2017-12-07 Vladimir Makarov <vmakarov@redhat.com>
[official-gcc.git] / gcc / lra.c
blob3fd15ee57943af3b31701f6bc5727526e489697e
1 /* LRA (local register allocator) driver and LRA utilities.
2 Copyright (C) 2010-2017 Free Software Foundation, Inc.
3 Contributed by Vladimir Makarov <vmakarov@redhat.com>.
5 This file is part of GCC.
7 GCC is free software; you can redistribute it and/or modify it under
8 the terms of the GNU General Public License as published by the Free
9 Software Foundation; either version 3, or (at your option) any later
10 version.
12 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
13 WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
15 for more details.
17 You should have received a copy of the GNU General Public License
18 along with GCC; see the file COPYING3. If not see
19 <http://www.gnu.org/licenses/>. */
22 /* The Local Register Allocator (LRA) is a replacement of former
23 reload pass. It is focused to simplify code solving the reload
24 pass tasks, to make the code maintenance easier, and to implement new
25 perspective optimizations.
27 The major LRA design solutions are:
28 o division small manageable, separated sub-tasks
29 o reflection of all transformations and decisions in RTL as more
30 as possible
31 o insn constraints as a primary source of the info (minimizing
32 number of target-depended macros/hooks)
34 In brief LRA works by iterative insn process with the final goal is
35 to satisfy all insn and address constraints:
36 o New reload insns (in brief reloads) and reload pseudos might be
37 generated;
38 o Some pseudos might be spilled to assign hard registers to
39 new reload pseudos;
40 o Recalculating spilled pseudo values (rematerialization);
41 o Changing spilled pseudos to stack memory or their equivalences;
42 o Allocation stack memory changes the address displacement and
43 new iteration is needed.
45 Here is block diagram of LRA passes:
47 ------------------------
48 --------------- | Undo inheritance for | ---------------
49 | Memory-memory | | spilled pseudos, | | New (and old) |
50 | move coalesce |<---| splits for pseudos got |<-- | pseudos |
51 --------------- | the same hard regs, | | assignment |
52 Start | | and optional reloads | ---------------
53 | | ------------------------ ^
54 V | ---------------- |
55 ----------- V | Update virtual | |
56 | Remove |----> ------------>| register | |
57 | scratches | ^ | displacements | |
58 ----------- | ---------------- |
59 | | |
60 | V New |
61 | ------------ pseudos -------------------
62 | |Constraints:| or insns | Inheritance/split |
63 | | RTL |--------->| transformations |
64 | | transfor- | | in EBB scope |
65 | substi- | mations | -------------------
66 | tutions ------------
67 | | No change
68 ---------------- V
69 | Spilled pseudo | -------------------
70 | to memory |<----| Rematerialization |
71 | substitution | -------------------
72 ----------------
73 | No susbtitions
75 -------------------------
76 | Hard regs substitution, |
77 | devirtalization, and |------> Finish
78 | restoring scratches got |
79 | memory |
80 -------------------------
82 To speed up the process:
83 o We process only insns affected by changes on previous
84 iterations;
85 o We don't use DFA-infrastructure because it results in much slower
86 compiler speed than a special IR described below does;
87 o We use a special insn representation for quick access to insn
88 info which is always *synchronized* with the current RTL;
89 o Insn IR is minimized by memory. It is divided on three parts:
90 o one specific for each insn in RTL (only operand locations);
91 o one common for all insns in RTL with the same insn code
92 (different operand attributes from machine descriptions);
93 o one oriented for maintenance of live info (list of pseudos).
94 o Pseudo data:
95 o all insns where the pseudo is referenced;
96 o live info (conflicting hard regs, live ranges, # of
97 references etc);
98 o data used for assigning (preferred hard regs, costs etc).
100 This file contains LRA driver, LRA utility functions and data, and
101 code for dealing with scratches. */
103 #include "config.h"
104 #include "system.h"
105 #include "coretypes.h"
106 #include "backend.h"
107 #include "target.h"
108 #include "rtl.h"
109 #include "tree.h"
110 #include "predict.h"
111 #include "df.h"
112 #include "memmodel.h"
113 #include "tm_p.h"
114 #include "optabs.h"
115 #include "regs.h"
116 #include "ira.h"
117 #include "recog.h"
118 #include "expr.h"
119 #include "cfgrtl.h"
120 #include "cfgbuild.h"
121 #include "lra.h"
122 #include "lra-int.h"
123 #include "print-rtl.h"
125 /* Dump bitmap SET with TITLE and BB INDEX. */
126 void
127 lra_dump_bitmap_with_title (const char *title, bitmap set, int index)
129 unsigned int i;
130 int count;
131 bitmap_iterator bi;
132 static const int max_nums_on_line = 10;
134 if (bitmap_empty_p (set))
135 return;
136 fprintf (lra_dump_file, " %s %d:", title, index);
137 fprintf (lra_dump_file, "\n");
138 count = max_nums_on_line + 1;
139 EXECUTE_IF_SET_IN_BITMAP (set, 0, i, bi)
141 if (count > max_nums_on_line)
143 fprintf (lra_dump_file, "\n ");
144 count = 0;
146 fprintf (lra_dump_file, " %4u", i);
147 count++;
149 fprintf (lra_dump_file, "\n");
152 /* Hard registers currently not available for allocation. It can
153 changed after some hard registers become not eliminable. */
154 HARD_REG_SET lra_no_alloc_regs;
156 static int get_new_reg_value (void);
157 static void expand_reg_info (void);
158 static void invalidate_insn_recog_data (int);
159 static int get_insn_freq (rtx_insn *);
160 static void invalidate_insn_data_regno_info (lra_insn_recog_data_t,
161 rtx_insn *, int);
163 /* Expand all regno related info needed for LRA. */
164 static void
165 expand_reg_data (int old)
167 resize_reg_info ();
168 expand_reg_info ();
169 ira_expand_reg_equiv ();
170 for (int i = (int) max_reg_num () - 1; i >= old; i--)
171 lra_change_class (i, ALL_REGS, " Set", true);
174 /* Create and return a new reg of ORIGINAL mode. If ORIGINAL is NULL
175 or of VOIDmode, use MD_MODE for the new reg. Initialize its
176 register class to RCLASS. Print message about assigning class
177 RCLASS containing new register name TITLE unless it is NULL. Use
178 attributes of ORIGINAL if it is a register. The created register
179 will have unique held value. */
181 lra_create_new_reg_with_unique_value (machine_mode md_mode, rtx original,
182 enum reg_class rclass, const char *title)
184 machine_mode mode;
185 rtx new_reg;
187 if (original == NULL_RTX || (mode = GET_MODE (original)) == VOIDmode)
188 mode = md_mode;
189 lra_assert (mode != VOIDmode);
190 new_reg = gen_reg_rtx (mode);
191 if (original == NULL_RTX || ! REG_P (original))
193 if (lra_dump_file != NULL)
194 fprintf (lra_dump_file, " Creating newreg=%i", REGNO (new_reg));
196 else
198 if (ORIGINAL_REGNO (original) >= FIRST_PSEUDO_REGISTER)
199 ORIGINAL_REGNO (new_reg) = ORIGINAL_REGNO (original);
200 REG_USERVAR_P (new_reg) = REG_USERVAR_P (original);
201 REG_POINTER (new_reg) = REG_POINTER (original);
202 REG_ATTRS (new_reg) = REG_ATTRS (original);
203 if (lra_dump_file != NULL)
204 fprintf (lra_dump_file, " Creating newreg=%i from oldreg=%i",
205 REGNO (new_reg), REGNO (original));
207 if (lra_dump_file != NULL)
209 if (title != NULL)
210 fprintf (lra_dump_file, ", assigning class %s to%s%s r%d",
211 reg_class_names[rclass], *title == '\0' ? "" : " ",
212 title, REGNO (new_reg));
213 fprintf (lra_dump_file, "\n");
215 expand_reg_data (max_reg_num ());
216 setup_reg_classes (REGNO (new_reg), rclass, NO_REGS, rclass);
217 return new_reg;
220 /* Analogous to the previous function but also inherits value of
221 ORIGINAL. */
223 lra_create_new_reg (machine_mode md_mode, rtx original,
224 enum reg_class rclass, const char *title)
226 rtx new_reg;
228 new_reg
229 = lra_create_new_reg_with_unique_value (md_mode, original, rclass, title);
230 if (original != NULL_RTX && REG_P (original))
231 lra_assign_reg_val (REGNO (original), REGNO (new_reg));
232 return new_reg;
235 /* Set up for REGNO unique hold value. */
236 void
237 lra_set_regno_unique_value (int regno)
239 lra_reg_info[regno].val = get_new_reg_value ();
242 /* Invalidate INSN related info used by LRA. The info should never be
243 used after that. */
244 void
245 lra_invalidate_insn_data (rtx_insn *insn)
247 lra_invalidate_insn_regno_info (insn);
248 invalidate_insn_recog_data (INSN_UID (insn));
251 /* Mark INSN deleted and invalidate the insn related info used by
252 LRA. */
253 void
254 lra_set_insn_deleted (rtx_insn *insn)
256 lra_invalidate_insn_data (insn);
257 SET_INSN_DELETED (insn);
260 /* Delete an unneeded INSN and any previous insns who sole purpose is
261 loading data that is dead in INSN. */
262 void
263 lra_delete_dead_insn (rtx_insn *insn)
265 rtx_insn *prev = prev_real_insn (insn);
266 rtx prev_dest;
268 /* If the previous insn sets a register that dies in our insn,
269 delete it too. */
270 if (prev && GET_CODE (PATTERN (prev)) == SET
271 && (prev_dest = SET_DEST (PATTERN (prev)), REG_P (prev_dest))
272 && reg_mentioned_p (prev_dest, PATTERN (insn))
273 && find_regno_note (insn, REG_DEAD, REGNO (prev_dest))
274 && ! side_effects_p (SET_SRC (PATTERN (prev))))
275 lra_delete_dead_insn (prev);
277 lra_set_insn_deleted (insn);
280 /* Emit insn x = y + z. Return NULL if we failed to do it.
281 Otherwise, return the insn. We don't use gen_add3_insn as it might
282 clobber CC. */
283 static rtx_insn *
284 emit_add3_insn (rtx x, rtx y, rtx z)
286 rtx_insn *last;
288 last = get_last_insn ();
290 if (have_addptr3_insn (x, y, z))
292 rtx_insn *insn = gen_addptr3_insn (x, y, z);
294 /* If the target provides an "addptr" pattern it hopefully does
295 for a reason. So falling back to the normal add would be
296 a bug. */
297 lra_assert (insn != NULL_RTX);
298 emit_insn (insn);
299 return insn;
302 rtx_insn *insn = emit_insn (gen_rtx_SET (x, gen_rtx_PLUS (GET_MODE (y),
303 y, z)));
304 if (recog_memoized (insn) < 0)
306 delete_insns_since (last);
307 insn = NULL;
309 return insn;
312 /* Emit insn x = x + y. Return the insn. We use gen_add2_insn as the
313 last resort. */
314 static rtx_insn *
315 emit_add2_insn (rtx x, rtx y)
317 rtx_insn *insn = emit_add3_insn (x, x, y);
318 if (insn == NULL_RTX)
320 insn = gen_add2_insn (x, y);
321 if (insn != NULL_RTX)
322 emit_insn (insn);
324 return insn;
327 /* Target checks operands through operand predicates to recognize an
328 insn. We should have a special precaution to generate add insns
329 which are frequent results of elimination.
