseudos as reload pseudos can die although original
690 pseudos still live where reload pseudos dies. */
691 if (REG_P (in_rtx) && (int) REGNO (in_rtx) < lra_new_regno_start
692 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx)))
693 lra_reg_info[REGNO (reg)].val = lra_reg_info[REGNO (in_rtx)].val;
695 else
697 reg = new_out_reg
698 = lra_create_new_reg_with_unique_value (outmode, out_rtx,
699 goal_class, "");
700 if (SCALAR_INT_MODE_P (outmode))
701 new_in_reg = gen_lowpart_SUBREG (inmode, reg);
702 else
703 new_in_reg = gen_rtx_SUBREG (inmode, reg, 0);
704 /* NEW_IN_REG is non-paradoxical subreg. We don't want
705 NEW_OUT_REG living above. We add clobber clause for
706 this. This is just a temporary clobber. We can remove
707 it at the end of LRA work. */
708 clobber = emit_clobber (new_out_reg);
709 LRA_TEMP_CLOBBER_P (PATTERN (clobber)) = 1;
710 if (GET_CODE (in_rtx) == SUBREG)
712 rtx subreg_reg = SUBREG_REG (in_rtx);
714 /* If SUBREG_REG is dying here and sub-registers IN_RTX
715 and NEW_IN_REG are similar, we can use the same hard
716 register for REG and SUBREG_REG. */
717 if (REG_P (subreg_reg)
718 && (int) REGNO (subreg_reg) < lra_new_regno_start
719 && GET_MODE (subreg_reg) == outmode
720 && SUBREG_BYTE (in_rtx) == SUBREG_BYTE (new_in_reg)
721 && find_regno_note (curr_insn, REG_DEAD, REGNO (subreg_reg)))
722 lra_reg_info[REGNO (reg)].val
723 = lra_reg_info[REGNO (subreg_reg)].val;
727 else
729 /* Pseudos have values -- see comments for lra_reg_info.
730 Different pseudos with the same value do not conflict even if
731 they live in the same place. When we create a pseudo we
732 assign value of original pseudo (if any) from which we
733 created the new pseudo. If we create the pseudo from the
734 input pseudo, the new pseudo will no conflict with the input
735 pseudo which is wrong when the input pseudo lives after the
736 insn and as the new pseudo value is changed by the insn
737 output. Therefore we create the new pseudo from the output.
739 We cannot reuse the current output register because we might
740 have a situation like "a <- a op b", where the constraints
741 force the second input operand ("b") to match the output
742 operand ("a"). "b" must then be copied into a new register
743 so that it doesn't clobber the current value of "a". */
745 new_in_reg = new_out_reg
746 = lra_create_new_reg_with_unique_value (outmode, out_rtx,
747 goal_class, "");
749 /* In and out operand can be got from transformations before
750 processing insn constraints. One example of such transformations
751 is subreg reloading (see function simplify_operand_subreg). The
752 new pseudos created by the transformations might have inaccurate
753 class (ALL_REGS) and we should make their classes more
754 accurate. */
755 narrow_reload_pseudo_class (in_rtx, goal_class);
756 narrow_reload_pseudo_class (out_rtx, goal_class);
757 lra_emit_move (copy_rtx (new_in_reg), in_rtx);
758 *before = get_insns ();
759 end_sequence ();
760 for (i = 0; (in = ins[i]) >= 0; i++)
762 lra_assert
763 (GET_MODE (*curr_id->operand_loc[in]) == VOIDmode
764 || GET_MODE (new_in_reg) == GET_MODE (*curr_id->operand_loc[in]));
765 *curr_id->operand_loc[in] = new_in_reg;
767 lra_update_dups (curr_id, ins);
768 if (find_reg_note (curr_insn, REG_UNUSED, out_rtx) == NULL_RTX)
770 start_sequence ();
771 lra_emit_move (out_rtx, copy_rtx (new_out_reg));
772 emit_insn (*after);
773 *after = get_insns ();
774 end_sequence ();
776 *curr_id->operand_loc[out] = new_out_reg;
777 lra_update_dup (curr_id, out);
780 /* Return register class which is union of all reg classes in insn
781 constraint alternative string starting with P. */
782 static enum reg_class
783 reg_class_from_constraints (const char *p)
785 int c, len;
786 enum reg_class op_class = NO_REGS;
789 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
791 case '#':
792 case ',':
793 return op_class;
795 case 'p':
796 op_class = (reg_class_subunion
797 [op_class][base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
798 ADDRESS, SCRATCH)]);
799 break;
801 case 'g':
802 case 'r':
803 op_class = reg_class_subunion[op_class][GENERAL_REGS];
804 break;
806 default:
807 if (REG_CLASS_FROM_CONSTRAINT (c, p) == NO_REGS)
809 #ifdef EXTRA_CONSTRAINT_STR
810 if (EXTRA_ADDRESS_CONSTRAINT (c, p))
811 op_class
812 = (reg_class_subunion
813 [op_class][base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
814 ADDRESS, SCRATCH)]);
815 #endif
816 break;
819 op_class
820 = reg_class_subunion[op_class][REG_CLASS_FROM_CONSTRAINT (c, p)];
821 break;
823 while ((p += len), c);
824 return op_class;
827 /* If OP is a register, return the class of the register as per
828 get_reg_class, otherwise return NO_REGS. */
829 static inline enum reg_class
830 get_op_class (rtx op)
832 return REG_P (op) ? get_reg_class (REGNO (op)) : NO_REGS;
835 /* Return generated insn mem_pseudo:=val if TO_P or val:=mem_pseudo
836 otherwise. If modes of MEM_PSEUDO and VAL are different, use
837 SUBREG for VAL to make them equal. */
838 static rtx
839 emit_spill_move (bool to_p, rtx mem_pseudo, rtx val)
841 if (GET_MODE (mem_pseudo) != GET_MODE (val))
842 val = gen_rtx_SUBREG (GET_MODE (mem_pseudo),
843 GET_CODE (val) == SUBREG ? SUBREG_REG (val) : val,
845 return (to_p
846 ? gen_move_insn (mem_pseudo, val)
847 : gen_move_insn (val, mem_pseudo));
850 /* Process a special case insn (register move), return true if we
851 don't need to process it anymore. Return that RTL was changed
852 through CHANGE_P and macro SECONDARY_MEMORY_NEEDED says to use
853 secondary memory through SEC_MEM_P. */
854 static bool
855 check_and_process_move (bool *change_p, bool *sec_mem_p)
857 int sregno, dregno;
858 rtx set, dest, src, dreg, sreg, old_sreg, new_reg, before, scratch_reg;
859 enum reg_class dclass, sclass, secondary_class;
860 enum machine_mode sreg_mode;
861 secondary_reload_info sri;
863 *sec_mem_p = *change_p = false;
864 if ((set = single_set (curr_insn)) == NULL)
865 return false;
866 dreg = dest = SET_DEST (set);
867 sreg = src = SET_SRC (set);
868 /* Quick check on the right move insn which does not need
869 reloads. */
870 if ((dclass = get_op_class (dest)) != NO_REGS
871 && (sclass = get_op_class (src)) != NO_REGS
872 /* The backend guarantees that register moves of cost 2 never
873 need reloads. */
874 && targetm.register_move_cost (GET_MODE (src), dclass, sclass) == 2)
875 return true;
876 if (GET_CODE (dest) == SUBREG)
877 dreg = SUBREG_REG (dest);
878 if (GET_CODE (src) == SUBREG)
879 sreg = SUBREG_REG (src);
880 if (! REG_P (dreg) || ! REG_P (sreg))
881 return false;
882 sclass = dclass = NO_REGS;
883 dreg = get_equiv_substitution (dreg);
884 if (REG_P (dreg))
885 dclass = get_reg_class (REGNO (dreg));
886 if (dclass == ALL_REGS)
887 /* ALL_REGS is used for new pseudos created by transformations
888 like reload of SUBREG_REG (see function
889 simplify_operand_subreg). We don't know their class yet. We
890 should figure out the class from processing the insn
891 constraints not in this fast path function. Even if ALL_REGS
892 were a right class for the pseudo, secondary_... hooks usually
893 are not define for ALL_REGS. */
894 return false;
895 sreg_mode = GET_MODE (sreg);
896 old_sreg = sreg;
897 sreg = get_equiv_substitution (sreg);
898 if (REG_P (sreg))
899 sclass = get_reg_class (REGNO (sreg));
900 if (sclass == ALL_REGS)
901 /* See comments above. */
902 return false;
903 #ifdef SECONDARY_MEMORY_NEEDED
904 if (dclass != NO_REGS && sclass != NO_REGS
905 && SECONDARY_MEMORY_NEEDED (sclass, dclass, GET_MODE (src)))
907 *sec_mem_p = true;
908 return false;
910 #endif
911 sri.prev_sri = NULL;
912 sri.icode = CODE_FOR_nothing;
913 sri.extra_cost = 0;
914 secondary_class = NO_REGS;
915 /* Set up hard register for a reload pseudo for hook
916 secondary_reload because some targets just ignore unassigned
917 pseudos in the hook. */
918 if (dclass != NO_REGS && lra_get_regno_hard_regno (REGNO (dreg)) < 0)
920 dregno = REGNO (dreg);
921 reg_renumber[dregno] = ira_class_hard_regs[dclass][0];
923 else
924 dregno = -1;
925 if (sclass != NO_REGS && lra_get_regno_hard_regno (REGNO (sreg)) < 0)
927 sregno = REGNO (sreg);
928 reg_renumber[sregno] = ira_class_hard_regs[sclass][0];
930 else
931 sregno = -1;
932 if (sclass != NO_REGS)
933 secondary_class
934 = (enum reg_class) targetm.secondary_reload (false, dest,
935 (reg_class_t) sclass,
936 GET_MODE (src), &sri);
937 if (sclass == NO_REGS
938 || ((secondary_class != NO_REGS || sri.icode != CODE_FOR_nothing)
939 && dclass != NO_REGS))
941 enum reg_class old_sclass = secondary_class;
942 secondary_reload_info old_sri = sri;
944 sri.prev_sri = NULL;
945 sri.icode = CODE_FOR_nothing;
946 sri.extra_cost = 0;
947 secondary_class
948 = (enum reg_class) targetm.secondary_reload (true, sreg,
949 (reg_class_t) dclass,
950 sreg_mode, &sri);
951 /* Check the target hook consistency. */
952 lra_assert
953 ((secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
954 || (old_sclass == NO_REGS && old_sri.icode == CODE_FOR_nothing)
955 || (secondary_class == old_sclass && sri.icode == old_sri.icode));
957 if (sregno >= 0)
958 reg_renumber [sregno] = -1;
959 if (dregno >= 0)
960 reg_renumber [dregno] = -1;
961 if (secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
962 return false;
963 *change_p = true;
964 new_reg = NULL_RTX;
965 if (secondary_class != NO_REGS)
966 new_reg = lra_create_new_reg_with_unique_value (sreg_mode, NULL_RTX,
967 secondary_class,
968 "secondary");
969 start_sequence ();
970 if (old_sreg != sreg)
971 sreg = copy_rtx (sreg);
972 if (sri.icode == CODE_FOR_nothing)
973 lra_emit_move (new_reg, sreg);
974 else
976 enum reg_class scratch_class;
978 scratch_class = (reg_class_from_constraints
979 (insn_data[sri.icode].operand[2].constraint));
980 scratch_reg = (lra_create_new_reg_with_unique_value
981 (insn_data[sri.icode].operand[2].mode, NULL_RTX,
982 scratch_class, "scratch"));
983 emit_insn (GEN_FCN (sri.icode) (new_reg != NULL_RTX ? new_reg : dest,
984 sreg, scratch_reg));
986 before = get_insns ();
987 end_sequence ();
988 lra_process_new_insns (curr_insn, before, NULL_RTX, "Inserting the move");
989 if (new_reg != NULL_RTX)
991 if (GET_CODE (src) == SUBREG)
992 SUBREG_REG (src) = new_reg;
993 else
994 SET_SRC (set) = new_reg;
996 else
998 if (lra_dump_file != NULL)
1000 fprintf (lra_dump_file, "Deleting move %u\n", INSN_UID (curr_insn));
1001 dump_insn_slim (lra_dump_file, curr_insn);
1003 lra_set_insn_deleted (curr_insn);
1004 return true;
1006 return false;
1009 /* The following data describe the result of process_alt_operands.
1010 The data are used in curr_insn_transform to generate reloads. */
1012 /* The chosen reg classes which should be used for the corresponding
1013 operands. */
1014 static enum reg_class goal_alt[MAX_RECOG_OPERANDS];
1015 /* True if the operand should be the same as another operand and that
1016 other operand does not need a reload. */
1017 static bool goal_alt_match_win[MAX_RECOG_OPERANDS];
1018 /* True if the operand does not need a reload. */
1019 static bool goal_alt_win[MAX_RECOG_OPERANDS];
1020 /* True if the operand can be offsetable memory. */
1021 static bool goal_alt_offmemok[MAX_RECOG_OPERANDS];
1022 /* The number of an operand to which given operand can be matched to. */
1023 static int goal_alt_matches[MAX_RECOG_OPERANDS];
1024 /* The number of elements in the following array. */
1025 static int goal_alt_dont_inherit_ops_num;
1026 /* Numbers of operands whose reload pseudos should not be inherited. */
1027 static int goal_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1028 /* True if the insn commutative operands should be swapped. */
1029 static bool goal_alt_swapped;
1030 /* The chosen insn alternative. */
1031 static int goal_alt_number;
1033 /* The following five variables are used to choose the best insn
1034 alternative. They reflect final characteristics of the best
1035 alternative. */
1037 /* Number of necessary reloads and overall cost reflecting the
1038 previous value and other unpleasantness of the best alternative. */
1039 static int best_losers, best_overall;
1040 /* Number of small register classes used for operands of the best
1041 alternative. */
1042 static int best_small_class_operands_num;
1043 /* Overall number hard registers used for reloads. For example, on
1044 some targets we need 2 general registers to reload DFmode and only
1045 one floating point register. */
1046 static int best_reload_nregs;
1047 /* Overall number reflecting distances of previous reloading the same
1048 value. The distances are counted from the current BB start. It is
1049 used to improve inheritance chances. */
1050 static int best_reload_sum;
1052 /* True if the current insn should have no correspondingly input or
1053 output reloads. */
1054 static bool no_input_reloads_p, no_output_reloads_p;
1056 /* True if we swapped the commutative operands in the current
1057 insn. */
1058 static int curr_swapped;
1060 /* Arrange for address element *LOC to be a register of class CL.
