PR tree-optimization/84740
[official-gcc.git] / gcc / lra.c
bloba64d8f1a3011eca2686fb210bcde77e658f3daee
1 /* LRA (local register allocator) driver and LRA utilities.
2 Copyright (C) 2010-2018 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 bind insns. If structure lra_static_insn_data is changed, the
606 initializer should be changed too. */
607 static struct lra_static_insn_data debug_bind_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 /* The following data are used as static insn data for all debug
622 marker insns. If structure lra_static_insn_data is changed, the
623 initializer should be changed too. */
624 static struct lra_static_insn_data debug_marker_static_data =
626 &debug_operand_data,
627 0, /* Duplication operands #. */
628 -1, /* Commutative operand #. */
629 0, /* Operands #. There isn't any operand. */
630 0, /* Duplications #. */
631 0, /* Alternatives #. We are not interesting in alternatives
632 because we does not proceed debug_insns for reloads. */
633 NULL, /* Hard registers referenced in machine description. */
634 NULL /* Descriptions of operands in alternatives. */
637 /* Called once per compiler work to initialize some LRA data related
638 to insns. */
639 static void
640 init_insn_code_data_once (void)
642 memset (insn_code_data, 0, sizeof (insn_code_data));
645 /* Called once per compiler work to finalize some LRA data related to
646 insns. */
647 static void
648 finish_insn_code_data_once (void)
650 for (unsigned int i = 0; i < NUM_INSN_CODES; i++)
652 if (insn_code_data[i] != NULL)
653 free (insn_code_data[i]);
657 /* Return static insn data, allocate and setup if necessary. Although
658 dup_num is static data (it depends only on icode), to set it up we
659 need to extract insn first. So recog_data should be valid for
660 normal insn (ICODE >= 0) before the call. */
661 static struct lra_static_insn_data *
662 get_static_insn_data (int icode, int nop, int ndup, int nalt)
664 struct lra_static_insn_data *data;
665 size_t n_bytes;
667 lra_assert (icode < (int) NUM_INSN_CODES);
668 if (icode >= 0 && (data = insn_code_data[icode]) != NULL)
669 return data;
670 lra_assert (nop >= 0 && ndup >= 0 && nalt >= 0);
671 n_bytes = sizeof (struct lra_static_insn_data)
672 + sizeof (struct lra_operand_data) * nop
673 + sizeof (int) * ndup;
674 data = XNEWVAR (struct lra_static_insn_data, n_bytes);
675 data->operand_alternative = NULL;
676 data->n_operands = nop;
677 data->n_dups = ndup;
678 data->n_alternatives = nalt;
679 data->operand = ((struct lra_operand_data *)
680 ((char *) data + sizeof (struct lra_static_insn_data)));
681 data->dup_num = ((int *) ((char *) data->operand
682 + sizeof (struct lra_operand_data) * nop));
683 if (icode >= 0)
685 int i;
687 insn_code_data[icode] = data;
688 for (i = 0; i < nop; i++)
690 data->operand[i].constraint
691 = insn_data[icode].operand[i].constraint;
692 data->operand[i].mode = insn_data[icode].operand[i].mode;
693 data->operand[i].strict_low = insn_data[icode].operand[i].strict_low;
694 data->operand[i].is_operator
695 = insn_data[icode].operand[i].is_operator;
696 data->operand[i].type
697 = (data->operand[i].constraint[0] == '=' ? OP_OUT
698 : data->operand[i].constraint[0] == '+' ? OP_INOUT
699 : OP_IN);
700 data->operand[i].is_address = false;
702 for (i = 0; i < ndup; i++)
703 data->dup_num[i] = recog_data.dup_num[i];
705 return data;
708 /* The current length of the following array. */
709 int lra_insn_recog_data_len;
711 /* Map INSN_UID -> the insn recog data (NULL if unknown). */
712 lra_insn_recog_data_t *lra_insn_recog_data;
714 /* Initialize LRA data about insns. */
715 static void
716 init_insn_recog_data (void)
718 lra_insn_recog_data_len = 0;
719 lra_insn_recog_data = NULL;
722 /* Expand, if necessary, LRA data about insns. */
723 static void
724 check_and_expand_insn_recog_data (int index)
726 int i, old;
728 if (lra_insn_recog_data_len > index)
729 return;
730 old = lra_insn_recog_data_len;
731 lra_insn_recog_data_len = index * 3 / 2 + 1;
732 lra_insn_recog_data = XRESIZEVEC (lra_insn_recog_data_t,
733 lra_insn_recog_data,
734 lra_insn_recog_data_len);
735 for (i = old; i < lra_insn_recog_data_len; i++)
736 lra_insn_recog_data[i] = NULL;
739 /* Finish LRA DATA about insn. */
740 static void
741 free_insn_recog_data (lra_insn_recog_data_t data)
743 if (data->operand_loc != NULL)
744 free (data->operand_loc);
745 if (data->dup_loc != NULL)
746 free (data->dup_loc);
747 if (data->arg_hard_regs != NULL)
748 free (data->arg_hard_regs);
749 if (data->icode < 0 && NONDEBUG_INSN_P (data->insn))
751 if (data->insn_static_data->operand_alternative != NULL)
752 free (const_cast <operand_alternative *>
753 (data->insn_static_data->operand_alternative));
754 free_insn_regs (data->insn_static_data->hard_regs);
755 free (data->insn_static_data);
757 free_insn_regs (data->regs);
758 data->regs = NULL;
759 free (data);
762 /* Pools for copies. */
763 static object_allocator<lra_copy> lra_copy_pool ("lra copies");
765 /* Finish LRA data about all insns. */
766 static void
767 finish_insn_recog_data (void)
769 int i;
770 lra_insn_recog_data_t data;
772 for (i = 0; i < lra_insn_recog_data_len; i++)
773 if ((data = lra_insn_recog_data[i]) != NULL)
774 free_insn_recog_data (data);
775 finish_insn_regs ();
776 lra_copy_pool.release ();
777 lra_insn_reg_pool.release ();
778 free (lra_insn_recog_data);
781 /* Setup info about operands in alternatives of LRA DATA of insn. */
782 static void
783 setup_operand_alternative (lra_insn_recog_data_t data,
784 const operand_alternative *op_alt)
786 int i, j, nop, nalt;
787 int icode = data->icode;
788 struct lra_static_insn_data *static_data = data->insn_static_data;
790 static_data->commutative = -1;
791 nop = static_data->n_operands;
792 nalt = static_data->n_alternatives;
793 static_data->operand_alternative = op_alt;
794 for (i = 0; i < nop; i++)
796 static_data->operand[i].early_clobber_alts = 0;
797 static_data->operand[i].early_clobber = false;
798 static_data->operand[i].is_address = false;
799 if (static_data->operand[i].constraint[0] == '%')
801 /* We currently only support one commutative pair of operands. */
802 if (static_data->commutative < 0)
803 static_data->commutative = i;
804 else
805 lra_assert (icode < 0); /* Asm */
806 /* The last operand should not be marked commutative. */
807 lra_assert (i != nop - 1);
810 for (j = 0; j < nalt; j++)
811 for (i = 0; i < nop; i++, op_alt++)
813 static_data->operand[i].early_clobber |= op_alt->earlyclobber;
814 if (op_alt->earlyclobber)
815 static_data->operand[i].early_clobber_alts |= (alternative_mask) 1 << j;
816 static_data->operand[i].is_address |= op_alt->is_address;
820 /* Recursively process X and collect info about registers, which are
821 not the insn operands, in X with TYPE (in/out/inout) and flag that
822 it is early clobbered in the insn (EARLY_CLOBBER) and add the info
823 to LIST. X is a part of insn given by DATA. Return the result
824 list. */
825 static struct lra_insn_reg *
826 collect_non_operand_hard_regs (rtx_insn *insn, rtx *x,
827 lra_insn_recog_data_t data,
828 struct lra_insn_reg *list,
829 enum op_type type, bool early_clobber)
831 int i, j, regno, last;
832 bool subreg_p;
833 machine_mode mode;
834 struct lra_insn_reg *curr;
835 rtx op = *x;
836 enum rtx_code code = GET_CODE (op);
837 const char *fmt = GET_RTX_FORMAT (code);
839 for (i = 0; i < data->insn_static_data->n_operands; i++)
840 if (! data->insn_static_data->operand[i].is_operator
841 && x == data->operand_loc[i])
842 /* It is an operand loc. Stop here. */