331 Emit insns for x = y + z. X can be used to store intermediate
332 values and should be not in Y and Z when we use X to store an
333 intermediate value. Y + Z should form [base] [+ index[ * scale]] [
334 + disp] where base and index are registers, disp and scale are
335 constants. Y should contain base if it is present, Z should
336 contain disp if any. index[*scale] can be part of Y or Z. */
337 void
338 lra_emit_add (rtx x, rtx y, rtx z)
340 int old;
341 rtx_insn *last;
342 rtx a1, a2, base, index, disp, scale, index_scale;
343 bool ok_p;
345 rtx_insn *add3_insn = emit_add3_insn (x, y, z);
346 old = max_reg_num ();
347 if (add3_insn != NULL)
349 else
351 disp = a2 = NULL_RTX;
352 if (GET_CODE (y) == PLUS)
354 a1 = XEXP (y, 0);
355 a2 = XEXP (y, 1);
356 disp = z;
358 else
360 a1 = y;
361 if (CONSTANT_P (z))
362 disp = z;
363 else
364 a2 = z;
366 index_scale = scale = NULL_RTX;
367 if (GET_CODE (a1) == MULT)
369 index_scale = a1;
370 index = XEXP (a1, 0);
371 scale = XEXP (a1, 1);
372 base = a2;
374 else if (a2 != NULL_RTX && GET_CODE (a2) == MULT)
376 index_scale = a2;
377 index = XEXP (a2, 0);
378 scale = XEXP (a2, 1);
379 base = a1;
381 else
383 base = a1;
384 index = a2;
386 if ((base != NULL_RTX && ! (REG_P (base) || GET_CODE (base) == SUBREG))
387 || (index != NULL_RTX
388 && ! (REG_P (index) || GET_CODE (index) == SUBREG))
389 || (disp != NULL_RTX && ! CONSTANT_P (disp))
390 || (scale != NULL_RTX && ! CONSTANT_P (scale)))
392 /* Probably we have no 3 op add. Last chance is to use 2-op
393 add insn. To succeed, don't move Z to X as an address
394 segment always comes in Y. Otherwise, we might fail when
395 adding the address segment to register. */
396 lra_assert (x != y && x != z);
397 emit_move_insn (x, y);
398 rtx_insn *insn = emit_add2_insn (x, z);
399 lra_assert (insn != NULL_RTX);
401 else
403 if (index_scale == NULL_RTX)
404 index_scale = index;
405 if (disp == NULL_RTX)
407 /* Generate x = index_scale; x = x + base. */
408 lra_assert (index_scale != NULL_RTX && base != NULL_RTX);
409 emit_move_insn (x, index_scale);
410 rtx_insn *insn = emit_add2_insn (x, base);
411 lra_assert (insn != NULL_RTX);
413 else if (scale == NULL_RTX)
415 /* Try x = base + disp. */
416 lra_assert (base != NULL_RTX);
417 last = get_last_insn ();
418 rtx_insn *move_insn =
419 emit_move_insn (x, gen_rtx_PLUS (GET_MODE (base), base, disp));
420 if (recog_memoized (move_insn) < 0)
422 delete_insns_since (last);
423 /* Generate x = disp; x = x + base. */
424 emit_move_insn (x, disp);
425 rtx_insn *add2_insn = emit_add2_insn (x, base);
426 lra_assert (add2_insn != NULL_RTX);
428 /* Generate x = x + index. */
429 if (index != NULL_RTX)
431 rtx_insn *insn = emit_add2_insn (x, index);
432 lra_assert (insn != NULL_RTX);
435 else
437 /* Try x = index_scale; x = x + disp; x = x + base. */
438 last = get_last_insn ();
439 rtx_insn *move_insn = emit_move_insn (x, index_scale);
440 ok_p = false;
441 if (recog_memoized (move_insn) >= 0)
443 rtx_insn *insn = emit_add2_insn (x, disp);
444 if (insn != NULL_RTX)
446 if (base == NULL_RTX)
447 ok_p = true;
448 else
450 insn = emit_add2_insn (x, base);
451 if (insn != NULL_RTX)
452 ok_p = true;
456 if (! ok_p)
458 rtx_insn *insn;
460 delete_insns_since (last);
461 /* Generate x = disp; x = x + base; x = x + index_scale. */
462 emit_move_insn (x, disp);
463 if (base != NULL_RTX)
465 insn = emit_add2_insn (x, base);
466 lra_assert (insn != NULL_RTX);
468 insn = emit_add2_insn (x, index_scale);
469 lra_assert (insn != NULL_RTX);
474 /* Functions emit_... can create pseudos -- so expand the pseudo
475 data. */
476 if (old != max_reg_num ())
477 expand_reg_data (old);
480 /* The number of emitted reload insns so far. */
481 int lra_curr_reload_num;
483 /* Emit x := y, processing special case when y = u + v or y = u + v *
484 scale + w through emit_add (Y can be an address which is base +
485 index reg * scale + displacement in general case). X may be used
486 as intermediate result therefore it should be not in Y. */
487 void
488 lra_emit_move (rtx x, rtx y)
490 int old;
492 if (GET_CODE (y) != PLUS)
494 if (rtx_equal_p (x, y))
495 return;
496 old = max_reg_num ();
497 emit_move_insn (x, y);
498 if (REG_P (x))
499 lra_reg_info[ORIGINAL_REGNO (x)].last_reload = ++lra_curr_reload_num;
500 /* Function emit_move can create pseudos -- so expand the pseudo
501 data. */
502 if (old != max_reg_num ())
503 expand_reg_data (old);
504 return;
506 lra_emit_add (x, XEXP (y, 0), XEXP (y, 1));
509 /* Update insn operands which are duplication of operands whose
510 numbers are in array of NOPS (with end marker -1). The insn is
511 represented by its LRA internal representation ID. */
512 void
513 lra_update_dups (lra_insn_recog_data_t id, signed char *nops)
515 int i, j, nop;
516 struct lra_static_insn_data *static_id = id->insn_static_data;
518 for (i = 0; i < static_id->n_dups; i++)
519 for (j = 0; (nop = nops[j]) >= 0; j++)
520 if (static_id->dup_num[i] == nop)
521 *id->dup_loc[i] = *id->operand_loc[nop];
526 /* This page contains code dealing with info about registers in the
527 insns. */
529 /* Pools for insn reg info. */
530 object_allocator<lra_insn_reg> lra_insn_reg_pool ("insn regs");
532 /* Create LRA insn related info about a reference to REGNO in INSN
533 with TYPE (in/out/inout), biggest reference mode MODE, flag that it
534 is reference through subreg (SUBREG_P), flag that is early
535 clobbered in the insn (EARLY_CLOBBER), and reference to the next
536 insn reg info (NEXT). If REGNO can be early clobbered,
537 alternatives in which it can be early clobbered are given by
538 EARLY_CLOBBER_ALTS. */
539 static struct lra_insn_reg *
540 new_insn_reg (rtx_insn *insn, int regno, enum op_type type,
541 machine_mode mode,
542 bool subreg_p, bool early_clobber,
543 alternative_mask early_clobber_alts,
544 struct lra_insn_reg *next)
546 lra_insn_reg *ir = lra_insn_reg_pool.allocate ();
547 ir->type = type;
548 ir->biggest_mode = mode;
549 if (NONDEBUG_INSN_P (insn)
550 && partial_subreg_p (lra_reg_info[regno].biggest_mode, mode))
551 lra_reg_info[regno].biggest_mode = mode;
552 ir->subreg_p = subreg_p;
553 ir->early_clobber = early_clobber;
554 ir->early_clobber_alts = early_clobber_alts;
555 ir->regno = regno;
556 ir->next = next;
557 return ir;
560 /* Free insn reg info list IR. */
561 static void
562 free_insn_regs (struct lra_insn_reg *ir)
564 struct lra_insn_reg *next_ir;
566 for (; ir != NULL; ir = next_ir)
568 next_ir = ir->next;
569 lra_insn_reg_pool.remove (ir);
573 /* Finish pool for insn reg info. */
574 static void
575 finish_insn_regs (void)
577 lra_insn_reg_pool.release ();
582 /* This page contains code dealing LRA insn info (or in other words
583 LRA internal insn representation). */
585 /* Map INSN_CODE -> the static insn data. This info is valid during
586 all translation unit. */
587 struct lra_static_insn_data *insn_code_data[NUM_INSN_CODES];
589 /* Debug insns are represented as a special insn with one input
590 operand which is RTL expression in var_location. */
592 /* The following data are used as static insn operand data for all
593 debug insns. If structure lra_operand_data is changed, the
594 initializer should be changed too. */
595 static struct lra_operand_data debug_operand_data =
597 NULL, /* alternative */
598 0, /* early_clobber_alts */
599 E_VOIDmode, /* We are not interesting in the operand mode. */
600 OP_IN,
601 0, 0, 0, 0
604 /* The following data are used as static insn data for all debug
605 insns. If structure lra_static_insn_data is changed, the
606 initializer should be changed too. */
607 static struct lra_static_insn_data debug_insn_static_data =
609 &debug_operand_data,
610 0, /* Duplication operands #. */
611 -1, /* Commutative operand #. */
612 1, /* Operands #. There is only one operand which is debug RTL
613 expression. */
614 0, /* Duplications #. */
615 0, /* Alternatives #. We are not interesting in alternatives
616 because we does not proceed debug_insns for reloads. */
617 NULL, /* Hard registers referenced in machine description. */
618 NULL /* Descriptions of operands in alternatives. */
621 /* Called once per compiler work to initialize some LRA data related
622 to insns. */
623 static void
624 init_insn_code_data_once (void)
626 memset (insn_code_data, 0, sizeof (insn_code_data));
629 /* Called once per compiler work to finalize some LRA data related to
630 insns. */
631 static void
632 finish_insn_code_data_once (void)
634 for (unsigned int i = 0; i < NUM_INSN_CODES; i++)
636 if (insn_code_data[i] != NULL)
637 free (insn_code_data[i]);
641 /* Return static insn data, allocate and setup if necessary. Although
642 dup_num is static data (it depends only on icode), to set it up we
643 need to extract insn first. So recog_data should be valid for
644 normal insn (ICODE >= 0) before the call. */
645 static struct lra_static_insn_data *
646 get_static_insn_data (int icode, int nop, int ndup, int nalt)
648 struct lra_static_insn_data *data;
649 size_t n_bytes;
651 lra_assert (icode < (int) NUM_INSN_CODES);
652 if (icode >= 0 && (data = insn_code_data[icode]) != NULL)
653 return data;
654 lra_assert (nop >= 0 && ndup >= 0 && nalt >= 0);
655 n_bytes = sizeof (struct lra_static_insn_data)
656 + sizeof (struct lra_operand_data) * nop
657 + sizeof (int) * ndup;
658 data = XNEWVAR (struct lra_static_insn_data, n_bytes);