1061 Add any input reloads to list BEFORE. AFTER is nonnull if *LOC is an
1062 automodified value; handle that case by adding the required output
1063 reloads to list AFTER. Return true if the RTL was changed. */
1064 static bool
1065 process_addr_reg (rtx *loc, rtx *before, rtx *after, enum reg_class cl)
1067 int regno;
1068 enum reg_class rclass, new_class;
1069 rtx reg;
1070 rtx new_reg;
1071 enum machine_mode mode;
1072 bool before_p = false;
1074 loc = strip_subreg (loc);
1075 reg = *loc;
1076 mode = GET_MODE (reg);
1077 if (! REG_P (reg))
1079 /* Always reload memory in an address even if the target supports
1080 such addresses. */
1081 new_reg = lra_create_new_reg_with_unique_value (mode, reg, cl, "address");
1082 before_p = true;
1084 else
1086 regno = REGNO (reg);
1087 rclass = get_reg_class (regno);
1088 if ((*loc = get_equiv_substitution (reg)) != reg)
1090 if (lra_dump_file != NULL)
1092 fprintf (lra_dump_file,
1093 "Changing pseudo %d in address of insn %u on equiv ",
1094 REGNO (reg), INSN_UID (curr_insn));
1095 dump_value_slim (lra_dump_file, *loc, 1);
1096 fprintf (lra_dump_file, "\n");
1098 *loc = copy_rtx (*loc);
1100 if (*loc != reg || ! in_class_p (reg, cl, &new_class))
1102 reg = *loc;
1103 if (get_reload_reg (after == NULL ? OP_IN : OP_INOUT,
1104 mode, reg, cl, "address", &new_reg))
1105 before_p = true;
1107 else if (new_class != NO_REGS && rclass != new_class)
1109 change_class (regno, new_class, " Change", true);
1110 return false;
1112 else
1113 return false;
1115 if (before_p)
1117 push_to_sequence (*before);
1118 lra_emit_move (new_reg, reg);
1119 *before = get_insns ();
1120 end_sequence ();
1122 *loc = new_reg;
1123 if (after != NULL)
1125 start_sequence ();
1126 lra_emit_move (reg, new_reg);
1127 emit_insn (*after);
1128 *after = get_insns ();
1129 end_sequence ();
1131 return true;
1134 /* Make reloads for subreg in operand NOP with internal subreg mode
1135 REG_MODE, add new reloads for further processing. Return true if
1136 any reload was generated. */
1137 static bool
1138 simplify_operand_subreg (int nop, enum machine_mode reg_mode)
1140 int hard_regno;
1141 rtx before, after;
1142 enum machine_mode mode;
1143 rtx reg, new_reg;
1144 rtx operand = *curr_id->operand_loc[nop];
1146 before = after = NULL_RTX;
1148 if (GET_CODE (operand) != SUBREG)
1149 return false;
1151 mode = GET_MODE (operand);
1152 reg = SUBREG_REG (operand);
1153 /* If we change address for paradoxical subreg of memory, the
1154 address might violate the necessary alignment or the access might
1155 be slow. So take this into consideration. We should not worry
1156 about access beyond allocated memory for paradoxical memory
1157 subregs as we don't substitute such equiv memory (see processing
1158 equivalences in function lra_constraints) and because for spilled
1159 pseudos we allocate stack memory enough for the biggest
1160 corresponding paradoxical subreg. */
1161 if ((MEM_P (reg)
1162 && (! SLOW_UNALIGNED_ACCESS (mode, MEM_ALIGN (reg))
1163 || MEM_ALIGN (reg) >= GET_MODE_ALIGNMENT (mode)))
1164 || (REG_P (reg) && REGNO (reg) < FIRST_PSEUDO_REGISTER))
1166 alter_subreg (curr_id->operand_loc[nop], false);
1167 return true;
1169 /* Put constant into memory when we have mixed modes. It generates
1170 a better code in most cases as it does not need a secondary
1171 reload memory. It also prevents LRA looping when LRA is using
1172 secondary reload memory again and again. */
1173 if (CONSTANT_P (reg) && CONST_POOL_OK_P (reg_mode, reg)
1174 && SCALAR_INT_MODE_P (reg_mode) != SCALAR_INT_MODE_P (mode))
1176 SUBREG_REG (operand) = force_const_mem (reg_mode, reg);
1177 alter_subreg (curr_id->operand_loc[nop], false);
1178 return true;
1180 /* Force a reload of the SUBREG_REG if this is a constant or PLUS or
1181 if there may be a problem accessing OPERAND in the outer
1182 mode. */
1183 if ((REG_P (reg)
1184 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1185 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1186 /* Don't reload paradoxical subregs because we could be looping
1187 having repeatedly final regno out of hard regs range. */
1188 && (hard_regno_nregs[hard_regno][GET_MODE (reg)]
1189 >= hard_regno_nregs[hard_regno][mode])
1190 && simplify_subreg_regno (hard_regno, GET_MODE (reg),
1191 SUBREG_BYTE (operand), mode) < 0)
1192 || CONSTANT_P (reg) || GET_CODE (reg) == PLUS || MEM_P (reg))
1194 enum op_type type = curr_static_id->operand[nop].type;
1195 /* The class will be defined later in curr_insn_transform. */
1196 enum reg_class rclass
1197 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1199 new_reg = lra_create_new_reg_with_unique_value (reg_mode, reg, rclass,
1200 "subreg reg");
1201 bitmap_set_bit (&lra_optional_reload_pseudos, REGNO (new_reg));
1202 if (type != OP_OUT
1203 || GET_MODE_SIZE (GET_MODE (reg)) > GET_MODE_SIZE (mode))
1205 push_to_sequence (before);
1206 lra_emit_move (new_reg, reg);
1207 before = get_insns ();
1208 end_sequence ();
1210 if (type != OP_IN)
1212 start_sequence ();
1213 lra_emit_move (reg, new_reg);
1214 emit_insn (after);
1215 after = get_insns ();
1216 end_sequence ();
1218 SUBREG_REG (operand) = new_reg;
1219 lra_process_new_insns (curr_insn, before, after,
1220 "Inserting subreg reload");
1221 return true;
1223 return false;
1226 /* Return TRUE if X refers for a hard register from SET. */
1227 static bool
1228 uses_hard_regs_p (rtx x, HARD_REG_SET set)
1230 int i, j, x_hard_regno;
1231 enum machine_mode mode;
1232 const char *fmt;
1233 enum rtx_code code;
1235 if (x == NULL_RTX)
1236 return false;
1237 code = GET_CODE (x);
1238 mode = GET_MODE (x);
1239 if (code == SUBREG)
1241 x = SUBREG_REG (x);
1242 code = GET_CODE (x);
1243 if (GET_MODE_SIZE (GET_MODE (x)) > GET_MODE_SIZE (mode))
1244 mode = GET_MODE (x);
1247 if (REG_P (x))
1249 x_hard_regno = get_hard_regno (x);
1250 return (x_hard_regno >= 0
1251 && overlaps_hard_reg_set_p (set, mode, x_hard_regno));
1253 if (MEM_P (x))
1255 struct address_info ad;
1257 decompose_mem_address (&ad, x);
1258 if (ad.base_term != NULL && uses_hard_regs_p (*ad.base_term, set))
1259 return true;
1260 if (ad.index_term != NULL && uses_hard_regs_p (*ad.index_term, set))
1261 return true;
1263 fmt = GET_RTX_FORMAT (code);
1264 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1266 if (fmt[i] == 'e')
1268 if (uses_hard_regs_p (XEXP (x, i), set))
1269 return true;
1271 else if (fmt[i] == 'E')
1273 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1274 if (uses_hard_regs_p (XVECEXP (x, i, j), set))
1275 return true;
1278 return false;
1281 /* Return true if OP is a spilled pseudo. */
1282 static inline bool
1283 spilled_pseudo_p (rtx op)
1285 return (REG_P (op)
1286 && REGNO (op) >= FIRST_PSEUDO_REGISTER && in_mem_p (REGNO (op)));
1289 /* Return true if X is a general constant. */
1290 static inline bool
1291 general_constant_p (rtx x)
1293 return CONSTANT_P (x) && (! flag_pic || LEGITIMATE_PIC_OPERAND_P (x));
1296 /* Major function to choose the current insn alternative and what
1297 operands should be reloaded and how. If ONLY_ALTERNATIVE is not
1298 negative we should consider only this alternative. Return false if
1299 we can not choose the alternative or find how to reload the
1300 operands. */
1301 static bool
1302 process_alt_operands (int only_alternative)
1304 bool ok_p = false;
1305 int nop, small_class_operands_num, overall, nalt;
1306 int n_alternatives = curr_static_id->n_alternatives;
1307 int n_operands = curr_static_id->n_operands;
1308 /* LOSERS counts the operands that don't fit this alternative and
1309 would require loading. */
1310 int losers;
1311 /* REJECT is a count of how undesirable this alternative says it is
1312 if any reloading is required. If the alternative matches exactly
1313 then REJECT is ignored, but otherwise it gets this much counted
1314 against it in addition to the reloading needed. */
1315 int reject;
1316 /* The number of elements in the following array. */
1317 int early_clobbered_regs_num;
1318 /* Numbers of operands which are early clobber registers. */
1319 int early_clobbered_nops[MAX_RECOG_OPERANDS];
1320 enum reg_class curr_alt[MAX_RECOG_OPERANDS];
1321 HARD_REG_SET curr_alt_set[MAX_RECOG_OPERANDS];
1322 bool curr_alt_match_win[MAX_RECOG_OPERANDS];
1323 bool curr_alt_win[MAX_RECOG_OPERANDS];
1324 bool curr_alt_offmemok[MAX_RECOG_OPERANDS];
1325 int curr_alt_matches[MAX_RECOG_OPERANDS];
1326 /* The number of elements in the following array. */
1327 int curr_alt_dont_inherit_ops_num;
1328 /* Numbers of operands whose reload pseudos should not be inherited. */
1329 int curr_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1330 rtx op;
1331 /* The register when the operand is a subreg of register, otherwise the
1332 operand itself. */
1333 rtx no_subreg_reg_operand[MAX_RECOG_OPERANDS];
1334 /* The register if the operand is a register or subreg of register,
1335 otherwise NULL. */
1336 rtx operand_reg[MAX_RECOG_OPERANDS];
1337 int hard_regno[MAX_RECOG_OPERANDS];
1338 enum machine_mode biggest_mode[MAX_RECOG_OPERANDS];
1339 int reload_nregs, reload_sum;
1340 bool costly_p;
1341 enum reg_class cl;
1343 /* Calculate some data common for all alternatives to speed up the
1344 function. */
1345 for (nop = 0; nop < n_operands; nop++)
1347 op = no_subreg_reg_operand[nop] = *curr_id->operand_loc[nop];
1348 /* The real hard regno of the operand after the allocation. */
1349 hard_regno[nop] = get_hard_regno (op);
1351 operand_reg[nop] = op;
1352 biggest_mode[nop] = GET_MODE (operand_reg[nop]);
1353 if (GET_CODE (operand_reg[nop]) == SUBREG)
1355 operand_reg[nop] = SUBREG_REG (operand_reg[nop]);
1356 if (GET_MODE_SIZE (biggest_mode[nop])
1357 < GET_MODE_SIZE (GET_MODE (operand_reg[nop])))
1358 biggest_mode[nop] = GET_MODE (operand_reg[nop]);
1360 if (REG_P (operand_reg[nop]))
1361 no_subreg_reg_operand[nop] = operand_reg[nop];
1362 else
1363 operand_reg[nop] = NULL_RTX;
1366 /* The constraints are made of several alternatives. Each operand's
1367 constraint looks like foo,bar,... with commas separating the
1368 alternatives. The first alternatives for all operands go
1369 together, the second alternatives go together, etc.
1371 First loop over alternatives. */
1372 for (nalt = 0; nalt < n_alternatives; nalt++)
1374 /* Loop over operands for one constraint alternative. */
1375 #ifdef HAVE_ATTR_enabled
1376 if (curr_id->alternative_enabled_p != NULL
1377 && ! curr_id->alternative_enabled_p[nalt])
1378 continue;
1379 #endif
1381 if (only_alternative >= 0 && nalt != only_alternative)
1382 continue;
1384 overall = losers = reject = reload_nregs = reload_sum = 0;
1385 for (nop = 0; nop < n_operands; nop++)
1386 reject += (curr_static_id
1387 ->operand_alternative[nalt * n_operands + nop].reject);
1388 early_clobbered_regs_num = 0;
1390 for (nop = 0; nop < n_operands; nop++)
1392 const char *p;
1393 char *end;
1394 int len, c, m, i, opalt_num, this_alternative_matches;
1395 bool win, did_match, offmemok, early_clobber_p;
1396 /* false => this operand can be reloaded somehow for this
1397 alternative. */
1398 bool badop;
1399 /* true => this operand can be reloaded if the alternative
1400 allows regs. */
1401 bool winreg;
1402 /* True if a constant forced into memory would be OK for
1403 this operand. */
1404 bool constmemok;
1405 enum reg_class this_alternative, this_costly_alternative;
1406 HARD_REG_SET this_alternative_set, this_costly_alternative_set;
1407 bool this_alternative_match_win, this_alternative_win;
1408 bool this_alternative_offmemok;
1409 enum machine_mode mode;
1411 opalt_num = nalt * n_operands + nop;
1412 if (curr_static_id->operand_alternative[opalt_num].anything_ok)
1414 /* Fast track for no constraints at all. */
1415 curr_alt[nop] = NO_REGS;
1416 CLEAR_HARD_REG_SET (curr_alt_set[nop]);
1417 curr_alt_win[nop] = true;
1418 curr_alt_match_win[nop] = false;
1419 curr_alt_offmemok[nop] = false;
1420 curr_alt_matches[nop] = -1;
1421 continue;
1424 op = no_subreg_reg_operand[nop];
1425 mode = curr_operand_mode[nop];
1427 win = did_match = winreg = offmemok = constmemok = false;
1428 badop = true;
1430 early_clobber_p = false;
1431 p = curr_static_id->operand_alternative[opalt_num].constraint;
1433 this_costly_alternative = this_alternative = NO_REGS;
1434 /* We update set of possible hard regs besides its class
1435 because reg class might be inaccurate. For example,
1436 union of LO_REGS (l), HI_REGS(h), and STACK_REG(k) in ARM
1437 is translated in HI_REGS because classes are merged by
1438 pairs and there is no accurate intermediate class. */
1439 CLEAR_HARD_REG_SET (this_alternative_set);
1440 CLEAR_HARD_REG_SET (this_costly_alternative_set);
1441 this_alternative_win = false;
1442 this_alternative_match_win = false;
1443 this_alternative_offmemok = false;
1444 this_alternative_matches = -1;
1446 /* An empty constraint should be excluded by the fast
1447 track. */
1448 lra_assert (*p != 0 && *p != ',');
1450 /* Scan this alternative's specs for this operand; set WIN
1451 if the operand fits any letter in this alternative.
1452 Otherwise, clear BADOP if this operand could fit some
1453 letter after reloads, or set WINREG if this operand could
1454 fit after reloads provided the constraint allows some
1455 registers. */
1456 costly_p = false;
1459 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
1461 case '\0':
1462 len = 0;
1463 break;
1464 case ',':
1465 c = '\0';
1466 break;
1468 case '=': case '+': case '?': case '*': case '!':
1469 case ' ': case '\t':
1470 break;
1472 case '%':
1473 /* We only support one commutative marker, the first
1474 one. We already set commutative above. */
1475 break;
1477 case '&':
1478 early_clobber_p = true;
1479 break;
1481 case '#':
1482 /* Ignore rest of this alternative. */
1483 c = '\0';
1484 break;
1486 case '0': case '1': case '2': case '3': case '4':
1487 case '5': case '6': case '7': case '8': case '9':
1489 int m_hregno;
1490 bool match_p;
1492 m = strtoul (p, &end, 10);
1493 p = end;
1494 len = 0;
1495 lra_assert (nop > m);
1497 this_alternative_matches = m;
1498 m_hregno = get_hard_regno (*curr_id->operand_loc[m]);
1499 /* We are supposed to match a previous operand.
1500 If we do, we win if that one did. If we do
1501 not, count both of the operands as losers.
1502 (This is too conservative, since most of the
1503 time only a single reload insn will be needed
1504 to make the two operands win. As a result,
1505 this alternative may be rejected when it is
1506 actually desirable.) */
1507 match_p = false;
1508 if (operands_match_p (*curr_id->operand_loc[nop],
1509 *curr_id->operand_loc[m], m_hregno))
1511 /* We should reject matching of an early
1512 clobber operand if the matching operand is
1513 not dying in the insn. */
1514 if (! curr_static_id->operand[m].early_clobber
1515 || operand_reg[nop] == NULL_RTX
1516 || (find_regno_note (curr_insn, REG_DEAD,
1517 REGNO (operand_reg[nop]))
1518 != NULL_RTX))
1519 match_p = true;
1521 if (match_p)
1523 /* If we are matching a non-offsettable
1524 address where an offsettable address was
1525 expected, then we must reject this
1526 combination, because we can't reload
1527 it. */
1528 if (curr_alt_offmemok[m]
1529 && MEM_P (*curr_id->operand_loc[m])
1530 && curr_alt[m] == NO_REGS && ! curr_alt_win[m])
1531 continue;
1534 else
1536 /* Operands don't match. Both operands must
1537 allow a reload register, otherwise we
1538 cannot make them match. */
1539 if (curr_alt[m] == NO_REGS)
1540 break;
1541 /* Retroactively mark the operand we had to
1542 match as a loser, if it wasn't already and
1543 it wasn't matched to a register constraint
1544 (e.g it might be matched by memory). */
1545 if (curr_alt_win[m]
1546 && (operand_reg[m] == NULL_RTX
1547 || hard_regno[m] < 0))
1549 losers++;
1550 reload_nregs
1551 += (ira_reg_class_max_nregs[curr_alt[m]]
1552 [GET_MODE (*curr_id->operand_loc[m])]);
1555 /* We prefer no matching alternatives because
1556 it gives more freedom in RA. */
1557 if (operand_reg[nop] == NULL_RTX
1558 || (find_regno_note (curr_insn, REG_DEAD,
1559 REGNO (operand_reg[nop]))
1560 == NULL_RTX))
1561 reject += 2;
1563 /* If we have to reload this operand and some
1564 previous operand also had to match the same
1565 thing as this operand, we don't know how to do
1566 that. */
1567 if (!match_p || !curr_alt_win[m])
1569 for (i = 0; i < nop; i++)
1570 if (curr_alt_matches[i] == m)
1571 break;
1572 if (i < nop)
1573 break;
1575 else
1576 did_match = true;
1578 /* This can be fixed with reloads if the operand
1579 we are supposed to match can be fixed with
1580 reloads. */
1581 badop = false;
1582 this_alternative = curr_alt[m];
1583 COPY_HARD_REG_SET (this_alternative_set, curr_alt_set[m]);
1584 winreg = this_alternative != NO_REGS;
1585 break;
1588 case 'p':
1589 cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
1590 ADDRESS, SCRATCH);
1591 this_alternative = reg_class_subunion[this_alternative][cl];
1592 IOR_HARD_REG_SET (this_alternative_set,
1593 reg_class_contents[cl]);
1594 if (costly_p)
1596 this_costly_alternative
1597 = reg_class_subunion[this_costly_alternative][cl];
1598 IOR_HARD_REG_SET (this_costly_alternative_set,
1599 reg_class_contents[cl]);
1601 win = true;
1602 badop = false;
1603 break;
1605 case TARGET_MEM_CONSTRAINT:
1606 if (MEM_P (op) || spilled_pseudo_p (op))
1607 win = true;
1608 /* We can put constant or pseudo value into memory
1609 to satisfy the constraint. */
1610 if (CONST_POOL_OK_P (mode, op) || REG_P (op))
1611 badop = false;
1612 constmemok = true;
1613 break;
1615 case '<':
1616 if (MEM_P (op)
1617 && (GET_CODE (XEXP (op, 0)) == PRE_DEC
1618 || GET_CODE (XEXP (op, 0)) == POST_DEC))
1619 win = true;
1620 break;
1622 case '>':
1623 if (MEM_P (op)
1624 && (GET_CODE (XEXP (op, 0)) == PRE_INC
1625 || GET_CODE (XEXP (op, 0)) == POST_INC))
1626 win = true;
1627 break;
1629 /* Memory op whose address is not offsettable. */
1630 case 'V':
1631 if (MEM_P (op)
1632 && ! offsettable_nonstrict_memref_p (op))
1633 win = true;
1634 break;
1636 /* Memory operand whose address is offsettable. */
1637 case 'o':
1638 if ((MEM_P (op)
1639 && offsettable_nonstrict_memref_p (op))
1640 || spilled_pseudo_p (op))
1641 win = true;
1642 /* We can put constant or pseudo value into memory
1643 or make memory address offsetable to satisfy the
1644 constraint. */
1645 if (CONST_POOL_OK_P (mode, op) || MEM_P (op) || REG_P (op))
1646 badop = false;
1647 constmemok = true;
1648 offmemok = true;
1649 break;
1651 case 'E':
1652 case 'F':
1653 if (GET_CODE (op) == CONST_DOUBLE
1654 || (GET_CODE (op) == CONST_VECTOR
1655 && (GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT)))
1656 win = true;
1657 break;
1659 case 'G':
1660 case 'H':
1661 if (GET_CODE (op) == CONST_DOUBLE
1662 && CONST_DOUBLE_OK_FOR_CONSTRAINT_P (op, c, p))
1663 win = true;
1664 break;
1666 case 's':
1667 if (CONST_INT_P (op)
1668 || (GET_CODE (op) == CONST_DOUBLE && mode == VOIDmode))
1669 break;
1671 case 'i':
1672 if (general_constant_p (op))
1673 win = true;
1674 break;
1676 case 'n':
1677 if (CONST_INT_P (op)
1678 || (GET_CODE (op) == CONST_DOUBLE && mode == VOIDmode))
1679 win = true;
1680 break;
1682 case 'I':
1683 case 'J':
1684 case 'K':
1685 case 'L':
1686 case 'M':
1687 case 'N':
1688 case 'O':
1689 case 'P':
1690 if (CONST_INT_P (op)
1691 && CONST_OK_FOR_CONSTRAINT_P (INTVAL (op), c, p))
1692 win = true;
1693 break;
1695 case 'X':
1696 /* This constraint should be excluded by the fast
1697 track. */
1698 gcc_unreachable ();
1699 break;
1701 case 'g':
1702 if (MEM_P (op)
1703 || general_constant_p (op)
1704 || spilled_pseudo_p (op))
1705 win = true;
1706 /* Drop through into 'r' case. */
1708 case 'r':
1709 this_alternative
1710 = reg_class_subunion[this_alternative][GENERAL_REGS];
1711 IOR_HARD_REG_SET (this_alternative_set,
1712 reg_class_contents[GENERAL_REGS]);
1713 if (costly_p)
1715 this_costly_alternative
1716 = (reg_class_subunion
1717 [this_costly_alternative][GENERAL_REGS]);
1718 IOR_HARD_REG_SET (this_costly_alternative_set,
1719 reg_class_contents[GENERAL_REGS]);
1721 goto reg;
1723 default:
1724 if (REG_CLASS_FROM_CONSTRAINT (c, p) == NO_REGS)
1726 #ifdef EXTRA_CONSTRAINT_STR
1727 if (EXTRA_MEMORY_CONSTRAINT (c, p))
1729 if (EXTRA_CONSTRAINT_STR (op, c, p))
1730 win = true;
1731 else if (spilled_pseudo_p (op))
1732 win = true;
1734 /* If we didn't already win, we can reload
1735 constants via force_const_mem or put the
1736 pseudo value into memory, or make other
1737 memory by reloading the address like for
1738 'o'. */
1739 if (CONST_POOL_OK_P (mode, op)
1740 || MEM_P (op) || REG_P (op))
1741 badop = false;
1742 constmemok = true;
1743 offmemok = true;
1744 break;
1746 if (EXTRA_ADDRESS_CONSTRAINT (c, p))
1748 if (EXTRA_CONSTRAINT_STR (op, c, p))
1749 win = true;
1751 /* If we didn't already win, we can reload
1752 the address into a base register. */
1753 cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
1754 ADDRESS, SCRATCH);
1755 this_alternative
1756 = reg_class_subunion[this_alternative][cl];
1757 IOR_HARD_REG_SET (this_alternative_set,
1758 reg_class_contents[cl]);
1759 if (costly_p)
1761 this_costly_alternative
1762 = (reg_class_subunion
1763 [this_costly_alternative][cl]);
1764 IOR_HARD_REG_SET (this_costly_alternative_set,
1765 reg_class_contents[cl]);
1767 badop = false;
1768 break;
1771 if (EXTRA_CONSTRAINT_STR (op, c, p))
1772 win = true;
1773 #endif
1774 break;
1777 cl = REG_CLASS_FROM_CONSTRAINT (c, p);
1778 this_alternative = reg_class_subunion[this_alternative][cl];
1779 IOR_HARD_REG_SET (this_alternative_set,
1780 reg_class_contents[cl]);
1781 if (costly_p)
1783 this_costly_alternative
1784 = reg_class_subunion[this_costly_alternative][cl];
1785 IOR_HARD_REG_SET (this_costly_alternative_set,
1786 reg_class_contents[cl]);
1788 reg:
1789 if (mode == BLKmode)
1790 break;
1791 winreg = true;
1792 if (REG_P (op))
1794 if (hard_regno[nop] >= 0
1795 && in_hard_reg_set_p (this_alternative_set,
1796 mode, hard_regno[nop]))
1797 win = true;
1798 else if (hard_regno[nop] < 0
1799 && in_class_p (op, this_alternative, NULL))
1800 win = true;
1802 break;
1804 if (c != ' ' && c != '\t')
1805 costly_p = c == '*';
1807 while ((p += len), c);
1809 /* Record which operands fit this alternative. */
1810 if (win)
1812 this_alternative_win = true;
1813 if (operand_reg[nop] != NULL_RTX)
1815 if (hard_regno[nop] >= 0)
1817 if (in_hard_reg_set_p (this_costly_alternative_set,
1818 mode, hard_regno[nop]))
1819 reject++;
1821 else
1823 /* Prefer won reg to spilled pseudo under other equal
1824 conditions. */
1825 reject++;
1826 if (in_class_p (operand_reg[nop],
1827 this_costly_alternative, NULL))
1828 reject++;
1830 /* We simulate the behaviour of old reload here.