843 return list;
844 for (i = 0; i < data->insn_static_data->n_dups; i++)
845 if (x == data->dup_loc[i])
846 /* It is a dup loc. Stop here. */
847 return list;
848 mode = GET_MODE (op);
849 subreg_p = false;
850 if (code == SUBREG)
852 mode = wider_subreg_mode (op);
853 if (read_modify_subreg_p (op))
854 subreg_p = true;
855 op = SUBREG_REG (op);
856 code = GET_CODE (op);
858 if (REG_P (op))
860 if ((regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER)
861 return list;
862 /* Process all regs even unallocatable ones as we need info
863 about all regs for rematerialization pass. */
864 for (last = end_hard_regno (mode, regno); regno < last; regno++)
866 for (curr = list; curr != NULL; curr = curr->next)
867 if (curr->regno == regno && curr->subreg_p == subreg_p
868 && curr->biggest_mode == mode)
870 if (curr->type != type)
871 curr->type = OP_INOUT;
872 if (early_clobber)
874 curr->early_clobber = true;
875 curr->early_clobber_alts = ALL_ALTERNATIVES;
877 break;
879 if (curr == NULL)
881 /* This is a new hard regno or the info can not be
882 integrated into the found structure. */
883 #ifdef STACK_REGS
884 early_clobber
885 = (early_clobber
886 /* This clobber is to inform popping floating
887 point stack only. */
888 && ! (FIRST_STACK_REG <= regno
889 && regno <= LAST_STACK_REG));
890 #endif
891 list = new_insn_reg (data->insn, regno, type, mode, subreg_p,
892 early_clobber,
893 early_clobber ? ALL_ALTERNATIVES : 0, list);
896 return list;
898 switch (code)
900 case SET:
901 list = collect_non_operand_hard_regs (insn, &SET_DEST (op), data,
902 list, OP_OUT, false);
903 list = collect_non_operand_hard_regs (insn, &SET_SRC (op), data,
904 list, OP_IN, false);
905 break;
906 case CLOBBER:
907 /* We treat clobber of non-operand hard registers as early clobber. */
908 list = collect_non_operand_hard_regs (insn, &XEXP (op, 0), data,
909 list, OP_OUT, true);
910 break;
911 case PRE_INC: case PRE_DEC: case POST_INC: case POST_DEC:
912 list = collect_non_operand_hard_regs (insn, &XEXP (op, 0), data,
913 list, OP_INOUT, false);
914 break;
915 case PRE_MODIFY: case POST_MODIFY:
916 list = collect_non_operand_hard_regs (insn, &XEXP (op, 0), data,
917 list, OP_INOUT, false);
918 list = collect_non_operand_hard_regs (insn, &XEXP (op, 1), data,
919 list, OP_IN, false);
920 break;
921 default:
922 fmt = GET_RTX_FORMAT (code);
923 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
925 if (fmt[i] == 'e')
926 list = collect_non_operand_hard_regs (insn, &XEXP (op, i), data,
927 list, OP_IN, false);
928 else if (fmt[i] == 'E')
929 for (j = XVECLEN (op, i) - 1; j >= 0; j--)
930 list = collect_non_operand_hard_regs (insn, &XVECEXP (op, i, j),
931 data, list, OP_IN, false);
934 return list;
937 /* Set up and return info about INSN. Set up the info if it is not set up
938 yet. */
939 lra_insn_recog_data_t
940 lra_set_insn_recog_data (rtx_insn *insn)
942 lra_insn_recog_data_t data;
943 int i, n, icode;
944 rtx **locs;
945 unsigned int uid = INSN_UID (insn);
946 struct lra_static_insn_data *insn_static_data;
948 check_and_expand_insn_recog_data (uid);
949 if (DEBUG_INSN_P (insn))
950 icode = -1;
951 else
953 icode = INSN_CODE (insn);
954 if (icode < 0)
955 /* It might be a new simple insn which is not recognized yet. */
956 INSN_CODE (insn) = icode = recog_memoized (insn);
958 data = XNEW (struct lra_insn_recog_data);
959 lra_insn_recog_data[uid] = data;
960 data->insn = insn;
961 data->used_insn_alternative = LRA_UNKNOWN_ALT;
962 data->icode = icode;
963 data->regs = NULL;
964 if (DEBUG_INSN_P (insn))
966 data->dup_loc = NULL;
967 data->arg_hard_regs = NULL;
968 data->preferred_alternatives = ALL_ALTERNATIVES;
969 if (DEBUG_BIND_INSN_P (insn))
971 data->insn_static_data = &debug_bind_static_data;
972 data->operand_loc = XNEWVEC (rtx *, 1);
973 data->operand_loc[0] = &INSN_VAR_LOCATION_LOC (insn);
975 else if (DEBUG_MARKER_INSN_P (insn))
977 data->insn_static_data = &debug_marker_static_data;
978 data->operand_loc = NULL;
980 return data;
982 if (icode < 0)
984 int nop, nalt;
985 machine_mode operand_mode[MAX_RECOG_OPERANDS];
986 const char *constraints[MAX_RECOG_OPERANDS];
988 nop = asm_noperands (PATTERN (insn));
989 data->operand_loc = data->dup_loc = NULL;
990 nalt = 1;
991 if (nop < 0)
993 /* It is a special insn like USE or CLOBBER. We should
994 recognize any regular insn otherwise LRA can do nothing
995 with this insn. */
996 gcc_assert (GET_CODE (PATTERN (insn)) == USE
997 || GET_CODE (PATTERN (insn)) == CLOBBER
998 || GET_CODE (PATTERN (insn)) == ASM_INPUT);
999 data->insn_static_data = insn_static_data
1000 = get_static_insn_data (-1, 0, 0, nalt);
1002 else
1004 /* expand_asm_operands makes sure there aren't too many
1005 operands. */
1006 lra_assert (nop <= MAX_RECOG_OPERANDS);
1007 if (nop != 0)
1008 data->operand_loc = XNEWVEC (rtx *, nop);
1009 /* Now get the operand values and constraints out of the
1010 insn. */
1011 decode_asm_operands (PATTERN (insn), NULL,
1012 data->operand_loc,
1013 constraints, operand_mode, NULL);
1014 if (nop > 0)
1016 const char *p = recog_data.constraints[0];
1018 for (p = constraints[0]; *p; p++)
1019 nalt += *p == ',';
1021 data->insn_static_data = insn_static_data
1022 = get_static_insn_data (-1, nop, 0, nalt);
1023 for (i = 0; i < nop; i++)
1025 insn_static_data->operand[i].mode = operand_mode[i];
1026 insn_static_data->operand[i].constraint = constraints[i];
1027 insn_static_data->operand[i].strict_low = false;
1028 insn_static_data->operand[i].is_operator = false;
1029 insn_static_data->operand[i].is_address = false;
1032 for (i = 0; i < insn_static_data->n_operands; i++)
1033 insn_static_data->operand[i].type
1034 = (insn_static_data->operand[i].constraint[0] == '=' ? OP_OUT
1035 : insn_static_data->operand[i].constraint[0] == '+' ? OP_INOUT
1036 : OP_IN);
1037 data->preferred_alternatives = ALL_ALTERNATIVES;
1038 if (nop > 0)
1040 operand_alternative *op_alt = XCNEWVEC (operand_alternative,
1041 nalt * nop);
1042 preprocess_constraints (nop, nalt, constraints, op_alt);
1043 setup_operand_alternative (data, op_alt);
1046 else
1048 insn_extract (insn);
1049 data->insn_static_data = insn_static_data
1050 = get_static_insn_data (icode, insn_data[icode].n_operands,
1051 insn_data[icode].n_dups,
1052 insn_data[icode].n_alternatives);
1053 n = insn_static_data->n_operands;
1054 if (n == 0)
1055 locs = NULL;
1056 else
1058 locs = XNEWVEC (rtx *, n);
1059 memcpy (locs, recog_data.operand_loc, n * sizeof (rtx *));
1061 data->operand_loc = locs;
1062 n = insn_static_data->n_dups;
1063 if (n == 0)
1064 locs = NULL;
1065 else
1067 locs = XNEWVEC (rtx *, n);
1068 memcpy (locs, recog_data.dup_loc, n * sizeof (rtx *));
1070 data->dup_loc = locs;
1071 data->preferred_alternatives = get_preferred_alternatives (insn);
1072 const operand_alternative *op_alt = preprocess_insn_constraints (icode);
1073 if (!insn_static_data->operand_alternative)
1074 setup_operand_alternative (data, op_alt);
1075 else if (op_alt != insn_static_data->operand_alternative)
1076 insn_static_data->operand_alternative = op_alt;
1078 if (GET_CODE (PATTERN (insn)) == CLOBBER || GET_CODE (PATTERN (insn)) == USE)
1079 insn_static_data->hard_regs = NULL;
1080 else
1081 insn_static_data->hard_regs
1082 = collect_non_operand_hard_regs (insn, &PATTERN (insn), data,
1083 NULL, OP_IN, false);
1084 data->arg_hard_regs = NULL;
1085 if (CALL_P (insn))
1087 bool use_p;
1088 rtx link;
1089 int n_hard_regs, regno, arg_hard_regs[FIRST_PSEUDO_REGISTER];
1091 n_hard_regs = 0;
1092 /* Finding implicit hard register usage. We believe it will be