659 data->operand_alternative = NULL;
660 data->n_operands = nop;
661 data->n_dups = ndup;
662 data->n_alternatives = nalt;
663 data->operand = ((struct lra_operand_data *)
664 ((char *) data + sizeof (struct lra_static_insn_data)));
665 data->dup_num = ((int *) ((char *) data->operand
666 + sizeof (struct lra_operand_data) * nop));
667 if (icode >= 0)
669 int i;
671 insn_code_data[icode] = data;
672 for (i = 0; i < nop; i++)
674 data->operand[i].constraint
675 = insn_data[icode].operand[i].constraint;
676 data->operand[i].mode = insn_data[icode].operand[i].mode;
677 data->operand[i].strict_low = insn_data[icode].operand[i].strict_low;
678 data->operand[i].is_operator
679 = insn_data[icode].operand[i].is_operator;
680 data->operand[i].type
681 = (data->operand[i].constraint[0] == '=' ? OP_OUT
682 : data->operand[i].constraint[0] == '+' ? OP_INOUT
683 : OP_IN);
684 data->operand[i].is_address = false;
686 for (i = 0; i < ndup; i++)
687 data->dup_num[i] = recog_data.dup_num[i];
689 return data;
692 /* The current length of the following array. */
693 int lra_insn_recog_data_len;
695 /* Map INSN_UID -> the insn recog data (NULL if unknown). */
696 lra_insn_recog_data_t *lra_insn_recog_data;
698 /* Initialize LRA data about insns. */
699 static void
700 init_insn_recog_data (void)
702 lra_insn_recog_data_len = 0;
703 lra_insn_recog_data = NULL;
706 /* Expand, if necessary, LRA data about insns. */
707 static void
708 check_and_expand_insn_recog_data (int index)
710 int i, old;
712 if (lra_insn_recog_data_len > index)
713 return;
714 old = lra_insn_recog_data_len;
715 lra_insn_recog_data_len = index * 3 / 2 + 1;
716 lra_insn_recog_data = XRESIZEVEC (lra_insn_recog_data_t,
717 lra_insn_recog_data,
718 lra_insn_recog_data_len);
719 for (i = old; i < lra_insn_recog_data_len; i++)
720 lra_insn_recog_data[i] = NULL;
723 /* Finish LRA DATA about insn. */
724 static void
725 free_insn_recog_data (lra_insn_recog_data_t data)
727 if (data->operand_loc != NULL)
728 free (data->operand_loc);
729 if (data->dup_loc != NULL)
730 free (data->dup_loc);
731 if (data->arg_hard_regs != NULL)
732 free (data->arg_hard_regs);
733 if (data->icode < 0 && NONDEBUG_INSN_P (data->insn))
735 if (data->insn_static_data->operand_alternative != NULL)
736 free (const_cast <operand_alternative *>
737 (data->insn_static_data->operand_alternative));
738 free_insn_regs (data->insn_static_data->hard_regs);
739 free (data->insn_static_data);
741 free_insn_regs (data->regs);
742 data->regs = NULL;
743 free (data);
746 /* Pools for copies. */
747 static object_allocator<lra_copy> lra_copy_pool ("lra copies");
749 /* Finish LRA data about all insns. */
750 static void
751 finish_insn_recog_data (void)
753 int i;
754 lra_insn_recog_data_t data;
756 for (i = 0; i < lra_insn_recog_data_len; i++)
757 if ((data = lra_insn_recog_data[i]) != NULL)
758 free_insn_recog_data (data);
759 finish_insn_regs ();
760 lra_copy_pool.release ();
761 lra_insn_reg_pool.release ();
762 free (lra_insn_recog_data);
765 /* Setup info about operands in alternatives of LRA DATA of insn. */
766 static void
767 setup_operand_alternative (lra_insn_recog_data_t data,
768 const operand_alternative *op_alt)
770 int i, j, nop, nalt;
771 int icode = data->icode;
772 struct lra_static_insn_data *static_data = data->insn_static_data;
774 static_data->commutative = -1;
775 nop = static_data->n_operands;
776 nalt = static_data->n_alternatives;
777 static_data->operand_alternative = op_alt;
778 for (i = 0; i < nop; i++)
780 static_data->operand[i].early_clobber_alts = 0;
781 static_data->operand[i].early_clobber = false;
782 static_data->operand[i].is_address = false;
783 if (static_data->operand[i].constraint[0] == '%')
785 /* We currently only support one commutative pair of operands. */
786 if (static_data->commutative < 0)
787 static_data->commutative = i;
788 else
789 lra_assert (icode < 0); /* Asm */
790 /* The last operand should not be marked commutative. */
791 lra_assert (i != nop - 1);
794 for (j = 0; j < nalt; j++)
795 for (i = 0; i < nop; i++, op_alt++)
797 static_data->operand[i].early_clobber |= op_alt->earlyclobber;
798 if (op_alt->earlyclobber)
799 static_data->operand[i].early_clobber_alts |= (alternative_mask) 1 << j;
800 static_data->operand[i].is_address |= op_alt->is_address;
804 /* Recursively process X and collect info about registers, which are
805 not the insn operands, in X with TYPE (in/out/inout) and flag that
806 it is early clobbered in the insn (EARLY_CLOBBER) and add the info
807 to LIST. X is a part of insn given by DATA. Return the result
808 list. */
809 static struct lra_insn_reg *
810 collect_non_operand_hard_regs (rtx_insn *insn, rtx *x,
811 lra_insn_recog_data_t data,
812 struct lra_insn_reg *list,
813 enum op_type type, bool early_clobber)
815 int i, j, regno, last;
816 bool subreg_p;
817 machine_mode mode;
818 struct lra_insn_reg *curr;
819 rtx op = *x;
820 enum rtx_code code = GET_CODE (op);
821 const char *fmt = GET_RTX_FORMAT (code);
823 for (i = 0; i < data->insn_static_data->n_operands; i++)
824 if (! data->insn_static_data->operand[i].is_operator
825 && x == data->operand_loc[i])
826 /* It is an operand loc. Stop here. */
827 return list;
828 for (i = 0; i < data->insn_static_data->n_dups; i++)
829 if (x == data->dup_loc[i])
830 /* It is a dup loc. Stop here. */
831 return list;
832 mode = GET_MODE (op);
833 subreg_p = false;
834 if (code == SUBREG)
836 mode = wider_subreg_mode (op);
837 if (read_modify_subreg_p (op))
838 subreg_p = true;
839 op = SUBREG_REG (op);
840 code = GET_CODE (op);
842 if (REG_P (op))
844 if ((regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER)
845 return list;
846 /* Process all regs even unallocatable ones as we need info
847 about all regs for rematerialization pass. */
848 for (last = end_hard_regno (mode, regno); regno < last; regno++)
850 for (curr = list; curr != NULL; curr = curr->next)
851 if (curr->regno == regno && curr->subreg_p == subreg_p
852 && curr->biggest_mode == mode)
854 if (curr->type != type)
855 curr->type = OP_INOUT;
856 if (early_clobber)
858 curr->early_clobber = true;
859 curr->early_clobber_alts = ALL_ALTERNATIVES;
861 break;
863 if (curr == NULL)
865 /* This is a new hard regno or the info can not be
866 integrated into the found structure. */
867 #ifdef STACK_REGS
868 early_clobber
869 = (early_clobber
870 /* This clobber is to inform popping floating
871 point stack only. */
872 && ! (FIRST_STACK_REG <= regno
873 && regno <= LAST_STACK_REG));
874 #endif
875 list = new_insn_reg (data->insn, regno, type, mode, subreg_p,
876 early_clobber,
877 early_clobber ? ALL_ALTERNATIVES : 0, list);
880 return list;
882 switch (code)
884 case SET:
885 list = collect_non_operand_hard_regs (insn, &SET_DEST (op), data,
886 list, OP_OUT, false);
887 list = collect_non_operand_hard_regs (insn, &SET_SRC (op), data,
888 list, OP_IN, false);
889 break;
890 case CLOBBER:
891 /* We treat clobber of non-operand hard registers as early clobber. */
892 list = collect_non_operand_hard_regs (insn, &XEXP (op, 0), data,
893 list, OP_OUT, true);
894 break;
895 case PRE_INC: case PRE_DEC: case POST_INC: case POST_DEC:
896 list = collect_non_operand_hard_regs (insn, &XEXP (op, 0), data,
897 list, OP_INOUT, false);
898 break;
899 case PRE_MODIFY: case POST_MODIFY:
900 list = collect_non_operand_hard_regs (insn, &XEXP (op, 0), data,
901 list, OP_INOUT, false);
902 list = collect_non_operand_hard_regs (insn, &XEXP (op, 1), data,
903 list, OP_IN, false);
904 break;
905 default:
906 fmt = GET_RTX_FORMAT (code);
907 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
909 if (fmt[i] == 'e')
910 list = collect_non_operand_hard_regs (insn, &XEXP (op, i), data,
911 list, OP_IN, false);
912 else if (fmt[i] == 'E')
913 for (j = XVECLEN (op, i) - 1; j >= 0; j--)
914 list = collect_non_operand_hard_regs (insn, &XVECEXP (op, i, j),
915 data, list, OP_IN, false);
918 return list;
921 /* Set up and return info about INSN. Set up the info if it is not set up
922 yet. */
923 lra_insn_recog_data_t
924 lra_set_insn_recog_data (rtx_insn *insn)
926 lra_insn_recog_data_t data;
927 int i, n, icode;
928 rtx **locs;
929 unsigned int uid = INSN_UID (insn);
930 struct lra_static_insn_data *insn_static_data;
932 check_and_expand_insn_recog_data (uid);
933 if (DEBUG_INSN_P (insn))
934 icode = -1;
935 else
937 icode = INSN_CODE (insn);
938 if (icode < 0)
939 /* It might be a new simple insn which is not recognized yet. */
940 INSN_CODE (insn) = icode = recog_memoized (insn);
942 data = XNEW (struct lra_insn_recog_data);
943 lra_insn_recog_data[uid] = data;
944 data->insn = insn;
945 data->used_insn_alternative = -1;
946 data->icode = icode;
947 data->regs = NULL;
948 if (DEBUG_INSN_P (insn))
950 data->insn_static_data = &debug_insn_static_data;
951 data->dup_loc = NULL;
952 data->arg_hard_regs = NULL;
953 data->preferred_alternatives = ALL_ALTERNATIVES;
954 data->operand_loc = XNEWVEC (rtx *, 1);
955 data->operand_loc[0] = &INSN_VAR_LOCATION_LOC (insn);
956 return data;
958 if (icode < 0)
960 int nop, nalt;
961 machine_mode operand_mode[MAX_RECOG_OPERANDS];
962 const char *constraints[MAX_RECOG_OPERANDS];
964 nop = asm_noperands (PATTERN (insn));
965 data->operand_loc = data->dup_loc = NULL;
966 nalt = 1;
967 if (nop < 0)
969 /* It is a special insn like USE or CLOBBER. We should
