debug/dwarf: support 64-bit DWARF in byte order check
[official-gcc.git] / gcc / lra.c
blob3122f2c25053a8b48b2ab277510bf18c4376cec9
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 *x, lra_insn_recog_data_t data,
811 struct lra_insn_reg *list,
812 enum op_type type, bool early_clobber)
814 int i, j, regno, last;
815 bool subreg_p;
816 machine_mode mode;
817 struct lra_insn_reg *curr;
818 rtx op = *x;
819 enum rtx_code code = GET_CODE (op);
820 const char *fmt = GET_RTX_FORMAT (code);
822 for (i = 0; i < data->insn_static_data->n_operands; i++)
823 if (! data->insn_static_data->operand[i].is_operator
824 && x == data->operand_loc[i])
825 /* It is an operand loc. Stop here. */
826 return list;
827 for (i = 0; i < data->insn_static_data->n_dups; i++)
828 if (x == data->dup_loc[i])
829 /* It is a dup loc. Stop here. */
830 return list;
831 mode = GET_MODE (op);
832 subreg_p = false;
833 if (code == SUBREG)
835 op = SUBREG_REG (op);
836 code = GET_CODE (op);
837 if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (op)))
839 mode = GET_MODE (op);
840 if (GET_MODE_SIZE (mode) > REGMODE_NATURAL_SIZE (mode))
841 subreg_p = true;
844 if (REG_P (op))
846 if ((regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER)
847 return list;
848 /* Process all regs even unallocatable ones as we need info
849 about all regs for rematerialization pass. */
850 for (last = end_hard_regno (mode, regno); regno < last; regno++)
852 for (curr = list; curr != NULL; curr = curr->next)
853 if (curr->regno == regno && curr->subreg_p == subreg_p
854 && curr->biggest_mode == mode)
856 if (curr->type != type)
857 curr->type = OP_INOUT;
858 if (early_clobber)
860 curr->early_clobber = true;
861 curr->early_clobber_alts = ALL_ALTERNATIVES;
863 break;
865 if (curr == NULL)
867 /* This is a new hard regno or the info can not be
868 integrated into the found structure. */
869 #ifdef STACK_REGS
870 early_clobber
871 = (early_clobber
872 /* This clobber is to inform popping floating
873 point stack only. */
874 && ! (FIRST_STACK_REG <= regno
875 && regno <= LAST_STACK_REG));
876 #endif
877 list = new_insn_reg (data->insn, regno, type, mode, subreg_p,
878 early_clobber,
879 early_clobber ? ALL_ALTERNATIVES : 0, list);
882 return list;
884 switch (code)
886 case SET:
887 list = collect_non_operand_hard_regs (&SET_DEST (op), data,
888 list, OP_OUT, false);
889 list = collect_non_operand_hard_regs (&SET_SRC (op), data,
890 list, OP_IN, false);
891 break;
892 case CLOBBER:
893 /* We treat clobber of non-operand hard registers as early
894 clobber (the behavior is expected from asm). */
895 list = collect_non_operand_hard_regs (&XEXP (op, 0), data,
896 list, OP_OUT, true);
897 break;
898 case PRE_INC: case PRE_DEC: case POST_INC: case POST_DEC:
899 list = collect_non_operand_hard_regs (&XEXP (op, 0), data,
900 list, OP_INOUT, false);
901 break;
902 case PRE_MODIFY: case POST_MODIFY:
903 list = collect_non_operand_hard_regs (&XEXP (op, 0), data,
904 list, OP_INOUT, false);
905 list = collect_non_operand_hard_regs (&XEXP (op, 1), data,
906 list, OP_IN, false);
907 break;
908 default:
909 fmt = GET_RTX_FORMAT (code);
910 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
912 if (fmt[i] == 'e')
913 list = collect_non_operand_hard_regs (&XEXP (op, i), data,
914 list, OP_IN, false);
915 else if (fmt[i] == 'E')
916 for (j = XVECLEN (op, i) - 1; j >= 0; j--)
917 list = collect_non_operand_hard_regs (&XVECEXP (op, i, j), data,
918 list, OP_IN, false);
921 return list;
924 /* Set up and return info about INSN. Set up the info if it is not set up
925 yet. */
926 lra_insn_recog_data_t
927 lra_set_insn_recog_data (rtx_insn *insn)
929 lra_insn_recog_data_t data;
930 int i, n, icode;
931 rtx **locs;
932 unsigned int uid = INSN_UID (insn);
933 struct lra_static_insn_data *insn_static_data;
935 check_and_expand_insn_recog_data (uid);
936 if (DEBUG_INSN_P (insn))
937 icode = -1;
938 else
940 icode = INSN_CODE (insn);
941 if (icode < 0)
942 /* It might be a new simple insn which is not recognized yet. */
943 INSN_CODE (insn) = icode = recog_memoized (insn);
945 data = XNEW (struct lra_insn_recog_data);
946 lra_insn_recog_data[uid] = data;
947 data->insn = insn;
948 data->used_insn_alternative = -1;
949 data->icode = icode;
950 data->regs = NULL;
951 if (DEBUG_INSN_P (insn))
953 data->insn_static_data = &debug_insn_static_data;
954 data->dup_loc = NULL;
955 data->arg_hard_regs = NULL;
956 data->preferred_alternatives = ALL_ALTERNATIVES;
957 data->operand_loc = XNEWVEC (rtx *, 1);
958 data->operand_loc[0] = &INSN_VAR_LOCATION_LOC (insn);
959 return data;
961 if (icode < 0)
963 int nop, nalt;
964 machine_mode operand_mode[MAX_RECOG_OPERANDS];
965 const char *constraints[MAX_RECOG_OPERANDS];
967 nop = asm_noperands (PATTERN (insn));
968 data->operand_loc = data->dup_loc = NULL;
969 nalt = 1;
970 if (nop < 0)
972 /* It is a special insn like USE or CLOBBER. We should
973 recognize any regular insn otherwise LRA can do nothing
974 with this insn. */
975 gcc_assert (GET_CODE (PATTERN (insn)) == USE
976 || GET_CODE (PATTERN (insn)) == CLOBBER
977 || GET_CODE (PATTERN (insn)) == ASM_INPUT);