1093 not changed whatever transformations are used. Call insns
1094 are such example. */
1095 for (link = CALL_INSN_FUNCTION_USAGE (insn);
1096 link != NULL_RTX;
1097 link = XEXP (link, 1))
1098 if (((use_p = GET_CODE (XEXP (link, 0)) == USE)
1099 || GET_CODE (XEXP (link, 0)) == CLOBBER)
1100 && REG_P (XEXP (XEXP (link, 0), 0)))
1102 regno = REGNO (XEXP (XEXP (link, 0), 0));
1103 lra_assert (regno < FIRST_PSEUDO_REGISTER);
1104 /* It is an argument register. */
1105 for (i = REG_NREGS (XEXP (XEXP (link, 0), 0)) - 1; i >= 0; i--)
1106 arg_hard_regs[n_hard_regs++]
1107 = regno + i + (use_p ? 0 : FIRST_PSEUDO_REGISTER);
1109 if (n_hard_regs != 0)
1111 arg_hard_regs[n_hard_regs++] = -1;
1112 data->arg_hard_regs = XNEWVEC (int, n_hard_regs);
1113 memcpy (data->arg_hard_regs, arg_hard_regs,
1114 sizeof (int) * n_hard_regs);
1117 /* Some output operand can be recognized only from the context not
1118 from the constraints which are empty in this case. Call insn may
1119 contain a hard register in set destination with empty constraint
1120 and extract_insn treats them as an input. */
1121 for (i = 0; i < insn_static_data->n_operands; i++)
1123 int j;
1124 rtx pat, set;
1125 struct lra_operand_data *operand = &insn_static_data->operand[i];
1127 /* ??? Should we treat 'X' the same way. It looks to me that
1128 'X' means anything and empty constraint means we do not
1129 care. */
1130 if (operand->type != OP_IN || *operand->constraint != '\0'
1131 || operand->is_operator)
1132 continue;
1133 pat = PATTERN (insn);
1134 if (GET_CODE (pat) == SET)
1136 if (data->operand_loc[i] != &SET_DEST (pat))
1137 continue;
1139 else if (GET_CODE (pat) == PARALLEL)
1141 for (j = XVECLEN (pat, 0) - 1; j >= 0; j--)
1143 set = XVECEXP (PATTERN (insn), 0, j);
1144 if (GET_CODE (set) == SET
1145 && &SET_DEST (set) == data->operand_loc[i])
1146 break;
1148 if (j < 0)
1149 continue;
1151 else
1152 continue;
1153 operand->type = OP_OUT;
1155 return data;
1158 /* Return info about insn give by UID. The info should be already set
1159 up. */
1160 static lra_insn_recog_data_t
1161 get_insn_recog_data_by_uid (int uid)
1163 lra_insn_recog_data_t data;
1165 data = lra_insn_recog_data[uid];
1166 lra_assert (data != NULL);
1167 return data;
1170 /* Invalidate all info about insn given by its UID. */
1171 static void
1172 invalidate_insn_recog_data (int uid)
1174 lra_insn_recog_data_t data;
1176 data = lra_insn_recog_data[uid];
1177 lra_assert (data != NULL);
1178 free_insn_recog_data (data);
1179 lra_insn_recog_data[uid] = NULL;
1182 /* Update all the insn info about INSN. It is usually called when
1183 something in the insn was changed. Return the updated info. */
1184 lra_insn_recog_data_t
1185 lra_update_insn_recog_data (rtx_insn *insn)
1187 lra_insn_recog_data_t data;
1188 int n;
1189 unsigned int uid = INSN_UID (insn);
1190 struct lra_static_insn_data *insn_static_data;
1191 poly_int64 sp_offset = 0;
1193 check_and_expand_insn_recog_data (uid);
1194 if ((data = lra_insn_recog_data[uid]) != NULL
1195 && data->icode != INSN_CODE (insn))
1197 sp_offset = data->sp_offset;
1198 invalidate_insn_data_regno_info (data, insn, get_insn_freq (insn));
1199 invalidate_insn_recog_data (uid);
1200 data = NULL;
1202 if (data == NULL)
1204 data = lra_get_insn_recog_data (insn);
1205 /* Initiate or restore SP offset. */
1206 data->sp_offset = sp_offset;
1207 return data;
1209 insn_static_data = data->insn_static_data;
1210 data->used_insn_alternative = LRA_UNKNOWN_ALT;
1211 if (DEBUG_INSN_P (insn))
1212 return data;
1213 if (data->icode < 0)
1215 int nop;
1216 machine_mode operand_mode[MAX_RECOG_OPERANDS];
1217 const char *constraints[MAX_RECOG_OPERANDS];
1219 nop = asm_noperands (PATTERN (insn));
1220 if (nop >= 0)
1222 lra_assert (nop == data->insn_static_data->n_operands);
1223 /* Now get the operand values and constraints out of the
1224 insn. */
1225 decode_asm_operands (PATTERN (insn), NULL,
1226 data->operand_loc,
1227 constraints, operand_mode, NULL);
1229 if (flag_checking)
1230 for (int i = 0; i < nop; i++)
1231 lra_assert
1232 (insn_static_data->operand[i].mode == operand_mode[i]
1233 && insn_static_data->operand[i].constraint == constraints[i]
1234 && ! insn_static_data->operand[i].is_operator);
1237 if (flag_checking)
1238 for (int i = 0; i < insn_static_data->n_operands; i++)
1239 lra_assert
1240 (insn_static_data->operand[i].type
1241 == (insn_static_data->operand[i].constraint[0] == '=' ? OP_OUT
1242 : insn_static_data->operand[i].constraint[0] == '+' ? OP_INOUT
1243 : OP_IN));
1245 else
1247 insn_extract (insn);
1248 n = insn_static_data->n_operands;
1249 if (n != 0)
1250 memcpy (data->operand_loc, recog_data.operand_loc, n * sizeof (rtx *));
1251 n = insn_static_data->n_dups;
1252 if (n != 0)
1253 memcpy (data->dup_loc, recog_data.dup_loc, n * sizeof (rtx *));
1254 lra_assert (check_bool_attrs (insn));
1256 return data;
1259 /* Set up that INSN is using alternative ALT now. */
1260 void
1261 lra_set_used_insn_alternative (rtx_insn *insn, int alt)
1263 lra_insn_recog_data_t data;
1265 data = lra_get_insn_recog_data (insn);
1266 data->used_insn_alternative = alt;
1269 /* Set up that insn with UID is using alternative ALT now. The insn
1270 info should be already set up. */
1271 void
1272 lra_set_used_insn_alternative_by_uid (int uid, int alt)
1274 lra_insn_recog_data_t data;
1276 check_and_expand_insn_recog_data (uid);
1277 data = lra_insn_recog_data[uid];
1278 lra_assert (data != NULL);
1279 data->used_insn_alternative = alt;
1284 /* This page contains code dealing with common register info and
1285 pseudo copies. */
1287 /* The size of the following array. */
1288 static int reg_info_size;
1289 /* Common info about each register. */
1290 struct lra_reg *lra_reg_info;
1292 HARD_REG_SET hard_regs_spilled_into;
1294 /* Last register value. */
1295 static int last_reg_value;
1297 /* Return new register value. */
1298 static int
1299 get_new_reg_value (void)
1301 return ++last_reg_value;
1304 /* Vec referring to pseudo copies. */
1305 static vec<lra_copy_t> copy_vec;
1307 /* Initialize I-th element of lra_reg_info. */
1308 static inline void
1309 initialize_lra_reg_info_element (int i)
1311 bitmap_initialize (&lra_reg_info[i].insn_bitmap, &reg_obstack);
1312 #ifdef STACK_REGS
1313 lra_reg_info[i].no_stack_p = false;
1314 #endif
1315 CLEAR_HARD_REG_SET (lra_reg_info[i].conflict_hard_regs);
1316 CLEAR_HARD_REG_SET (lra_reg_info[i].actual_call_used_reg_set);
1317 lra_reg_info[i].preferred_hard_regno1 = -1;
1318 lra_reg_info[i].preferred_hard_regno2 = -1;
1319 lra_reg_info[i].preferred_hard_regno_profit1 = 0;
1320 lra_reg_info[i].preferred_hard_regno_profit2 = 0;
1321 lra_reg_info[i].biggest_mode = VOIDmode;
1322 lra_reg_info[i].live_ranges = NULL;
1323 lra_reg_info[i].nrefs = lra_reg_info[i].freq = 0;
1324 lra_reg_info[i].last_reload = 0;
1325 lra_reg_info[i].restore_rtx = NULL_RTX;
1326 lra_reg_info[i].val = get_new_reg_value ();
1327 lra_reg_info[i].offset = 0;
1328 lra_reg_info[i].copies = NULL;
1331 /* Initialize common reg info and copies. */
1332 static void
1333 init_reg_info (void)
1335 int i;
1337 last_reg_value = 0;
1338 reg_info_size = max_reg_num () * 3 / 2 + 1;
1339 lra_reg_info = XNEWVEC (struct lra_reg, reg_info_size);
1340 for (i = 0; i < reg_info_size; i++)
1341 initialize_lra_reg_info_element (i);
1342 copy_vec.truncate (0);
1343 CLEAR_HARD_REG_SET (hard_regs_spilled_into);
1347 /* Finish common reg info and copies. */
1348 static void
1349 finish_reg_info (void)
1351 int i;
1353 for (i = 0; i < reg_info_size; i++)