1831 Although scratches need hard registers and it
1832 might result in spilling other pseudos, no reload
1833 insns are generated for the scratches. So it
1834 might cost something but probably less than old
1835 reload pass believes. */
1836 if (lra_former_scratch_p (REGNO (operand_reg[nop])))
1837 reject += LRA_LOSER_COST_FACTOR;
1840 else if (did_match)
1841 this_alternative_match_win = true;
1842 else
1844 int const_to_mem = 0;
1845 bool no_regs_p;
1847 no_regs_p
1848 = (this_alternative == NO_REGS
1849 || (hard_reg_set_subset_p
1850 (reg_class_contents[this_alternative],
1851 lra_no_alloc_regs)));
1852 /* If this operand accepts a register, and if the
1853 register class has at least one allocatable register,
1854 then this operand can be reloaded. */
1855 if (winreg && !no_regs_p)
1856 badop = false;
1858 if (badop)
1859 goto fail;
1861 this_alternative_offmemok = offmemok;
1862 if (this_costly_alternative != NO_REGS)
1863 reject++;
1864 /* If the operand is dying, has a matching constraint,
1865 and satisfies constraints of the matched operand
1866 which failed to satisfy the own constraints, we do
1867 not need to generate a reload insn for this
1868 operand. */
1869 if (!(this_alternative_matches >= 0
1870 && !curr_alt_win[this_alternative_matches]
1871 && REG_P (op)
1872 && find_regno_note (curr_insn, REG_DEAD, REGNO (op))
1873 && (hard_regno[nop] >= 0
1874 ? in_hard_reg_set_p (this_alternative_set,
1875 mode, hard_regno[nop])
1876 : in_class_p (op, this_alternative, NULL))))
1877 losers++;
1878 if (operand_reg[nop] != NULL_RTX
1879 /* Output operands and matched input operands are
1880 not inherited. The following conditions do not
1881 exactly describe the previous statement but they
1882 are pretty close. */
1883 && curr_static_id->operand[nop].type != OP_OUT
1884 && (this_alternative_matches < 0
1885 || curr_static_id->operand[nop].type != OP_IN))
1887 int last_reload = (lra_reg_info[ORIGINAL_REGNO
1888 (operand_reg[nop])]
1889 .last_reload);
1891 if (last_reload > bb_reload_num)
1892 reload_sum += last_reload - bb_reload_num;
1894 /* If this is a constant that is reloaded into the
1895 desired class by copying it to memory first, count
1896 that as another reload. This is consistent with
1897 other code and is required to avoid choosing another
1898 alternative when the constant is moved into memory.
1899 Note that the test here is precisely the same as in
1900 the code below that calls force_const_mem. */
1901 if (CONST_POOL_OK_P (mode, op)
1902 && ((targetm.preferred_reload_class
1903 (op, this_alternative) == NO_REGS)
1904 || no_input_reloads_p))
1906 const_to_mem = 1;
1907 if (! no_regs_p)
1908 losers++;
1911 /* Alternative loses if it requires a type of reload not
1912 permitted for this insn. We can always reload
1913 objects with a REG_UNUSED note. */
1914 if ((curr_static_id->operand[nop].type != OP_IN
1915 && no_output_reloads_p
1916 && ! find_reg_note (curr_insn, REG_UNUSED, op))
1917 || (curr_static_id->operand[nop].type != OP_OUT
1918 && no_input_reloads_p && ! const_to_mem))
1919 goto fail;
1921 /* Check strong discouragement of reload of non-constant
1922 into class THIS_ALTERNATIVE. */
1923 if (! CONSTANT_P (op) && ! no_regs_p
1924 && (targetm.preferred_reload_class
1925 (op, this_alternative) == NO_REGS
1926 || (curr_static_id->operand[nop].type == OP_OUT
1927 && (targetm.preferred_output_reload_class
1928 (op, this_alternative) == NO_REGS))))
1929 reject += LRA_MAX_REJECT;
1931 if (! ((const_to_mem && constmemok)
1932 || (MEM_P (op) && offmemok)))
1934 /* We prefer to reload pseudos over reloading other
1935 things, since such reloads may be able to be
1936 eliminated later. So bump REJECT in other cases.
1937 Don't do this in the case where we are forcing a
1938 constant into memory and it will then win since
1939 we don't want to have a different alternative
1940 match then. */
1941 if (! (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER))
1942 reject += 2;
1944 if (! no_regs_p)
1945 reload_nregs
1946 += ira_reg_class_max_nregs[this_alternative][mode];
1949 /* We are trying to spill pseudo into memory. It is
1950 usually more costly than moving to a hard register
1951 although it might takes the same number of
1952 reloads. */
1953 if (no_regs_p && REG_P (op))
1954 reject++;
1956 #ifdef SECONDARY_MEMORY_NEEDED
1957 /* If reload requires moving value through secondary
1958 memory, it will need one more insn at least. */
1959 if (this_alternative != NO_REGS
1960 && REG_P (op) && (cl = get_reg_class (REGNO (op))) != NO_REGS
1961 && ((curr_static_id->operand[nop].type != OP_OUT
1962 && SECONDARY_MEMORY_NEEDED (cl, this_alternative,
1963 GET_MODE (op)))
1964 || (curr_static_id->operand[nop].type != OP_IN
1965 && SECONDARY_MEMORY_NEEDED (this_alternative, cl,
1966 GET_MODE (op)))))
1967 losers++;
1968 #endif
1969 /* Input reloads can be inherited more often than output
1970 reloads can be removed, so penalize output
1971 reloads. */
1972 if (!REG_P (op) || curr_static_id->operand[nop].type != OP_IN)
1973 reject++;
1976 if (early_clobber_p)
1977 reject++;
1978 /* ??? We check early clobbers after processing all operands
1979 (see loop below) and there we update the costs more.
1980 Should we update the cost (may be approximately) here
1981 because of early clobber register reloads or it is a rare
1982 or non-important thing to be worth to do it. */
1983 overall = losers * LRA_LOSER_COST_FACTOR + reject;
1984 if ((best_losers == 0 || losers != 0) && best_overall < overall)
1985 goto fail;
1987 curr_alt[nop] = this_alternative;
1988 COPY_HARD_REG_SET (curr_alt_set[nop], this_alternative_set);
1989 curr_alt_win[nop] = this_alternative_win;
1990 curr_alt_match_win[nop] = this_alternative_match_win;
1991 curr_alt_offmemok[nop] = this_alternative_offmemok;
1992 curr_alt_matches[nop] = this_alternative_matches;
1994 if (this_alternative_matches >= 0
1995 && !did_match && !this_alternative_win)
1996 curr_alt_win[this_alternative_matches] = false;
1998 if (early_clobber_p && operand_reg[nop] != NULL_RTX)
1999 early_clobbered_nops[early_clobbered_regs_num++] = nop;
2001 ok_p = true;
2002 curr_alt_dont_inherit_ops_num = 0;
2003 for (nop = 0; nop < early_clobbered_regs_num; nop++)
2005 int i, j, clobbered_hard_regno;
2006 HARD_REG_SET temp_set;
2008 i = early_clobbered_nops[nop];
2009 if ((! curr_alt_win[i] && ! curr_alt_match_win[i])
2010 || hard_regno[i] < 0)
2011 continue;
2012 clobbered_hard_regno = hard_regno[i];
2013 CLEAR_HARD_REG_SET (temp_set);
2014 add_to_hard_reg_set (&temp_set, biggest_mode[i], clobbered_hard_regno);
2015 for (j = 0; j < n_operands; j++)
2016 if (j == i
2017 /* We don't want process insides of match_operator and
2018 match_parallel because otherwise we would process
2019 their operands once again generating a wrong
2020 code. */
2021 || curr_static_id->operand[j].is_operator)
2022 continue;
2023 else if ((curr_alt_matches[j] == i && curr_alt_match_win[j])
2024 || (curr_alt_matches[i] == j && curr_alt_match_win[i]))
2025 continue;
2026 else if (uses_hard_regs_p (*curr_id->operand_loc[j], temp_set))
2027 break;
2028 if (j >= n_operands)
2029 continue;
2030 /* We need to reload early clobbered register. */
2031 for (j = 0; j < n_operands; j++)
2032 if (curr_alt_matches[j] == i)
2034 curr_alt_match_win[j] = false;
2035 losers++;
2036 overall += LRA_LOSER_COST_FACTOR;
2038 if (! curr_alt_match_win[i])
2039 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++] = i;
2040 else
2042 /* Remember pseudos used for match reloads are never
2043 inherited. */
2044 lra_assert (curr_alt_matches[i] >= 0);
2045 curr_alt_win[curr_alt_matches[i]] = false;
2047 curr_alt_win[i] = curr_alt_match_win[i] = false;
2048 losers++;
2049 overall += LRA_LOSER_COST_FACTOR;
2051 small_class_operands_num = 0;
2052 for (nop = 0; nop < n_operands; nop++)
2053 small_class_operands_num
2054 += SMALL_REGISTER_CLASS_P (curr_alt[nop]) ? 1 : 0;
2056 /* If this alternative can be made to work by reloading, and it
2057 needs less reloading than the others checked so far, record
2058 it as the chosen goal for reloading. */
2059 if ((best_losers != 0 && losers == 0)
2060 || (((best_losers == 0 && losers == 0)
2061 || (best_losers != 0 && losers != 0))
2062 && (best_overall > overall
2063 || (best_overall == overall
2064 /* If the cost of the reloads is the same,
2065 prefer alternative which requires minimal
2066 number of small register classes for the
2067 operands. This improves chances of reloads
2068 for insn requiring small register
2069 classes. */
2070 && (small_class_operands_num
2071 < best_small_class_operands_num
2072 || (small_class_operands_num
2073 == best_small_class_operands_num
2074 && (reload_nregs < best_reload_nregs
2075 || (reload_nregs == best_reload_nregs
2076 && best_reload_sum < reload_sum))))))))
2078 for (nop = 0; nop < n_operands; nop++)
2080 goal_alt_win[nop] = curr_alt_win[nop];
2081 goal_alt_match_win[nop] = curr_alt_match_win[nop];
2082 goal_alt_matches[nop] = curr_alt_matches[nop];
2083 goal_alt[nop] = curr_alt[nop];
2084 goal_alt_offmemok[nop] = curr_alt_offmemok[nop];
2086 goal_alt_dont_inherit_ops_num = curr_alt_dont_inherit_ops_num;
2087 for (nop = 0; nop < curr_alt_dont_inherit_ops_num; nop++)
2088 goal_alt_dont_inherit_ops[nop] = curr_alt_dont_inherit_ops[nop];
2089 goal_alt_swapped = curr_swapped;
2090 best_overall = overall;
2091 best_losers = losers;
2092 best_small_class_operands_num = small_class_operands_num;
2093 best_reload_nregs = reload_nregs;
2094 best_reload_sum = reload_sum;
2095 goal_alt_number = nalt;
2097 if (losers == 0)
2098 /* Everything is satisfied. Do not process alternatives
2099 anymore. */
2100 break;
2101 fail:
2104 return ok_p;
2107 /* Return 1 if ADDR is a valid memory address for mode MODE in address
2108 space AS, and check that each pseudo has the proper kind of hard
2109 reg. */
2110 static int
2111 valid_address_p (enum machine_mode mode ATTRIBUTE_UNUSED,
2112 rtx addr, addr_space_t as)
2114 #ifdef GO_IF_LEGITIMATE_ADDRESS
2115 lra_assert (ADDR_SPACE_GENERIC_P (as));
2116 GO_IF_LEGITIMATE_ADDRESS (mode, addr, win);
2117 return 0;
2119 win:
2120 return 1;
2121 #else
2122 return targetm.addr_space.legitimate_address_p (mode, addr, 0, as);
2123 #endif
2126 /* Return whether address AD is valid. */
2128 static bool
2129 valid_address_p (struct address_info *ad)
2131 /* Some ports do not check displacements for eliminable registers,
2132 so we replace them temporarily with the elimination target. */
2133 rtx saved_base_reg = NULL_RTX;
2134 rtx saved_index_reg = NULL_RTX;
2135 rtx *base_term = strip_subreg (ad->base_term);
2136 rtx *index_term = strip_subreg (ad->index_term);
2137 if (base_term != NULL)
2139 saved_base_reg = *base_term;
2140 lra_eliminate_reg_if_possible (base_term);
2141 if (ad->base_term2 != NULL)
2142 *ad->base_term2 = *ad->base_term;
2144 if (index_term != NULL)
2146 saved_index_reg = *index_term;
2147 lra_eliminate_reg_if_possible (index_term);
2149 bool ok_p = valid_address_p (ad->mode, *ad->outer, ad->as);
2150 if (saved_base_reg != NULL_RTX)
2152 *base_term = saved_base_reg;
2153 if (ad->base_term2 != NULL)
2154 *ad->base_term2 = *ad->base_term;
2156 if (saved_index_reg != NULL_RTX)
2157 *index_term = saved_index_reg;
2158 return ok_p;
2161 /* Make reload base reg + disp from address AD. Return the new pseudo. */
2162 static rtx
2163 base_plus_disp_to_reg (struct address_info *ad)
2165 enum reg_class cl;
2166 rtx new_reg;
2168 lra_assert (ad->base == ad->base_term && ad->disp == ad->disp_term);
2169 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
2170 get_index_code (ad));
2171 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
2172 cl, "base + disp");
2173 lra_emit_add (new_reg, *ad->base_term, *ad->disp_term);
2174 return new_reg;
2177 /* Return true if we can add a displacement to address AD, even if that
2178 makes the address invalid. The fix-up code requires any new address
2179 to be the sum of the BASE_TERM, INDEX and DISP_TERM fields. */
2180 static bool
2181 can_add_disp_p (struct address_info *ad)
2183 return (!ad->autoinc_p
2184 && ad->segment == NULL
2185 && ad->base == ad->base_term
2186 && ad->disp == ad->disp_term);
2189 /* Make equiv substitution in address AD. Return true if a substitution
2190 was made. */
2191 static bool
2192 equiv_address_substitution (struct address_info *ad)
2194 rtx base_reg, new_base_reg, index_reg, new_index_reg, *base_term, *index_term;
2195 HOST_WIDE_INT disp, scale;
2196 bool change_p;
2198 base_term = strip_subreg (ad->base_term);
2199 if (base_term == NULL)
2200 base_reg = new_base_reg = NULL_RTX;
2201 else
2203 base_reg = *base_term;
2204 new_base_reg = get_equiv_substitution (base_reg);
2206 index_term = strip_subreg (ad->index_term);
2207 if (index_term == NULL)
2208 index_reg = new_index_reg = NULL_RTX;
2209 else
2211 index_reg = *index_term;
2212 new_index_reg = get_equiv_substitution (index_reg);
2214 if (base_reg == new_base_reg && index_reg == new_index_reg)
2215 return false;
2216 disp = 0;
2217 change_p = false;
2218 if (lra_dump_file != NULL)
2220 fprintf (lra_dump_file, "Changing address in insn %d ",
2221 INSN_UID (curr_insn));
2222 dump_value_slim (lra_dump_file, *ad->outer, 1);
2224 if (base_reg != new_base_reg)
2226 if (REG_P (new_base_reg))
2228 *base_term = new_base_reg;
2229 change_p = true;
2231 else if (GET_CODE (new_base_reg) == PLUS
2232 && REG_P (XEXP (new_base_reg, 0))
2233 && CONST_INT_P (XEXP (new_base_reg, 1))
2234 && can_add_disp_p (ad))
2236 disp += INTVAL (XEXP (new_base_reg, 1));
2237 *base_term = XEXP (new_base_reg, 0);
2238 change_p = true;
2240 if (ad->base_term2 != NULL)
2241 *ad->base_term2 = *ad->base_term;
2243 if (index_reg != new_index_reg)
2245 if (REG_P (new_index_reg))
2247 *index_term = new_index_reg;
2248 change_p = true;
2250 else if (GET_CODE (new_index_reg) == PLUS
2251 && REG_P (XEXP (new_index_reg, 0))
2252 && CONST_INT_P (XEXP (new_index_reg, 1))
2253 && can_add_disp_p (ad)
2254 && (scale = get_index_scale (ad)))
2256 disp += INTVAL (XEXP (new_index_reg, 1)) * scale;
2257 *index_term = XEXP (new_index_reg, 0);
2258 change_p = true;
2261 if (disp != 0)
2263 if (ad->disp != NULL)
2264 *ad->disp = plus_constant (GET_MODE (*ad->inner), *ad->disp, disp);
2265 else
2267 *ad->inner = plus_constant (GET_MODE (*ad->inner), *ad->inner, disp);
2268 update_address (ad);
2270 change_p = true;
2272 if (lra_dump_file != NULL)
2274 if (! change_p)
2275 fprintf (lra_dump_file, " -- no change\n");
2276 else
2278 fprintf (lra_dump_file, " on equiv ");
2279 dump_value_slim (lra_dump_file, *ad->outer, 1);
2280 fprintf (lra_dump_file, "\n");
2283 return change_p;
2286 /* Major function to make reloads for an address in operand NOP.
2287 The supported cases are:
2289 1) an address that existed before LRA started, at which point it must
2290 have been valid. These addresses are subject to elimination and
2291 may have become invalid due to the elimination offset being out
2292 of range.
2294 2) an address created by forcing a constant to memory (force_const_to_mem).
2295 The initial form of these addresses might not be valid, and it is this
2296 function's job to make them valid.
2298 3) a frame address formed from a register and a (possibly zero)
2299 constant offset. As above, these addresses might not be valid
2300 and this function must make them so.
2302 Add reloads to the lists *BEFORE and *AFTER. We might need to add
2303 reloads to *AFTER because of inc/dec, {pre, post} modify in the
2304 address. Return true for any RTL change. */
2305 static bool
2306 process_address (int nop, rtx *before, rtx *after)
2308 struct address_info ad;
2309 rtx new_reg;
2310 rtx op = *curr_id->operand_loc[nop];
2311 const char *constraint = curr_static_id->operand[nop].constraint;
2312 bool change_p;
2314 if (constraint[0] == 'p'
2315 || EXTRA_ADDRESS_CONSTRAINT (constraint[0], constraint))
2316 decompose_lea_address (&ad, curr_id->operand_loc[nop]);
2317 else if (MEM_P (op))
2318 decompose_mem_address (&ad, op);
2319 else if (GET_CODE (op) == SUBREG
2320 && MEM_P (SUBREG_REG (op)))
2321 decompose_mem_address (&ad, SUBREG_REG (op));
2322 else
2323 return false;
2324 change_p = equiv_address_substitution (&ad);
2325 if (ad.base_term != NULL
2326 && (process_addr_reg
2327 (ad.base_term, before,
2328 (ad.autoinc_p
2329 && !(REG_P (*ad.base_term)
2330 && find_regno_note (curr_insn, REG_DEAD,
2331 REGNO (*ad.base_term)) != NULL_RTX)
2332 ? after : NULL),
2333 base_reg_class (ad.mode, ad.as, ad.base_outer_code,
2334 get_index_code (&ad)))))
2336 change_p = true;
2337 if (ad.base_term2 != NULL)
2338 *ad.base_term2 = *ad.base_term;
2340 if (ad.index_term != NULL
2341 && process_addr_reg (ad.index_term, before, NULL, INDEX_REG_CLASS))
2342 change_p = true;
2344 /* There are three cases where the shape of *AD.INNER may now be invalid:
2346 1) the original address was valid, but either elimination or
2347 equiv_address_substitution applied a displacement that made
2348 it invalid.
2350 2) the address is an invalid symbolic address created by
2351 force_const_to_mem.
2353 3) the address is a frame address with an invalid offset.