978 data->insn_static_data = insn_static_data
979 = get_static_insn_data (-1, 0, 0, nalt);
981 else
983 /* expand_asm_operands makes sure there aren't too many
984 operands. */
985 lra_assert (nop <= MAX_RECOG_OPERANDS);
986 if (nop != 0)
987 data->operand_loc = XNEWVEC (rtx *, nop);
988 /* Now get the operand values and constraints out of the
989 insn. */
990 decode_asm_operands (PATTERN (insn), NULL,
991 data->operand_loc,
992 constraints, operand_mode, NULL);
993 if (nop > 0)
995 const char *p = recog_data.constraints[0];
997 for (p = constraints[0]; *p; p++)
998 nalt += *p == ',';
1000 data->insn_static_data = insn_static_data
1001 = get_static_insn_data (-1, nop, 0, nalt);
1002 for (i = 0; i < nop; i++)
1004 insn_static_data->operand[i].mode = operand_mode[i];
1005 insn_static_data->operand[i].constraint = constraints[i];
1006 insn_static_data->operand[i].strict_low = false;
1007 insn_static_data->operand[i].is_operator = false;
1008 insn_static_data->operand[i].is_address = false;
1011 for (i = 0; i < insn_static_data->n_operands; i++)
1012 insn_static_data->operand[i].type
1013 = (insn_static_data->operand[i].constraint[0] == '=' ? OP_OUT
1014 : insn_static_data->operand[i].constraint[0] == '+' ? OP_INOUT
1015 : OP_IN);
1016 data->preferred_alternatives = ALL_ALTERNATIVES;
1017 if (nop > 0)
1019 operand_alternative *op_alt = XCNEWVEC (operand_alternative,
1020 nalt * nop);
1021 preprocess_constraints (nop, nalt, constraints, op_alt);
1022 setup_operand_alternative (data, op_alt);
1025 else
1027 insn_extract (insn);
1028 data->insn_static_data = insn_static_data
1029 = get_static_insn_data (icode, insn_data[icode].n_operands,
1030 insn_data[icode].n_dups,
1031 insn_data[icode].n_alternatives);
1032 n = insn_static_data->n_operands;
1033 if (n == 0)
1034 locs = NULL;
1035 else
1037 locs = XNEWVEC (rtx *, n);
1038 memcpy (locs, recog_data.operand_loc, n * sizeof (rtx *));
1040 data->operand_loc = locs;
1041 n = insn_static_data->n_dups;
1042 if (n == 0)
1043 locs = NULL;
1044 else
1046 locs = XNEWVEC (rtx *, n);
1047 memcpy (locs, recog_data.dup_loc, n * sizeof (rtx *));
1049 data->dup_loc = locs;
1050 data->preferred_alternatives = get_preferred_alternatives (insn);
1051 const operand_alternative *op_alt = preprocess_insn_constraints (icode);
1052 if (!insn_static_data->operand_alternative)
1053 setup_operand_alternative (data, op_alt);
1054 else if (op_alt != insn_static_data->operand_alternative)
1055 insn_static_data->operand_alternative = op_alt;
1057 if (GET_CODE (PATTERN (insn)) == CLOBBER || GET_CODE (PATTERN (insn)) == USE)
1058 insn_static_data->hard_regs = NULL;
1059 else
1060 insn_static_data->hard_regs
1061 = collect_non_operand_hard_regs (&PATTERN (insn), data,
1062 NULL, OP_IN, false);
1063 data->arg_hard_regs = NULL;
1064 if (CALL_P (insn))
1066 bool use_p;
1067 rtx link;
1068 int n_hard_regs, regno, arg_hard_regs[FIRST_PSEUDO_REGISTER];
1070 n_hard_regs = 0;
1071 /* Finding implicit hard register usage. We believe it will be
1072 not changed whatever transformations are used. Call insns
1073 are such example. */
1074 for (link = CALL_INSN_FUNCTION_USAGE (insn);
1075 link != NULL_RTX;
1076 link = XEXP (link, 1))
1077 if (((use_p = GET_CODE (XEXP (link, 0)) == USE)
1078 || GET_CODE (XEXP (link, 0)) == CLOBBER)
1079 && REG_P (XEXP (XEXP (link, 0), 0)))
1081 regno = REGNO (XEXP (XEXP (link, 0), 0));
1082 lra_assert (regno < FIRST_PSEUDO_REGISTER);
1083 /* It is an argument register. */
1084 for (i = REG_NREGS (XEXP (XEXP (link, 0), 0)) - 1; i >= 0; i--)
1085 arg_hard_regs[n_hard_regs++]
1086 = regno + i + (use_p ? 0 : FIRST_PSEUDO_REGISTER);
1088 if (n_hard_regs != 0)
1090 arg_hard_regs[n_hard_regs++] = -1;
1091 data->arg_hard_regs = XNEWVEC (int, n_hard_regs);
1092 memcpy (data->arg_hard_regs, arg_hard_regs,
1093 sizeof (int) * n_hard_regs);
1096 /* Some output operand can be recognized only from the context not
1097 from the constraints which are empty in this case. Call insn may
1098 contain a hard register in set destination with empty constraint
1099 and extract_insn treats them as an input. */
1100 for (i = 0; i < insn_static_data->n_operands; i++)
1102 int j;
1103 rtx pat, set;
1104 struct lra_operand_data *operand = &insn_static_data->operand[i];
1106 /* ??? Should we treat 'X' the same way. It looks to me that
1107 'X' means anything and empty constraint means we do not
1108 care. */
1109 if (operand->type != OP_IN || *operand->constraint != '\0'
1110 || operand->is_operator)
1111 continue;
1112 pat = PATTERN (insn);
1113 if (GET_CODE (pat) == SET)
1115 if (data->operand_loc[i] != &SET_DEST (pat))
1116 continue;
1118 else if (GET_CODE (pat) == PARALLEL)
1120 for (j = XVECLEN (pat, 0) - 1; j >= 0; j--)
1122 set = XVECEXP (PATTERN (insn), 0, j);
1123 if (GET_CODE (set) == SET
1124 && &SET_DEST (set) == data->operand_loc[i])
1125 break;
1127 if (j < 0)
1128 continue;
1130 else
1131 continue;
1132 operand->type = OP_OUT;
1134 return data;
1137 /* Return info about insn give by UID. The info should be already set
1138 up. */
1139 static lra_insn_recog_data_t
1140 get_insn_recog_data_by_uid (int uid)