1354 bitmap_clear (&lra_reg_info[i].insn_bitmap);
1355 free (lra_reg_info);
1356 reg_info_size = 0;
1359 /* Expand common reg info if it is necessary. */
1360 static void
1361 expand_reg_info (void)
1363 int i, old = reg_info_size;
1365 if (reg_info_size > max_reg_num ())
1366 return;
1367 reg_info_size = max_reg_num () * 3 / 2 + 1;
1368 lra_reg_info = XRESIZEVEC (struct lra_reg, lra_reg_info, reg_info_size);
1369 for (i = old; i < reg_info_size; i++)
1370 initialize_lra_reg_info_element (i);
1373 /* Free all copies. */
1374 void
1375 lra_free_copies (void)
1377 lra_copy_t cp;
1379 while (copy_vec.length () != 0)
1381 cp = copy_vec.pop ();
1382 lra_reg_info[cp->regno1].copies = lra_reg_info[cp->regno2].copies = NULL;
1383 lra_copy_pool.remove (cp);
1387 /* Create copy of two pseudos REGNO1 and REGNO2. The copy execution
1388 frequency is FREQ. */
1389 void
1390 lra_create_copy (int regno1, int regno2, int freq)
1392 bool regno1_dest_p;
1393 lra_copy_t cp;
1395 lra_assert (regno1 != regno2);
1396 regno1_dest_p = true;
1397 if (regno1 > regno2)
1399 std::swap (regno1, regno2);
1400 regno1_dest_p = false;
1402 cp = lra_copy_pool.allocate ();
1403 copy_vec.safe_push (cp);
1404 cp->regno1_dest_p = regno1_dest_p;
1405 cp->freq = freq;
1406 cp->regno1 = regno1;
1407 cp->regno2 = regno2;
1408 cp->regno1_next = lra_reg_info[regno1].copies;
1409 lra_reg_info[regno1].copies = cp;
1410 cp->regno2_next = lra_reg_info[regno2].copies;
1411 lra_reg_info[regno2].copies = cp;
1412 if (lra_dump_file != NULL)
1413 fprintf (lra_dump_file, " Creating copy r%d%sr%d@%d\n",
1414 regno1, regno1_dest_p ? "<-" : "->", regno2, freq);
1417 /* Return N-th (0, 1, ...) copy. If there is no copy, return
1418 NULL. */
1419 lra_copy_t
1420 lra_get_copy (int n)
1422 if (n >= (int) copy_vec.length ())
1423 return NULL;
1424 return copy_vec[n];
1429 /* This page contains code dealing with info about registers in
1430 insns. */
1432 /* Process X of INSN recursively and add info (operand type is
1433 given by TYPE, flag of that it is early clobber is EARLY_CLOBBER)
1434 about registers in X to the insn DATA. If X can be early clobbered,
1435 alternatives in which it can be early clobbered are given by
1436 EARLY_CLOBBER_ALTS. */
1437 static void
1438 add_regs_to_insn_regno_info (lra_insn_recog_data_t data, rtx x,
1439 rtx_insn *insn,
1440 enum op_type type, bool early_clobber,
1441 alternative_mask early_clobber_alts)
1443 int i, j, regno;
1444 bool subreg_p;
1445 machine_mode mode;
1446 const char *fmt;
1447 enum rtx_code code;
1448 struct lra_insn_reg *curr;
1450 code = GET_CODE (x);
1451 mode = GET_MODE (x);
1452 subreg_p = false;
1453 if (GET_CODE (x) == SUBREG)
1455 mode = wider_subreg_mode (x);
1456 if (read_modify_subreg_p (x))
1457 subreg_p = true;
1458 x = SUBREG_REG (x);
1459 code = GET_CODE (x);
1461 if (REG_P (x))
1463 regno = REGNO (x);
1464 /* Process all regs even unallocatable ones as we need info about
1465 all regs for rematerialization pass. */
1466 expand_reg_info ();
1467 if (bitmap_set_bit (&lra_reg_info[regno].insn_bitmap, INSN_UID (insn)))
1469 data->regs = new_insn_reg (data->insn, regno, type, mode, subreg_p,
1470 early_clobber, early_clobber_alts,
1471 data->regs);
1472 return;
1474 else
1476 for (curr = data->regs; curr != NULL; curr = curr->next)
1477 if (curr->regno == regno)
1479 if (curr->subreg_p != subreg_p || curr->biggest_mode != mode)
1480 /* The info can not be integrated into the found
1481 structure. */
1482 data->regs = new_insn_reg (data->insn, regno, type, mode,
1483 subreg_p, early_clobber,
1484 early_clobber_alts, data->regs);
1485 else
1487 if (curr->type != type)
1488 curr->type = OP_INOUT;
1489 if (curr->early_clobber != early_clobber)
1490 curr->early_clobber = true;
1491 curr->early_clobber_alts |= early_clobber_alts;
1493 return;
1495 gcc_unreachable ();
1499 switch (code)
1501 case SET:
1502 add_regs_to_insn_regno_info (data, SET_DEST (x), insn, OP_OUT, false, 0);
1503 add_regs_to_insn_regno_info (data, SET_SRC (x), insn, OP_IN, false, 0);
1504 break;
1505 case CLOBBER:
1506 /* We treat clobber of non-operand hard registers as early
1507 clobber. */
1508 add_regs_to_insn_regno_info (data, XEXP (x, 0), insn, OP_OUT,
1509 true, ALL_ALTERNATIVES);
1510 break;
1511 case PRE_INC: case PRE_DEC: case POST_INC: case POST_DEC:
1512 add_regs_to_insn_regno_info (data, XEXP (x, 0), insn, OP_INOUT, false, 0);
1513 break;
1514 case PRE_MODIFY: case POST_MODIFY:
1515 add_regs_to_insn_regno_info (data, XEXP (x, 0), insn, OP_INOUT, false, 0);
1516 add_regs_to_insn_regno_info (data, XEXP (x, 1), insn, OP_IN, false, 0);
1517 break;
1518 default:
1519 if ((code != PARALLEL && code != EXPR_LIST) || type != OP_OUT)
1520 /* Some targets place small structures in registers for return
1521 values of functions, and those registers are wrapped in
1522 PARALLEL that we may see as the destination of a SET. Here
1523 is an example:
1525 (call_insn 13 12 14 2 (set (parallel:BLK [
1526 (expr_list:REG_DEP_TRUE (reg:DI 0 ax)
1527 (const_int 0 [0]))
1528 (expr_list:REG_DEP_TRUE (reg:DI 1 dx)
1529 (const_int 8 [0x8]))
1531 (call (mem:QI (symbol_ref:DI (... */
1532 type = OP_IN;
1533 fmt = GET_RTX_FORMAT (code);
1534 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1536 if (fmt[i] == 'e')
1537 add_regs_to_insn_regno_info (data, XEXP (x, i), insn, type, false, 0);
1538 else if (fmt[i] == 'E')
1540 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1541 add_regs_to_insn_regno_info (data, XVECEXP (x, i, j), insn,
1542 type, false, 0);
1548 /* Return execution frequency of INSN. */
1549 static int
1550 get_insn_freq (rtx_insn *insn)
1552 basic_block bb = BLOCK_FOR_INSN (insn);
1554 gcc_checking_assert (bb != NULL);
1555 return REG_FREQ_FROM_BB (bb);
1558 /* Invalidate all reg info of INSN with DATA and execution frequency
1559 FREQ. Update common info about the invalidated registers. */
1560 static void
1561 invalidate_insn_data_regno_info (lra_insn_recog_data_t data, rtx_insn *insn,
1562 int freq)
1564 int uid;
1565 bool debug_p;
1566 unsigned int i;
1567 struct lra_insn_reg *ir, *next_ir;
1569 uid = INSN_UID (insn);
1570 debug_p = DEBUG_INSN_P (insn);
1571 for (ir = data->regs; ir != NULL; ir = next_ir)
1573 i = ir->regno;
1574 next_ir = ir->next;
1575 lra_insn_reg_pool.remove (ir);
1576 bitmap_clear_bit (&lra_reg_info[i].insn_bitmap, uid);
1577 if (i >= FIRST_PSEUDO_REGISTER && ! debug_p)
1579 lra_reg_info[i].nrefs--;
1580 lra_reg_info[i].freq -= freq;
1581 lra_assert (lra_reg_info[i].nrefs >= 0 && lra_reg_info[i].freq >= 0);
1584 data->regs = NULL;
1587 /* Invalidate all reg info of INSN. Update common info about the
1588 invalidated registers. */
1589 void
1590 lra_invalidate_insn_regno_info (rtx_insn *insn)
1592 invalidate_insn_data_regno_info (lra_get_insn_recog_data (insn), insn,
1593 get_insn_freq (insn));
1596 /* Update common reg info from reg info of insn given by its DATA and
1597 execution frequency FREQ. */
1598 static void
1599 setup_insn_reg_info (lra_insn_recog_data_t data, int freq)
1601 unsigned int i;
1602 struct lra_insn_reg *ir;
1604 for (ir = data->regs; ir != NULL; ir = ir->next)
1605 if ((i = ir->regno) >= FIRST_PSEUDO_REGISTER)
1607 lra_reg_info[i].nrefs++;
1608 lra_reg_info[i].freq += freq;
1612 /* Set up insn reg info of INSN. Update common reg info from reg info
1613 of INSN. */
1614 void
1615 lra_update_insn_regno_info (rtx_insn *insn)
1617 int i, freq;
1618 lra_insn_recog_data_t data;
1619 struct lra_static_insn_data *static_data;
1620 enum rtx_code code;
1621 rtx link;
1623 if (! INSN_P (insn))