2355 All these cases involve a displacement and a non-autoinc address,
2356 so there is no point revalidating other types. */
2357 if (ad.disp == NULL || ad.autoinc_p || valid_address_p (&ad))
2358 return change_p;
2360 /* Any index existed before LRA started, so we can assume that the
2361 presence and shape of the index is valid. */
2362 push_to_sequence (*before);
2363 gcc_assert (ad.segment == NULL);
2364 gcc_assert (ad.disp == ad.disp_term);
2365 if (ad.base == NULL)
2367 if (ad.index == NULL)
2369 int code = -1;
2370 enum reg_class cl = base_reg_class (ad.mode, ad.as,
2371 SCRATCH, SCRATCH);
2372 rtx disp = *ad.disp;
2374 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp");
2375 #ifdef HAVE_lo_sum
2377 rtx insn;
2378 rtx last = get_last_insn ();
2380 /* disp => lo_sum (new_base, disp), case (2) above. */
2381 insn = emit_insn (gen_rtx_SET
2382 (VOIDmode, new_reg,
2383 gen_rtx_HIGH (Pmode, copy_rtx (disp))));
2384 code = recog_memoized (insn);
2385 if (code >= 0)
2387 *ad.disp = gen_rtx_LO_SUM (Pmode, new_reg, disp);
2388 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
2390 *ad.disp = disp;
2391 code = -1;
2394 if (code < 0)
2395 delete_insns_since (last);
2397 #endif
2398 if (code < 0)
2400 /* disp => new_base, case (2) above. */
2401 lra_emit_move (new_reg, disp);
2402 *ad.disp = new_reg;
2405 else
2407 /* index * scale + disp => new base + index * scale,
2408 case (1) above. */
2409 enum reg_class cl = base_reg_class (ad.mode, ad.as, PLUS,
2410 GET_CODE (*ad.index));
2412 lra_assert (INDEX_REG_CLASS != NO_REGS);
2413 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp");
2414 lra_emit_move (new_reg, *ad.disp);
2415 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
2416 new_reg, *ad.index);
2419 else if (ad.index == NULL)
2421 /* base + disp => new base, cases (1) and (3) above. */
2422 /* Another option would be to reload the displacement into an
2423 index register. However, postreload has code to optimize
2424 address reloads that have the same base and different
2425 displacements, so reloading into an index register would
2426 not necessarily be a win. */
2427 new_reg = base_plus_disp_to_reg (&ad);
2428 *ad.inner = new_reg;
2430 else
2432 /* base + scale * index + disp => new base + scale * index,
2433 case (1) above. */
2434 new_reg = base_plus_disp_to_reg (&ad);
2435 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
2436 new_reg, *ad.index);
2438 *before = get_insns ();
2439 end_sequence ();
2440 return true;
2443 /* Emit insns to reload VALUE into a new register. VALUE is an
2444 auto-increment or auto-decrement RTX whose operand is a register or
2445 memory location; so reloading involves incrementing that location.
2446 IN is either identical to VALUE, or some cheaper place to reload
2447 value being incremented/decremented from.
2449 INC_AMOUNT is the number to increment or decrement by (always
2450 positive and ignored for POST_MODIFY/PRE_MODIFY).
2452 Return pseudo containing the result. */
2453 static rtx
2454 emit_inc (enum reg_class new_rclass, rtx in, rtx value, int inc_amount)
2456 /* REG or MEM to be copied and incremented. */
2457 rtx incloc = XEXP (value, 0);
2458 /* Nonzero if increment after copying. */
2459 int post = (GET_CODE (value) == POST_DEC || GET_CODE (value) == POST_INC
2460 || GET_CODE (value) == POST_MODIFY);
2461 rtx last;
2462 rtx inc;
2463 rtx add_insn;
2464 int code;
2465 rtx real_in = in == value ? incloc : in;
2466 rtx result;
2467 bool plus_p = true;
2469 if (GET_CODE (value) == PRE_MODIFY || GET_CODE (value) == POST_MODIFY)
2471 lra_assert (GET_CODE (XEXP (value, 1)) == PLUS
2472 || GET_CODE (XEXP (value, 1)) == MINUS);
2473 lra_assert (rtx_equal_p (XEXP (XEXP (value, 1), 0), XEXP (value, 0)));
2474 plus_p = GET_CODE (XEXP (value, 1)) == PLUS;
2475 inc = XEXP (XEXP (value, 1), 1);
2477 else
2479 if (GET_CODE (value) == PRE_DEC || GET_CODE (value) == POST_DEC)
2480 inc_amount = -inc_amount;
2482 inc = GEN_INT (inc_amount);
2485 if (! post && REG_P (incloc))
2486 result = incloc;
2487 else
2488 result = lra_create_new_reg (GET_MODE (value), value, new_rclass,
2489 "INC/DEC result");
2491 if (real_in != result)
2493 /* First copy the location to the result register. */
2494 lra_assert (REG_P (result));
2495 emit_insn (gen_move_insn (result, real_in));
2498 /* We suppose that there are insns to add/sub with the constant
2499 increment permitted in {PRE/POST)_{DEC/INC/MODIFY}. At least the
2500 old reload worked with this assumption. If the assumption
2501 becomes wrong, we should use approach in function
2502 base_plus_disp_to_reg. */
2503 if (in == value)
2505 /* See if we can directly increment INCLOC. */
2506 last = get_last_insn ();
2507 add_insn = emit_insn (plus_p
2508 ? gen_add2_insn (incloc, inc)
2509 : gen_sub2_insn (incloc, inc));
2511 code = recog_memoized (add_insn);
2512 if (code >= 0)
2514 if (! post && result != incloc)
2515 emit_insn (gen_move_insn (result, incloc));
2516 return result;
2518 delete_insns_since (last);
2521 /* If couldn't do the increment directly, must increment in RESULT.
2522 The way we do this depends on whether this is pre- or
2523 post-increment. For pre-increment, copy INCLOC to the reload
2524 register, increment it there, then save back. */
2525 if (! post)
2527 if (real_in != result)
2528 emit_insn (gen_move_insn (result, real_in));
2529 if (plus_p)
2530 emit_insn (gen_add2_insn (result, inc));
2531 else
2532 emit_insn (gen_sub2_insn (result, inc));
2533 if (result != incloc)
2534 emit_insn (gen_move_insn (incloc, result));
2536 else
2538 /* Post-increment.
2540 Because this might be a jump insn or a compare, and because
2541 RESULT may not be available after the insn in an input
2542 reload, we must do the incrementing before the insn being
2543 reloaded for.
2545 We have already copied IN to RESULT. Increment the copy in
2546 RESULT, save that back, then decrement RESULT so it has
2547 the original value. */
2548 if (plus_p)
2549 emit_insn (gen_add2_insn (result, inc));
2550 else
2551 emit_insn (gen_sub2_insn (result, inc));
2552 emit_insn (gen_move_insn (incloc, result));
2553 /* Restore non-modified value for the result. We prefer this
2554 way because it does not require an additional hard
2555 register. */
2556 if (plus_p)
2558 if (CONST_INT_P (inc))
2559 emit_insn (gen_add2_insn (result, GEN_INT (-INTVAL (inc))));
2560 else
2561 emit_insn (gen_sub2_insn (result, inc));
2563 else
2564 emit_insn (gen_add2_insn (result, inc));
2566 return result;
2569 /* Swap operands NOP and NOP + 1. */
2570 static inline void
2571 swap_operands (int nop)
2573 enum machine_mode mode = curr_operand_mode[nop];
2574 curr_operand_mode[nop] = curr_operand_mode[nop + 1];
2575 curr_operand_mode[nop + 1] = mode;
2576 rtx x = *curr_id->operand_loc[nop];
2577 *curr_id->operand_loc[nop] = *curr_id->operand_loc[nop + 1];
2578 *curr_id->operand_loc[nop + 1] = x;
2579 /* Swap the duplicates too. */
2580 lra_update_dup (curr_id, nop);
2581 lra_update_dup (curr_id, nop + 1);
2584 /* Main entry point of the constraint code: search the body of the
2585 current insn to choose the best alternative. It is mimicking insn
2586 alternative cost calculation model of former reload pass. That is
2587 because machine descriptions were written to use this model. This
2588 model can be changed in future. Make commutative operand exchange
2589 if it is chosen.
2591 Return true if some RTL changes happened during function call. */
2592 static bool
2593 curr_insn_transform (void)
2595 int i, j, k;
2596 int n_operands;
2597 int n_alternatives;
2598 int commutative;
2599 signed char goal_alt_matched[MAX_RECOG_OPERANDS][MAX_RECOG_OPERANDS];
2600 rtx before, after;
2601 bool alt_p = false;
2602 /* Flag that the insn has been changed through a transformation. */
2603 bool change_p;
2604 bool sec_mem_p;
2605 #ifdef SECONDARY_MEMORY_NEEDED
2606 bool use_sec_mem_p;
2607 #endif
2608 int max_regno_before;
2609 int reused_alternative_num;
2611 no_input_reloads_p = no_output_reloads_p = false;
2612 goal_alt_number = -1;
2614 if (check_and_process_move (&change_p, &sec_mem_p))
2615 return change_p;
2617 /* JUMP_INSNs and CALL_INSNs are not allowed to have any output
2618 reloads; neither are insns that SET cc0. Insns that use CC0 are
2619 not allowed to have any input reloads. */
2620 if (JUMP_P (curr_insn) || CALL_P (curr_insn))
2621 no_output_reloads_p = true;
2623 #ifdef HAVE_cc0
2624 if (reg_referenced_p (cc0_rtx, PATTERN (curr_insn)))
2625 no_input_reloads_p = true;
2626 if (reg_set_p (cc0_rtx, PATTERN (curr_insn)))
2627 no_output_reloads_p = true;
2628 #endif
2630 n_operands = curr_static_id->n_operands;
2631 n_alternatives = curr_static_id->n_alternatives;
2633 /* Just return "no reloads" if insn has no operands with
2634 constraints. */
2635 if (n_operands == 0 || n_alternatives == 0)
2636 return false;
2638 max_regno_before = max_reg_num ();
2640 for (i = 0; i < n_operands; i++)
2642 goal_alt_matched[i][0] = -1;
2643 goal_alt_matches[i] = -1;
2646 commutative = curr_static_id->commutative;
2648 /* Now see what we need for pseudos that didn't get hard regs or got
2649 the wrong kind of hard reg. For this, we must consider all the
2650 operands together against the register constraints. */
2652 best_losers = best_overall = INT_MAX;
2653 best_small_class_operands_num = best_reload_sum = 0;
2655 curr_swapped = false;
2656 goal_alt_swapped = false;
2658 /* Make equivalence substitution and memory subreg elimination
2659 before address processing because an address legitimacy can
2660 depend on memory mode. */
2661 for (i = 0; i < n_operands; i++)
2663 rtx op = *curr_id->operand_loc[i];
2664 rtx subst, old = op;
2665 bool op_change_p = false;
2667 if (GET_CODE (old) == SUBREG)
2668 old = SUBREG_REG (old);
2669 subst = get_equiv_substitution (old);
2670 if (subst != old)
2672 subst = copy_rtx (subst);
2673 lra_assert (REG_P (old));
2674 if (GET_CODE (op) == SUBREG)
2675 SUBREG_REG (op) = subst;
2676 else
2677 *curr_id->operand_loc[i] = subst;
2678 if (lra_dump_file != NULL)
2680 fprintf (lra_dump_file,
2681 "Changing pseudo %d in operand %i of insn %u on equiv ",
2682 REGNO (old), i, INSN_UID (curr_insn));
2683 dump_value_slim (lra_dump_file, subst, 1);
2684 fprintf (lra_dump_file, "\n");
2686 op_change_p = change_p = true;
2688 if (simplify_operand_subreg (i, GET_MODE (old)) || op_change_p)
2690 change_p = true;
2691 lra_update_dup (curr_id, i);
2695 /* Reload address registers and displacements. We do it before
2696 finding an alternative because of memory constraints. */
2697 before = after = NULL_RTX;
2698 for (i = 0; i < n_operands; i++)
2699 if (! curr_static_id->operand[i].is_operator
2700 && process_address (i, &before, &after))
2702 change_p = true;
2703 lra_update_dup (curr_id, i);
2706 if (change_p)
2707 /* If we've changed the instruction then any alternative that
2708 we chose previously may no longer be valid. */
2709 lra_set_used_insn_alternative (curr_insn, -1);
2711 try_swapped:
2713 reused_alternative_num = curr_id->used_insn_alternative;
2714 if (lra_dump_file != NULL && reused_alternative_num >= 0)
2715 fprintf (lra_dump_file, "Reusing alternative %d for insn #%u\n",
2716 reused_alternative_num, INSN_UID (curr_insn));
2718 if (process_alt_operands (reused_alternative_num))
2719 alt_p = true;
2721 /* If insn is commutative (it's safe to exchange a certain pair of
2722 operands) then we need to try each alternative twice, the second
2723 time matching those two operands as if we had exchanged them. To
2724 do this, really exchange them in operands.
2726 If we have just tried the alternatives the second time, return
2727 operands to normal and drop through. */
2729 if (reused_alternative_num < 0 && commutative >= 0)
2731 curr_swapped = !curr_swapped;
2732 if (curr_swapped)
2734 swap_operands (commutative);
2735 goto try_swapped;
2737 else
2738 swap_operands (commutative);
2741 /* The operands don't meet the constraints. goal_alt describes the
2742 alternative that we could reach by reloading the fewest operands.
2743 Reload so as to fit it. */
2745 if (! alt_p && ! sec_mem_p)
2747 /* No alternative works with reloads?? */
2748 if (INSN_CODE (curr_insn) >= 0)
2749 fatal_insn ("unable to generate reloads for:", curr_insn);
2750 error_for_asm (curr_insn,
2751 "inconsistent operand constraints in an %<asm%>");
2752 /* Avoid further trouble with this insn. */
2753 PATTERN (curr_insn) = gen_rtx_USE (VOIDmode, const0_rtx);
2754 lra_invalidate_insn_data (curr_insn);
2755 return true;
2758 /* If the best alternative is with operands 1 and 2 swapped, swap
2759 them. Update the operand numbers of any reloads already
2760 pushed. */
2762 if (goal_alt_swapped)
2764 if (lra_dump_file != NULL)
2765 fprintf (lra_dump_file, " Commutative operand exchange in insn %u\n",
2766 INSN_UID (curr_insn));
2768 /* Swap the duplicates too. */
2769 swap_operands (commutative);
2770 change_p = true;
2773 #ifdef SECONDARY_MEMORY_NEEDED
2774 /* Some target macros SECONDARY_MEMORY_NEEDED (e.g. x86) are defined
2775 too conservatively. So we use the secondary memory only if there
2776 is no any alternative without reloads. */
2777 use_sec_mem_p = false;
2778 if (! alt_p)
2779 use_sec_mem_p = true;
2780 else if (sec_mem_p)
2782 for (i = 0; i < n_operands; i++)
2783 if (! goal_alt_win[i] && ! goal_alt_match_win[i])
2784 break;
2785 use_sec_mem_p = i < n_operands;
2788 if (use_sec_mem_p)
2790 rtx new_reg, set, src, dest;
2791 enum machine_mode sec_mode;
2793 lra_assert (sec_mem_p);
2794 set = single_set (curr_insn);
2795 lra_assert (set != NULL_RTX && ! side_effects_p (set));
2796 dest = SET_DEST (set);
2797 src = SET_SRC (set);
2798 #ifdef SECONDARY_MEMORY_NEEDED_MODE
2799 sec_mode = SECONDARY_MEMORY_NEEDED_MODE (GET_MODE (src));
2800 #else
2801 sec_mode = GET_MODE (src);
2802 #endif
2803 new_reg = lra_create_new_reg (sec_mode, NULL_RTX,
2804 NO_REGS, "secondary");
2805 /* If the mode is changed, it should be wider. */
2806 lra_assert (GET_MODE_SIZE (GET_MODE (new_reg))
2807 >= GET_MODE_SIZE (GET_MODE (src)));
2808 after = emit_spill_move (false, new_reg, dest);
2809 lra_process_new_insns (curr_insn, NULL_RTX, after,
2810 "Inserting the sec. move");
2811 before = emit_spill_move (true, new_reg, src);
2812 lra_process_new_insns (curr_insn, before, NULL_RTX, "Changing on");
2813 lra_set_insn_deleted (curr_insn);
2814 return true;
2816 #endif
2818 lra_assert (goal_alt_number >= 0);
2819 lra_set_used_insn_alternative (curr_insn, goal_alt_number);
2821 if (lra_dump_file != NULL)
2823 const char *p;
2825 fprintf (lra_dump_file, " Choosing alt %d in insn %u:",
2826 goal_alt_number, INSN_UID (curr_insn));
2827 for (i = 0; i < n_operands; i++)
2829 p = (curr_static_id->operand_alternative
2830 [goal_alt_number * n_operands + i].constraint);
2831 if (*p == '\0')
2832 continue;
2833 fprintf (lra_dump_file, " (%d) ", i);
2834 for (; *p != '\0' && *p != ',' && *p != '#'; p++)
2835 fputc (*p, lra_dump_file);
2837 fprintf (lra_dump_file, "\n");
2840 /* Right now, for any pair of operands I and J that are required to
2841 match, with J < I, goal_alt_matches[I] is J. Add I to
2842 goal_alt_matched[J]. */
2844 for (i = 0; i < n_operands; i++)
2845 if ((j = goal_alt_matches[i]) >= 0)
2847 for (k = 0; goal_alt_matched[j][k] >= 0; k++)
2849 /* We allow matching one output operand and several input
2850 operands. */
2851 lra_assert (k == 0
2852 || (curr_static_id->operand[j].type == OP_OUT
2853 && curr_static_id->operand[i].type == OP_IN
2854 && (curr_static_id->operand
2855 [goal_alt_matched[j][0]].type == OP_IN)));
2856 goal_alt_matched[j][k] = i;
2857 goal_alt_matched[j][k + 1] = -1;
2860 for (i = 0; i < n_operands; i++)
2861 goal_alt_win[i] |= goal_alt_match_win[i];
2863 /* Any constants that aren't allowed and can't be reloaded into
2864 registers are here changed into memory references. */
2865 for (i = 0; i < n_operands; i++)
2866 if (goal_alt_win[i])
2868 int regno;
2869 enum reg_class new_class;
2870 rtx reg = *curr_id->operand_loc[i];
2872 if (GET_CODE (reg) == SUBREG)
2873 reg = SUBREG_REG (reg);
2875 if (REG_P (reg) && (regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER)
2877 bool ok_p = in_class_p (reg, goal_alt[i], &new_class);
2879 if (new_class != NO_REGS && get_reg_class (regno) != new_class)
2881 lra_assert (ok_p);
2882 change_class (regno, new_class, " Change", true);
2886 else
2888 const char *constraint;
2889 char c;
2890 rtx op = *curr_id->operand_loc[i];
2891 rtx subreg = NULL_RTX;
2892 enum machine_mode mode = curr_operand_mode[i];
2894 if (GET_CODE (op) == SUBREG)