1142 lra_insn_recog_data_t data;
1144 data = lra_insn_recog_data[uid];
1145 lra_assert (data != NULL);
1146 return data;
1149 /* Invalidate all info about insn given by its UID. */
1150 static void
1151 invalidate_insn_recog_data (int uid)
1153 lra_insn_recog_data_t data;
1155 data = lra_insn_recog_data[uid];
1156 lra_assert (data != NULL);
1157 free_insn_recog_data (data);
1158 lra_insn_recog_data[uid] = NULL;
1161 /* Update all the insn info about INSN. It is usually called when
1162 something in the insn was changed. Return the updated info. */
1163 lra_insn_recog_data_t
1164 lra_update_insn_recog_data (rtx_insn *insn)
1166 lra_insn_recog_data_t data;
1167 int n;
1168 unsigned int uid = INSN_UID (insn);
1169 struct lra_static_insn_data *insn_static_data;
1170 HOST_WIDE_INT sp_offset = 0;
1172 check_and_expand_insn_recog_data (uid);
1173 if ((data = lra_insn_recog_data[uid]) != NULL
1174 && data->icode != INSN_CODE (insn))
1176 sp_offset = data->sp_offset;
1177 invalidate_insn_data_regno_info (data, insn, get_insn_freq (insn));
1178 invalidate_insn_recog_data (uid);
1179 data = NULL;
1181 if (data == NULL)
1183 data = lra_get_insn_recog_data (insn);
1184 /* Initiate or restore SP offset. */
1185 data->sp_offset = sp_offset;
1186 return data;
1188 insn_static_data = data->insn_static_data;
1189 data->used_insn_alternative = -1;
1190 if (DEBUG_INSN_P (insn))
1191 return data;
1192 if (data->icode < 0)
1194 int nop;
1195 machine_mode operand_mode[MAX_RECOG_OPERANDS];
1196 const char *constraints[MAX_RECOG_OPERANDS];
1198 nop = asm_noperands (PATTERN (insn));
1199 if (nop >= 0)
1201 lra_assert (nop == data->insn_static_data->n_operands);
1202 /* Now get the operand values and constraints out of the
1203 insn. */
1204 decode_asm_operands (PATTERN (insn), NULL,
1205 data->operand_loc,
1206 constraints, operand_mode, NULL);
1208 if (flag_checking)
1209 for (int i = 0; i < nop; i++)
1210 lra_assert
1211 (insn_static_data->operand[i].mode == operand_mode[i]
1212 && insn_static_data->operand[i].constraint == constraints[i]
1213 && ! insn_static_data->operand[i].is_operator);
1216 if (flag_checking)
1217 for (int i = 0; i < insn_static_data->n_operands; i++)
1218 lra_assert
1219 (insn_static_data->operand[i].type
1220 == (insn_static_data->operand[i].constraint[0] == '=' ? OP_OUT
1221 : insn_static_data->operand[i].constraint[0] == '+' ? OP_INOUT
1222 : OP_IN));
1224 else
1226 insn_extract (insn);
1227 n = insn_static_data->n_operands;
1228 if (n != 0)
1229 memcpy (data->operand_loc, recog_data.operand_loc, n * sizeof (rtx *));
1230 n = insn_static_data->n_dups;
1231 if (n != 0)
1232 memcpy (data->dup_loc, recog_data.dup_loc, n * sizeof (rtx *));
1233 lra_assert (check_bool_attrs (insn));
1235 return data;
1238 /* Set up that INSN is using alternative ALT now. */
1239 void
1240 lra_set_used_insn_alternative (rtx_insn *insn, int alt)
1242 lra_insn_recog_data_t data;
1244 data = lra_get_insn_recog_data (insn);
1245 data->used_insn_alternative = alt;
1248 /* Set up that insn with UID is using alternative ALT now. The insn
1249 info should be already set up. */
1250 void
1251 lra_set_used_insn_alternative_by_uid (int uid, int alt)
1253 lra_insn_recog_data_t data;
1255 check_and_expand_insn_recog_data (uid);
1256 data = lra_insn_recog_data[uid];
1257 lra_assert (data != NULL);
1258 data->used_insn_alternative = alt;
1263 /* This page contains code dealing with common register info and
1264 pseudo copies. */
1266 /* The size of the following array. */
1267 static int reg_info_size;
1268 /* Common info about each register. */
1269 struct lra_reg *lra_reg_info;
1271 /* Last register value. */
1272 static int last_reg_value;
1274 /* Return new register value. */
1275 static int
1276 get_new_reg_value (void)
1278 return ++last_reg_value;
1281 /* Vec referring to pseudo copies. */
1282 static vec<lra_copy_t> copy_vec;
1284 /* Initialize I-th element of lra_reg_info. */
1285 static inline void
1286 initialize_lra_reg_info_element (int i)
1288 bitmap_initialize (&lra_reg_info[i].insn_bitmap, &reg_obstack);
1289 #ifdef STACK_REGS
1290 lra_reg_info[i].no_stack_p = false;
1291 #endif
1292 CLEAR_HARD_REG_SET (lra_reg_info[i].conflict_hard_regs);
1293 CLEAR_HARD_REG_SET (lra_reg_info[i].actual_call_used_reg_set);
1294 lra_reg_info[i].preferred_hard_regno1 = -1;
1295 lra_reg_info[i].preferred_hard_regno2 = -1;
1296 lra_reg_info[i].preferred_hard_regno_profit1 = 0;
1297 lra_reg_info[i].preferred_hard_regno_profit2 = 0;
1298 lra_reg_info[i].biggest_mode = VOIDmode;
1299 lra_reg_info[i].live_ranges = NULL;
1300 lra_reg_info[i].nrefs = lra_reg_info[i].freq = 0;
1301 lra_reg_info[i].last_reload = 0;
1302 lra_reg_info[i].restore_rtx = NULL_RTX;
1303 lra_reg_info[i].val = get_new_reg_value ();
1304 lra_reg_info[i].offset = 0;
1305 lra_reg_info[i].copies = NULL;
1308 /* Initialize common reg info and copies. */
1309 static void
1310 init_reg_info (void)
1312 int i;
1314 last_reg_value = 0;
1315 reg_info_size = max_reg_num () * 3 / 2 + 1;
1316 lra_reg_info = XNEWVEC (struct lra_reg, reg_info_size);
1317 for (i = 0; i < reg_info_size; i++)