1624 return;
1625 data = lra_get_insn_recog_data (insn);
1626 static_data = data->insn_static_data;
1627 freq = NONDEBUG_INSN_P (insn) ? get_insn_freq (insn) : 0;
1628 invalidate_insn_data_regno_info (data, insn, freq);
1629 for (i = static_data->n_operands - 1; i >= 0; i--)
1630 add_regs_to_insn_regno_info (data, *data->operand_loc[i], insn,
1631 static_data->operand[i].type,
1632 static_data->operand[i].early_clobber,
1633 static_data->operand[i].early_clobber_alts);
1634 if ((code = GET_CODE (PATTERN (insn))) == CLOBBER || code == USE)
1635 add_regs_to_insn_regno_info (data, XEXP (PATTERN (insn), 0), insn,
1636 code == USE ? OP_IN : OP_OUT, false, 0);
1637 if (CALL_P (insn))
1638 /* On some targets call insns can refer to pseudos in memory in
1639 CALL_INSN_FUNCTION_USAGE list. Process them in order to
1640 consider their occurrences in calls for different
1641 transformations (e.g. inheritance) with given pseudos. */
1642 for (link = CALL_INSN_FUNCTION_USAGE (insn);
1643 link != NULL_RTX;
1644 link = XEXP (link, 1))
1645 if (((code = GET_CODE (XEXP (link, 0))) == USE || code == CLOBBER)
1646 && MEM_P (XEXP (XEXP (link, 0), 0)))
1647 add_regs_to_insn_regno_info (data, XEXP (XEXP (link, 0), 0), insn,
1648 code == USE ? OP_IN : OP_OUT, false, 0);
1649 if (NONDEBUG_INSN_P (insn))
1650 setup_insn_reg_info (data, freq);
1653 /* Return reg info of insn given by it UID. */
1654 struct lra_insn_reg *
1655 lra_get_insn_regs (int uid)
1657 lra_insn_recog_data_t data;
1659 data = get_insn_recog_data_by_uid (uid);
1660 return data->regs;
1665 /* Recursive hash function for RTL X. */
1666 hashval_t
1667 lra_rtx_hash (rtx x)
1669 int i, j;
1670 enum rtx_code code;
1671 const char *fmt;
1672 hashval_t val = 0;
1674 if (x == 0)
1675 return val;
1677 code = GET_CODE (x);
1678 val += (int) code + 4095;
1680 /* Some RTL can be compared nonrecursively. */
1681 switch (code)
1683 case REG:
1684 return val + REGNO (x);
1686 case LABEL_REF:
1687 return iterative_hash_object (XEXP (x, 0), val);
1689 case SYMBOL_REF:
1690 return iterative_hash_object (XSTR (x, 0), val);
1692 case SCRATCH:
1693 case CONST_DOUBLE:
1694 case CONST_INT:
1695 case CONST_VECTOR:
1696 return val;
1698 default:
1699 break;
1702 /* Hash the elements. */
1703 fmt = GET_RTX_FORMAT (code);
1704 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1706 switch (fmt[i])
1708 case 'w':
1709 val += XWINT (x, i);
1710 break;
1712 case 'n':
1713 case 'i':
1714 val += XINT (x, i);
1715 break;
1717 case 'V':
1718 case 'E':
1719 val += XVECLEN (x, i);
1721 for (j = 0; j < XVECLEN (x, i); j++)
1722 val += lra_rtx_hash (XVECEXP (x, i, j));
1723 break;
1725 case 'e':
1726 val += lra_rtx_hash (XEXP (x, i));
1727 break;
1729 case 'S':
1730 case 's':
1731 val += htab_hash_string (XSTR (x, i));
1732 break;
1734 case 'u':
1735 case '0':
1736 case 't':
1737 break;
1739 /* It is believed that rtx's at this level will never
1740 contain anything but integers and other rtx's, except for
1741 within LABEL_REFs and SYMBOL_REFs. */
1742 default:
1743 abort ();
1746 return val;
1751 /* This page contains code dealing with stack of the insns which
1752 should be processed by the next constraint pass. */
1754 /* Bitmap used to put an insn on the stack only in one exemplar. */
1755 static sbitmap lra_constraint_insn_stack_bitmap;
1757 /* The stack itself. */
1758 vec<rtx_insn *> lra_constraint_insn_stack;
1760 /* Put INSN on the stack. If ALWAYS_UPDATE is true, always update the reg
1761 info for INSN, otherwise only update it if INSN is not already on the
1762 stack. */
1763 static inline void
1764 lra_push_insn_1 (rtx_insn *insn, bool always_update)
1766 unsigned int uid = INSN_UID (insn);
1767 if (always_update)
1768 lra_update_insn_regno_info (insn);
1769 if (uid >= SBITMAP_SIZE (lra_constraint_insn_stack_bitmap))
1770 lra_constraint_insn_stack_bitmap =
1771 sbitmap_resize (lra_constraint_insn_stack_bitmap, 3 * uid / 2, 0);
1772 if (bitmap_bit_p (lra_constraint_insn_stack_bitmap, uid))
1773 return;
1774 bitmap_set_bit (lra_constraint_insn_stack_bitmap, uid);
1775 if (! always_update)
1776 lra_update_insn_regno_info (insn);
1777 lra_constraint_insn_stack.safe_push (insn);
1780 /* Put INSN on the stack. */
1781 void
1782 lra_push_insn (rtx_insn *insn)
1784 lra_push_insn_1 (insn, false);
1787 /* Put INSN on the stack and update its reg info. */
1788 void
1789 lra_push_insn_and_update_insn_regno_info (rtx_insn *insn)
1791 lra_push_insn_1 (insn, true);
1794 /* Put insn with UID on the stack. */
1795 void
1796 lra_push_insn_by_uid (unsigned int uid)
1798 lra_push_insn (lra_insn_recog_data[uid]->insn);
1801 /* Take the last-inserted insns off the stack and return it. */
1802 rtx_insn *
1803 lra_pop_insn (void)
1805 rtx_insn *insn = lra_constraint_insn_stack.pop ();
1806 bitmap_clear_bit (lra_constraint_insn_stack_bitmap, INSN_UID (insn));
1807 return insn;
1810 /* Return the current size of the insn stack. */
1811 unsigned int
1812 lra_insn_stack_length (void)
1814 return lra_constraint_insn_stack.length ();
1817 /* Push insns FROM to TO (excluding it) going in reverse order. */
1818 static void
1819 push_insns (rtx_insn *from, rtx_insn *to)
1821 rtx_insn *insn;
1823 if (from == NULL_RTX)
1824 return;
1825 for (insn = from; insn != to; insn = PREV_INSN (insn))
1826 if (INSN_P (insn))
1827 lra_push_insn (insn);
1830 /* Set up sp offset for insn in range [FROM, LAST]. The offset is
1831 taken from the next BB insn after LAST or zero if there in such
1832 insn. */
1833 static void
1834 setup_sp_offset (rtx_insn *from, rtx_insn *last)
1836 rtx_insn *before = next_nonnote_nondebug_insn_bb (last);
1837 poly_int64 offset = (before == NULL_RTX || ! INSN_P (before)
1838 ? 0 : lra_get_insn_recog_data (before)->sp_offset);
1840 for (rtx_insn *insn = from; insn != NEXT_INSN (last); insn = NEXT_INSN (insn))
1841 lra_get_insn_recog_data (insn)->sp_offset = offset;
1844 /* Emit insns BEFORE before INSN and insns AFTER after INSN. Put the
1845 insns onto the stack. Print about emitting the insns with
1846 TITLE. */
1847 void
1848 lra_process_new_insns (rtx_insn *insn, rtx_insn *before, rtx_insn *after,
1849 const char *title)
1851 rtx_insn *last;
1853 if (before == NULL_RTX && after == NULL_RTX)
1854 return;
1855 if (lra_dump_file != NULL)
1857 dump_insn_slim (lra_dump_file, insn);
1858 if (before != NULL_RTX)
1860 fprintf (lra_dump_file," %s before:\n", title);
1861 dump_rtl_slim (lra_dump_file, before, NULL, -1, 0);
1863 if (after != NULL_RTX)
1865 fprintf (lra_dump_file, " %s after:\n", title);
1866 dump_rtl_slim (lra_dump_file, after, NULL, -1, 0);
1868 fprintf (lra_dump_file, "\n");
1870 if (before != NULL_RTX)
1872 if (cfun->can_throw_non_call_exceptions)
1873 copy_reg_eh_region_note_forward (insn, before, NULL);
1874 emit_insn_before (before, insn);
1875 push_insns (PREV_INSN (insn), PREV_INSN (before));
1876 setup_sp_offset (before, PREV_INSN (insn));
1878 if (after != NULL_RTX)
1880 if (cfun->can_throw_non_call_exceptions)
1881 copy_reg_eh_region_note_forward (insn, after, NULL);
1882 for (last = after; NEXT_INSN (last) != NULL_RTX; last = NEXT_INSN (last))
1884 emit_insn_after (after, insn);
1885 push_insns (last, insn);
1886 setup_sp_offset (after, last);
1888 if (cfun->can_throw_non_call_exceptions)
1890 rtx note = find_reg_note (insn, REG_EH_REGION, NULL_RTX);
1891 if (note && !insn_could_throw_p (insn))
1892 remove_note (insn, note);
1897 /* Replace all references to register OLD_REGNO in *LOC with pseudo
1898 register NEW_REG. Try to simplify subreg of constant if SUBREG_P.