970 recognize any regular insn otherwise LRA can do nothing
971 with this insn. */
972 gcc_assert (GET_CODE (PATTERN (insn)) == USE
973 || GET_CODE (PATTERN (insn)) == CLOBBER
974 || GET_CODE (PATTERN (insn)) == ASM_INPUT);
975 data->insn_static_data = insn_static_data
976 = get_static_insn_data (-1, 0, 0, nalt);
978 else
980 /* expand_asm_operands makes sure there aren't too many
981 operands. */
982 lra_assert (nop <= MAX_RECOG_OPERANDS);
983 if (nop != 0)
984 data->operand_loc = XNEWVEC (rtx *, nop);
985 /* Now get the operand values and constraints out of the
986 insn. */
987 decode_asm_operands (PATTERN (insn), NULL,
988 data->operand_loc,
989 constraints, operand_mode, NULL);
990 if (nop > 0)
992 const char *p = recog_data.constraints[0];
994 for (p = constraints[0]; *p; p++)
995 nalt += *p == ',';
997 data->insn_static_data = insn_static_data
998 = get_static_insn_data (-1, nop, 0, nalt);
999 for (i = 0; i < nop; i++)
1001 insn_static_data->operand[i].mode = operand_mode[i];
1002 insn_static_data->operand[i].constraint = constraints[i];
1003 insn_static_data->operand[i].strict_low = false;
1004 insn_static_data->operand[i].is_operator = false;
1005 insn_static_data->operand[i].is_address = false;
1008 for (i = 0; i < insn_static_data->n_operands; i++)
1009 insn_static_data->operand[i].type
1010 = (insn_static_data->operand[i].constraint[0] == '=' ? OP_OUT
1011 : insn_static_data->operand[i].constraint[0] == '+' ? OP_INOUT
1012 : OP_IN);
1013 data->preferred_alternatives = ALL_ALTERNATIVES;
1014 if (nop > 0)
1016 operand_alternative *op_alt = XCNEWVEC (operand_alternative,
1017 nalt * nop);
1018 preprocess_constraints (nop, nalt, constraints, op_alt);
1019 setup_operand_alternative (data, op_alt);
1022 else
1024 insn_extract (insn);
1025 data->insn_static_data = insn_static_data
1026 = get_static_insn_data (icode, insn_data[icode].n_operands,
1027 insn_data[icode].n_dups,
1028 insn_data[icode].n_alternatives);
1029 n = insn_static_data->n_operands;
1030 if (n == 0)
1031 locs = NULL;
1032 else
1034 locs = XNEWVEC (rtx *, n);
1035 memcpy (locs, recog_data.operand_loc, n * sizeof (rtx *));
1037 data->operand_loc = locs;
1038 n = insn_static_data->n_dups;
1039 if (n == 0)
1040 locs = NULL;
1041 else
1043 locs = XNEWVEC (rtx *, n);
1044 memcpy (locs, recog_data.dup_loc, n * sizeof (rtx *));
1046 data->dup_loc = locs;
1047 data->preferred_alternatives = get_preferred_alternatives (insn);
1048 const operand_alternative *op_alt = preprocess_insn_constraints (icode);
1049 if (!insn_static_data->operand_alternative)
1050 setup_operand_alternative (data, op_alt);
1051 else if (op_alt != insn_static_data->operand_alternative)
1052 insn_static_data->operand_alternative = op_alt;
1054 if (GET_CODE (PATTERN (insn)) == CLOBBER || GET_CODE (PATTERN (insn)) == USE)
1055 insn_static_data->hard_regs = NULL;
1056 else
1057 insn_static_data->hard_regs
1058 = collect_non_operand_hard_regs (insn, &PATTERN (insn), data,
1059 NULL, OP_IN, false);
1060 data->arg_hard_regs = NULL;
1061 if (CALL_P (insn))
1063 bool use_p;
1064 rtx link;
1065 int n_hard_regs, regno, arg_hard_regs[FIRST_PSEUDO_REGISTER];
1067 n_hard_regs = 0;
1068 /* Finding implicit hard register usage. We believe it will be
1069 not changed whatever transformations are used. Call insns
1070 are such example. */
1071 for (link = CALL_INSN_FUNCTION_USAGE (insn);
1072 link != NULL_RTX;
1073 link = XEXP (link, 1))
1074 if (((use_p = GET_CODE (XEXP (link, 0)) == USE)
1075 || GET_CODE (XEXP (link, 0)) == CLOBBER)
1076 && REG_P (XEXP (XEXP (link, 0), 0)))
1078 regno = REGNO (XEXP (XEXP (link, 0), 0));
1079 lra_assert (regno < FIRST_PSEUDO_REGISTER);
1080 /* It is an argument register. */
1081 for (i = REG_NREGS (XEXP (XEXP (link, 0), 0)) - 1; i >= 0; i--)
1082 arg_hard_regs[n_hard_regs++]
1083 = regno + i + (use_p ? 0 : FIRST_PSEUDO_REGISTER);
1085 if (n_hard_regs != 0)
1087 arg_hard_regs[n_hard_regs++] = -1;
1088 data->arg_hard_regs = XNEWVEC (int, n_hard_regs);
1089 memcpy (data->arg_hard_regs, arg_hard_regs,
1090 sizeof (int) * n_hard_regs);
1093 /* Some output operand can be recognized only from the context not
1094 from the constraints which are empty in this case. Call insn may
1095 contain a hard register in set destination with empty constraint
1096 and extract_insn treats them as an input. */
1097 for (i = 0; i < insn_static_data->n_operands; i++)
1099 int j;
1100 rtx pat, set;
1101 struct lra_operand_data *operand = &insn_static_data->operand[i];
1103 /* ??? Should we treat 'X' the same way. It looks to me that
1104 'X' means anything and empty constraint means we do not
1105 care. */
1106 if (operand->type != OP_IN || *operand->constraint != '\0'
1107 || operand->is_operator)
1108 continue;
1109 pat = PATTERN (insn);
1110 if (GET_CODE (pat) == SET)
1112 if (data->operand_loc[i] != &SET_DEST (pat))
1113 continue;
1115 else if (GET_CODE (pat) == PARALLEL)
1117 for (j = XVECLEN (pat, 0) - 1; j >= 0; j--)
1119 set = XVECEXP (PATTERN (insn), 0, j);
1120 if (GET_CODE (set) == SET
1121 && &SET_DEST (set) == data->operand_loc[i])
1122 break;
1124 if (j < 0)
1125 continue;
1127 else
1128 continue;
1129 operand->type = OP_OUT;
1131 return data;
1134 /* Return info about insn give by UID. The info should be already set
1135 up. */
1136 static lra_insn_recog_data_t
1137 get_insn_recog_data_by_uid (int uid)
1139 lra_insn_recog_data_t data;
1141 data = lra_insn_recog_data[uid];
1142 lra_assert (data != NULL);
1143 return data;
1146 /* Invalidate all info about insn given by its UID. */
1147 static void
1148 invalidate_insn_recog_data (int uid)
1150 lra_insn_recog_data_t data;
1152 data = lra_insn_recog_data[uid];
1153 lra_assert (data != NULL);
1154 free_insn_recog_data (data);
1155 lra_insn_recog_data[uid] = NULL;
1158 /* Update all the insn info about INSN. It is usually called when
1159 something in the insn was changed. Return the updated info. */
1160 lra_insn_recog_data_t
1161 lra_update_insn_recog_data (rtx_insn *insn)
1163 lra_insn_recog_data_t data;
1164 int n;
1165 unsigned int uid = INSN_UID (insn);
1166 struct lra_static_insn_data *insn_static_data;
1167 HOST_WIDE_INT sp_offset = 0;
1169 check_and_expand_insn_recog_data (uid);
1170 if ((data = lra_insn_recog_data[uid]) != NULL
1171 && data->icode != INSN_CODE (insn))
1173 sp_offset = data->sp_offset;
1174 invalidate_insn_data_regno_info (data, insn, get_insn_freq (insn));
1175 invalidate_insn_recog_data (uid);
1176 data = NULL;
1178 if (data == NULL)
1180 data = lra_get_insn_recog_data (insn);
1181 /* Initiate or restore SP offset. */
1182 data->sp_offset = sp_offset;
1183 return data;
1185 insn_static_data = data->insn_static_data;
1186 data->used_insn_alternative = -1;
1187 if (DEBUG_INSN_P (insn))
1188 return data;
1189 if (data->icode < 0)
1191 int nop;
1192 machine_mode operand_mode[MAX_RECOG_OPERANDS];
1193 const char *constraints[MAX_RECOG_OPERANDS];
1195 nop = asm_noperands (PATTERN (insn));
1196 if (nop >= 0)
1198 lra_assert (nop == data->insn_static_data->n_operands);
1199 /* Now get the operand values and constraints out of the
1200 insn. */
1201 decode_asm_operands (PATTERN (insn), NULL,
1202 data->operand_loc,
1203 constraints, operand_mode, NULL);
1205 if (flag_checking)
1206 for (int i = 0; i < nop; i++)
1207 lra_assert
1208 (insn_static_data->operand[i].mode == operand_mode[i]
1209 && insn_static_data->operand[i].constraint == constraints[i]
1210 && ! insn_static_data->operand[i].is_operator);
1213 if (flag_checking)
1214 for (int i = 0; i < insn_static_data->n_operands; i++)
1215 lra_assert
1216 (insn_static_data->operand[i].type
1217 == (insn_static_data->operand[i].constraint[0] == '=' ? OP_OUT
1218 : insn_static_data->operand[i].constraint[0] == '+' ? OP_INOUT
1219 : OP_IN));
1221 else
1223 insn_extract (insn);
1224 n = insn_static_data->n_operands;
1225 if (n != 0)
1226 memcpy (data->operand_loc, recog_data.operand_loc, n * sizeof (rtx *));
1227 n = insn_static_data->n_dups;
1228 if (n != 0)
1229 memcpy (data->dup_loc, recog_data.dup_loc, n * sizeof (rtx *));
1230 lra_assert (check_bool_attrs (insn));
1232 return data;
1235 /* Set up that INSN is using alternative ALT now. */
1236 void
1237 lra_set_used_insn_alternative (rtx_insn *insn, int alt)
1239 lra_insn_recog_data_t data;
1241 data = lra_get_insn_recog_data (insn);
1242 data->used_insn_alternative = alt;
1245 /* Set up that insn with UID is using alternative ALT now. The insn
1246 info should be already set up. */
1247 void
1248 lra_set_used_insn_alternative_by_uid (int uid, int alt)
1250 lra_insn_recog_data_t data;
1252 check_and_expand_insn_recog_data (uid);
1253 data = lra_insn_recog_data[uid];
1254 lra_assert (data != NULL);
1255 data->used_insn_alternative = alt;
1260 /* This page contains code dealing with common register info and
1261 pseudo copies. */
1263 /* The size of the following array. */
1264 static int reg_info_size;
1265 /* Common info about each register. */
1266 struct lra_reg *lra_reg_info;
1268 /* Last register value. */
1269 static int last_reg_value;
1271 /* Return new register value. */
1272 static int
1273 get_new_reg_value (void)
1275 return ++last_reg_value;
1278 /* Vec referring to pseudo copies. */
1279 static vec<lra_copy_t> copy_vec;
1281 /* Initialize I-th element of lra_reg_info. */
1282 static inline void