1318 initialize_lra_reg_info_element (i);
1319 copy_vec.truncate (0);
1323 /* Finish common reg info and copies. */
1324 static void
1325 finish_reg_info (void)
1327 int i;
1329 for (i = 0; i < reg_info_size; i++)
1330 bitmap_clear (&lra_reg_info[i].insn_bitmap);
1331 free (lra_reg_info);
1332 reg_info_size = 0;
1335 /* Expand common reg info if it is necessary. */
1336 static void
1337 expand_reg_info (void)
1339 int i, old = reg_info_size;
1341 if (reg_info_size > max_reg_num ())
1342 return;
1343 reg_info_size = max_reg_num () * 3 / 2 + 1;
1344 lra_reg_info = XRESIZEVEC (struct lra_reg, lra_reg_info, reg_info_size);
1345 for (i = old; i < reg_info_size; i++)
1346 initialize_lra_reg_info_element (i);
1349 /* Free all copies. */
1350 void
1351 lra_free_copies (void)
1353 lra_copy_t cp;
1355 while (copy_vec.length () != 0)
1357 cp = copy_vec.pop ();
1358 lra_reg_info[cp->regno1].copies = lra_reg_info[cp->regno2].copies = NULL;
1359 lra_copy_pool.remove (cp);
1363 /* Create copy of two pseudos REGNO1 and REGNO2. The copy execution
1364 frequency is FREQ. */
1365 void
1366 lra_create_copy (int regno1, int regno2, int freq)
1368 bool regno1_dest_p;
1369 lra_copy_t cp;
1371 lra_assert (regno1 != regno2);
1372 regno1_dest_p = true;
1373 if (regno1 > regno2)
1375 std::swap (regno1, regno2);
1376 regno1_dest_p = false;
1378 cp = lra_copy_pool.allocate ();
1379 copy_vec.safe_push (cp);
1380 cp->regno1_dest_p = regno1_dest_p;
1381 cp->freq = freq;
1382 cp->regno1 = regno1;
1383 cp->regno2 = regno2;
1384 cp->regno1_next = lra_reg_info[regno1].copies;
1385 lra_reg_info[regno1].copies = cp;
1386 cp->regno2_next = lra_reg_info[regno2].copies;
1387 lra_reg_info[regno2].copies = cp;
1388 if (lra_dump_file != NULL)
1389 fprintf (lra_dump_file, " Creating copy r%d%sr%d@%d\n",
1390 regno1, regno1_dest_p ? "<-" : "->", regno2, freq);
1393 /* Return N-th (0, 1, ...) copy. If there is no copy, return
1394 NULL. */
1395 lra_copy_t
1396 lra_get_copy (int n)
1398 if (n >= (int) copy_vec.length ())
1399 return NULL;
1400 return copy_vec[n];
1405 /* This page contains code dealing with info about registers in
1406 insns. */
1408 /* Process X of insn UID recursively and add info (operand type is
1409 given by TYPE, flag of that it is early clobber is EARLY_CLOBBER)
1410 about registers in X to the insn DATA. If X can be early clobbered,
1411 alternatives in which it can be early clobbered are given by
1412 EARLY_CLOBBER_ALTS. */
1413 static void
1414 add_regs_to_insn_regno_info (lra_insn_recog_data_t data, rtx x, int uid,
1415 enum op_type type, bool early_clobber,
1416 alternative_mask early_clobber_alts)
1418 int i, j, regno;
1419 bool subreg_p;
1420 machine_mode mode;
1421 const char *fmt;
1422 enum rtx_code code;
1423 struct lra_insn_reg *curr;
1425 code = GET_CODE (x);
1426 mode = GET_MODE (x);
1427 subreg_p = false;
1428 if (GET_CODE (x) == SUBREG)
1430 x = SUBREG_REG (x);
1431 code = GET_CODE (x);
1432 if (GET_MODE_SIZE (mode) < GET_MODE_SIZE (GET_MODE (x)))
1434 mode = GET_MODE (x);
1435 if (GET_MODE_SIZE (mode) > REGMODE_NATURAL_SIZE (mode))
1436 subreg_p = true;
1439 if (REG_P (x))
1441 regno = REGNO (x);
1442 /* Process all regs even unallocatable ones as we need info about
1443 all regs for rematerialization pass. */
1444 expand_reg_info ();
1445 if (bitmap_set_bit (&lra_reg_info[regno].insn_bitmap, uid))
1447 data->regs = new_insn_reg (data->insn, regno, type, mode, subreg_p,
1448 early_clobber, early_clobber_alts,
1449 data->regs);
1450 return;
1452 else
1454 for (curr = data->regs; curr != NULL; curr = curr->next)
1455 if (curr->regno == regno)
1457 if (curr->subreg_p != subreg_p || curr->biggest_mode != mode)
1458 /* The info can not be integrated into the found
1459 structure. */
1460 data->regs = new_insn_reg (data->insn, regno, type, mode,
1461 subreg_p, early_clobber,
1462 early_clobber_alts, data->regs);
1463 else
1465 if (curr->type != type)
1466 curr->type = OP_INOUT;
1467 if (curr->early_clobber != early_clobber)
1468 curr->early_clobber = true;
1469 curr->early_clobber_alts |= early_clobber_alts;
1471 return;
1473 gcc_unreachable ();
1477 switch (code)
1479 case SET:
1480 add_regs_to_insn_regno_info (data, SET_DEST (x), uid, OP_OUT, false, 0);
1481 add_regs_to_insn_regno_info (data, SET_SRC (x), uid, OP_IN, false, 0);
1482 break;
1483 case CLOBBER:
1484 /* We treat clobber of non-operand hard registers as early
1485 clobber (the behavior is expected from asm). */
1486 add_regs_to_insn_regno_info (data, XEXP (x, 0), uid, OP_OUT, true, ALL_ALTERNATIVES);
1487 break;
1488 case PRE_INC: case PRE_DEC: case POST_INC: case POST_DEC:
1489 add_regs_to_insn_regno_info (data, XEXP (x, 0), uid, OP_INOUT, false, 0);
1490 break;
1491 case PRE_MODIFY: case POST_MODIFY:
1492 add_regs_to_insn_regno_info (data, XEXP (x, 0), uid, OP_INOUT, false, 0);
1493 add_regs_to_insn_regno_info (data, XEXP (x, 1), uid, OP_IN, false, 0);
1494 break;
1495 default:
1496 if ((code != PARALLEL && code != EXPR_LIST) || type != OP_OUT)
1497 /* Some targets place small structures in registers for return
1498 values of functions, and those registers are wrapped in