2896 subreg = op;
2897 op = SUBREG_REG (op);
2898 mode = GET_MODE (op);
2901 if (CONST_POOL_OK_P (mode, op)
2902 && ((targetm.preferred_reload_class
2903 (op, (enum reg_class) goal_alt[i]) == NO_REGS)
2904 || no_input_reloads_p))
2906 rtx tem = force_const_mem (mode, op);
2908 change_p = true;
2909 if (subreg != NULL_RTX)
2910 tem = gen_rtx_SUBREG (mode, tem, SUBREG_BYTE (subreg));
2912 *curr_id->operand_loc[i] = tem;
2913 lra_update_dup (curr_id, i);
2914 process_address (i, &before, &after);
2916 /* If the alternative accepts constant pool refs directly
2917 there will be no reload needed at all. */
2918 if (subreg != NULL_RTX)
2919 continue;
2920 /* Skip alternatives before the one requested. */
2921 constraint = (curr_static_id->operand_alternative
2922 [goal_alt_number * n_operands + i].constraint);
2923 for (;
2924 (c = *constraint) && c != ',' && c != '#';
2925 constraint += CONSTRAINT_LEN (c, constraint))
2927 if (c == TARGET_MEM_CONSTRAINT || c == 'o')
2928 break;
2929 #ifdef EXTRA_CONSTRAINT_STR
2930 if (EXTRA_MEMORY_CONSTRAINT (c, constraint)
2931 && EXTRA_CONSTRAINT_STR (tem, c, constraint))
2932 break;
2933 #endif
2935 if (c == '\0' || c == ',' || c == '#')
2936 continue;
2938 goal_alt_win[i] = true;
2942 for (i = 0; i < n_operands; i++)
2944 rtx old, new_reg;
2945 rtx op = *curr_id->operand_loc[i];
2947 if (goal_alt_win[i])
2949 if (goal_alt[i] == NO_REGS
2950 && REG_P (op)
2951 /* When we assign NO_REGS it means that we will not
2952 assign a hard register to the scratch pseudo by
2953 assigment pass and the scratch pseudo will be
2954 spilled. Spilled scratch pseudos are transformed
2955 back to scratches at the LRA end. */
2956 && lra_former_scratch_operand_p (curr_insn, i))
2957 change_class (REGNO (op), NO_REGS, " Change", true);
2958 continue;
2961 /* Operands that match previous ones have already been handled. */
2962 if (goal_alt_matches[i] >= 0)
2963 continue;
2965 /* We should not have an operand with a non-offsettable address
2966 appearing where an offsettable address will do. It also may
2967 be a case when the address should be special in other words
2968 not a general one (e.g. it needs no index reg). */
2969 if (goal_alt_matched[i][0] == -1 && goal_alt_offmemok[i] && MEM_P (op))
2971 enum reg_class rclass;
2972 rtx *loc = &XEXP (op, 0);
2973 enum rtx_code code = GET_CODE (*loc);
2975 push_to_sequence (before);
2976 rclass = base_reg_class (GET_MODE (op), MEM_ADDR_SPACE (op),
2977 MEM, SCRATCH);
2978 if (GET_RTX_CLASS (code) == RTX_AUTOINC)
2979 new_reg = emit_inc (rclass, *loc, *loc,
2980 /* This value does not matter for MODIFY. */
2981 GET_MODE_SIZE (GET_MODE (op)));
2982 else if (get_reload_reg (OP_IN, Pmode, *loc, rclass,
2983 "offsetable address", &new_reg))
2984 lra_emit_move (new_reg, *loc);
2985 before = get_insns ();
2986 end_sequence ();
2987 *loc = new_reg;
2988 lra_update_dup (curr_id, i);
2990 else if (goal_alt_matched[i][0] == -1)
2992 enum machine_mode mode;
2993 rtx reg, *loc;
2994 int hard_regno, byte;
2995 enum op_type type = curr_static_id->operand[i].type;
2997 loc = curr_id->operand_loc[i];
2998 mode = curr_operand_mode[i];
2999 if (GET_CODE (*loc) == SUBREG)
3001 reg = SUBREG_REG (*loc);
3002 byte = SUBREG_BYTE (*loc);
3003 if (REG_P (reg)
3004 /* Strict_low_part requires reload the register not
3005 the sub-register. */
3006 && (curr_static_id->operand[i].strict_low
3007 || (GET_MODE_SIZE (mode)
3008 <= GET_MODE_SIZE (GET_MODE (reg))
3009 && (hard_regno
3010 = get_try_hard_regno (REGNO (reg))) >= 0
3011 && (simplify_subreg_regno
3012 (hard_regno,
3013 GET_MODE (reg), byte, mode) < 0)
3014 && (goal_alt[i] == NO_REGS
3015 || (simplify_subreg_regno
3016 (ira_class_hard_regs[goal_alt[i]][0],
3017 GET_MODE (reg), byte, mode) >= 0)))))
3019 loc = &SUBREG_REG (*loc);
3020 mode = GET_MODE (*loc);
3023 old = *loc;
3024 if (get_reload_reg (type, mode, old, goal_alt[i], "", &new_reg)
3025 && type != OP_OUT)
3027 push_to_sequence (before);
3028 lra_emit_move (new_reg, old);
3029 before = get_insns ();
3030 end_sequence ();
3032 *loc = new_reg;
3033 if (type != OP_IN
3034 && find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX)
3036 start_sequence ();
3037 lra_emit_move (type == OP_INOUT ? copy_rtx (old) : old, new_reg);
3038 emit_insn (after);
3039 after = get_insns ();
3040 end_sequence ();
3041 *loc = new_reg;
3043 for (j = 0; j < goal_alt_dont_inherit_ops_num; j++)
3044 if (goal_alt_dont_inherit_ops[j] == i)
3046 lra_set_regno_unique_value (REGNO (new_reg));
3047 break;
3049 lra_update_dup (curr_id, i);
3051 else if (curr_static_id->operand[i].type == OP_IN
3052 && (curr_static_id->operand[goal_alt_matched[i][0]].type
3053 == OP_OUT))
3055 signed char arr[2];
3057 arr[0] = i;
3058 arr[1] = -1;
3059 match_reload (goal_alt_matched[i][0], arr,
3060 goal_alt[i], &before, &after);
3062 else if (curr_static_id->operand[i].type == OP_OUT
3063 && (curr_static_id->operand[goal_alt_matched[i][0]].type
3064 == OP_IN))
3065 match_reload (i, goal_alt_matched[i], goal_alt[i], &before, &after);
3066 else
3067 /* We must generate code in any case when function
3068 process_alt_operands decides that it is possible. */
3069 gcc_unreachable ();
3071 if (before != NULL_RTX || after != NULL_RTX
3072 || max_regno_before != max_reg_num ())
3073 change_p = true;
3074 if (change_p)
3076 lra_update_operator_dups (curr_id);
3077 /* Something changes -- process the insn. */
3078 lra_update_insn_regno_info (curr_insn);
3080 lra_process_new_insns (curr_insn, before, after, "Inserting insn reload");
3081 return change_p;
3084 /* Return true if X is in LIST. */
3085 static bool
3086 in_list_p (rtx x, rtx list)
3088 for (; list != NULL_RTX; list = XEXP (list, 1))
3089 if (XEXP (list, 0) == x)
3090 return true;
3091 return false;
3094 /* Return true if X contains an allocatable hard register (if
3095 HARD_REG_P) or a (spilled if SPILLED_P) pseudo. */
3096 static bool
3097 contains_reg_p (rtx x, bool hard_reg_p, bool spilled_p)
3099 int i, j;
3100 const char *fmt;
3101 enum rtx_code code;
3103 code = GET_CODE (x);
3104 if (REG_P (x))
3106 int regno = REGNO (x);
3107 HARD_REG_SET alloc_regs;
3109 if (hard_reg_p)
3111 if (regno >= FIRST_PSEUDO_REGISTER)
3112 regno = lra_get_regno_hard_regno (regno);
3113 if (regno < 0)
3114 return false;
3115 COMPL_HARD_REG_SET (alloc_regs, lra_no_alloc_regs);
3116 return overlaps_hard_reg_set_p (alloc_regs, GET_MODE (x), regno);
3118 else
3120 if (regno < FIRST_PSEUDO_REGISTER)
3121 return false;
3122 if (! spilled_p)
3123 return true;
3124 return lra_get_regno_hard_regno (regno) < 0;
3127 fmt = GET_RTX_FORMAT (code);
3128 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3130 if (fmt[i] == 'e')
3132 if (contains_reg_p (XEXP (x, i), hard_reg_p, spilled_p))
3133 return true;
3135 else if (fmt[i] == 'E')
3137 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3138 if (contains_reg_p (XVECEXP (x, i, j), hard_reg_p, spilled_p))
3139 return true;
3142 return false;
3145 /* Process all regs in location *LOC and change them on equivalent
3146 substitution. Return true if any change was done. */
3147 static bool
3148 loc_equivalence_change_p (rtx *loc)
3150 rtx subst, reg, x = *loc;
3151 bool result = false;
3152 enum rtx_code code = GET_CODE (x);
3153 const char *fmt;
3154 int i, j;
3156 if (code == SUBREG)
3158 reg = SUBREG_REG (x);
3159 if ((subst = get_equiv_substitution (reg)) != reg
3160 && GET_MODE (subst) == VOIDmode)
3162 /* We cannot reload debug location. Simplify subreg here
3163 while we know the inner mode. */
3164 *loc = simplify_gen_subreg (GET_MODE (x), subst,
3165 GET_MODE (reg), SUBREG_BYTE (x));
3166 return true;
3169 if (code == REG && (subst = get_equiv_substitution (x)) != x)
3171 *loc = subst;
3172 return true;
3175 /* Scan all the operand sub-expressions. */
3176 fmt = GET_RTX_FORMAT (code);
3177 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3179 if (fmt[i] == 'e')
3180 result = loc_equivalence_change_p (&XEXP (x, i)) || result;
3181 else if (fmt[i] == 'E')
3182 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3183 result
3184 = loc_equivalence_change_p (&XVECEXP (x, i, j)) || result;
3186 return result;
3189 /* Maximum allowed number of constraint pass iterations after the last
3190 spill pass. It is for preventing LRA cycling in a bug case. */
3191 #define MAX_CONSTRAINT_ITERATION_NUMBER 30
3193 /* Maximum number of generated reload insns per an insn. It is for
3194 preventing this pass cycling in a bug case. */
3195 #define MAX_RELOAD_INSNS_NUMBER LRA_MAX_INSN_RELOADS
3197 /* The current iteration number of this LRA pass. */
3198 int lra_constraint_iter;
3200 /* The current iteration number of this LRA pass after the last spill
3201 pass. */
3202 int lra_constraint_iter_after_spill;
3204 /* True if we substituted equiv which needs checking register
3205 allocation correctness because the equivalent value contains
3206 allocatable hard registers or when we restore multi-register
3207 pseudo. */
3208 bool lra_risky_transformations_p;
3210 /* Return true if REGNO is referenced in more than one block. */
3211 static bool
3212 multi_block_pseudo_p (int regno)
3214 basic_block bb = NULL;
3215 unsigned int uid;
3216 bitmap_iterator bi;
3218 if (regno < FIRST_PSEUDO_REGISTER)
3219 return false;
3221 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi)
3222 if (bb == NULL)
3223 bb = BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn);
3224 else if (BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn) != bb)
3225 return true;
3226 return false;
3229 /* Return true if LIST contains a deleted insn. */
3230 static bool
3231 contains_deleted_insn_p (rtx list)
3233 for (; list != NULL_RTX; list = XEXP (list, 1))
3234 if (NOTE_P (XEXP (list, 0))
3235 && NOTE_KIND (XEXP (list, 0)) == NOTE_INSN_DELETED)
3236 return true;
3237 return false;
3240 /* Return true if X contains a pseudo dying in INSN. */
3241 static bool
3242 dead_pseudo_p (rtx x, rtx insn)
3244 int i, j;
3245 const char *fmt;
3246 enum rtx_code code;
3248 if (REG_P (x))
3249 return (insn != NULL_RTX
3250 && find_regno_note (insn, REG_DEAD, REGNO (x)) != NULL_RTX);
3251 code = GET_CODE (x);
3252 fmt = GET_RTX_FORMAT (code);
3253 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3255 if (fmt[i] == 'e')
3257 if (dead_pseudo_p (XEXP (x, i), insn))
3258 return true;
3260 else if (fmt[i] == 'E')
3262 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3263 if (dead_pseudo_p (XVECEXP (x, i, j), insn))
3264 return true;
3267 return false;
3270 /* Return true if INSN contains a dying pseudo in INSN right hand
3271 side. */
3272 static bool
3273 insn_rhs_dead_pseudo_p (rtx insn)
3275 rtx set = single_set (insn);
3277 gcc_assert (set != NULL);
3278 return dead_pseudo_p (SET_SRC (set), insn);
3281 /* Return true if any init insn of REGNO contains a dying pseudo in
3282 insn right hand side. */
3283 static bool
3284 init_insn_rhs_dead_pseudo_p (int regno)
3286 rtx insns = ira_reg_equiv[regno].init_insns;
3288 if (insns == NULL)
3289 return false;
3290 if (INSN_P (insns))
3291 return insn_rhs_dead_pseudo_p (insns);
3292 for (; insns != NULL_RTX; insns = XEXP (insns, 1))
3293 if (insn_rhs_dead_pseudo_p (XEXP (insns, 0)))
3294 return true;
3295 return false;
3298 /* Entry function of LRA constraint pass. Return true if the
3299 constraint pass did change the code. */
3300 bool
3301 lra_constraints (bool first_p)
3303 bool changed_p;
3304 int i, hard_regno, new_insns_num;
3305 unsigned int min_len, new_min_len, uid;
3306 rtx set, x, reg, dest_reg;
3307 basic_block last_bb;
3308 bitmap_head equiv_insn_bitmap;
3309 bitmap_iterator bi;
3311 lra_constraint_iter++;
3312 if (lra_dump_file != NULL)
3313 fprintf (lra_dump_file, "\n********** Local #%d: **********\n\n",
3314 lra_constraint_iter);
3315 lra_constraint_iter_after_spill++;
3316 if (lra_constraint_iter_after_spill > MAX_CONSTRAINT_ITERATION_NUMBER)
3317 internal_error
3318 ("Maximum number of LRA constraint passes is achieved (%d)\n",
3319 MAX_CONSTRAINT_ITERATION_NUMBER);
3320 changed_p = false;
3321 lra_risky_transformations_p = false;
3322 new_insn_uid_start = get_max_uid ();
3323 new_regno_start = first_p ? lra_constraint_new_regno_start : max_reg_num ();
3324 bitmap_initialize (&equiv_insn_bitmap, &reg_obstack);
3325 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
3326 if (lra_reg_info[i].nrefs != 0)
3328 ira_reg_equiv[i].profitable_p = true;
3329 reg = regno_reg_rtx[i];
3330 if ((hard_regno = lra_get_regno_hard_regno (i)) >= 0)
3332 int j, nregs = hard_regno_nregs[hard_regno][PSEUDO_REGNO_MODE (i)];
3334 for (j = 0; j < nregs; j++)
3335 df_set_regs_ever_live (hard_regno + j, true);
3337 else if ((x = get_equiv_substitution (reg)) != reg)
3339 bool pseudo_p = contains_reg_p (x, false, false);
3340 rtx set, insn;
3342 /* After RTL transformation, we can not guarantee that
3343 pseudo in the substitution was not reloaded which might
3344 make equivalence invalid. For example, in reverse
3345 equiv of p0
3347 p0 <- ...
3349 equiv_mem <- p0
3351 the memory address register was reloaded before the 2nd
3352 insn. */
3353 if ((! first_p && pseudo_p)
3354 /* We don't use DF for compilation speed sake. So it
3355 is problematic to update live info when we use an
3356 equivalence containing pseudos in more than one
3357 BB. */
3358 || (pseudo_p && multi_block_pseudo_p (i))
3359 /* If an init insn was deleted for some reason, cancel
3360 the equiv. We could update the equiv insns after
3361 transformations including an equiv insn deletion
3362 but it is not worthy as such cases are extremely
3363 rare. */
3364 || contains_deleted_insn_p (ira_reg_equiv[i].init_insns)
3365 /* If it is not a reverse equivalence, we check that a
3366 pseudo in rhs of the init insn is not dying in the
3367 insn. Otherwise, the live info at the beginning of
3368 the corresponding BB might be wrong after we
3369 removed the insn. When the equiv can be a
3370 constant, the right hand side of the init insn can
3371 be a pseudo. */
3372 || (! ((insn = ira_reg_equiv[i].init_insns) != NULL_RTX
3373 && INSN_P (insn)
3374 && (set = single_set (insn)) != NULL_RTX
3375 && REG_P (SET_DEST (set))
3376 && (int) REGNO (SET_DEST (set)) == i)
3377 && init_insn_rhs_dead_pseudo_p (i))
3378 /* Prevent access beyond equivalent memory for
3379 paradoxical subregs. */
3380 || (MEM_P (x)
3381 && (GET_MODE_SIZE (lra_reg_info[i].biggest_mode)
3382 > GET_MODE_SIZE (GET_MODE (x)))))
3383 ira_reg_equiv[i].defined_p = false;
3384 if (contains_reg_p (x, false, true))
3385 ira_reg_equiv[i].profitable_p = false;
3386 if (get_equiv_substitution (reg) != reg)
3387 bitmap_ior_into (&equiv_insn_bitmap, &lra_reg_info[i].insn_bitmap);
3390 /* We should add all insns containing pseudos which should be
3391 substituted by their equivalences. */
3392 EXECUTE_IF_SET_IN_BITMAP (&equiv_insn_bitmap, 0, uid, bi)
3393 lra_push_insn_by_uid (uid);
3394 lra_eliminate (false);
3395 min_len = lra_insn_stack_length ();
3396 new_insns_num = 0;
3397 last_bb = NULL;
3398 changed_p = false;
3399 while ((new_min_len = lra_insn_stack_length ()) != 0)
3401 curr_insn = lra_pop_insn ();
3402 --new_min_len;
3403 curr_bb = BLOCK_FOR_INSN (curr_insn);
3404 if (curr_bb != last_bb)
3406 last_bb = curr_bb;
3407 bb_reload_num = lra_curr_reload_num;
3409 if (min_len > new_min_len)
3411 min_len = new_min_len;
3412 new_insns_num = 0;
3414 if (new_insns_num > MAX_RELOAD_INSNS_NUMBER)
3415 internal_error
3416 ("Max. number of generated reload insns per insn is achieved (%d)\n",
3417 MAX_RELOAD_INSNS_NUMBER);
3418 new_insns_num++;
3419 if (DEBUG_INSN_P (curr_insn))
3421 /* We need to check equivalence in debug insn and change
3422 pseudo to the equivalent value if necessary. */
3423 curr_id = lra_get_insn_recog_data (curr_insn);
3424 if (bitmap_bit_p (&equiv_insn_bitmap, INSN_UID (curr_insn))
3425 && loc_equivalence_change_p (curr_id->operand_loc[0]))
3427 lra_update_insn_regno_info (curr_insn);
3428 changed_p = true;
3431 else if (INSN_P (curr_insn))
3433 if ((set = single_set (curr_insn)) != NULL_RTX)
3435 dest_reg = SET_DEST (set);
3436 /* The equivalence pseudo could be set up as SUBREG in a
3437 case when it is a call restore insn in a mode
3438 different from the pseudo mode. */
3439 if (GET_CODE (dest_reg) == SUBREG)
3440 dest_reg = SUBREG_REG (dest_reg);
3441 if ((REG_P (dest_reg)
3442 && (x = get_equiv_substitution (dest_reg)) != dest_reg
3443 /* Remove insns which set up a pseudo whose value
3444 can not be changed. Such insns might be not in
3445 init_insns because we don't update equiv data
3446 during insn transformations.