1283 initialize_lra_reg_info_element (int i)
1285 bitmap_initialize (&lra_reg_info[i].insn_bitmap, &reg_obstack);
1286 #ifdef STACK_REGS
1287 lra_reg_info[i].no_stack_p = false;
1288 #endif
1289 CLEAR_HARD_REG_SET (lra_reg_info[i].conflict_hard_regs);
1290 CLEAR_HARD_REG_SET (lra_reg_info[i].actual_call_used_reg_set);
1291 lra_reg_info[i].preferred_hard_regno1 = -1;
1292 lra_reg_info[i].preferred_hard_regno2 = -1;
1293 lra_reg_info[i].preferred_hard_regno_profit1 = 0;
1294 lra_reg_info[i].preferred_hard_regno_profit2 = 0;
1295 lra_reg_info[i].biggest_mode = VOIDmode;
1296 lra_reg_info[i].live_ranges = NULL;
1297 lra_reg_info[i].nrefs = lra_reg_info[i].freq = 0;
1298 lra_reg_info[i].last_reload = 0;
1299 lra_reg_info[i].restore_rtx = NULL_RTX;
1300 lra_reg_info[i].val = get_new_reg_value ();
1301 lra_reg_info[i].offset = 0;
1302 lra_reg_info[i].copies = NULL;
1305 /* Initialize common reg info and copies. */
1306 static void
1307 init_reg_info (void)
1309 int i;
1311 last_reg_value = 0;
1312 reg_info_size = max_reg_num () * 3 / 2 + 1;
1313 lra_reg_info = XNEWVEC (struct lra_reg, reg_info_size);
1314 for (i = 0; i < reg_info_size; i++)
1315 initialize_lra_reg_info_element (i);
1316 copy_vec.truncate (0);
1320 /* Finish common reg info and copies. */
1321 static void
1322 finish_reg_info (void)
1324 int i;
1326 for (i = 0; i < reg_info_size; i++)
1327 bitmap_clear (&lra_reg_info[i].insn_bitmap);
1328 free (lra_reg_info);
1329 reg_info_size = 0;
1332 /* Expand common reg info if it is necessary. */
1333 static void
1334 expand_reg_info (void)
1336 int i, old = reg_info_size;
1338 if (reg_info_size > max_reg_num ())
1339 return;
1340 reg_info_size = max_reg_num () * 3 / 2 + 1;
1341 lra_reg_info = XRESIZEVEC (struct lra_reg, lra_reg_info, reg_info_size);
1342 for (i = old; i < reg_info_size; i++)
1343 initialize_lra_reg_info_element (i);
1346 /* Free all copies. */
1347 void
1348 lra_free_copies (void)
1350 lra_copy_t cp;
1352 while (copy_vec.length () != 0)
1354 cp = copy_vec.pop ();
1355 lra_reg_info[cp->regno1].copies = lra_reg_info[cp->regno2].copies = NULL;
1356 lra_copy_pool.remove (cp);
1360 /* Create copy of two pseudos REGNO1 and REGNO2. The copy execution
1361 frequency is FREQ. */
1362 void
1363 lra_create_copy (int regno1, int regno2, int freq)
1365 bool regno1_dest_p;
1366 lra_copy_t cp;
1368 lra_assert (regno1 != regno2);
1369 regno1_dest_p = true;
1370 if (regno1 > regno2)
1372 std::swap (regno1, regno2);
1373 regno1_dest_p = false;
1375 cp = lra_copy_pool.allocate ();
1376 copy_vec.safe_push (cp);
1377 cp->regno1_dest_p = regno1_dest_p;
1378 cp->freq = freq;
1379 cp->regno1 = regno1;
1380 cp->regno2 = regno2;
1381 cp->regno1_next = lra_reg_info[regno1].copies;
1382 lra_reg_info[regno1].copies = cp;
1383 cp->regno2_next = lra_reg_info[regno2].copies;
1384 lra_reg_info[regno2].copies = cp;
1385 if (lra_dump_file != NULL)
1386 fprintf (lra_dump_file, " Creating copy r%d%sr%d@%d\n",
1387 regno1, regno1_dest_p ? "<-" : "->", regno2, freq);
1390 /* Return N-th (0, 1, ...) copy. If there is no copy, return
1391 NULL. */
1392 lra_copy_t
1393 lra_get_copy (int n)
1395 if (n >= (int) copy_vec.length ())
1396 return NULL;
1397 return copy_vec[n];
1402 /* This page contains code dealing with info about registers in
1403 insns. */
1405 /* Process X of INSN recursively and add info (operand type is
1406 given by TYPE, flag of that it is early clobber is EARLY_CLOBBER)
1407 about registers in X to the insn DATA. If X can be early clobbered,
1408 alternatives in which it can be early clobbered are given by
1409 EARLY_CLOBBER_ALTS. */
1410 static void
1411 add_regs_to_insn_regno_info (lra_insn_recog_data_t data, rtx x,
1412 rtx_insn *insn,
1413 enum op_type type, bool early_clobber,
1414 alternative_mask early_clobber_alts)
1416 int i, j, regno;
1417 bool subreg_p;
1418 machine_mode mode;
1419 const char *fmt;
1420 enum rtx_code code;
1421 struct lra_insn_reg *curr;
1423 code = GET_CODE (x);
1424 mode = GET_MODE (x);
1425 subreg_p = false;
1426 if (GET_CODE (x) == SUBREG)
1428 mode = wider_subreg_mode (x);
1429 if (read_modify_subreg_p (x))
1430 subreg_p = true;
1431 x = SUBREG_REG (x);
1432 code = GET_CODE (x);
1434 if (REG_P (x))
1436 regno = REGNO (x);
1437 /* Process all regs even unallocatable ones as we need info about
1438 all regs for rematerialization pass. */
1439 expand_reg_info ();
1440 if (bitmap_set_bit (&lra_reg_info[regno].insn_bitmap, INSN_UID (insn)))
1442 data->regs = new_insn_reg (data->insn, regno, type, mode, subreg_p,
1443 early_clobber, early_clobber_alts,
1444 data->regs);
1445 return;
1447 else
1449 for (curr = data->regs; curr != NULL; curr = curr->next)
1450 if (curr->regno == regno)
1452 if (curr->subreg_p != subreg_p || curr->biggest_mode != mode)
1453 /* The info can not be integrated into the found
1454 structure. */
1455 data->regs = new_insn_reg (data->insn, regno, type, mode,
1456 subreg_p, early_clobber,
1457 early_clobber_alts, data->regs);
1458 else
1460 if (curr->type != type)
1461 curr->type = OP_INOUT;
1462 if (curr->early_clobber != early_clobber)
1463 curr->early_clobber = true;
1464 curr->early_clobber_alts |= early_clobber_alts;
1466 return;
1468 gcc_unreachable ();
1472 switch (code)
1474 case SET:
1475 add_regs_to_insn_regno_info (data, SET_DEST (x), insn, OP_OUT, false, 0);
1476 add_regs_to_insn_regno_info (data, SET_SRC (x), insn, OP_IN, false, 0);
1477 break;
1478 case CLOBBER:
1479 /* We treat clobber of non-operand hard registers as early
1480 clobber. */
1481 add_regs_to_insn_regno_info (data, XEXP (x, 0), insn, OP_OUT,
1482 true, ALL_ALTERNATIVES);
1483 break;
1484 case PRE_INC: case PRE_DEC: case POST_INC: case POST_DEC:
1485 add_regs_to_insn_regno_info (data, XEXP (x, 0), insn, OP_INOUT, false, 0);
1486 break;
1487 case PRE_MODIFY: case POST_MODIFY:
1488 add_regs_to_insn_regno_info (data, XEXP (x, 0), insn, OP_INOUT, false, 0);
1489 add_regs_to_insn_regno_info (data, XEXP (x, 1), insn, OP_IN, false, 0);
1490 break;
1491 default:
1492 if ((code != PARALLEL && code != EXPR_LIST) || type != OP_OUT)
1493 /* Some targets place small structures in registers for return
1494 values of functions, and those registers are wrapped in
1495 PARALLEL that we may see as the destination of a SET. Here
1496 is an example:
1498 (call_insn 13 12 14 2 (set (parallel:BLK [
1499 (expr_list:REG_DEP_TRUE (reg:DI 0 ax)
1500 (const_int 0 [0]))
1501 (expr_list:REG_DEP_TRUE (reg:DI 1 dx)
1502 (const_int 8 [0x8]))
1504 (call (mem:QI (symbol_ref:DI (... */
1505 type = OP_IN;
1506 fmt = GET_RTX_FORMAT (code);
1507 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1509 if (fmt[i] == 'e')
1510 add_regs_to_insn_regno_info (data, XEXP (x, i), insn, type, false, 0);
1511 else if (fmt[i] == 'E')
1513 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1514 add_regs_to_insn_regno_info (data, XVECEXP (x, i, j), insn,
1515 type, false, 0);
1521 /* Return execution frequency of INSN. */
1522 static int
1523 get_insn_freq (rtx_insn *insn)
1525 basic_block bb = BLOCK_FOR_INSN (insn);
1527 gcc_checking_assert (bb != NULL);
1528 return REG_FREQ_FROM_BB (bb);
1531 /* Invalidate all reg info of INSN with DATA and execution frequency
1532 FREQ. Update common info about the invalidated registers. */
1533 static void
1534 invalidate_insn_data_regno_info (lra_insn_recog_data_t data, rtx_insn *insn,
1535 int freq)
1537 int uid;
1538 bool debug_p;
1539 unsigned int i;
1540 struct lra_insn_reg *ir, *next_ir;
1542 uid = INSN_UID (insn);
1543 debug_p = DEBUG_INSN_P (insn);
1544 for (ir = data->regs; ir != NULL; ir = next_ir)
1546 i = ir->regno;
1547 next_ir = ir->next;
1548 lra_insn_reg_pool.remove (ir);
1549 bitmap_clear_bit (&lra_reg_info[i].insn_bitmap, uid);
1550 if (i >= FIRST_PSEUDO_REGISTER && ! debug_p)
1552 lra_reg_info[i].nrefs--;
1553 lra_reg_info[i].freq -= freq;
1554 lra_assert (lra_reg_info[i].nrefs >= 0 && lra_reg_info[i].freq >= 0);
1557 data->regs = NULL;
1560 /* Invalidate all reg info of INSN. Update common info about the
1561 invalidated registers. */
1562 void
1563 lra_invalidate_insn_regno_info (rtx_insn *insn)
1565 invalidate_insn_data_regno_info (lra_get_insn_recog_data (insn), insn,
1566 get_insn_freq (insn));
1569 /* Update common reg info from reg info of insn given by its DATA and
1570 execution frequency FREQ. */
1571 static void
1572 setup_insn_reg_info (lra_insn_recog_data_t data, int freq)
1574 unsigned int i;
1575 struct lra_insn_reg *ir;
1577 for (ir = data->regs; ir != NULL; ir = ir->next)
1578 if ((i = ir->regno) >= FIRST_PSEUDO_REGISTER)
1580 lra_reg_info[i].nrefs++;
1581 lra_reg_info[i].freq += freq;
1585 /* Set up insn reg info of INSN. Update common reg info from reg info
1586 of INSN. */
1587 void
1588 lra_update_insn_regno_info (rtx_insn *insn)
1590 int i, freq;
1591 lra_insn_recog_data_t data;
1592 struct lra_static_insn_data *static_data;
1593 enum rtx_code code;
1594 rtx link;
1596 if (! INSN_P (insn))
1597 return;
1598 data = lra_get_insn_recog_data (insn);
1599 static_data = data->insn_static_data;
1600 freq = get_insn_freq (insn);
1601 invalidate_insn_data_regno_info (data, insn, freq);
1602 for (i = static_data->n_operands - 1; i >= 0; i--)
1603 add_regs_to_insn_regno_info (data, *data->operand_loc[i], insn,
1604 static_data->operand[i].type,
1605 static_data->operand[i].early_clobber,