1499 PARALLEL that we may see as the destination of a SET. Here
1500 is an example:
1502 (call_insn 13 12 14 2 (set (parallel:BLK [
1503 (expr_list:REG_DEP_TRUE (reg:DI 0 ax)
1504 (const_int 0 [0]))
1505 (expr_list:REG_DEP_TRUE (reg:DI 1 dx)
1506 (const_int 8 [0x8]))
1508 (call (mem:QI (symbol_ref:DI (... */
1509 type = OP_IN;
1510 fmt = GET_RTX_FORMAT (code);
1511 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1513 if (fmt[i] == 'e')
1514 add_regs_to_insn_regno_info (data, XEXP (x, i), uid, type, false, 0);
1515 else if (fmt[i] == 'E')
1517 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1518 add_regs_to_insn_regno_info (data, XVECEXP (x, i, j), uid,
1519 type, false, 0);
1525 /* Return execution frequency of INSN. */
1526 static int
1527 get_insn_freq (rtx_insn *insn)
1529 basic_block bb = BLOCK_FOR_INSN (insn);
1531 gcc_checking_assert (bb != NULL);
1532 return REG_FREQ_FROM_BB (bb);
1535 /* Invalidate all reg info of INSN with DATA and execution frequency
1536 FREQ. Update common info about the invalidated registers. */
1537 static void
1538 invalidate_insn_data_regno_info (lra_insn_recog_data_t data, rtx_insn *insn,
1539 int freq)
1541 int uid;
1542 bool debug_p;
1543 unsigned int i;
1544 struct lra_insn_reg *ir, *next_ir;
1546 uid = INSN_UID (insn);
1547 debug_p = DEBUG_INSN_P (insn);
1548 for (ir = data->regs; ir != NULL; ir = next_ir)
1550 i = ir->regno;
1551 next_ir = ir->next;
1552 lra_insn_reg_pool.remove (ir);
1553 bitmap_clear_bit (&lra_reg_info[i].insn_bitmap, uid);
1554 if (i >= FIRST_PSEUDO_REGISTER && ! debug_p)
1556 lra_reg_info[i].nrefs--;
1557 lra_reg_info[i].freq -= freq;
1558 lra_assert (lra_reg_info[i].nrefs >= 0 && lra_reg_info[i].freq >= 0);
1561 data->regs = NULL;
1564 /* Invalidate all reg info of INSN. Update common info about the
1565 invalidated registers. */
1566 void
1567 lra_invalidate_insn_regno_info (rtx_insn *insn)
1569 invalidate_insn_data_regno_info (lra_get_insn_recog_data (insn), insn,
1570 get_insn_freq (insn));
1573 /* Update common reg info from reg info of insn given by its DATA and
1574 execution frequency FREQ. */
1575 static void
1576 setup_insn_reg_info (lra_insn_recog_data_t data, int freq)
1578 unsigned int i;
1579 struct lra_insn_reg *ir;
1581 for (ir = data->regs; ir != NULL; ir = ir->next)
1582 if ((i = ir->regno) >= FIRST_PSEUDO_REGISTER)
1584 lra_reg_info[i].nrefs++;
1585 lra_reg_info[i].freq += freq;
1589 /* Set up insn reg info of INSN. Update common reg info from reg info
1590 of INSN. */
1591 void
1592 lra_update_insn_regno_info (rtx_insn *insn)
1594 int i, uid, freq;
1595 lra_insn_recog_data_t data;
1596 struct lra_static_insn_data *static_data;
1597 enum rtx_code code;
1598 rtx link;
1600 if (! INSN_P (insn))
1601 return;
1602 data = lra_get_insn_recog_data (insn);
1603 static_data = data->insn_static_data;
1604 freq = get_insn_freq (insn);
1605 invalidate_insn_data_regno_info (data, insn, freq);
1606 uid = INSN_UID (insn);
1607 for (i = static_data->n_operands - 1; i >= 0; i--)
1608 add_regs_to_insn_regno_info (data, *data->operand_loc[i], uid,
1609 static_data->operand[i].type,
1610 static_data->operand[i].early_clobber,
1611 static_data->operand[i].early_clobber_alts);
1612 if ((code = GET_CODE (PATTERN (insn))) == CLOBBER || code == USE)
1613 add_regs_to_insn_regno_info (data, XEXP (PATTERN (insn), 0), uid,
1614 code == USE ? OP_IN : OP_OUT, false, 0);
1615 if (CALL_P (insn))
1616 /* On some targets call insns can refer to pseudos in memory in
1617 CALL_INSN_FUNCTION_USAGE list. Process them in order to
1618 consider their occurrences in calls for different
1619 transformations (e.g. inheritance) with given pseudos. */
1620 for (link = CALL_INSN_FUNCTION_USAGE (insn);
1621 link != NULL_RTX;
1622 link = XEXP (link, 1))
1623 if (((code = GET_CODE (XEXP (link, 0))) == USE || code == CLOBBER)
1624 && MEM_P (XEXP (XEXP (link, 0), 0)))
1625 add_regs_to_insn_regno_info (data, XEXP (XEXP (link, 0), 0), uid,
1626 code == USE ? OP_IN : OP_OUT, false, 0);
1627 if (NONDEBUG_INSN_P (insn))
1628 setup_insn_reg_info (data, freq);
1631 /* Return reg info of insn given by it UID. */
1632 struct lra_insn_reg *
1633 lra_get_insn_regs (int uid)
1635 lra_insn_recog_data_t data;
1637 data = get_insn_recog_data_by_uid (uid);
1638 return data->regs;
1643 /* Recursive hash function for RTL X. */
1644 hashval_t
1645 lra_rtx_hash (rtx x)
1647 int i, j;
1648 enum rtx_code code;
1649 const char *fmt;
1650 hashval_t val = 0;
1652 if (x == 0)
1653 return val;
1655 code = GET_CODE (x);
1656 val += (int) code + 4095;
1658 /* Some RTL can be compared nonrecursively. */
1659 switch (code)
1661 case REG:
1662 return val + REGNO (x);
1664 case LABEL_REF:
1665 return iterative_hash_object (XEXP (x, 0), val);
1667 case SYMBOL_REF:
1668 return iterative_hash_object (XSTR (x, 0), val);
1670 case SCRATCH:
1671 case CONST_DOUBLE:
1672 case CONST_INT:
1673 case CONST_VECTOR:
1674 return val;
1676 default:
1677 break;
1680 /* Hash the elements. */
1681 fmt = GET_RTX_FORMAT (code);
1682 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1684 switch (fmt[i])
1686 case 'w':
1687 val += XWINT (x, i);
1688 break;