1899 DEBUG_P is if LOC is within a DEBUG_INSN. Return true if any
1900 change was made. */
1901 bool
1902 lra_substitute_pseudo (rtx *loc, int old_regno, rtx new_reg, bool subreg_p,
1903 bool debug_p)
1905 rtx x = *loc;
1906 bool result = false;
1907 enum rtx_code code;
1908 const char *fmt;
1909 int i, j;
1911 if (x == NULL_RTX)
1912 return false;
1914 code = GET_CODE (x);
1915 if (code == SUBREG && subreg_p)
1917 rtx subst, inner = SUBREG_REG (x);
1918 /* Transform subreg of constant while we still have inner mode
1919 of the subreg. The subreg internal should not be an insn
1920 operand. */
1921 if (REG_P (inner) && (int) REGNO (inner) == old_regno
1922 && CONSTANT_P (new_reg)
1923 && (subst = simplify_subreg (GET_MODE (x), new_reg, GET_MODE (inner),
1924 SUBREG_BYTE (x))) != NULL_RTX)
1926 *loc = subst;
1927 return true;
1931 else if (code == REG && (int) REGNO (x) == old_regno)
1933 machine_mode mode = GET_MODE (x);
1934 machine_mode inner_mode = GET_MODE (new_reg);
1936 if (mode != inner_mode
1937 && ! (CONST_INT_P (new_reg) && SCALAR_INT_MODE_P (mode)))
1939 poly_uint64 offset = 0;
1940 if (partial_subreg_p (mode, inner_mode)
1941 && SCALAR_INT_MODE_P (inner_mode))
1942 offset = subreg_lowpart_offset (mode, inner_mode);
1943 if (debug_p)
1944 new_reg = gen_rtx_raw_SUBREG (mode, new_reg, offset);
1945 else
1946 new_reg = gen_rtx_SUBREG (mode, new_reg, offset);
1948 *loc = new_reg;
1949 return true;
1952 /* Scan all the operand sub-expressions. */
1953 fmt = GET_RTX_FORMAT (code);
1954 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1956 if (fmt[i] == 'e')
1958 if (lra_substitute_pseudo (&XEXP (x, i), old_regno,
1959 new_reg, subreg_p, debug_p))
1960 result = true;
1962 else if (fmt[i] == 'E')
1964 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1965 if (lra_substitute_pseudo (&XVECEXP (x, i, j), old_regno,
1966 new_reg, subreg_p, debug_p))
1967 result = true;
1970 return result;
1973 /* Call lra_substitute_pseudo within an insn. Try to simplify subreg
1974 of constant if SUBREG_P. This won't update the insn ptr, just the
1975 contents of the insn. */
1976 bool
1977 lra_substitute_pseudo_within_insn (rtx_insn *insn, int old_regno,
1978 rtx new_reg, bool subreg_p)
1980 rtx loc = insn;
1981 return lra_substitute_pseudo (&loc, old_regno, new_reg, subreg_p,
1982 DEBUG_INSN_P (insn));
1987 /* This page contains code dealing with scratches (changing them onto
1988 pseudos and restoring them from the pseudos).
1990 We change scratches into pseudos at the beginning of LRA to
1991 simplify dealing with them (conflicts, hard register assignments).
1993 If the pseudo denoting scratch was spilled it means that we do need
1994 a hard register for it. Such pseudos are transformed back to
1995 scratches at the end of LRA. */
1997 /* Description of location of a former scratch operand. */
1998 struct sloc
2000 rtx_insn *insn; /* Insn where the scratch was. */
2001 int nop; /* Number of the operand which was a scratch. */
2004 typedef struct sloc *sloc_t;
2006 /* Locations of the former scratches. */
2007 static vec<sloc_t> scratches;
2009 /* Bitmap of scratch regnos. */
2010 static bitmap_head scratch_bitmap;
2012 /* Bitmap of scratch operands. */
2013 static bitmap_head scratch_operand_bitmap;
2015 /* Return true if pseudo REGNO is made of SCRATCH. */
2016 bool
2017 lra_former_scratch_p (int regno)
2019 return bitmap_bit_p (&scratch_bitmap, regno);
2022 /* Return true if the operand NOP of INSN is a former scratch. */
2023 bool
2024 lra_former_scratch_operand_p (rtx_insn *insn, int nop)
2026 return bitmap_bit_p (&scratch_operand_bitmap,
2027 INSN_UID (insn) * MAX_RECOG_OPERANDS + nop) != 0;
2030 /* Register operand NOP in INSN as a former scratch. It will be
2031 changed to scratch back, if it is necessary, at the LRA end. */
2032 void
2033 lra_register_new_scratch_op (rtx_insn *insn, int nop)
2035 lra_insn_recog_data_t id = lra_get_insn_recog_data (insn);
2036 rtx op = *id->operand_loc[nop];
2037 sloc_t loc = XNEW (struct sloc);
2038 lra_assert (REG_P (op));
2039 loc->insn = insn;
2040 loc->nop = nop;
2041 scratches.safe_push (loc);
2042 bitmap_set_bit (&scratch_bitmap, REGNO (op));
2043 bitmap_set_bit (&scratch_operand_bitmap,
2044 INSN_UID (insn) * MAX_RECOG_OPERANDS + nop);
2045 add_reg_note (insn, REG_UNUSED, op);
2048 /* Change scratches onto pseudos and save their location. */
2049 static void
2050 remove_scratches (void)
2052 int i;
2053 bool insn_changed_p;
2054 basic_block bb;
2055 rtx_insn *insn;
2056 rtx reg;
2057 lra_insn_recog_data_t id;
2058 struct lra_static_insn_data *static_id;
2060 scratches.create (get_max_uid ());
2061 bitmap_initialize (&scratch_bitmap, &reg_obstack);
2062 bitmap_initialize (&scratch_operand_bitmap, &reg_obstack);
2063 FOR_EACH_BB_FN (bb, cfun)
2064 FOR_BB_INSNS (bb, insn)
2065 if (INSN_P (insn))
2067 id = lra_get_insn_recog_data (insn);
2068 static_id = id->insn_static_data;
2069 insn_changed_p = false;
2070 for (i = 0; i < static_id->n_operands; i++)
2071 if (GET_CODE (*id->operand_loc[i]) == SCRATCH
2072 && GET_MODE (*id->operand_loc[i]) != VOIDmode)
2074 insn_changed_p = true;
2075 *id->operand_loc[i] = reg
2076 = lra_create_new_reg (static_id->operand[i].mode,
2077 *id->operand_loc[i], ALL_REGS, NULL);
2078 lra_register_new_scratch_op (insn, i);
2079 if (lra_dump_file != NULL)
2080 fprintf (lra_dump_file,
2081 "Removing SCRATCH in insn #%u (nop %d)\n",
2082 INSN_UID (insn), i);
2084 if (insn_changed_p)
2085 /* Because we might use DF right after caller-saves sub-pass
2086 we need to keep DF info up to date. */
2087 df_insn_rescan (insn);
2091 /* Changes pseudos created by function remove_scratches onto scratches. */
2092 static void
2093 restore_scratches (void)
2095 int regno;
2096 unsigned i;
2097 sloc_t loc;
2098 rtx_insn *last = NULL;
2099 lra_insn_recog_data_t id = NULL;
2101 for (i = 0; scratches.iterate (i, &loc); i++)
2103 /* Ignore already deleted insns. */
2104 if (NOTE_P (loc->insn)
2105 && NOTE_KIND (loc->insn) == NOTE_INSN_DELETED)
2106 continue;
2107 if (last != loc->insn)
2109 last = loc->insn;
2110 id = lra_get_insn_recog_data (last);
2112 if (REG_P (*id->operand_loc[loc->nop])
2113 && ((regno = REGNO (*id->operand_loc[loc->nop]))
2114 >= FIRST_PSEUDO_REGISTER)
2115 && lra_get_regno_hard_regno (regno) < 0)
2117 /* It should be only case when scratch register with chosen
2118 constraint 'X' did not get memory or hard register. */
2119 lra_assert (lra_former_scratch_p (regno));
2120 *id->operand_loc[loc->nop]
2121 = gen_rtx_SCRATCH (GET_MODE (*id->operand_loc[loc->nop]));
2122 lra_update_dup (id, loc->nop);
2123 if (lra_dump_file != NULL)
2124 fprintf (lra_dump_file, "Restoring SCRATCH in insn #%u(nop %d)\n",
2125 INSN_UID (loc->insn), loc->nop);
2128 for (i = 0; scratches.iterate (i, &loc); i++)
2129 free (loc);
2130 scratches.release ();
2131 bitmap_clear (&scratch_bitmap);
2132 bitmap_clear (&scratch_operand_bitmap);
2137 /* Function checks RTL for correctness. If FINAL_P is true, it is