3448 As an example, let suppose that a pseudo got
3449 hard register and on the 1st pass was not
3450 changed to equivalent constant. We generate an
3451 additional insn setting up the pseudo because of
3452 secondary memory movement. Then the pseudo is
3453 spilled and we use the equiv constant. In this
3454 case we should remove the additional insn and
3455 this insn is not init_insns list. */
3456 && (! MEM_P (x) || MEM_READONLY_P (x)
3457 || in_list_p (curr_insn,
3458 ira_reg_equiv
3459 [REGNO (dest_reg)].init_insns)))
3460 || (((x = get_equiv_substitution (SET_SRC (set)))
3461 != SET_SRC (set))
3462 && in_list_p (curr_insn,
3463 ira_reg_equiv
3464 [REGNO (SET_SRC (set))].init_insns)))
3466 /* This is equiv init insn of pseudo which did not get a
3467 hard register -- remove the insn. */
3468 if (lra_dump_file != NULL)
3470 fprintf (lra_dump_file,
3471 " Removing equiv init insn %i (freq=%d)\n",
3472 INSN_UID (curr_insn),
3473 BLOCK_FOR_INSN (curr_insn)->frequency);
3474 dump_insn_slim (lra_dump_file, curr_insn);
3476 if (contains_reg_p (x, true, false))
3477 lra_risky_transformations_p = true;
3478 lra_set_insn_deleted (curr_insn);
3479 continue;
3482 curr_id = lra_get_insn_recog_data (curr_insn);
3483 curr_static_id = curr_id->insn_static_data;
3484 init_curr_insn_input_reloads ();
3485 init_curr_operand_mode ();
3486 if (curr_insn_transform ())
3487 changed_p = true;
3488 /* Check non-transformed insns too for equiv change as USE
3489 or CLOBBER don't need reloads but can contain pseudos
3490 being changed on their equivalences. */
3491 else if (bitmap_bit_p (&equiv_insn_bitmap, INSN_UID (curr_insn))
3492 && loc_equivalence_change_p (&PATTERN (curr_insn)))
3494 lra_update_insn_regno_info (curr_insn);
3495 changed_p = true;
3499 bitmap_clear (&equiv_insn_bitmap);
3500 /* If we used a new hard regno, changed_p should be true because the
3501 hard reg is assigned to a new pseudo. */
3502 #ifdef ENABLE_CHECKING
3503 if (! changed_p)
3505 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
3506 if (lra_reg_info[i].nrefs != 0
3507 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
3509 int j, nregs = hard_regno_nregs[hard_regno][PSEUDO_REGNO_MODE (i)];
3511 for (j = 0; j < nregs; j++)
3512 lra_assert (df_regs_ever_live_p (hard_regno + j));
3515 #endif
3516 return changed_p;
3519 /* Initiate the LRA constraint pass. It is done once per
3520 function. */
3521 void
3522 lra_constraints_init (void)
3526 /* Finalize the LRA constraint pass. It is done once per
3527 function. */
3528 void
3529 lra_constraints_finish (void)
3535 /* This page contains code to do inheritance/split
3536 transformations. */
3538 /* Number of reloads passed so far in current EBB. */
3539 static int reloads_num;
3541 /* Number of calls passed so far in current EBB. */
3542 static int calls_num;
3544 /* Current reload pseudo check for validity of elements in
3545 USAGE_INSNS. */
3546 static int curr_usage_insns_check;
3548 /* Info about last usage of registers in EBB to do inheritance/split
3549 transformation. Inheritance transformation is done from a spilled
3550 pseudo and split transformations from a hard register or a pseudo
3551 assigned to a hard register. */
3552 struct usage_insns
3554 /* If the value is equal to CURR_USAGE_INSNS_CHECK, then the member
3555 value INSNS is valid. The insns is chain of optional debug insns
3556 and a finishing non-debug insn using the corresponding reg. */
3557 int check;
3558 /* Value of global reloads_num at the last insn in INSNS. */
3559 int reloads_num;
3560 /* Value of global reloads_nums at the last insn in INSNS. */
3561 int calls_num;
3562 /* It can be true only for splitting. And it means that the restore
3563 insn should be put after insn given by the following member. */
3564 bool after_p;
3565 /* Next insns in the current EBB which use the original reg and the
3566 original reg value is not changed between the current insn and
3567 the next insns. In order words, e.g. for inheritance, if we need
3568 to use the original reg value again in the next insns we can try
3569 to use the value in a hard register from a reload insn of the
3570 current insn. */
3571 rtx insns;
3574 /* Map: regno -> corresponding pseudo usage insns. */
3575 static struct usage_insns *usage_insns;
3577 static void
3578 setup_next_usage_insn (int regno, rtx insn, int reloads_num, bool after_p)
3580 usage_insns[regno].check = curr_usage_insns_check;
3581 usage_insns[regno].insns = insn;
3582 usage_insns[regno].reloads_num = reloads_num;
3583 usage_insns[regno].calls_num = calls_num;
3584 usage_insns[regno].after_p = after_p;
3587 /* The function is used to form list REGNO usages which consists of
3588 optional debug insns finished by a non-debug insn using REGNO.
3589 RELOADS_NUM is current number of reload insns processed so far. */
3590 static void
3591 add_next_usage_insn (int regno, rtx insn, int reloads_num)
3593 rtx next_usage_insns;
3595 if (usage_insns[regno].check == curr_usage_insns_check
3596 && (next_usage_insns = usage_insns[regno].insns) != NULL_RTX
3597 && DEBUG_INSN_P (insn))
3599 /* Check that we did not add the debug insn yet. */
3600 if (next_usage_insns != insn
3601 && (GET_CODE (next_usage_insns) != INSN_LIST
3602 || XEXP (next_usage_insns, 0) != insn))
3603 usage_insns[regno].insns = gen_rtx_INSN_LIST (VOIDmode, insn,
3604 next_usage_insns);
3606 else if (NONDEBUG_INSN_P (insn))
3607 setup_next_usage_insn (regno, insn, reloads_num, false);
3608 else
3609 usage_insns[regno].check = 0;
3612 /* Replace all references to register OLD_REGNO in *LOC with pseudo
3613 register NEW_REG. Return true if any change was made. */
3614 static bool
3615 substitute_pseudo (rtx *loc, int old_regno, rtx new_reg)
3617 rtx x = *loc;
3618 bool result = false;
3619 enum rtx_code code;
3620 const char *fmt;
3621 int i, j;
3623 if (x == NULL_RTX)
3624 return false;
3626 code = GET_CODE (x);
3627 if (code == REG && (int) REGNO (x) == old_regno)
3629 enum machine_mode mode = GET_MODE (*loc);
3630 enum machine_mode inner_mode = GET_MODE (new_reg);
3632 if (mode != inner_mode)
3634 if (GET_MODE_SIZE (mode) >= GET_MODE_SIZE (inner_mode)
3635 || ! SCALAR_INT_MODE_P (inner_mode))
3636 new_reg = gen_rtx_SUBREG (mode, new_reg, 0);
3637 else
3638 new_reg = gen_lowpart_SUBREG (mode, new_reg);
3640 *loc = new_reg;
3641 return true;
3644 /* Scan all the operand sub-expressions. */
3645 fmt = GET_RTX_FORMAT (code);
3646 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3648 if (fmt[i] == 'e')
3650 if (substitute_pseudo (&XEXP (x, i), old_regno, new_reg))
3651 result = true;
3653 else if (fmt[i] == 'E')
3655 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3656 if (substitute_pseudo (&XVECEXP (x, i, j), old_regno, new_reg))
3657 result = true;
3660 return result;
3663 /* Return first non-debug insn in list USAGE_INSNS. */
3664 static rtx
3665 skip_usage_debug_insns (rtx usage_insns)
3667 rtx insn;
3669 /* Skip debug insns. */
3670 for (insn = usage_insns;
3671 insn != NULL_RTX && GET_CODE (insn) == INSN_LIST;
3672 insn = XEXP (insn, 1))
3674 return insn;
3677 /* Return true if we need secondary memory moves for insn in
3678 USAGE_INSNS after inserting inherited pseudo of class INHER_CL
3679 into the insn. */
3680 static bool
3681 check_secondary_memory_needed_p (enum reg_class inher_cl ATTRIBUTE_UNUSED,
3682 rtx usage_insns ATTRIBUTE_UNUSED)
3684 #ifndef SECONDARY_MEMORY_NEEDED
3685 return false;
3686 #else
3687 rtx insn, set, dest;
3688 enum reg_class cl;
3690 if (inher_cl == ALL_REGS
3691 || (insn = skip_usage_debug_insns (usage_insns)) == NULL_RTX)
3692 return false;
3693 lra_assert (INSN_P (insn));
3694 if ((set = single_set (insn)) == NULL_RTX || ! REG_P (SET_DEST (set)))
3695 return false;
3696 dest = SET_DEST (set);
3697 if (! REG_P (dest))
3698 return false;
3699 lra_assert (inher_cl != NO_REGS);
3700 cl = get_reg_class (REGNO (dest));
3701 return (cl != NO_REGS && cl != ALL_REGS
3702 && SECONDARY_MEMORY_NEEDED (inher_cl, cl, GET_MODE (dest)));
3703 #endif
3706 /* Registers involved in inheritance/split in the current EBB
3707 (inheritance/split pseudos and original registers). */
3708 static bitmap_head check_only_regs;
3710 /* Do inheritance transformations for insn INSN, which defines (if
3711 DEF_P) or uses ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which
3712 instruction in the EBB next uses ORIGINAL_REGNO; it has the same
3713 form as the "insns" field of usage_insns. Return true if we
3714 succeed in such transformation.
3716 The transformations look like:
3718 p <- ... i <- ...
3719 ... p <- i (new insn)
3720 ... =>
3721 <- ... p ... <- ... i ...
3723 ... i <- p (new insn)
3724 <- ... p ... <- ... i ...
3725 ... =>
3726 <- ... p ... <- ... i ...
3727 where p is a spilled original pseudo and i is a new inheritance pseudo.
3730 The inheritance pseudo has the smallest class of two classes CL and
3731 class of ORIGINAL REGNO. */
3732 static bool
3733 inherit_reload_reg (bool def_p, int original_regno,
3734 enum reg_class cl, rtx insn, rtx next_usage_insns)
3736 enum reg_class rclass = lra_get_allocno_class (original_regno);
3737 rtx original_reg = regno_reg_rtx[original_regno];
3738 rtx new_reg, new_insns, usage_insn;
3740 lra_assert (! usage_insns[original_regno].after_p);
3741 if (lra_dump_file != NULL)
3742 fprintf (lra_dump_file,
3743 " <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n");
3744 if (! ira_reg_classes_intersect_p[cl][rclass])
3746 if (lra_dump_file != NULL)
3748 fprintf (lra_dump_file,
3749 " Rejecting inheritance for %d "
3750 "because of disjoint classes %s and %s\n",
3751 original_regno, reg_class_names[cl],
3752 reg_class_names[rclass]);
3753 fprintf (lra_dump_file,
3754 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
3756 return false;
3758 if ((ira_class_subset_p[cl][rclass] && cl != rclass)
3759 /* We don't use a subset of two classes because it can be
3760 NO_REGS. This transformation is still profitable in most
3761 cases even if the classes are not intersected as register
3762 move is probably cheaper than a memory load. */
3763 || ira_class_hard_regs_num[cl] < ira_class_hard_regs_num[rclass])
3765 if (lra_dump_file != NULL)
3766 fprintf (lra_dump_file, " Use smallest class of %s and %s\n",
3767 reg_class_names[cl], reg_class_names[rclass]);
3769 rclass = cl;
3771 if (check_secondary_memory_needed_p (cl, next_usage_insns))
3773 /* Reject inheritance resulting in secondary memory moves.
3774 Otherwise, there is a danger in LRA cycling. Also such
3775 transformation will be unprofitable. */
3776 if (lra_dump_file != NULL)
3778 rtx insn = skip_usage_debug_insns (next_usage_insns);
3779 rtx set = single_set (insn);
3781 lra_assert (set != NULL_RTX);
3783 rtx dest = SET_DEST (set);
3785 lra_assert (REG_P (dest));
3786 fprintf (lra_dump_file,
3787 " Rejecting inheritance for insn %d(%s)<-%d(%s) "
3788 "as secondary mem is needed\n",
3789 REGNO (dest), reg_class_names[get_reg_class (REGNO (dest))],
3790 original_regno, reg_class_names[cl]);
3791 fprintf (lra_dump_file,
3792 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
3794 return false;
3796 new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg,
3797 rclass, "inheritance");
3798 start_sequence ();
3799 if (def_p)
3800 emit_move_insn (original_reg, new_reg);
3801 else
3802 emit_move_insn (new_reg, original_reg);
3803 new_insns = get_insns ();
3804 end_sequence ();
3805 if (NEXT_INSN (new_insns) != NULL_RTX)
3807 if (lra_dump_file != NULL)
3809 fprintf (lra_dump_file,
3810 " Rejecting inheritance %d->%d "
3811 "as it results in 2 or more insns:\n",
3812 original_regno, REGNO (new_reg));
3813 dump_rtl_slim (lra_dump_file, new_insns, NULL_RTX, -1, 0);
3814 fprintf (lra_dump_file,
3815 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
3817 return false;
3819 substitute_pseudo (&insn, original_regno, new_reg);
3820 lra_update_insn_regno_info (insn);
3821 if (! def_p)
3822 /* We now have a new usage insn for original regno. */
3823 setup_next_usage_insn (original_regno, new_insns, reloads_num, false);
3824 if (lra_dump_file != NULL)
3825 fprintf (lra_dump_file, " Original reg change %d->%d (bb%d):\n",
3826 original_regno, REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
3827 lra_reg_info[REGNO (new_reg)].restore_regno = original_regno;
3828 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
3829 bitmap_set_bit (&check_only_regs, original_regno);
3830 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
3831 if (def_p)
3832 lra_process_new_insns (insn, NULL_RTX, new_insns,
3833 "Add original<-inheritance");
3834 else
3835 lra_process_new_insns (insn, new_insns, NULL_RTX,
3836 "Add inheritance<-original");
3837 while (next_usage_insns != NULL_RTX)
3839 if (GET_CODE (next_usage_insns) != INSN_LIST)
3841 usage_insn = next_usage_insns;
3842 lra_assert (NONDEBUG_INSN_P (usage_insn));
3843 next_usage_insns = NULL;
3845 else
3847 usage_insn = XEXP (next_usage_insns, 0);
3848 lra_assert (DEBUG_INSN_P (usage_insn));
3849 next_usage_insns = XEXP (next_usage_insns, 1);
3851 substitute_pseudo (&usage_insn, original_regno, new_reg);
3852 lra_update_insn_regno_info (usage_insn);
3853 if (lra_dump_file != NULL)
3855 fprintf (lra_dump_file,
3856 " Inheritance reuse change %d->%d (bb%d):\n",
3857 original_regno, REGNO (new_reg),
3858 BLOCK_FOR_INSN (usage_insn)->index);
3859 dump_insn_slim (lra_dump_file, usage_insn);
3862 if (lra_dump_file != NULL)
3863 fprintf (lra_dump_file,
3864 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
3865 return true;
3868 /* Return true if we need a caller save/restore for pseudo REGNO which
3869 was assigned to a hard register. */
3870 static inline bool
3871 need_for_call_save_p (int regno)
3873 lra_assert (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] >= 0);
3874 return (usage_insns[regno].calls_num < calls_num
3875 && (overlaps_hard_reg_set_p
3876 (call_used_reg_set,
3877 PSEUDO_REGNO_MODE (regno), reg_renumber[regno])));
3880 /* Global registers occuring in the current EBB. */
3881 static bitmap_head ebb_global_regs;
3883 /* Return true if we need a split for hard register REGNO or pseudo
3884 REGNO which was assigned to a hard register.
3885 POTENTIAL_RELOAD_HARD_REGS contains hard registers which might be
3886 used for reloads since the EBB end. It is an approximation of the
3887 used hard registers in the split range. The exact value would
3888 require expensive calculations. If we were aggressive with
3889 splitting because of the approximation, the split pseudo will save
3890 the same hard register assignment and will be removed in the undo
3891 pass. We still need the approximation because too aggressive
3892 splitting would result in too inaccurate cost calculation in the
3893 assignment pass because of too many generated moves which will be
3894 probably removed in the undo pass. */
3895 static inline bool
3896 need_for_split_p (HARD_REG_SET potential_reload_hard_regs, int regno)
3898 int hard_regno = regno < FIRST_PSEUDO_REGISTER ? regno : reg_renumber[regno];
3900 lra_assert (hard_regno >= 0);
3901 return ((TEST_HARD_REG_BIT (potential_reload_hard_regs, hard_regno)
3902 /* Don't split eliminable hard registers, otherwise we can
3903 split hard registers like hard frame pointer, which
3904 lives on BB start/end according to DF-infrastructure,
3905 when there is a pseudo assigned to the register and
3906 living in the same BB. */
3907 && (regno >= FIRST_PSEUDO_REGISTER
3908 || ! TEST_HARD_REG_BIT (eliminable_regset, hard_regno))
3909 && ! TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno)
3910 /* We need at least 2 reloads to make pseudo splitting
3911 profitable. We should provide hard regno splitting in
3912 any case to solve 1st insn scheduling problem when
3913 moving hard register definition up might result in
3914 impossibility to find hard register for reload pseudo of
3915 small register class. */
3916 && (usage_insns[regno].reloads_num
3917 + (regno < FIRST_PSEUDO_REGISTER ? 0 : 2) < reloads_num)
3918 && (regno < FIRST_PSEUDO_REGISTER
3919 /* For short living pseudos, spilling + inheritance can
3920 be considered a substitution for splitting.
3921 Therefore we do not splitting for local pseudos. It
3922 decreases also aggressiveness of splitting. The
3923 minimal number of references is chosen taking into
3924 account that for 2 references splitting has no sense
3925 as we can just spill the pseudo. */
3926 || (regno >= FIRST_PSEUDO_REGISTER
3927 && lra_reg_info[regno].nrefs > 3
3928 && bitmap_bit_p (&ebb_global_regs, regno))))
3929 || (regno >= FIRST_PSEUDO_REGISTER && need_for_call_save_p (regno)));
3932 /* Return class for the split pseudo created from original pseudo with
3933 ALLOCNO_CLASS and MODE which got a hard register HARD_REGNO. We
3934 choose subclass of ALLOCNO_CLASS which contains HARD_REGNO and
3935 results in no secondary memory movements. */
3936 static enum reg_class
3937 choose_split_class (enum reg_class allocno_class,
3938 int hard_regno ATTRIBUTE_UNUSED,
3939 enum machine_mode mode ATTRIBUTE_UNUSED)
3941 #ifndef SECONDARY_MEMORY_NEEDED
3942 return allocno_class;
3943 #else
3944 int i;
3945 enum reg_class cl, best_cl = NO_REGS;
3946 enum reg_class hard_reg_class ATTRIBUTE_UNUSED
3947 = REGNO_REG_CLASS (hard_regno);
3949 if (! SECONDARY_MEMORY_NEEDED (allocno_class, allocno_class, mode)
3950 && TEST_HARD_REG_BIT (reg_class_contents[allocno_class], hard_regno))
3951 return allocno_class;
3952 for (i = 0;
3953 (cl = reg_class_subclasses[allocno_class][i]) != LIM_REG_CLASSES;
3954 i++)
3955 if (! SECONDARY_MEMORY_NEEDED (cl, hard_reg_class, mode)
3956 && ! SECONDARY_MEMORY_NEEDED (hard_reg_class, cl, mode)
3957 && TEST_HARD_REG_BIT (reg_class_contents[cl], hard_regno)
3958 && (best_cl == NO_REGS
3959 || ira_class_hard_regs_num[best_cl] < ira_class_hard_regs_num[cl]))
3960 best_cl = cl;
3961 return best_cl;
3962 #endif
3965 /* Do split transformations for insn INSN, which defines or uses
3966 ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which instruction in
3967 the EBB next uses ORIGINAL_REGNO; it has the same form as the
3968 "insns" field of usage_insns.
3970 The transformations look like:
3972 p <- ... p <- ...
3973 ... s <- p (new insn -- save)
3974 ... =>
3975 ... p <- s (new insn -- restore)
3976 <- ... p ... <- ... p ...
3978 <- ... p ... <- ... p ...
3979 ... s <- p (new insn -- save)
3980 ... =>
3981 ... p <- s (new insn -- restore)
3982 <- ... p ... <- ... p ...