1690 case 'n':
1691 case 'i':
1692 val += XINT (x, i);
1693 break;
1695 case 'V':
1696 case 'E':
1697 val += XVECLEN (x, i);
1699 for (j = 0; j < XVECLEN (x, i); j++)
1700 val += lra_rtx_hash (XVECEXP (x, i, j));
1701 break;
1703 case 'e':
1704 val += lra_rtx_hash (XEXP (x, i));
1705 break;
1707 case 'S':
1708 case 's':
1709 val += htab_hash_string (XSTR (x, i));
1710 break;
1712 case 'u':
1713 case '0':
1714 case 't':
1715 break;
1717 /* It is believed that rtx's at this level will never
1718 contain anything but integers and other rtx's, except for
1719 within LABEL_REFs and SYMBOL_REFs. */
1720 default:
1721 abort ();
1724 return val;
1729 /* This page contains code dealing with stack of the insns which
1730 should be processed by the next constraint pass. */
1732 /* Bitmap used to put an insn on the stack only in one exemplar. */
1733 static sbitmap lra_constraint_insn_stack_bitmap;
1735 /* The stack itself. */
1736 vec<rtx_insn *> lra_constraint_insn_stack;
1738 /* Put INSN on the stack. If ALWAYS_UPDATE is true, always update the reg
1739 info for INSN, otherwise only update it if INSN is not already on the
1740 stack. */
1741 static inline void
1742 lra_push_insn_1 (rtx_insn *insn, bool always_update)
1744 unsigned int uid = INSN_UID (insn);
1745 if (always_update)
1746 lra_update_insn_regno_info (insn);
1747 if (uid >= SBITMAP_SIZE (lra_constraint_insn_stack_bitmap))
1748 lra_constraint_insn_stack_bitmap =
1749 sbitmap_resize (lra_constraint_insn_stack_bitmap, 3 * uid / 2, 0);
1750 if (bitmap_bit_p (lra_constraint_insn_stack_bitmap, uid))
1751 return;
1752 bitmap_set_bit (lra_constraint_insn_stack_bitmap, uid);
1753 if (! always_update)
1754 lra_update_insn_regno_info (insn);
1755 lra_constraint_insn_stack.safe_push (insn);
1758 /* Put INSN on the stack. */
1759 void
1760 lra_push_insn (rtx_insn *insn)
1762 lra_push_insn_1 (insn, false);
1765 /* Put INSN on the stack and update its reg info. */
1766 void
1767 lra_push_insn_and_update_insn_regno_info (rtx_insn *insn)
1769 lra_push_insn_1 (insn, true);
1772 /* Put insn with UID on the stack. */
1773 void
1774 lra_push_insn_by_uid (unsigned int uid)
1776 lra_push_insn (lra_insn_recog_data[uid]->insn);
1779 /* Take the last-inserted insns off the stack and return it. */
1780 rtx_insn *
1781 lra_pop_insn (void)
1783 rtx_insn *insn = lra_constraint_insn_stack.pop ();
1784 bitmap_clear_bit (lra_constraint_insn_stack_bitmap, INSN_UID (insn));
1785 return insn;
1788 /* Return the current size of the insn stack. */
1789 unsigned int
1790 lra_insn_stack_length (void)
1792 return lra_constraint_insn_stack.length ();
1795 /* Push insns FROM to TO (excluding it) going in reverse order. */
1796 static void
1797 push_insns (rtx_insn *from, rtx_insn *to)
1799 rtx_insn *insn;
1801 if (from == NULL_RTX)
1802 return;
1803 for (insn = from; insn != to; insn = PREV_INSN (insn))
1804 if (INSN_P (insn))
1805 lra_push_insn (insn);
1808 /* Set up sp offset for insn in range [FROM, LAST]. The offset is
1809 taken from the next BB insn after LAST or zero if there in such
1810 insn. */
1811 static void
1812 setup_sp_offset (rtx_insn *from, rtx_insn *last)
1814 rtx_insn *before = next_nonnote_insn_bb (last);
1815 HOST_WIDE_INT offset = (before == NULL_RTX || ! INSN_P (before)
1816 ? 0 : lra_get_insn_recog_data (before)->sp_offset);
1818 for (rtx_insn *insn = from; insn != NEXT_INSN (last); insn = NEXT_INSN (insn))
1819 lra_get_insn_recog_data (insn)->sp_offset = offset;
1822 /* Emit insns BEFORE before INSN and insns AFTER after INSN. Put the
1823 insns onto the stack. Print about emitting the insns with
1824 TITLE. */
1825 void
1826 lra_process_new_insns (rtx_insn *insn, rtx_insn *before, rtx_insn *after,
1827 const char *title)
1829 rtx_insn *last;
1831 if (before == NULL_RTX && after == NULL_RTX)
1832 return;
1833 if (lra_dump_file != NULL)
1835 dump_insn_slim (lra_dump_file, insn);
1836 if (before != NULL_RTX)
1838 fprintf (lra_dump_file," %s before:\n", title);
1839 dump_rtl_slim (lra_dump_file, before, NULL, -1, 0);
1841 if (after != NULL_RTX)
1843 fprintf (lra_dump_file, " %s after:\n", title);
1844 dump_rtl_slim (lra_dump_file, after, NULL, -1, 0);
1846 fprintf (lra_dump_file, "\n");
1848 if (before != NULL_RTX)
1850 if (cfun->can_throw_non_call_exceptions)
1851 copy_reg_eh_region_note_forward (insn, before, NULL);
1852 emit_insn_before (before, insn);
1853 push_insns (PREV_INSN (insn), PREV_INSN (before));
1854 setup_sp_offset (before, PREV_INSN (insn));
1856 if (after != NULL_RTX)
1858 if (cfun->can_throw_non_call_exceptions)
1859 copy_reg_eh_region_note_forward (insn, after, NULL);
1860 for (last = after; NEXT_INSN (last) != NULL_RTX; last = NEXT_INSN (last))
1862 emit_insn_after (after, insn);
1863 push_insns (last, insn);
1864 setup_sp_offset (after, last);
1866 if (cfun->can_throw_non_call_exceptions)
1868 rtx note = find_reg_note (insn, REG_EH_REGION, NULL_RTX);
1869 if (note && !insn_could_throw_p (insn))
1870 remove_note (insn, note);
1875 /* Replace all references to register OLD_REGNO in *LOC with pseudo
1876 register NEW_REG. Try to simplify subreg of constant if SUBREG_P.