2138 done at the end of LRA and the check is more rigorous. */
2139 static void
2140 check_rtl (bool final_p)
2142 basic_block bb;
2143 rtx_insn *insn;
2145 lra_assert (! final_p || reload_completed);
2146 FOR_EACH_BB_FN (bb, cfun)
2147 FOR_BB_INSNS (bb, insn)
2148 if (NONDEBUG_INSN_P (insn)
2149 && GET_CODE (PATTERN (insn)) != USE
2150 && GET_CODE (PATTERN (insn)) != CLOBBER
2151 && GET_CODE (PATTERN (insn)) != ASM_INPUT)
2153 if (final_p)
2155 extract_constrain_insn (insn);
2156 continue;
2158 /* LRA code is based on assumption that all addresses can be
2159 correctly decomposed. LRA can generate reloads for
2160 decomposable addresses. The decomposition code checks the
2161 correctness of the addresses. So we don't need to check
2162 the addresses here. Don't call insn_invalid_p here, it can
2163 change the code at this stage. */
2164 if (recog_memoized (insn) < 0 && asm_noperands (PATTERN (insn)) < 0)
2165 fatal_insn_not_found (insn);
2169 /* Determine if the current function has an exception receiver block
2170 that reaches the exit block via non-exceptional edges */
2171 static bool
2172 has_nonexceptional_receiver (void)
2174 edge e;
2175 edge_iterator ei;
2176 basic_block *tos, *worklist, bb;
2178 /* If we're not optimizing, then just err on the safe side. */
2179 if (!optimize)
2180 return true;
2182 /* First determine which blocks can reach exit via normal paths. */
2183 tos = worklist = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun) + 1);
2185 FOR_EACH_BB_FN (bb, cfun)
2186 bb->flags &= ~BB_REACHABLE;
2188 /* Place the exit block on our worklist. */
2189 EXIT_BLOCK_PTR_FOR_FN (cfun)->flags |= BB_REACHABLE;
2190 *tos++ = EXIT_BLOCK_PTR_FOR_FN (cfun);
2192 /* Iterate: find everything reachable from what we've already seen. */
2193 while (tos != worklist)
2195 bb = *--tos;
2197 FOR_EACH_EDGE (e, ei, bb->preds)
2198 if (e->flags & EDGE_ABNORMAL)
2200 free (worklist);
2201 return true;
2203 else
2205 basic_block src = e->src;
2207 if (!(src->flags & BB_REACHABLE))
2209 src->flags |= BB_REACHABLE;
2210 *tos++ = src;
2214 free (worklist);
2215 /* No exceptional block reached exit unexceptionally. */
2216 return false;
2220 /* Process recursively X of INSN and add REG_INC notes if necessary. */
2221 static void
2222 add_auto_inc_notes (rtx_insn *insn, rtx x)
2224 enum rtx_code code = GET_CODE (x);
2225 const char *fmt;
2226 int i, j;
2228 if (code == MEM && auto_inc_p (XEXP (x, 0)))
2230 add_reg_note (insn, REG_INC, XEXP (XEXP (x, 0), 0));
2231 return;
2234 /* Scan all X sub-expressions. */
2235 fmt = GET_RTX_FORMAT (code);
2236 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2238 if (fmt[i] == 'e')
2239 add_auto_inc_notes (insn, XEXP (x, i));
2240 else if (fmt[i] == 'E')
2241 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
2242 add_auto_inc_notes (insn, XVECEXP (x, i, j));
2247 /* Remove all REG_DEAD and REG_UNUSED notes and regenerate REG_INC.
2248 We change pseudos by hard registers without notification of DF and
2249 that can make the notes obsolete. DF-infrastructure does not deal
2250 with REG_INC notes -- so we should regenerate them here. */
2251 static void
2252 update_inc_notes (void)
2254 rtx *pnote;
2255 basic_block bb;
2256 rtx_insn *insn;
2258 FOR_EACH_BB_FN (bb, cfun)
2259 FOR_BB_INSNS (bb, insn)
2260 if (NONDEBUG_INSN_P (insn))
2262 pnote = &REG_NOTES (insn);
2263 while (*pnote != 0)
2265 if (REG_NOTE_KIND (*pnote) == REG_DEAD
2266 || REG_NOTE_KIND (*pnote) == REG_UNUSED
2267 || REG_NOTE_KIND (*pnote) == REG_INC)
2268 *pnote = XEXP (*pnote, 1);
2269 else
2270 pnote = &XEXP (*pnote, 1);
2273 if (AUTO_INC_DEC)
2274 add_auto_inc_notes (insn, PATTERN (insn));
2278 /* Set to 1 while in lra. */
2279 int lra_in_progress;
2281 /* Start of pseudo regnos before the LRA. */
2282 int lra_new_regno_start;
2284 /* Start of reload pseudo regnos before the new spill pass. */
2285 int lra_constraint_new_regno_start;
2287 /* Avoid spilling pseudos with regno more than the following value if
2288 it is possible. */
2289 int lra_bad_spill_regno_start;
2291 /* Inheritance pseudo regnos before the new spill pass. */
2292 bitmap_head lra_inheritance_pseudos;
2294 /* Split regnos before the new spill pass. */
2295 bitmap_head lra_split_regs;
2297 /* Reload pseudo regnos before the new assignment pass which still can
2298 be spilled after the assignment pass as memory is also accepted in
2299 insns for the reload pseudos. */
2300 bitmap_head lra_optional_reload_pseudos;
2302 /* Pseudo regnos used for subreg reloads before the new assignment
2303 pass. Such pseudos still can be spilled after the assignment
2304 pass. */
2305 bitmap_head lra_subreg_reload_pseudos;
2307 /* File used for output of LRA debug information. */
2308 FILE *lra_dump_file;
2310 /* True if we should try spill into registers of different classes
2311 instead of memory. */
2312 bool lra_reg_spill_p;
2314 /* Set up value LRA_REG_SPILL_P. */
2315 static void
2316 setup_reg_spill_flag (void)
2318 int cl, mode;
2320 if (targetm.spill_class != NULL)
2321 for (cl = 0; cl < (int) LIM_REG_CLASSES; cl++)
2322 for (mode = 0; mode < MAX_MACHINE_MODE; mode++)
2323 if (targetm.spill_class ((enum reg_class) cl,
2324 (machine_mode) mode) != NO_REGS)
2326 lra_reg_spill_p = true;
2327 return;
2329 lra_reg_spill_p = false;
2332 /* True if the current function is too big to use regular algorithms
2333 in LRA. In other words, we should use simpler and faster algorithms
2334 in LRA. It also means we should not worry about generation code
2335 for caller saves. The value is set up in IRA. */
2336 bool lra_simple_p;
2338 /* Major LRA entry function. F is a file should be used to dump LRA
2339 debug info. */
2340 void
2341 lra (FILE *f)
2343 int i;
2344 bool live_p, inserted_p;
2346 lra_dump_file = f;
2348 timevar_push (TV_LRA);
2350 /* Make sure that the last insn is a note. Some subsequent passes
2351 need it. */
2352 emit_note (NOTE_INSN_DELETED);
2354 COPY_HARD_REG_SET (lra_no_alloc_regs, ira_no_alloc_regs);
2356 init_reg_info ();
2357 expand_reg_info ();
2359 init_insn_recog_data ();
2361 /* Some quick check on RTL generated by previous passes. */
2362 if (flag_checking)
2363 check_rtl (false);
2365 lra_in_progress = 1;
2367 lra_live_range_iter = lra_coalesce_iter = lra_constraint_iter = 0;
2368 lra_assignment_iter = lra_assignment_iter_after_spill = 0;
2369 lra_inheritance_iter = lra_undo_inheritance_iter = 0;
2370 lra_rematerialization_iter = 0;
2372 setup_reg_spill_flag ();
2374 /* Function remove_scratches can creates new pseudos for clobbers --
2375 so set up lra_constraint_new_regno_start before its call to
2376 permit changing reg classes for pseudos created by this
2377 simplification. */
2378 lra_constraint_new_regno_start = lra_new_regno_start = max_reg_num ();
2379 lra_bad_spill_regno_start = INT_MAX;
2380 remove_scratches ();
2382 /* A function that has a non-local label that can reach the exit
2383 block via non-exceptional paths must save all call-saved