1606 static_data->operand[i].early_clobber_alts);
1607 if ((code = GET_CODE (PATTERN (insn))) == CLOBBER || code == USE)
1608 add_regs_to_insn_regno_info (data, XEXP (PATTERN (insn), 0), insn,
1609 code == USE ? OP_IN : OP_OUT, false, 0);
1610 if (CALL_P (insn))
1611 /* On some targets call insns can refer to pseudos in memory in
1612 CALL_INSN_FUNCTION_USAGE list. Process them in order to
1613 consider their occurrences in calls for different
1614 transformations (e.g. inheritance) with given pseudos. */
1615 for (link = CALL_INSN_FUNCTION_USAGE (insn);
1616 link != NULL_RTX;
1617 link = XEXP (link, 1))
1618 if (((code = GET_CODE (XEXP (link, 0))) == USE || code == CLOBBER)
1619 && MEM_P (XEXP (XEXP (link, 0), 0)))
1620 add_regs_to_insn_regno_info (data, XEXP (XEXP (link, 0), 0), insn,
1621 code == USE ? OP_IN : OP_OUT, false, 0);
1622 if (NONDEBUG_INSN_P (insn))
1623 setup_insn_reg_info (data, freq);
1626 /* Return reg info of insn given by it UID. */
1627 struct lra_insn_reg *
1628 lra_get_insn_regs (int uid)
1630 lra_insn_recog_data_t data;
1632 data = get_insn_recog_data_by_uid (uid);
1633 return data->regs;
1638 /* Recursive hash function for RTL X. */
1639 hashval_t
1640 lra_rtx_hash (rtx x)
1642 int i, j;
1643 enum rtx_code code;
1644 const char *fmt;
1645 hashval_t val = 0;
1647 if (x == 0)
1648 return val;
1650 code = GET_CODE (x);
1651 val += (int) code + 4095;
1653 /* Some RTL can be compared nonrecursively. */
1654 switch (code)
1656 case REG:
1657 return val + REGNO (x);
1659 case LABEL_REF:
1660 return iterative_hash_object (XEXP (x, 0), val);
1662 case SYMBOL_REF:
1663 return iterative_hash_object (XSTR (x, 0), val);
1665 case SCRATCH:
1666 case CONST_DOUBLE:
1667 case CONST_INT:
1668 case CONST_VECTOR:
1669 return val;
1671 default:
1672 break;
1675 /* Hash the elements. */
1676 fmt = GET_RTX_FORMAT (code);
1677 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1679 switch (fmt[i])
1681 case 'w':
1682 val += XWINT (x, i);
1683 break;
1685 case 'n':
1686 case 'i':
1687 val += XINT (x, i);
1688 break;
1690 case 'V':
1691 case 'E':
1692 val += XVECLEN (x, i);
1694 for (j = 0; j < XVECLEN (x, i); j++)
1695 val += lra_rtx_hash (XVECEXP (x, i, j));
1696 break;
1698 case 'e':
1699 val += lra_rtx_hash (XEXP (x, i));
1700 break;
1702 case 'S':
1703 case 's':
1704 val += htab_hash_string (XSTR (x, i));
1705 break;
1707 case 'u':
1708 case '0':
1709 case 't':
1710 break;
1712 /* It is believed that rtx's at this level will never
1713 contain anything but integers and other rtx's, except for
1714 within LABEL_REFs and SYMBOL_REFs. */
1715 default:
1716 abort ();
1719 return val;
1724 /* This page contains code dealing with stack of the insns which
1725 should be processed by the next constraint pass. */
1727 /* Bitmap used to put an insn on the stack only in one exemplar. */
1728 static sbitmap lra_constraint_insn_stack_bitmap;
1730 /* The stack itself. */
1731 vec<rtx_insn *> lra_constraint_insn_stack;
1733 /* Put INSN on the stack. If ALWAYS_UPDATE is true, always update the reg
1734 info for INSN, otherwise only update it if INSN is not already on the
1735 stack. */
1736 static inline void
1737 lra_push_insn_1 (rtx_insn *insn, bool always_update)
1739 unsigned int uid = INSN_UID (insn);
1740 if (always_update)
1741 lra_update_insn_regno_info (insn);
1742 if (uid >= SBITMAP_SIZE (lra_constraint_insn_stack_bitmap))
1743 lra_constraint_insn_stack_bitmap =
1744 sbitmap_resize (lra_constraint_insn_stack_bitmap, 3 * uid / 2, 0);
1745 if (bitmap_bit_p (lra_constraint_insn_stack_bitmap, uid))
1746 return;
1747 bitmap_set_bit (lra_constraint_insn_stack_bitmap, uid);
1748 if (! always_update)
1749 lra_update_insn_regno_info (insn);
1750 lra_constraint_insn_stack.safe_push (insn);
1753 /* Put INSN on the stack. */
1754 void
1755 lra_push_insn (rtx_insn *insn)
1757 lra_push_insn_1 (insn, false);
1760 /* Put INSN on the stack and update its reg info. */
1761 void
1762 lra_push_insn_and_update_insn_regno_info (rtx_insn *insn)
1764 lra_push_insn_1 (insn, true);
1767 /* Put insn with UID on the stack. */
1768 void
1769 lra_push_insn_by_uid (unsigned int uid)
1771 lra_push_insn (lra_insn_recog_data[uid]->insn);
1774 /* Take the last-inserted insns off the stack and return it. */
1775 rtx_insn *
1776 lra_pop_insn (void)
1778 rtx_insn *insn = lra_constraint_insn_stack.pop ();
1779 bitmap_clear_bit (lra_constraint_insn_stack_bitmap, INSN_UID (insn));
1780 return insn;
1783 /* Return the current size of the insn stack. */
1784 unsigned int
1785 lra_insn_stack_length (void)
1787 return lra_constraint_insn_stack.length ();
1790 /* Push insns FROM to TO (excluding it) going in reverse order. */
1791 static void
1792 push_insns (rtx_insn *from, rtx_insn *to)
1794 rtx_insn *insn;
1796 if (from == NULL_RTX)
1797 return;
1798 for (insn = from; insn != to; insn = PREV_INSN (insn))
1799 if (INSN_P (insn))
1800 lra_push_insn (insn);
1803 /* Set up sp offset for insn in range [FROM, LAST]. The offset is
1804 taken from the next BB insn after LAST or zero if there in such
1805 insn. */
1806 static void
1807 setup_sp_offset (rtx_insn *from, rtx_insn *last)
1809 rtx_insn *before = next_nonnote_insn_bb (last);
1810 HOST_WIDE_INT offset = (before == NULL_RTX || ! INSN_P (before)
1811 ? 0 : lra_get_insn_recog_data (before)->sp_offset);
1813 for (rtx_insn *insn = from; insn != NEXT_INSN (last); insn = NEXT_INSN (insn))
1814 lra_get_insn_recog_data (insn)->sp_offset = offset;
1817 /* Emit insns BEFORE before INSN and insns AFTER after INSN. Put the
1818 insns onto the stack. Print about emitting the insns with
1819 TITLE. */
1820 void
1821 lra_process_new_insns (rtx_insn *insn, rtx_insn *before, rtx_insn *after,
1822 const char *title)
1824 rtx_insn *last;
1826 if (before == NULL_RTX && after == NULL_RTX)
1827 return;
1828 if (lra_dump_file != NULL)
1830 dump_insn_slim (lra_dump_file, insn);
1831 if (before != NULL_RTX)
1833 fprintf (lra_dump_file," %s before:\n", title);
1834 dump_rtl_slim (lra_dump_file, before, NULL, -1, 0);
1836 if (after != NULL_RTX)
1838 fprintf (lra_dump_file, " %s after:\n", title);
1839 dump_rtl_slim (lra_dump_file, after, NULL, -1, 0);
1841 fprintf (lra_dump_file, "\n");
1843 if (before != NULL_RTX)
1845 if (cfun->can_throw_non_call_exceptions)
1846 copy_reg_eh_region_note_forward (insn, before, NULL);
1847 emit_insn_before (before, insn);
1848 push_insns (PREV_INSN (insn), PREV_INSN (before));
1849 setup_sp_offset (before, PREV_INSN (insn));
1851 if (after != NULL_RTX)
1853 if (cfun->can_throw_non_call_exceptions)
1854 copy_reg_eh_region_note_forward (insn, after, NULL);
1855 for (last = after; NEXT_INSN (last) != NULL_RTX; last = NEXT_INSN (last))
1857 emit_insn_after (after, insn);
1858 push_insns (last, insn);
1859 setup_sp_offset (after, last);
1861 if (cfun->can_throw_non_call_exceptions)
1863 rtx note = find_reg_note (insn, REG_EH_REGION, NULL_RTX);
1864 if (note && !insn_could_throw_p (insn))
1865 remove_note (insn, note);
1870 /* Replace all references to register OLD_REGNO in *LOC with pseudo
1871 register NEW_REG. Try to simplify subreg of constant if SUBREG_P.
1872 Return true if any change was made. */
1873 bool
1874 lra_substitute_pseudo (rtx *loc, int old_regno, rtx new_reg, bool subreg_p)
1876 rtx x = *loc;
1877 bool result = false;
1878 enum rtx_code code;
1879 const char *fmt;
1880 int i, j;
1882 if (x == NULL_RTX)
1883 return false;
1885 code = GET_CODE (x);
1886 if (code == SUBREG && subreg_p)
1888 rtx subst, inner = SUBREG_REG (x);
1889 /* Transform subreg of constant while we still have inner mode
1890 of the subreg. The subreg internal should not be an insn
1891 operand. */
1892 if (REG_P (inner) && (int) REGNO (inner) == old_regno
1893 && CONSTANT_P (new_reg)
1894 && (subst = simplify_subreg (GET_MODE (x), new_reg, GET_MODE (inner),
1895 SUBREG_BYTE (x))) != NULL_RTX)
1897 *loc = subst;
1898 return true;
1902 else if (code == REG && (int) REGNO (x) == old_regno)
1904 machine_mode mode = GET_MODE (x);
1905 machine_mode inner_mode = GET_MODE (new_reg);
1907 if (mode != inner_mode
1908 && ! (CONST_INT_P (new_reg) && SCALAR_INT_MODE_P (mode)))
1910 if (!partial_subreg_p (mode, inner_mode)
1911 || ! SCALAR_INT_MODE_P (inner_mode))
1912 new_reg = gen_rtx_SUBREG (mode, new_reg, 0);
1913 else
1914 new_reg = gen_lowpart_SUBREG (mode, new_reg);
1916 *loc = new_reg;
1917 return true;
1920 /* Scan all the operand sub-expressions. */
1921 fmt = GET_RTX_FORMAT (code);
1922 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1924 if (fmt[i] == 'e')
1926 if (lra_substitute_pseudo (&XEXP (x, i), old_regno,
1927 new_reg, subreg_p))
1928 result = true;
1930 else if (fmt[i] == 'E')
1932 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1933 if (lra_substitute_pseudo (&XVECEXP (x, i, j), old_regno,
1934 new_reg, subreg_p))
1935 result = true;
1938 return result;
1941 /* Call lra_substitute_pseudo within an insn. Try to simplify subreg
1942 of constant if SUBREG_P. This won't update the insn ptr, just the
1943 contents of the insn. */
1944 bool
1945 lra_substitute_pseudo_within_insn (rtx_insn *insn, int old_regno,
1946 rtx new_reg, bool subreg_p)
1948 rtx loc = insn;
1949 return lra_substitute_pseudo (&loc, old_regno, new_reg, subreg_p);
1954 /* This page contains code dealing with scratches (changing them onto
1955 pseudos and restoring them from the pseudos).