1877 Return true if any change was made. */
1878 bool
1879 lra_substitute_pseudo (rtx *loc, int old_regno, rtx new_reg, bool subreg_p)
1881 rtx x = *loc;
1882 bool result = false;
1883 enum rtx_code code;
1884 const char *fmt;
1885 int i, j;
1887 if (x == NULL_RTX)
1888 return false;
1890 code = GET_CODE (x);
1891 if (code == SUBREG && subreg_p)
1893 rtx subst, inner = SUBREG_REG (x);
1894 /* Transform subreg of constant while we still have inner mode
1895 of the subreg. The subreg internal should not be an insn
1896 operand. */
1897 if (REG_P (inner) && (int) REGNO (inner) == old_regno
1898 && CONSTANT_P (new_reg)
1899 && (subst = simplify_subreg (GET_MODE (x), new_reg, GET_MODE (inner),
1900 SUBREG_BYTE (x))) != NULL_RTX)
1902 *loc = subst;
1903 return true;
1907 else if (code == REG && (int) REGNO (x) == old_regno)
1909 machine_mode mode = GET_MODE (x);
1910 machine_mode inner_mode = GET_MODE (new_reg);
1912 if (mode != inner_mode
1913 && ! (CONST_INT_P (new_reg) && SCALAR_INT_MODE_P (mode)))
1915 if (!partial_subreg_p (mode, inner_mode)
1916 || ! SCALAR_INT_MODE_P (inner_mode))
1917 new_reg = gen_rtx_SUBREG (mode, new_reg, 0);
1918 else
1919 new_reg = gen_lowpart_SUBREG (mode, new_reg);
1921 *loc = new_reg;
1922 return true;
1925 /* Scan all the operand sub-expressions. */
1926 fmt = GET_RTX_FORMAT (code);
1927 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1929 if (fmt[i] == 'e')
1931 if (lra_substitute_pseudo (&XEXP (x, i), old_regno,
1932 new_reg, subreg_p))
1933 result = true;
1935 else if (fmt[i] == 'E')
1937 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1938 if (lra_substitute_pseudo (&XVECEXP (x, i, j), old_regno,
1939 new_reg, subreg_p))
1940 result = true;
1943 return result;
1946 /* Call lra_substitute_pseudo within an insn. Try to simplify subreg
1947 of constant if SUBREG_P. This won't update the insn ptr, just the
1948 contents of the insn. */
1949 bool
1950 lra_substitute_pseudo_within_insn (rtx_insn *insn, int old_regno,
1951 rtx new_reg, bool subreg_p)
1953 rtx loc = insn;
1954 return lra_substitute_pseudo (&loc, old_regno, new_reg, subreg_p);
1959 /* This page contains code dealing with scratches (changing them onto
1960 pseudos and restoring them from the pseudos).
1962 We change scratches into pseudos at the beginning of LRA to
1963 simplify dealing with them (conflicts, hard register assignments).
1965 If the pseudo denoting scratch was spilled it means that we do need
1966 a hard register for it. Such pseudos are transformed back to
1967 scratches at the end of LRA. */
1969 /* Description of location of a former scratch operand. */
1970 struct sloc
1972 rtx_insn *insn; /* Insn where the scratch was. */
1973 int nop; /* Number of the operand which was a scratch. */
1976 typedef struct sloc *sloc_t;
1978 /* Locations of the former scratches. */
1979 static vec<sloc_t> scratches;
1981 /* Bitmap of scratch regnos. */
1982 static bitmap_head scratch_bitmap;
1984 /* Bitmap of scratch operands. */
1985 static bitmap_head scratch_operand_bitmap;
1987 /* Return true if pseudo REGNO is made of SCRATCH. */
1988 bool
1989 lra_former_scratch_p (int regno)
1991 return bitmap_bit_p (&scratch_bitmap, regno);
1994 /* Return true if the operand NOP of INSN is a former scratch. */
1995 bool
1996 lra_former_scratch_operand_p (rtx_insn *insn, int nop)
1998 return bitmap_bit_p (&scratch_operand_bitmap,
1999 INSN_UID (insn) * MAX_RECOG_OPERANDS + nop) != 0;
2002 /* Register operand NOP in INSN as a former scratch. It will be
2003 changed to scratch back, if it is necessary, at the LRA end. */
2004 void
2005 lra_register_new_scratch_op (rtx_insn *insn, int nop)
2007 lra_insn_recog_data_t id = lra_get_insn_recog_data (insn);
2008 rtx op = *id->operand_loc[nop];
2009 sloc_t loc = XNEW (struct sloc);
2010 lra_assert (REG_P (op));
2011 loc->insn = insn;
2012 loc->nop = nop;
2013 scratches.safe_push (loc);
2014 bitmap_set_bit (&scratch_bitmap, REGNO (op));
2015 bitmap_set_bit (&scratch_operand_bitmap,
2016 INSN_UID (insn) * MAX_RECOG_OPERANDS + nop);
2017 add_reg_note (insn, REG_UNUSED, op);
2020 /* Change scratches onto pseudos and save their location. */
2021 static void
2022 remove_scratches (void)
2024 int i;
2025 bool insn_changed_p;
2026 basic_block bb;
2027 rtx_insn *insn;
2028 rtx reg;
2029 lra_insn_recog_data_t id;
2030 struct lra_static_insn_data *static_id;
2032 scratches.create (get_max_uid ());
2033 bitmap_initialize (&scratch_bitmap, &reg_obstack);
2034 bitmap_initialize (&scratch_operand_bitmap, &reg_obstack);
2035 FOR_EACH_BB_FN (bb, cfun)
2036 FOR_BB_INSNS (bb, insn)
2037 if (INSN_P (insn))
2039 id = lra_get_insn_recog_data (insn);
2040 static_id = id->insn_static_data;
2041 insn_changed_p = false;
2042 for (i = 0; i < static_id->n_operands; i++)
2043 if (GET_CODE (*id->operand_loc[i]) == SCRATCH
2044 && GET_MODE (*id->operand_loc[i]) != VOIDmode)
2046 insn_changed_p = true;
2047 *id->operand_loc[i] = reg
2048 = lra_create_new_reg (static_id->operand[i].mode,
2049 *id->operand_loc[i], ALL_REGS, NULL);
2050 lra_register_new_scratch_op (insn, i);
2051 if (lra_dump_file != NULL)
2052 fprintf (lra_dump_file,
2053 "Removing SCRATCH in insn #%u (nop %d)\n",
2054 INSN_UID (insn), i);
2056 if (insn_changed_p)
2057 /* Because we might use DF right after caller-saves sub-pass
2058 we need to keep DF info up to date. */
2059 df_insn_rescan (insn);
2063 /* Changes pseudos created by function remove_scratches onto scratches. */
2064 static void
2065 restore_scratches (void)
2067 int regno;
2068 unsigned i;
2069 sloc_t loc;
2070 rtx_insn *last = NULL;
2071 lra_insn_recog_data_t id = NULL;
2073 for (i = 0; scratches.iterate (i, &loc); i++)
2075 /* Ignore already deleted insns. */