3984 where p is an original pseudo got a hard register or a hard
3985 register and s is a new split pseudo. The save is put before INSN
3986 if BEFORE_P is true. Return true if we succeed in such
3987 transformation. */
3988 static bool
3989 split_reg (bool before_p, int original_regno, rtx insn, rtx next_usage_insns)
3991 enum reg_class rclass;
3992 rtx original_reg;
3993 int hard_regno;
3994 rtx new_reg, save, restore, usage_insn;
3995 bool after_p;
3996 bool call_save_p;
3998 if (original_regno < FIRST_PSEUDO_REGISTER)
4000 rclass = ira_allocno_class_translate[REGNO_REG_CLASS (original_regno)];
4001 hard_regno = original_regno;
4002 call_save_p = false;
4004 else
4006 hard_regno = reg_renumber[original_regno];
4007 rclass = lra_get_allocno_class (original_regno);
4008 original_reg = regno_reg_rtx[original_regno];
4009 call_save_p = need_for_call_save_p (original_regno);
4011 original_reg = regno_reg_rtx[original_regno];
4012 lra_assert (hard_regno >= 0);
4013 if (lra_dump_file != NULL)
4014 fprintf (lra_dump_file,
4015 " ((((((((((((((((((((((((((((((((((((((((((((((((\n");
4016 if (call_save_p)
4018 enum machine_mode sec_mode;
4020 #ifdef SECONDARY_MEMORY_NEEDED_MODE
4021 sec_mode = SECONDARY_MEMORY_NEEDED_MODE (GET_MODE (original_reg));
4022 #else
4023 sec_mode = GET_MODE (original_reg);
4024 #endif
4025 new_reg = lra_create_new_reg (sec_mode, NULL_RTX,
4026 NO_REGS, "save");
4028 else
4030 rclass = choose_split_class (rclass, hard_regno,
4031 GET_MODE (original_reg));
4032 if (rclass == NO_REGS)
4034 if (lra_dump_file != NULL)
4036 fprintf (lra_dump_file,
4037 " Rejecting split of %d(%s): "
4038 "no good reg class for %d(%s)\n",
4039 original_regno,
4040 reg_class_names[lra_get_allocno_class (original_regno)],
4041 hard_regno,
4042 reg_class_names[REGNO_REG_CLASS (hard_regno)]);
4043 fprintf
4044 (lra_dump_file,
4045 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
4047 return false;
4049 new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg,
4050 rclass, "split");
4051 reg_renumber[REGNO (new_reg)] = hard_regno;
4053 save = emit_spill_move (true, new_reg, original_reg);
4054 if (NEXT_INSN (save) != NULL_RTX)
4056 lra_assert (! call_save_p);
4057 if (lra_dump_file != NULL)
4059 fprintf
4060 (lra_dump_file,
4061 " Rejecting split %d->%d resulting in > 2 %s save insns:\n",
4062 original_regno, REGNO (new_reg), call_save_p ? "call" : "");
4063 dump_rtl_slim (lra_dump_file, save, NULL_RTX, -1, 0);
4064 fprintf (lra_dump_file,
4065 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
4067 return false;
4069 restore = emit_spill_move (false, new_reg, original_reg);
4070 if (NEXT_INSN (restore) != NULL_RTX)
4072 lra_assert (! call_save_p);
4073 if (lra_dump_file != NULL)
4075 fprintf (lra_dump_file,
4076 " Rejecting split %d->%d "
4077 "resulting in > 2 %s restore insns:\n",
4078 original_regno, REGNO (new_reg), call_save_p ? "call" : "");
4079 dump_rtl_slim (lra_dump_file, restore, NULL_RTX, -1, 0);
4080 fprintf (lra_dump_file,
4081 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
4083 return false;
4085 after_p = usage_insns[original_regno].after_p;
4086 lra_reg_info[REGNO (new_reg)].restore_regno = original_regno;
4087 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
4088 bitmap_set_bit (&check_only_regs, original_regno);
4089 bitmap_set_bit (&lra_split_regs, REGNO (new_reg));
4090 for (;;)
4092 if (GET_CODE (next_usage_insns) != INSN_LIST)
4094 usage_insn = next_usage_insns;
4095 break;
4097 usage_insn = XEXP (next_usage_insns, 0);
4098 lra_assert (DEBUG_INSN_P (usage_insn));
4099 next_usage_insns = XEXP (next_usage_insns, 1);
4100 substitute_pseudo (&usage_insn, original_regno, new_reg);
4101 lra_update_insn_regno_info (usage_insn);
4102 if (lra_dump_file != NULL)
4104 fprintf (lra_dump_file, " Split reuse change %d->%d:\n",
4105 original_regno, REGNO (new_reg));
4106 dump_insn_slim (lra_dump_file, usage_insn);
4109 lra_assert (NOTE_P (usage_insn) || NONDEBUG_INSN_P (usage_insn));
4110 lra_assert (usage_insn != insn || (after_p && before_p));
4111 lra_process_new_insns (usage_insn, after_p ? NULL_RTX : restore,
4112 after_p ? restore : NULL_RTX,
4113 call_save_p
4114 ? "Add reg<-save" : "Add reg<-split");
4115 lra_process_new_insns (insn, before_p ? save : NULL_RTX,
4116 before_p ? NULL_RTX : save,
4117 call_save_p
4118 ? "Add save<-reg" : "Add split<-reg");
4119 if (lra_dump_file != NULL)
4120 fprintf (lra_dump_file,
4121 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
4122 return true;
4125 /* Recognize that we need a split transformation for insn INSN, which
4126 defines or uses REGNO in its insn biggest MODE (we use it only if
4127 REGNO is a hard register). POTENTIAL_RELOAD_HARD_REGS contains
4128 hard registers which might be used for reloads since the EBB end.
4129 Put the save before INSN if BEFORE_P is true. MAX_UID is maximla
4130 uid before starting INSN processing. Return true if we succeed in
4131 such transformation. */
4132 static bool
4133 split_if_necessary (int regno, enum machine_mode mode,
4134 HARD_REG_SET potential_reload_hard_regs,
4135 bool before_p, rtx insn, int max_uid)
4137 bool res = false;
4138 int i, nregs = 1;
4139 rtx next_usage_insns;
4141 if (regno < FIRST_PSEUDO_REGISTER)
4142 nregs = hard_regno_nregs[regno][mode];
4143 for (i = 0; i < nregs; i++)
4144 if (usage_insns[regno + i].check == curr_usage_insns_check
4145 && (next_usage_insns = usage_insns[regno + i].insns) != NULL_RTX
4146 /* To avoid processing the register twice or more. */
4147 && ((GET_CODE (next_usage_insns) != INSN_LIST
4148 && INSN_UID (next_usage_insns) < max_uid)
4149 || (GET_CODE (next_usage_insns) == INSN_LIST
4150 && (INSN_UID (XEXP (next_usage_insns, 0)) < max_uid)))
4151 && need_for_split_p (potential_reload_hard_regs, regno + i)
4152 && split_reg (before_p, regno + i, insn, next_usage_insns))
4153 res = true;
4154 return res;
4157 /* Check only registers living at the current program point in the
4158 current EBB. */
4159 static bitmap_head live_regs;
4161 /* Update live info in EBB given by its HEAD and TAIL insns after
4162 inheritance/split transformation. The function removes dead moves
4163 too. */
4164 static void
4165 update_ebb_live_info (rtx head, rtx tail)
4167 unsigned int j;
4168 int regno;
4169 bool live_p;
4170 rtx prev_insn, set;
4171 bool remove_p;
4172 basic_block last_bb, prev_bb, curr_bb;
4173 bitmap_iterator bi;
4174 struct lra_insn_reg *reg;
4175 edge e;
4176 edge_iterator ei;
4178 last_bb = BLOCK_FOR_INSN (tail);
4179 prev_bb = NULL;
4180 for (curr_insn = tail;
4181 curr_insn != PREV_INSN (head);
4182 curr_insn = prev_insn)
4184 prev_insn = PREV_INSN (curr_insn);
4185 /* We need to process empty blocks too. They contain
4186 NOTE_INSN_BASIC_BLOCK referring for the basic block. */
4187 if (NOTE_P (curr_insn) && NOTE_KIND (curr_insn) != NOTE_INSN_BASIC_BLOCK)
4188 continue;
4189 curr_bb = BLOCK_FOR_INSN (curr_insn);
4190 if (curr_bb != prev_bb)
4192 if (prev_bb != NULL)
4194 /* Update df_get_live_in (prev_bb): */
4195 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
4196 if (bitmap_bit_p (&live_regs, j))
4197 bitmap_set_bit (df_get_live_in (prev_bb), j);
4198 else
4199 bitmap_clear_bit (df_get_live_in (prev_bb), j);
4201 if (curr_bb != last_bb)
4203 /* Update df_get_live_out (curr_bb): */
4204 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
4206 live_p = bitmap_bit_p (&live_regs, j);
4207 if (! live_p)
4208 FOR_EACH_EDGE (e, ei, curr_bb->succs)
4209 if (bitmap_bit_p (df_get_live_in (e->dest), j))
4211 live_p = true;
4212 break;
4214 if (live_p)
4215 bitmap_set_bit (df_get_live_out (curr_bb), j);
4216 else
4217 bitmap_clear_bit (df_get_live_out (curr_bb), j);
4220 prev_bb = curr_bb;
4221 bitmap_and (&live_regs, &check_only_regs, df_get_live_out (curr_bb));
4223 if (! NONDEBUG_INSN_P (curr_insn))
4224 continue;
4225 curr_id = lra_get_insn_recog_data (curr_insn);
4226 remove_p = false;
4227 if ((set = single_set (curr_insn)) != NULL_RTX && REG_P (SET_DEST (set))
4228 && (regno = REGNO (SET_DEST (set))) >= FIRST_PSEUDO_REGISTER
4229 && bitmap_bit_p (&check_only_regs, regno)
4230 && ! bitmap_bit_p (&live_regs, regno))
4231 remove_p = true;
4232 /* See which defined values die here. */
4233 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
4234 if (reg->type == OP_OUT && ! reg->subreg_p)
4235 bitmap_clear_bit (&live_regs, reg->regno);
4236 /* Mark each used value as live. */
4237 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
4238 if (reg->type == OP_IN
4239 && bitmap_bit_p (&check_only_regs, reg->regno))
4240 bitmap_set_bit (&live_regs, reg->regno);
4241 /* It is quite important to remove dead move insns because it
4242 means removing dead store. We don't need to process them for
4243 constraints. */
4244 if (remove_p)
4246 if (lra_dump_file != NULL)
4248 fprintf (lra_dump_file, " Removing dead insn:\n ");
4249 dump_insn_slim (lra_dump_file, curr_insn);
4251 lra_set_insn_deleted (curr_insn);
4256 /* The structure describes info to do an inheritance for the current
4257 insn. We need to collect such info first before doing the
4258 transformations because the transformations change the insn
4259 internal representation. */
4260 struct to_inherit
4262 /* Original regno. */
4263 int regno;
4264 /* Subsequent insns which can inherit original reg value. */
4265 rtx insns;
4268 /* Array containing all info for doing inheritance from the current
4269 insn. */
4270 static struct to_inherit to_inherit[LRA_MAX_INSN_RELOADS];
4272 /* Number elements in the previous array. */
4273 static int to_inherit_num;
4275 /* Add inheritance info REGNO and INSNS. Their meaning is described in
4276 structure to_inherit. */
4277 static void
4278 add_to_inherit (int regno, rtx insns)
4280 int i;
4282 for (i = 0; i < to_inherit_num; i++)
4283 if (to_inherit[i].regno == regno)
4284 return;
4285 lra_assert (to_inherit_num < LRA_MAX_INSN_RELOADS);
4286 to_inherit[to_inherit_num].regno = regno;
4287 to_inherit[to_inherit_num++].insns = insns;
4290 /* Return the last non-debug insn in basic block BB, or the block begin
4291 note if none. */
4292 static rtx
4293 get_last_insertion_point (basic_block bb)
4295 rtx insn;
4297 FOR_BB_INSNS_REVERSE (bb, insn)
4298 if (NONDEBUG_INSN_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn))
4299 return insn;
4300 gcc_unreachable ();
4303 /* Set up RES by registers living on edges FROM except the edge (FROM,
4304 TO) or by registers set up in a jump insn in BB FROM. */
4305 static void
4306 get_live_on_other_edges (basic_block from, basic_block to, bitmap res)
4308 rtx last;
4309 struct lra_insn_reg *reg;
4310 edge e;
4311 edge_iterator ei;
4313 lra_assert (to != NULL);
4314 bitmap_clear (res);
4315 FOR_EACH_EDGE (e, ei, from->succs)
4316 if (e->dest != to)
4317 bitmap_ior_into (res, df_get_live_in (e->dest));
4318 last = get_last_insertion_point (from);
4319 if (! JUMP_P (last))
4320 return;
4321 curr_id = lra_get_insn_recog_data (last);
4322 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
4323 if (reg->type != OP_IN)
4324 bitmap_set_bit (res, reg->regno);
4327 /* Used as a temporary results of some bitmap calculations. */
4328 static bitmap_head temp_bitmap;
4330 /* Do inheritance/split transformations in EBB starting with HEAD and
4331 finishing on TAIL. We process EBB insns in the reverse order.
4332 Return true if we did any inheritance/split transformation in the
4333 EBB.
4335 We should avoid excessive splitting which results in worse code
4336 because of inaccurate cost calculations for spilling new split
4337 pseudos in such case. To achieve this we do splitting only if
4338 register pressure is high in given basic block and there are reload
4339 pseudos requiring hard registers. We could do more register
4340 pressure calculations at any given program point to avoid necessary
4341 splitting even more but it is to expensive and the current approach
4342 works well enough. */
4343 static bool
4344 inherit_in_ebb (rtx head, rtx tail)
4346 int i, src_regno, dst_regno, nregs;
4347 bool change_p, succ_p;
4348 rtx prev_insn, next_usage_insns, set, last_insn;
4349 enum reg_class cl;
4350 struct lra_insn_reg *reg;
4351 basic_block last_processed_bb, curr_bb = NULL;
4352 HARD_REG_SET potential_reload_hard_regs, live_hard_regs;
4353 bitmap to_process;
4354 unsigned int j;
4355 bitmap_iterator bi;
4356 bool head_p, after_p;
4358 change_p = false;
4359 curr_usage_insns_check++;
4360 reloads_num = calls_num = 0;
4361 bitmap_clear (&check_only_regs);
4362 last_processed_bb = NULL;
4363 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
4364 CLEAR_HARD_REG_SET (live_hard_regs);
4365 /* We don't process new insns generated in the loop. */
4366 for (curr_insn = tail; curr_insn != PREV_INSN (head); curr_insn = prev_insn)
4368 prev_insn = PREV_INSN (curr_insn);
4369 if (BLOCK_FOR_INSN (curr_insn) != NULL)
4370 curr_bb = BLOCK_FOR_INSN (curr_insn);
4371 if (last_processed_bb != curr_bb)
4373 /* We are at the end of BB. Add qualified living
4374 pseudos for potential splitting. */
4375 to_process = df_get_live_out (curr_bb);
4376 if (last_processed_bb != NULL)
4378 /* We are somewhere in the middle of EBB. */
4379 get_live_on_other_edges (curr_bb, last_processed_bb,
4380 &temp_bitmap);
4381 to_process = &temp_bitmap;
4383 last_processed_bb = curr_bb;
4384 last_insn = get_last_insertion_point (curr_bb);
4385 after_p = (! JUMP_P (last_insn)
4386 && (! CALL_P (last_insn)
4387 || (find_reg_note (last_insn,
4388 REG_NORETURN, NULL_RTX) == NULL_RTX
4389 && ! SIBLING_CALL_P (last_insn))));
4390 REG_SET_TO_HARD_REG_SET (live_hard_regs, df_get_live_out (curr_bb));
4391 IOR_HARD_REG_SET (live_hard_regs, eliminable_regset);
4392 IOR_HARD_REG_SET (live_hard_regs, lra_no_alloc_regs);
4393 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
4394 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
4396 if ((int) j >= lra_constraint_new_regno_start)
4397 break;
4398 if (j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
4400 if (j < FIRST_PSEUDO_REGISTER)
4401 SET_HARD_REG_BIT (live_hard_regs, j);
4402 else
4403 add_to_hard_reg_set (&live_hard_regs,
4404 PSEUDO_REGNO_MODE (j),
4405 reg_renumber[j]);
4406 setup_next_usage_insn (j, last_insn, reloads_num, after_p);
4410 src_regno = dst_regno = -1;
4411 if (NONDEBUG_INSN_P (curr_insn)
4412 && (set = single_set (curr_insn)) != NULL_RTX
4413 && REG_P (SET_DEST (set)) && REG_P (SET_SRC (set)))
4415 src_regno = REGNO (SET_SRC (set));
4416 dst_regno = REGNO (SET_DEST (set));
4418 if (src_regno < lra_constraint_new_regno_start
4419 && src_regno >= FIRST_PSEUDO_REGISTER
4420 && reg_renumber[src_regno] < 0
4421 && dst_regno >= lra_constraint_new_regno_start
4422 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS)
4424 /* 'reload_pseudo <- original_pseudo'. */
4425 reloads_num++;
4426 succ_p = false;
4427 if (usage_insns[src_regno].check == curr_usage_insns_check
4428 && (next_usage_insns = usage_insns[src_regno].insns) != NULL_RTX)
4429 succ_p = inherit_reload_reg (false, src_regno, cl,
4430 curr_insn, next_usage_insns);
4431 if (succ_p)
4432 change_p = true;
4433 else
4434 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
4435 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
4436 IOR_HARD_REG_SET (potential_reload_hard_regs,
4437 reg_class_contents[cl]);
4439 else if (src_regno >= lra_constraint_new_regno_start
4440 && dst_regno < lra_constraint_new_regno_start
4441 && dst_regno >= FIRST_PSEUDO_REGISTER
4442 && reg_renumber[dst_regno] < 0
4443 && (cl = lra_get_allocno_class (src_regno)) != NO_REGS
4444 && usage_insns[dst_regno].check == curr_usage_insns_check
4445 && (next_usage_insns
4446 = usage_insns[dst_regno].insns) != NULL_RTX)
4448 reloads_num++;
4449 /* 'original_pseudo <- reload_pseudo'. */
4450 if (! JUMP_P (curr_insn)
4451 && inherit_reload_reg (true, dst_regno, cl,
4452 curr_insn, next_usage_insns))
4453 change_p = true;
4454 /* Invalidate. */
4455 usage_insns[dst_regno].check = 0;
4456 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
4457 IOR_HARD_REG_SET (potential_reload_hard_regs,
4458 reg_class_contents[cl]);
4460 else if (INSN_P (curr_insn))
4462 int max_uid = get_max_uid ();
4464 curr_id = lra_get_insn_recog_data (curr_insn);
4465 to_inherit_num = 0;
4466 /* Process insn definitions. */
4467 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
4468 if (reg->type != OP_IN
4469 && (dst_regno = reg->regno) < lra_constraint_new_regno_start)
4471 if (dst_regno >= FIRST_PSEUDO_REGISTER && reg->type == OP_OUT
4472 && reg_renumber[dst_regno] < 0 && ! reg->subreg_p
4473 && usage_insns[dst_regno].check == curr_usage_insns_check
4474 && (next_usage_insns
4475 = usage_insns[dst_regno].insns) != NULL_RTX)
4477 struct lra_insn_reg *r;
4479 for (r = curr_id->regs; r != NULL; r = r->next)
4480 if (r->type != OP_OUT && r->regno == dst_regno)
4481 break;
4482 /* Don't do inheritance if the pseudo is also
4483 used in the insn. */
4484 if (r == NULL)
4485 /* We can not do inheritance right now
4486 because the current insn reg info (chain
4487 regs) can change after that. */
4488 add_to_inherit (dst_regno, next_usage_insns);
4490 /* We can not process one reg twice here because of
4491 usage_insns invalidation. */
4492 if ((dst_regno < FIRST_PSEUDO_REGISTER
4493 || reg_renumber[dst_regno] >= 0)
4494 && ! reg->subreg_p && reg->type == OP_OUT)
4496 HARD_REG_SET s;
4498 if (split_if_necessary (dst_regno, reg->biggest_mode,
4499 potential_reload_hard_regs,
4500 false, curr_insn, max_uid))