2384 registers. */
2385 if (cfun->has_nonlocal_label && has_nonexceptional_receiver ())
2386 crtl->saves_all_registers = 1;
2388 if (crtl->saves_all_registers)
2389 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2390 if (! call_used_regs[i] && ! fixed_regs[i] && ! LOCAL_REGNO (i))
2391 df_set_regs_ever_live (i, true);
2393 /* We don't DF from now and avoid its using because it is to
2394 expensive when a lot of RTL changes are made. */
2395 df_set_flags (DF_NO_INSN_RESCAN);
2396 lra_constraint_insn_stack.create (get_max_uid ());
2397 lra_constraint_insn_stack_bitmap = sbitmap_alloc (get_max_uid ());
2398 bitmap_clear (lra_constraint_insn_stack_bitmap);
2399 lra_live_ranges_init ();
2400 lra_constraints_init ();
2401 lra_curr_reload_num = 0;
2402 push_insns (get_last_insn (), NULL);
2403 /* It is needed for the 1st coalescing. */
2404 bitmap_initialize (&lra_inheritance_pseudos, &reg_obstack);
2405 bitmap_initialize (&lra_split_regs, &reg_obstack);
2406 bitmap_initialize (&lra_optional_reload_pseudos, &reg_obstack);
2407 bitmap_initialize (&lra_subreg_reload_pseudos, &reg_obstack);
2408 live_p = false;
2409 if (maybe_ne (get_frame_size (), 0) && crtl->stack_alignment_needed)
2410 /* If we have a stack frame, we must align it now. The stack size
2411 may be a part of the offset computation for register
2412 elimination. */
2413 assign_stack_local (BLKmode, 0, crtl->stack_alignment_needed);
2414 lra_init_equiv ();
2415 for (;;)
2417 for (;;)
2419 bool reloads_p = lra_constraints (lra_constraint_iter == 0);
2420 /* Constraint transformations may result in that eliminable
2421 hard regs become uneliminable and pseudos which use them
2422 should be spilled. It is better to do it before pseudo
2423 assignments.
2425 For example, rs6000 can make
2426 RS6000_PIC_OFFSET_TABLE_REGNUM uneliminable if we started
2427 to use a constant pool. */
2428 lra_eliminate (false, false);
2429 /* We should try to assign hard registers to scratches even
2430 if there were no RTL transformations in lra_constraints.
2431 Also we should check IRA assignments on the first
2432 iteration as they can be wrong because of early clobbers
2433 operands which are ignored in IRA. */
2434 if (! reloads_p && lra_constraint_iter > 1)
2436 /* Stack is not empty here only when there are changes
2437 during the elimination sub-pass. */
2438 if (bitmap_empty_p (lra_constraint_insn_stack_bitmap))
2439 break;
2440 else
2441 /* If there are no reloads but changing due
2442 elimination, restart the constraint sub-pass
2443 first. */
2444 continue;
2446 /* Do inheritance only for regular algorithms. */
2447 if (! lra_simple_p)
2449 if (flag_ipa_ra)
2451 if (live_p)
2452 lra_clear_live_ranges ();
2453 /* As a side-effect of lra_create_live_ranges, we calculate
2454 actual_call_used_reg_set, which is needed during
2455 lra_inheritance. */
2456 lra_create_live_ranges (true, true);
2457 live_p = true;
2459 lra_inheritance ();
2461 if (live_p)
2462 lra_clear_live_ranges ();
2463 /* We need live ranges for lra_assign -- so build them. But
2464 don't remove dead insns or change global live info as we
2465 can undo inheritance transformations after inheritance
2466 pseudo assigning. */
2467 lra_create_live_ranges (true, false);
2468 live_p = true;
2469 /* If we don't spill non-reload and non-inheritance pseudos,
2470 there is no sense to run memory-memory move coalescing.
2471 If inheritance pseudos were spilled, the memory-memory
2472 moves involving them will be removed by pass undoing
2473 inheritance. */
2474 if (lra_simple_p)
2475 lra_assign ();
2476 else
2478 bool spill_p = !lra_assign ();
2480 if (lra_undo_inheritance ())
2481 live_p = false;
2482 if (spill_p)
2484 if (! live_p)
2486 lra_create_live_ranges (true, true);
2487 live_p = true;
2489 if (lra_coalesce ())
2490 live_p = false;
2492 if (! live_p)
2493 lra_clear_live_ranges ();
2496 /* Don't clear optional reloads bitmap until all constraints are
2497 satisfied as we need to differ them from regular reloads. */
2498 bitmap_clear (&lra_optional_reload_pseudos);
2499 bitmap_clear (&lra_subreg_reload_pseudos);
2500 bitmap_clear (&lra_inheritance_pseudos);
2501 bitmap_clear (&lra_split_regs);
2502 if (! live_p)
2504 /* We need full live info for spilling pseudos into
2505 registers instead of memory. */
2506 lra_create_live_ranges (lra_reg_spill_p, true);
2507 live_p = true;
2509 /* We should check necessity for spilling here as the above live
2510 range pass can remove spilled pseudos. */
2511 if (! lra_need_for_spills_p ())
2512 break;
2513 /* Now we know what pseudos should be spilled. Try to
2514 rematerialize them first. */
2515 if (lra_remat ())
2517 /* We need full live info -- see the comment above. */
2518 lra_create_live_ranges (lra_reg_spill_p, true);
2519 live_p = true;
2520 if (! lra_need_for_spills_p ())
2521 break;
2523 lra_spill ();
2524 /* Assignment of stack slots changes elimination offsets for
2525 some eliminations. So update the offsets here. */
2526 lra_eliminate (false, false);
2527 lra_constraint_new_regno_start = max_reg_num ();
2528 if (lra_bad_spill_regno_start == INT_MAX
2529 && lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES
2530 && lra_rematerialization_iter > LRA_MAX_REMATERIALIZATION_PASSES)
2531 /* After switching off inheritance and rematerialization
2532 passes, avoid spilling reload pseudos will be created to
2533 prevent LRA cycling in some complicated cases. */
2534 lra_bad_spill_regno_start = lra_constraint_new_regno_start;
2535 lra_assignment_iter_after_spill = 0;
2537 restore_scratches ();
2538 lra_eliminate (true, false);
2539 lra_final_code_change ();
2540 lra_in_progress = 0;
2541 if (live_p)
2542 lra_clear_live_ranges ();
2543 lra_live_ranges_finish ();
2544 lra_constraints_finish ();
2545 finish_reg_info ();
2546 sbitmap_free (lra_constraint_insn_stack_bitmap);
2547 lra_constraint_insn_stack.release ();
2548 finish_insn_recog_data ();
2549 regstat_free_n_sets_and_refs ();
2550 regstat_free_ri ();
2551 reload_completed = 1;
2552 update_inc_notes ();
2554 inserted_p = fixup_abnormal_edges ();
2556 /* We've possibly turned single trapping insn into multiple ones. */
2557 if (cfun->can_throw_non_call_exceptions)
2559 auto_sbitmap blocks (last_basic_block_for_fn (cfun));
2560 bitmap_ones (blocks);
2561 find_many_sub_basic_blocks (blocks);
2564 if (inserted_p)
2565 commit_edge_insertions ();
2567 /* Replacing pseudos with their memory equivalents might have
2568 created shared rtx. Subsequent passes would get confused
2569 by this, so unshare everything here. */
2570 unshare_all_rtl_again (get_insns ());
2572 if (flag_checking)
2573 check_rtl (true);
2575 timevar_pop (TV_LRA);
2578 /* Called once per compiler to initialize LRA data once. */
2579 void
2580 lra_init_once (void)
2582 init_insn_code_data_once ();
2585 /* Called once per compiler to finish LRA data which are initialize
2586 once. */
2587 void
2588 lra_finish_once (void)
2590 finish_insn_code_data_once ();