1957 We change scratches into pseudos at the beginning of LRA to
1958 simplify dealing with them (conflicts, hard register assignments).
1960 If the pseudo denoting scratch was spilled it means that we do need
1961 a hard register for it. Such pseudos are transformed back to
1962 scratches at the end of LRA. */
1964 /* Description of location of a former scratch operand. */
1965 struct sloc
1967 rtx_insn *insn; /* Insn where the scratch was. */
1968 int nop; /* Number of the operand which was a scratch. */
1971 typedef struct sloc *sloc_t;
1973 /* Locations of the former scratches. */
1974 static vec<sloc_t> scratches;
1976 /* Bitmap of scratch regnos. */
1977 static bitmap_head scratch_bitmap;
1979 /* Bitmap of scratch operands. */
1980 static bitmap_head scratch_operand_bitmap;
1982 /* Return true if pseudo REGNO is made of SCRATCH. */
1983 bool
1984 lra_former_scratch_p (int regno)
1986 return bitmap_bit_p (&scratch_bitmap, regno);
1989 /* Return true if the operand NOP of INSN is a former scratch. */
1990 bool
1991 lra_former_scratch_operand_p (rtx_insn *insn, int nop)
1993 return bitmap_bit_p (&scratch_operand_bitmap,
1994 INSN_UID (insn) * MAX_RECOG_OPERANDS + nop) != 0;
1997 /* Register operand NOP in INSN as a former scratch. It will be
1998 changed to scratch back, if it is necessary, at the LRA end. */
1999 void
2000 lra_register_new_scratch_op (rtx_insn *insn, int nop)
2002 lra_insn_recog_data_t id = lra_get_insn_recog_data (insn);
2003 rtx op = *id->operand_loc[nop];
2004 sloc_t loc = XNEW (struct sloc);
2005 lra_assert (REG_P (op));
2006 loc->insn = insn;
2007 loc->nop = nop;
2008 scratches.safe_push (loc);
2009 bitmap_set_bit (&scratch_bitmap, REGNO (op));
2010 bitmap_set_bit (&scratch_operand_bitmap,
2011 INSN_UID (insn) * MAX_RECOG_OPERANDS + nop);
2012 add_reg_note (insn, REG_UNUSED, op);
2015 /* Change scratches onto pseudos and save their location. */
2016 static void
2017 remove_scratches (void)
2019 int i;
2020 bool insn_changed_p;
2021 basic_block bb;
2022 rtx_insn *insn;
2023 rtx reg;
2024 lra_insn_recog_data_t id;
2025 struct lra_static_insn_data *static_id;
2027 scratches.create (get_max_uid ());
2028 bitmap_initialize (&scratch_bitmap, &reg_obstack);
2029 bitmap_initialize (&scratch_operand_bitmap, &reg_obstack);
2030 FOR_EACH_BB_FN (bb, cfun)
2031 FOR_BB_INSNS (bb, insn)
2032 if (INSN_P (insn))
2034 id = lra_get_insn_recog_data (insn);
2035 static_id = id->insn_static_data;
2036 insn_changed_p = false;
2037 for (i = 0; i < static_id->n_operands; i++)
2038 if (GET_CODE (*id->operand_loc[i]) == SCRATCH
2039 && GET_MODE (*id->operand_loc[i]) != VOIDmode)
2041 insn_changed_p = true;
2042 *id->operand_loc[i] = reg
2043 = lra_create_new_reg (static_id->operand[i].mode,
2044 *id->operand_loc[i], ALL_REGS, NULL);
2045 lra_register_new_scratch_op (insn, i);
2046 if (lra_dump_file != NULL)
2047 fprintf (lra_dump_file,
2048 "Removing SCRATCH in insn #%u (nop %d)\n",
2049 INSN_UID (insn), i);
2051 if (insn_changed_p)
2052 /* Because we might use DF right after caller-saves sub-pass
2053 we need to keep DF info up to date. */
2054 df_insn_rescan (insn);
2058 /* Changes pseudos created by function remove_scratches onto scratches. */
2059 static void
2060 restore_scratches (void)
2062 int regno;
2063 unsigned i;
2064 sloc_t loc;
2065 rtx_insn *last = NULL;
2066 lra_insn_recog_data_t id = NULL;
2068 for (i = 0; scratches.iterate (i, &loc); i++)
2070 /* Ignore already deleted insns. */
2071 if (NOTE_P (loc->insn)
2072 && NOTE_KIND (loc->insn) == NOTE_INSN_DELETED)
2073 continue;
2074 if (last != loc->insn)
2076 last = loc->insn;
2077 id = lra_get_insn_recog_data (last);
2079 if (REG_P (*id->operand_loc[loc->nop])
2080 && ((regno = REGNO (*id->operand_loc[loc->nop]))
2081 >= FIRST_PSEUDO_REGISTER)
2082 && lra_get_regno_hard_regno (regno) < 0)
2084 /* It should be only case when scratch register with chosen
2085 constraint 'X' did not get memory or hard register. */
2086 lra_assert (lra_former_scratch_p (regno));
2087 *id->operand_loc[loc->nop]
2088 = gen_rtx_SCRATCH (GET_MODE (*id->operand_loc[loc->nop]));
2089 lra_update_dup (id, loc->nop);
2090 if (lra_dump_file != NULL)
2091 fprintf (lra_dump_file, "Restoring SCRATCH in insn #%u(nop %d)\n",
2092 INSN_UID (loc->insn), loc->nop);
2095 for (i = 0; scratches.iterate (i, &loc); i++)
2096 free (loc);
2097 scratches.release ();
2098 bitmap_clear (&scratch_bitmap);
2099 bitmap_clear (&scratch_operand_bitmap);
2104 /* Function checks RTL for correctness. If FINAL_P is true, it is
2105 done at the end of LRA and the check is more rigorous. */
2106 static void
2107 check_rtl (bool final_p)
2109 basic_block bb;
2110 rtx_insn *insn;
2112 lra_assert (! final_p || reload_completed);
2113 FOR_EACH_BB_FN (bb, cfun)
2114 FOR_BB_INSNS (bb, insn)
2115 if (NONDEBUG_INSN_P (insn)
2116 && GET_CODE (PATTERN (insn)) != USE
2117 && GET_CODE (PATTERN (insn)) != CLOBBER
2118 && GET_CODE (PATTERN (insn)) != ASM_INPUT)
2120 if (final_p)
2122 extract_constrain_insn (insn);
2123 continue;
2125 /* LRA code is based on assumption that all addresses can be
2126 correctly decomposed. LRA can generate reloads for
2127 decomposable addresses. The decomposition code checks the
2128 correctness of the addresses. So we don't need to check
2129 the addresses here. Don't call insn_invalid_p here, it can
2130 change the code at this stage. */
2131 if (recog_memoized (insn) < 0 && asm_noperands (PATTERN (insn)) < 0)
2132 fatal_insn_not_found (insn);
2136 /* Determine if the current function has an exception receiver block
2137 that reaches the exit block via non-exceptional edges */
2138 static bool
2139 has_nonexceptional_receiver (void)
2141 edge e;
2142 edge_iterator ei;
2143 basic_block *tos, *worklist, bb;
2145 /* If we're not optimizing, then just err on the safe side. */
2146 if (!optimize)
2147 return true;
2149 /* First determine which blocks can reach exit via normal paths. */
2150 tos = worklist = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun) + 1);
2152 FOR_EACH_BB_FN (bb, cfun)
2153 bb->flags &= ~BB_REACHABLE;
2155 /* Place the exit block on our worklist. */
2156 EXIT_BLOCK_PTR_FOR_FN (cfun)->flags |= BB_REACHABLE;
2157 *tos++ = EXIT_BLOCK_PTR_FOR_FN (cfun);
2159 /* Iterate: find everything reachable from what we've already seen. */
2160 while (tos != worklist)
2162 bb = *--tos;
2164 FOR_EACH_EDGE (e, ei, bb->preds)
2165 if (e->flags & EDGE_ABNORMAL)
2167 free (worklist);
2168 return true;
2170 else
2172 basic_block src = e->src;
2174 if (!(src->flags & BB_REACHABLE))
2176 src->flags |= BB_REACHABLE;
2177 *tos++ = src;
2181 free (worklist);
2182 /* No exceptional block reached exit unexceptionally. */
2183 return false;
2187 /* Process recursively X of INSN and add REG_INC notes if necessary. */
2188 static void
2189 add_auto_inc_notes (rtx_insn *insn, rtx x)
2191 enum rtx_code code = GET_CODE (x);
2192 const char *fmt;
2193 int i, j;
2195 if (code == MEM && auto_inc_p (XEXP (x, 0)))
2197 add_reg_note (insn, REG_INC, XEXP (XEXP (x, 0), 0));
2198 return;
2201 /* Scan all X sub-expressions. */
2202 fmt = GET_RTX_FORMAT (code);
2203 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2205 if (fmt[i] == 'e')
2206 add_auto_inc_notes (insn, XEXP (x, i));
2207 else if (fmt[i] == 'E')
2208 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
2209 add_auto_inc_notes (insn, XVECEXP (x, i, j));
2214 /* Remove all REG_DEAD and REG_UNUSED notes and regenerate REG_INC.