4501 change_p = true;
4502 CLEAR_HARD_REG_SET (s);
4503 if (dst_regno < FIRST_PSEUDO_REGISTER)
4504 add_to_hard_reg_set (&s, reg->biggest_mode, dst_regno);
4505 else
4506 add_to_hard_reg_set (&s, PSEUDO_REGNO_MODE (dst_regno),
4507 reg_renumber[dst_regno]);
4508 AND_COMPL_HARD_REG_SET (live_hard_regs, s);
4510 /* We should invalidate potential inheritance or
4511 splitting for the current insn usages to the next
4512 usage insns (see code below) as the output pseudo
4513 prevents this. */
4514 if ((dst_regno >= FIRST_PSEUDO_REGISTER
4515 && reg_renumber[dst_regno] < 0)
4516 || (reg->type == OP_OUT && ! reg->subreg_p
4517 && (dst_regno < FIRST_PSEUDO_REGISTER
4518 || reg_renumber[dst_regno] >= 0)))
4520 /* Invalidate. */
4521 if (dst_regno >= FIRST_PSEUDO_REGISTER)
4522 usage_insns[dst_regno].check = 0;
4523 else
4525 nregs = hard_regno_nregs[dst_regno][reg->biggest_mode];
4526 for (i = 0; i < nregs; i++)
4527 usage_insns[dst_regno + i].check = 0;
4531 if (! JUMP_P (curr_insn))
4532 for (i = 0; i < to_inherit_num; i++)
4533 if (inherit_reload_reg (true, to_inherit[i].regno,
4534 ALL_REGS, curr_insn,
4535 to_inherit[i].insns))
4536 change_p = true;
4537 if (CALL_P (curr_insn))
4539 rtx cheap, pat, dest, restore;
4540 int regno, hard_regno;
4542 calls_num++;
4543 if ((cheap = find_reg_note (curr_insn,
4544 REG_RETURNED, NULL_RTX)) != NULL_RTX
4545 && ((cheap = XEXP (cheap, 0)), true)
4546 && (regno = REGNO (cheap)) >= FIRST_PSEUDO_REGISTER
4547 && (hard_regno = reg_renumber[regno]) >= 0
4548 /* If there are pending saves/restores, the
4549 optimization is not worth. */
4550 && usage_insns[regno].calls_num == calls_num - 1
4551 && TEST_HARD_REG_BIT (call_used_reg_set, hard_regno))
4553 /* Restore the pseudo from the call result as
4554 REG_RETURNED note says that the pseudo value is
4555 in the call result and the pseudo is an argument
4556 of the call. */
4557 pat = PATTERN (curr_insn);
4558 if (GET_CODE (pat) == PARALLEL)
4559 pat = XVECEXP (pat, 0, 0);
4560 dest = SET_DEST (pat);
4561 start_sequence ();
4562 emit_move_insn (cheap, copy_rtx (dest));
4563 restore = get_insns ();
4564 end_sequence ();
4565 lra_process_new_insns (curr_insn, NULL, restore,
4566 "Inserting call parameter restore");
4567 /* We don't need to save/restore of the pseudo from
4568 this call. */
4569 usage_insns[regno].calls_num = calls_num;
4570 bitmap_set_bit (&check_only_regs, regno);
4573 to_inherit_num = 0;
4574 /* Process insn usages. */
4575 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
4576 if ((reg->type != OP_OUT
4577 || (reg->type == OP_OUT && reg->subreg_p))
4578 && (src_regno = reg->regno) < lra_constraint_new_regno_start)
4580 if (src_regno >= FIRST_PSEUDO_REGISTER
4581 && reg_renumber[src_regno] < 0 && reg->type == OP_IN)
4583 if (usage_insns[src_regno].check == curr_usage_insns_check
4584 && (next_usage_insns
4585 = usage_insns[src_regno].insns) != NULL_RTX
4586 && NONDEBUG_INSN_P (curr_insn))
4587 add_to_inherit (src_regno, next_usage_insns);
4588 else
4589 /* Add usages. */
4590 add_next_usage_insn (src_regno, curr_insn, reloads_num);
4592 else if (src_regno < FIRST_PSEUDO_REGISTER
4593 || reg_renumber[src_regno] >= 0)
4595 bool before_p;
4596 rtx use_insn = curr_insn;
4598 before_p = (JUMP_P (curr_insn)
4599 || (CALL_P (curr_insn) && reg->type == OP_IN));
4600 if (NONDEBUG_INSN_P (curr_insn)
4601 && split_if_necessary (src_regno, reg->biggest_mode,
4602 potential_reload_hard_regs,
4603 before_p, curr_insn, max_uid))
4605 if (reg->subreg_p)
4606 lra_risky_transformations_p = true;
4607 change_p = true;
4608 /* Invalidate. */
4609 usage_insns[src_regno].check = 0;
4610 if (before_p)
4611 use_insn = PREV_INSN (curr_insn);
4613 if (NONDEBUG_INSN_P (curr_insn))
4615 if (src_regno < FIRST_PSEUDO_REGISTER)
4616 add_to_hard_reg_set (&live_hard_regs,
4617 reg->biggest_mode, src_regno);
4618 else
4619 add_to_hard_reg_set (&live_hard_regs,
4620 PSEUDO_REGNO_MODE (src_regno),
4621 reg_renumber[src_regno]);
4623 add_next_usage_insn (src_regno, use_insn, reloads_num);
4626 for (i = 0; i < to_inherit_num; i++)
4628 src_regno = to_inherit[i].regno;
4629 if (inherit_reload_reg (false, src_regno, ALL_REGS,
4630 curr_insn, to_inherit[i].insns))
4631 change_p = true;
4632 else
4633 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
4636 /* We reached the start of the current basic block. */
4637 if (prev_insn == NULL_RTX || prev_insn == PREV_INSN (head)
4638 || BLOCK_FOR_INSN (prev_insn) != curr_bb)
4640 /* We reached the beginning of the current block -- do
4641 rest of spliting in the current BB. */
4642 to_process = df_get_live_in (curr_bb);
4643 if (BLOCK_FOR_INSN (head) != curr_bb)
4645 /* We are somewhere in the middle of EBB. */
4646 get_live_on_other_edges (EDGE_PRED (curr_bb, 0)->src,
4647 curr_bb, &temp_bitmap);
4648 to_process = &temp_bitmap;
4650 head_p = true;
4651 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
4653 if ((int) j >= lra_constraint_new_regno_start)
4654 break;
4655 if (((int) j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
4656 && usage_insns[j].check == curr_usage_insns_check
4657 && (next_usage_insns = usage_insns[j].insns) != NULL_RTX)
4659 if (need_for_split_p (potential_reload_hard_regs, j))
4661 if (lra_dump_file != NULL && head_p)
4663 fprintf (lra_dump_file,
4664 " ----------------------------------\n");
4665 head_p = false;
4667 if (split_reg (false, j, bb_note (curr_bb),
4668 next_usage_insns))
4669 change_p = true;
4671 usage_insns[j].check = 0;
4676 return change_p;
4679 /* The maximal number of inheritance/split passes in LRA. It should
4680 be more 1 in order to perform caller saves transformations and much
4681 less MAX_CONSTRAINT_ITERATION_NUMBER to prevent LRA to do as many
4682 as permitted constraint passes in some complicated cases. The
4683 first inheritance/split pass has a biggest impact on generated code
4684 quality. Each subsequent affects generated code in less degree.
4685 For example, the 3rd pass does not change generated SPEC2000 code
4686 at all on x86-64. */
4687 #define MAX_INHERITANCE_PASSES 2
4689 #if MAX_INHERITANCE_PASSES <= 0 \
4690 || MAX_INHERITANCE_PASSES >= MAX_CONSTRAINT_ITERATION_NUMBER - 8
4691 #error wrong MAX_INHERITANCE_PASSES value
4692 #endif
4694 /* This value affects EBB forming. If probability of edge from EBB to
4695 a BB is not greater than the following value, we don't add the BB
4696 to EBB. */
4697 #define EBB_PROBABILITY_CUTOFF (REG_BR_PROB_BASE / 2)
4699 /* Current number of inheritance/split iteration. */
4700 int lra_inheritance_iter;
4702 /* Entry function for inheritance/split pass. */
4703 void
4704 lra_inheritance (void)
4706 int i;
4707 basic_block bb, start_bb;
4708 edge e;
4710 lra_inheritance_iter++;
4711 if (lra_inheritance_iter > MAX_INHERITANCE_PASSES)
4712 return;
4713 timevar_push (TV_LRA_INHERITANCE);
4714 if (lra_dump_file != NULL)
4715 fprintf (lra_dump_file, "\n********** Inheritance #%d: **********\n\n",
4716 lra_inheritance_iter);
4717 curr_usage_insns_check = 0;
4718 usage_insns = XNEWVEC (struct usage_insns, lra_constraint_new_regno_start);
4719 for (i = 0; i < lra_constraint_new_regno_start; i++)
4720 usage_insns[i].check = 0;
4721 bitmap_initialize (&check_only_regs, &reg_obstack);
4722 bitmap_initialize (&live_regs, &reg_obstack);
4723 bitmap_initialize (&temp_bitmap, &reg_obstack);
4724 bitmap_initialize (&ebb_global_regs, &reg_obstack);
4725 FOR_EACH_BB (bb)
4727 start_bb = bb;
4728 if (lra_dump_file != NULL)
4729 fprintf (lra_dump_file, "EBB");
4730 /* Form a EBB starting with BB. */
4731 bitmap_clear (&ebb_global_regs);
4732 bitmap_ior_into (&ebb_global_regs, df_get_live_in (bb));
4733 for (;;)
4735 if (lra_dump_file != NULL)
4736 fprintf (lra_dump_file, " %d", bb->index);
4737 if (bb->next_bb == EXIT_BLOCK_PTR || LABEL_P (BB_HEAD (bb->next_bb)))
4738 break;
4739 e = find_fallthru_edge (bb->succs);
4740 if (! e)
4741 break;
4742 if (e->probability <= EBB_PROBABILITY_CUTOFF)
4743 break;
4744 bb = bb->next_bb;
4746 bitmap_ior_into (&ebb_global_regs, df_get_live_out (bb));
4747 if (lra_dump_file != NULL)
4748 fprintf (lra_dump_file, "\n");
4749 if (inherit_in_ebb (BB_HEAD (start_bb), BB_END (bb)))
4750 /* Remember that the EBB head and tail can change in
4751 inherit_in_ebb. */
4752 update_ebb_live_info (BB_HEAD (start_bb), BB_END (bb));
4754 bitmap_clear (&ebb_global_regs);
4755 bitmap_clear (&temp_bitmap);
4756 bitmap_clear (&live_regs);
4757 bitmap_clear (&check_only_regs);
4758 free (usage_insns);
4760 timevar_pop (TV_LRA_INHERITANCE);
4765 /* This page contains code to undo failed inheritance/split
4766 transformations. */
4768 /* Current number of iteration undoing inheritance/split. */
4769 int lra_undo_inheritance_iter;
4771 /* Fix BB live info LIVE after removing pseudos created on pass doing
4772 inheritance/split which are REMOVED_PSEUDOS. */
4773 static void
4774 fix_bb_live_info (bitmap live, bitmap removed_pseudos)
4776 unsigned int regno;
4777 bitmap_iterator bi;
4779 EXECUTE_IF_SET_IN_BITMAP (removed_pseudos, 0, regno, bi)
4780 if (bitmap_clear_bit (live, regno))
4781 bitmap_set_bit (live, lra_reg_info[regno].restore_regno);
4784 /* Return regno of the (subreg of) REG. Otherwise, return a negative
4785 number. */
4786 static int
4787 get_regno (rtx reg)
4789 if (GET_CODE (reg) == SUBREG)
4790 reg = SUBREG_REG (reg);
4791 if (REG_P (reg))
4792 return REGNO (reg);
4793 return -1;
4796 /* Remove inheritance/split pseudos which are in REMOVE_PSEUDOS and
4797 return true if we did any change. The undo transformations for
4798 inheritance looks like
4799 i <- i2
4800 p <- i => p <- i2
4801 or removing
4802 p <- i, i <- p, and i <- i3
4803 where p is original pseudo from which inheritance pseudo i was
4804 created, i and i3 are removed inheritance pseudos, i2 is another
4805 not removed inheritance pseudo. All split pseudos or other
4806 occurrences of removed inheritance pseudos are changed on the
4807 corresponding original pseudos.
4809 The function also schedules insns changed and created during
4810 inheritance/split pass for processing by the subsequent constraint
4811 pass. */
4812 static bool
4813 remove_inheritance_pseudos (bitmap remove_pseudos)
4815 basic_block bb;
4816 int regno, sregno, prev_sregno, dregno, restore_regno;
4817 rtx set, prev_set, prev_insn;
4818 bool change_p, done_p;
4820 change_p = ! bitmap_empty_p (remove_pseudos);
4821 /* We can not finish the function right away if CHANGE_P is true
4822 because we need to marks insns affected by previous
4823 inheritance/split pass for processing by the subsequent
4824 constraint pass. */
4825 FOR_EACH_BB (bb)
4827 fix_bb_live_info (df_get_live_in (bb), remove_pseudos);
4828 fix_bb_live_info (df_get_live_out (bb), remove_pseudos);
4829 FOR_BB_INSNS_REVERSE (bb, curr_insn)
4831 if (! INSN_P (curr_insn))
4832 continue;
4833 done_p = false;
4834 sregno = dregno = -1;
4835 if (change_p && NONDEBUG_INSN_P (curr_insn)
4836 && (set = single_set (curr_insn)) != NULL_RTX)
4838 dregno = get_regno (SET_DEST (set));
4839 sregno = get_regno (SET_SRC (set));
4842 if (sregno >= 0 && dregno >= 0)
4844 if ((bitmap_bit_p (remove_pseudos, sregno)
4845 && (lra_reg_info[sregno].restore_regno == dregno
4846 || (bitmap_bit_p (remove_pseudos, dregno)
4847 && (lra_reg_info[sregno].restore_regno
4848 == lra_reg_info[dregno].restore_regno))))
4849 || (bitmap_bit_p (remove_pseudos, dregno)
4850 && lra_reg_info[dregno].restore_regno == sregno))
4851 /* One of the following cases:
4852 original <- removed inheritance pseudo
4853 removed inherit pseudo <- another removed inherit pseudo
4854 removed inherit pseudo <- original pseudo
4856 removed_split_pseudo <- original_reg
4857 original_reg <- removed_split_pseudo */
4859 if (lra_dump_file != NULL)
4861 fprintf (lra_dump_file, " Removing %s:\n",
4862 bitmap_bit_p (&lra_split_regs, sregno)
4863 || bitmap_bit_p (&lra_split_regs, dregno)
4864 ? "split" : "inheritance");
4865 dump_insn_slim (lra_dump_file, curr_insn);
4867 lra_set_insn_deleted (curr_insn);
4868 done_p = true;
4870 else if (bitmap_bit_p (remove_pseudos, sregno)
4871 && bitmap_bit_p (&lra_inheritance_pseudos, sregno))
4873 /* Search the following pattern:
4874 inherit_or_split_pseudo1 <- inherit_or_split_pseudo2
4875 original_pseudo <- inherit_or_split_pseudo1
4876 where the 2nd insn is the current insn and
4877 inherit_or_split_pseudo2 is not removed. If it is found,
4878 change the current insn onto:
4879 original_pseudo <- inherit_or_split_pseudo2. */
4880 for (prev_insn = PREV_INSN (curr_insn);
4881 prev_insn != NULL_RTX && ! NONDEBUG_INSN_P (prev_insn);
4882 prev_insn = PREV_INSN (prev_insn))
4884 if (prev_insn != NULL_RTX && BLOCK_FOR_INSN (prev_insn) == bb
4885 && (prev_set = single_set (prev_insn)) != NULL_RTX
4886 /* There should be no subregs in insn we are
4887 searching because only the original reg might
4888 be in subreg when we changed the mode of
4889 load/store for splitting. */
4890 && REG_P (SET_DEST (prev_set))
4891 && REG_P (SET_SRC (prev_set))
4892 && (int) REGNO (SET_DEST (prev_set)) == sregno
4893 && ((prev_sregno = REGNO (SET_SRC (prev_set)))
4894 >= FIRST_PSEUDO_REGISTER)
4895 /* As we consider chain of inheritance or
4896 splitting described in above comment we should
4897 check that sregno and prev_sregno were
4898 inheritance/split pseudos created from the
4899 same original regno. */
4900 && (lra_reg_info[sregno].restore_regno
4901 == lra_reg_info[prev_sregno].restore_regno)
4902 && ! bitmap_bit_p (remove_pseudos, prev_sregno))
4904 lra_assert (GET_MODE (SET_SRC (prev_set))
4905 == GET_MODE (regno_reg_rtx[sregno]));
4906 if (GET_CODE (SET_SRC (set)) == SUBREG)
4907 SUBREG_REG (SET_SRC (set)) = SET_SRC (prev_set);
4908 else
4909 SET_SRC (set) = SET_SRC (prev_set);
4910 lra_push_insn_and_update_insn_regno_info (curr_insn);
4911 lra_set_used_insn_alternative_by_uid
4912 (INSN_UID (curr_insn), -1);
4913 done_p = true;
4914 if (lra_dump_file != NULL)
4916 fprintf (lra_dump_file, " Change reload insn:\n");
4917 dump_insn_slim (lra_dump_file, curr_insn);
4922 if (! done_p)
4924 struct lra_insn_reg *reg;
4925 bool restored_regs_p = false;
4926 bool kept_regs_p = false;
4928 curr_id = lra_get_insn_recog_data (curr_insn);
4929 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
4931 regno = reg->regno;
4932 restore_regno = lra_reg_info[regno].restore_regno;
4933 if (restore_regno >= 0)
4935 if (change_p && bitmap_bit_p (remove_pseudos, regno))
4937 substitute_pseudo (&curr_insn, regno,
4938 regno_reg_rtx[restore_regno]);
4939 restored_regs_p = true;
4941 else
4942 kept_regs_p = true;
4945 if (NONDEBUG_INSN_P (curr_insn) && kept_regs_p)
4947 /* The instruction has changed since the previous
4948 constraints pass. */
4949 lra_push_insn_and_update_insn_regno_info (curr_insn);
4950 lra_set_used_insn_alternative_by_uid
4951 (INSN_UID (curr_insn), -1);
4953 else if (restored_regs_p)
4954 /* The instruction has been restored to the form that
4955 it had during the previous constraints pass. */
4956 lra_update_insn_regno_info (curr_insn);
4957 if (restored_regs_p && lra_dump_file != NULL)
4959 fprintf (lra_dump_file, " Insn after restoring regs:\n");
4960 dump_insn_slim (lra_dump_file, curr_insn);
4965 return change_p;
4968 /* Entry function for undoing inheritance/split transformation. Return true
4969 if we did any RTL change in this pass. */
4970 bool
4971 lra_undo_inheritance (void)
4973 unsigned int regno;
4974 int restore_regno, hard_regno;
4975 int n_all_inherit, n_inherit, n_all_split, n_split;
4976 bitmap_head remove_pseudos;
4977 bitmap_iterator bi;
4978 bool change_p;
4980 lra_undo_inheritance_iter++;
4981 if (lra_undo_inheritance_iter > MAX_INHERITANCE_PASSES)
4982 return false;
4983 if (lra_dump_file != NULL)
4984 fprintf (lra_dump_file,
4985 "\n********** Undoing inheritance #%d: **********\n\n",
4986 lra_undo_inheritance_iter);
4987 bitmap_initialize (&remove_pseudos, &reg_obstack);
4988 n_inherit = n_all_inherit = 0;
4989 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
4990 if (lra_reg_info[regno].restore_regno >= 0)
4992 n_all_inherit++;
4993 if (reg_renumber[regno] < 0)
4994 bitmap_set_bit (&remove_pseudos, regno);
4995 else
4996 n_inherit++;
4998 if (lra_dump_file != NULL && n_all_inherit != 0)
4999 fprintf (lra_dump_file, "Inherit %d out of %d (%.2f%%)\n",
5000 n_inherit, n_all_inherit,
5001 (double) n_inherit / n_all_inherit * 100);
5002 n_split = n_all_split = 0;
5003 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
5004 if ((restore_regno = lra_reg_info[regno].restore_regno) >= 0)
5006 n_all_split++;
5007 hard_regno = (restore_regno >= FIRST_PSEUDO_REGISTER
5008 ? reg_renumber[restore_regno] : restore_regno);
5009 if (hard_regno < 0 || reg_renumber[regno] == hard_regno)
5010 bitmap_set_bit (&remove_pseudos, regno);
5011 else
5013 n_split++;
5014 if (lra_dump_file != NULL)
5015 fprintf (lra_dump_file, " Keep split r%d (orig=r%d)\n",
5016 regno, restore_regno);
5019 if (lra_dump_file != NULL && n_all_split != 0)
5020 fprintf (lra_dump_file, "Split %d out of %d (%.2f%%)\n",
5021 n_split, n_all_split,
5022 (double) n_split / n_all_split * 100);
5023 change_p = remove_inheritance_pseudos (&remove_pseudos);
5024 bitmap_clear (&remove_pseudos);
5025 /* Clear restore_regnos. */
5026 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
5027 lra_reg_info[regno].restore_regno = -1;
5028 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
5029 lra_reg_info[regno].restore_regno = -1;
5030 return change_p;