2215 We change pseudos by hard registers without notification of DF and
2216 that can make the notes obsolete. DF-infrastructure does not deal
2217 with REG_INC notes -- so we should regenerate them here. */
2218 static void
2219 update_inc_notes (void)
2221 rtx *pnote;
2222 basic_block bb;
2223 rtx_insn *insn;
2225 FOR_EACH_BB_FN (bb, cfun)
2226 FOR_BB_INSNS (bb, insn)
2227 if (NONDEBUG_INSN_P (insn))
2229 pnote = &REG_NOTES (insn);
2230 while (*pnote != 0)
2232 if (REG_NOTE_KIND (*pnote) == REG_DEAD
2233 || REG_NOTE_KIND (*pnote) == REG_UNUSED
2234 || REG_NOTE_KIND (*pnote) == REG_INC)
2235 *pnote = XEXP (*pnote, 1);
2236 else
2237 pnote = &XEXP (*pnote, 1);
2240 if (AUTO_INC_DEC)
2241 add_auto_inc_notes (insn, PATTERN (insn));
2245 /* Set to 1 while in lra. */
2246 int lra_in_progress;
2248 /* Start of pseudo regnos before the LRA. */
2249 int lra_new_regno_start;
2251 /* Start of reload pseudo regnos before the new spill pass. */
2252 int lra_constraint_new_regno_start;
2254 /* Avoid spilling pseudos with regno more than the following value if
2255 it is possible. */
2256 int lra_bad_spill_regno_start;
2258 /* Inheritance pseudo regnos before the new spill pass. */
2259 bitmap_head lra_inheritance_pseudos;
2261 /* Split regnos before the new spill pass. */
2262 bitmap_head lra_split_regs;
2264 /* Reload pseudo regnos before the new assignment pass which still can
2265 be spilled after the assignment pass as memory is also accepted in
2266 insns for the reload pseudos. */
2267 bitmap_head lra_optional_reload_pseudos;
2269 /* Pseudo regnos used for subreg reloads before the new assignment
2270 pass. Such pseudos still can be spilled after the assignment
2271 pass. */
2272 bitmap_head lra_subreg_reload_pseudos;
2274 /* File used for output of LRA debug information. */
2275 FILE *lra_dump_file;
2277 /* True if we should try spill into registers of different classes
2278 instead of memory. */
2279 bool lra_reg_spill_p;
2281 /* Set up value LRA_REG_SPILL_P. */
2282 static void
2283 setup_reg_spill_flag (void)
2285 int cl, mode;
2287 if (targetm.spill_class != NULL)
2288 for (cl = 0; cl < (int) LIM_REG_CLASSES; cl++)
2289 for (mode = 0; mode < MAX_MACHINE_MODE; mode++)
2290 if (targetm.spill_class ((enum reg_class) cl,
2291 (machine_mode) mode) != NO_REGS)
2293 lra_reg_spill_p = true;
2294 return;
2296 lra_reg_spill_p = false;
2299 /* True if the current function is too big to use regular algorithms
2300 in LRA. In other words, we should use simpler and faster algorithms
2301 in LRA. It also means we should not worry about generation code
2302 for caller saves. The value is set up in IRA. */
2303 bool lra_simple_p;
2305 /* Major LRA entry function. F is a file should be used to dump LRA
2306 debug info. */
2307 void
2308 lra (FILE *f)
2310 int i;
2311 bool live_p, inserted_p;
2313 lra_dump_file = f;
2315 timevar_push (TV_LRA);
2317 /* Make sure that the last insn is a note. Some subsequent passes
2318 need it. */
2319 emit_note (NOTE_INSN_DELETED);
2321 COPY_HARD_REG_SET (lra_no_alloc_regs, ira_no_alloc_regs);
2323 init_reg_info ();
2324 expand_reg_info ();
2326 init_insn_recog_data ();
2328 /* Some quick check on RTL generated by previous passes. */
2329 if (flag_checking)
2330 check_rtl (false);
2332 lra_in_progress = 1;
2334 lra_live_range_iter = lra_coalesce_iter = lra_constraint_iter = 0;
2335 lra_assignment_iter = lra_assignment_iter_after_spill = 0;
2336 lra_inheritance_iter = lra_undo_inheritance_iter = 0;
2337 lra_rematerialization_iter = 0;
2339 setup_reg_spill_flag ();
2341 /* Function remove_scratches can creates new pseudos for clobbers --
2342 so set up lra_constraint_new_regno_start before its call to
2343 permit changing reg classes for pseudos created by this
2344 simplification. */
2345 lra_constraint_new_regno_start = lra_new_regno_start = max_reg_num ();
2346 lra_bad_spill_regno_start = INT_MAX;
2347 remove_scratches ();
2349 /* A function that has a non-local label that can reach the exit
2350 block via non-exceptional paths must save all call-saved
2351 registers. */
2352 if (cfun->has_nonlocal_label && has_nonexceptional_receiver ())
2353 crtl->saves_all_registers = 1;
2355 if (crtl->saves_all_registers)
2356 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2357 if (! call_used_regs[i] && ! fixed_regs[i] && ! LOCAL_REGNO (i))
2358 df_set_regs_ever_live (i, true);
2360 /* We don't DF from now and avoid its using because it is to
2361 expensive when a lot of RTL changes are made. */
2362 df_set_flags (DF_NO_INSN_RESCAN);
2363 lra_constraint_insn_stack.create (get_max_uid ());
2364 lra_constraint_insn_stack_bitmap = sbitmap_alloc (get_max_uid ());
2365 bitmap_clear (lra_constraint_insn_stack_bitmap);
2366 lra_live_ranges_init ();
2367 lra_constraints_init ();
2368 lra_curr_reload_num = 0;
2369 push_insns (get_last_insn (), NULL);
2370 /* It is needed for the 1st coalescing. */
2371 bitmap_initialize (&lra_inheritance_pseudos, &reg_obstack);
2372 bitmap_initialize (&lra_split_regs, &reg_obstack);
2373 bitmap_initialize (&lra_optional_reload_pseudos, &reg_obstack);
2374 bitmap_initialize (&lra_subreg_reload_pseudos, &reg_obstack);
2375 live_p = false;
2376 if (get_frame_size () != 0 && crtl->stack_alignment_needed)
2377 /* If we have a stack frame, we must align it now. The stack size
2378 may be a part of the offset computation for register
2379 elimination. */
2380 assign_stack_local (BLKmode, 0, crtl->stack_alignment_needed);
2381 lra_init_equiv ();
2382 for (;;)
2384 for (;;)
2386 bool reloads_p = lra_constraints (lra_constraint_iter == 0);
2387 /* Constraint transformations may result in that eliminable
2388 hard regs become uneliminable and pseudos which use them
2389 should be spilled. It is better to do it before pseudo
2390 assignments.
2392 For example, rs6000 can make
2393 RS6000_PIC_OFFSET_TABLE_REGNUM uneliminable if we started
2394 to use a constant pool. */
2395 lra_eliminate (false, false);
2396 /* We should try to assign hard registers to scratches even
2397 if there were no RTL transformations in lra_constraints.
2398 Also we should check IRA assignments on the first
2399 iteration as they can be wrong because of early clobbers
2400 operands which are ignored in IRA. */
2401 if (! reloads_p && lra_constraint_iter > 1)
2403 /* Stack is not empty here only when there are changes
2404 during the elimination sub-pass. */
2405 if (bitmap_empty_p (lra_constraint_insn_stack_bitmap))
2406 break;
2407 else
2408 /* If there are no reloads but changing due
2409 elimination, restart the constraint sub-pass
2410 first. */
2411 continue;
2413 /* Do inheritance only for regular algorithms. */
2414 if (! lra_simple_p)
2416 if (flag_ipa_ra)
2418 if (live_p)
2419 lra_clear_live_ranges ();
2420 /* As a side-effect of lra_create_live_ranges, we calculate
2421 actual_call_used_reg_set, which is needed during
2422 lra_inheritance. */
2423 lra_create_live_ranges (true, true);
2424 live_p = true;
2426 lra_inheritance ();
2428 if (live_p)
2429 lra_clear_live_ranges ();
2430 /* We need live ranges for lra_assign -- so build them. But
2431 don't remove dead insns or change global live info as we
2432 can undo inheritance transformations after inheritance
2433 pseudo assigning. */
2434 lra_create_live_ranges (true, false);
2435 live_p = true;
2436 /* If we don't spill non-reload and non-inheritance pseudos,
2437 there is no sense to run memory-memory move coalescing.
2438 If inheritance pseudos were spilled, the memory-memory
2439 moves involving them will be removed by pass undoing
2440 inheritance. */
2441 if (lra_simple_p)
2442 lra_assign ();
2443 else
2445 bool spill_p = !lra_assign ();
2447 if (lra_undo_inheritance ())
2448 live_p = false;
2449 if (spill_p)
2451 if (! live_p)
2453 lra_create_live_ranges (true, true);
2454 live_p = true;
2456 if (lra_coalesce ())
2457 live_p = false;
2459 if (! live_p)
2460 lra_clear_live_ranges ();
2463 /* Don't clear optional reloads bitmap until all constraints are
2464 satisfied as we need to differ them from regular reloads. */
2465 bitmap_clear (&lra_optional_reload_pseudos);
2466 bitmap_clear (&lra_subreg_reload_pseudos);
2467 bitmap_clear (&lra_inheritance_pseudos);
2468 bitmap_clear (&lra_split_regs);
2469 if (! live_p)
2471 /* We need full live info for spilling pseudos into
2472 registers instead of memory. */
2473 lra_create_live_ranges (lra_reg_spill_p, true);
2474 live_p = true;
2476 /* We should check necessity for spilling here as the above live
2477 range pass can remove spilled pseudos. */
2478 if (! lra_need_for_spills_p ())
2479 break;
2480 /* Now we know what pseudos should be spilled. Try to
2481 rematerialize them first. */
2482 if (lra_remat ())
2484 /* We need full live info -- see the comment above. */
2485 lra_create_live_ranges (lra_reg_spill_p, true);
2486 live_p = true;
2487 if (! lra_need_for_spills_p ())
2488 break;
2490 lra_spill ();
2491 /* Assignment of stack slots changes elimination offsets for
2492 some eliminations. So update the offsets here. */
2493 lra_eliminate (false, false);
2494 lra_constraint_new_regno_start = max_reg_num ();
2495 if (lra_bad_spill_regno_start == INT_MAX
2496 && lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES
2497 && lra_rematerialization_iter > LRA_MAX_REMATERIALIZATION_PASSES)
2498 /* After switching off inheritance and rematerialization
2499 passes, avoid spilling reload pseudos will be created to
2500 prevent LRA cycling in some complicated cases. */
2501 lra_bad_spill_regno_start = lra_constraint_new_regno_start;
2502 lra_assignment_iter_after_spill = 0;
2504 restore_scratches ();
2505 lra_eliminate (true, false);
2506 lra_final_code_change ();
2507 lra_in_progress = 0;
2508 if (live_p)
2509 lra_clear_live_ranges ();
2510 lra_live_ranges_finish ();
2511 lra_constraints_finish ();
2512 finish_reg_info ();
2513 sbitmap_free (lra_constraint_insn_stack_bitmap);
2514 lra_constraint_insn_stack.release ();
2515 finish_insn_recog_data ();
2516 regstat_free_n_sets_and_refs ();
2517 regstat_free_ri ();
2518 reload_completed = 1;
2519 update_inc_notes ();
2521 inserted_p = fixup_abnormal_edges ();
2523 /* We've possibly turned single trapping insn into multiple ones. */
2524 if (cfun->can_throw_non_call_exceptions)
2526 auto_sbitmap blocks (last_basic_block_for_fn (cfun));
2527 bitmap_ones (blocks);
2528 find_many_sub_basic_blocks (blocks);
2531 if (inserted_p)
2532 commit_edge_insertions ();
2534 /* Replacing pseudos with their memory equivalents might have
2535 created shared rtx. Subsequent passes would get confused
2536 by this, so unshare everything here. */
2537 unshare_all_rtl_again (get_insns ());
2539 if (flag_checking)
2540 check_rtl (true);
2542 timevar_pop (TV_LRA);
2545 /* Called once per compiler to initialize LRA data once. */
2546 void
2547 lra_init_once (void)
2549 init_insn_code_data_once ();
2552 /* Called once per compiler to finish LRA data which are initialize
2553 once. */
2554 void
2555 lra_finish_once (void)
2557 finish_insn_code_data_once ();