2076 if (NOTE_P (loc->insn)
2077 && NOTE_KIND (loc->insn) == NOTE_INSN_DELETED)
2078 continue;
2079 if (last != loc->insn)
2081 last = loc->insn;
2082 id = lra_get_insn_recog_data (last);
2084 if (REG_P (*id->operand_loc[loc->nop])
2085 && ((regno = REGNO (*id->operand_loc[loc->nop]))
2086 >= FIRST_PSEUDO_REGISTER)
2087 && lra_get_regno_hard_regno (regno) < 0)
2089 /* It should be only case when scratch register with chosen
2090 constraint 'X' did not get memory or hard register. */
2091 lra_assert (lra_former_scratch_p (regno));
2092 *id->operand_loc[loc->nop]
2093 = gen_rtx_SCRATCH (GET_MODE (*id->operand_loc[loc->nop]));
2094 lra_update_dup (id, loc->nop);
2095 if (lra_dump_file != NULL)
2096 fprintf (lra_dump_file, "Restoring SCRATCH in insn #%u(nop %d)\n",
2097 INSN_UID (loc->insn), loc->nop);
2100 for (i = 0; scratches.iterate (i, &loc); i++)
2101 free (loc);
2102 scratches.release ();
2103 bitmap_clear (&scratch_bitmap);
2104 bitmap_clear (&scratch_operand_bitmap);
2109 /* Function checks RTL for correctness. If FINAL_P is true, it is
2110 done at the end of LRA and the check is more rigorous. */
2111 static void
2112 check_rtl (bool final_p)
2114 basic_block bb;
2115 rtx_insn *insn;
2117 lra_assert (! final_p || reload_completed);
2118 FOR_EACH_BB_FN (bb, cfun)
2119 FOR_BB_INSNS (bb, insn)
2120 if (NONDEBUG_INSN_P (insn)
2121 && GET_CODE (PATTERN (insn)) != USE
2122 && GET_CODE (PATTERN (insn)) != CLOBBER
2123 && GET_CODE (PATTERN (insn)) != ASM_INPUT)
2125 if (final_p)
2127 extract_constrain_insn (insn);
2128 continue;
2130 /* LRA code is based on assumption that all addresses can be
2131 correctly decomposed. LRA can generate reloads for
2132 decomposable addresses. The decomposition code checks the
2133 correctness of the addresses. So we don't need to check
2134 the addresses here. Don't call insn_invalid_p here, it can
2135 change the code at this stage. */
2136 if (recog_memoized (insn) < 0 && asm_noperands (PATTERN (insn)) < 0)
2137 fatal_insn_not_found (insn);
2141 /* Determine if the current function has an exception receiver block
2142 that reaches the exit block via non-exceptional edges */
2143 static bool
2144 has_nonexceptional_receiver (void)
2146 edge e;
2147 edge_iterator ei;
2148 basic_block *tos, *worklist, bb;
2150 /* If we're not optimizing, then just err on the safe side. */
2151 if (!optimize)
2152 return true;
2154 /* First determine which blocks can reach exit via normal paths. */
2155 tos = worklist = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun) + 1);
2157 FOR_EACH_BB_FN (bb, cfun)
2158 bb->flags &= ~BB_REACHABLE;
2160 /* Place the exit block on our worklist. */
2161 EXIT_BLOCK_PTR_FOR_FN (cfun)->flags |= BB_REACHABLE;
2162 *tos++ = EXIT_BLOCK_PTR_FOR_FN (cfun);
2164 /* Iterate: find everything reachable from what we've already seen. */
2165 while (tos != worklist)
2167 bb = *--tos;
2169 FOR_EACH_EDGE (e, ei, bb->preds)
2170 if (e->flags & EDGE_ABNORMAL)
2172 free (worklist);
2173 return true;
2175 else
2177 basic_block src = e->src;
2179 if (!(src->flags & BB_REACHABLE))
2181 src->flags |= BB_REACHABLE;
2182 *tos++ = src;
2186 free (worklist);
2187 /* No exceptional block reached exit unexceptionally. */
2188 return false;
2192 /* Process recursively X of INSN and add REG_INC notes if necessary. */
2193 static void
2194 add_auto_inc_notes (rtx_insn *insn, rtx x)
2196 enum rtx_code code = GET_CODE (x);
2197 const char *fmt;
2198 int i, j;
2200 if (code == MEM && auto_inc_p (XEXP (x, 0)))
2202 add_reg_note (insn, REG_INC, XEXP (XEXP (x, 0), 0));
2203 return;
2206 /* Scan all X sub-expressions. */
2207 fmt = GET_RTX_FORMAT (code);
2208 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2210 if (fmt[i] == 'e')
2211 add_auto_inc_notes (insn, XEXP (x, i));
2212 else if (fmt[i] == 'E')
2213 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
2214 add_auto_inc_notes (insn, XVECEXP (x, i, j));
2219 /* Remove all REG_DEAD and REG_UNUSED notes and regenerate REG_INC.
2220 We change pseudos by hard registers without notification of DF and
2221 that can make the notes obsolete. DF-infrastructure does not deal
2222 with REG_INC notes -- so we should regenerate them here. */
2223 static void
2224 update_inc_notes (void)
2226 rtx *pnote;
2227 basic_block bb;
2228 rtx_insn *insn;
2230 FOR_EACH_BB_FN (bb, cfun)
2231 FOR_BB_INSNS (bb, insn)
2232 if (NONDEBUG_INSN_P (insn))
2234 pnote = &REG_NOTES (insn);
2235 while (*pnote != 0)
2237 if (REG_NOTE_KIND (*pnote) == REG_DEAD
2238 || REG_NOTE_KIND (*pnote) == REG_UNUSED
2239 || REG_NOTE_KIND (*pnote) == REG_INC)
2240 *pnote = XEXP (*pnote, 1);
2241 else
2242 pnote = &XEXP (*pnote, 1);
2245 if (AUTO_INC_DEC)
2246 add_auto_inc_notes (insn, PATTERN (insn));
2250 /* Set to 1 while in lra. */
2251 int lra_in_progress;
2253 /* Start of pseudo regnos before the LRA. */
2254 int lra_new_regno_start;
2256 /* Start of reload pseudo regnos before the new spill pass. */
2257 int lra_constraint_new_regno_start;
2259 /* Avoid spilling pseudos with regno more than the following value if
2260 it is possible. */
2261 int lra_bad_spill_regno_start;
2263 /* Inheritance pseudo regnos before the new spill pass. */
2264 bitmap_head lra_inheritance_pseudos;
2266 /* Split regnos before the new spill pass. */
2267 bitmap_head lra_split_regs;
2269 /* Reload pseudo regnos before the new assignment pass which still can
2270 be spilled after the assignment pass as memory is also accepted in
2271 insns for the reload pseudos. */
2272 bitmap_head lra_optional_reload_pseudos;
2274 /* Pseudo regnos used for subreg reloads before the new assignment
2275 pass. Such pseudos still can be spilled after the assignment
2276 pass. */
2277 bitmap_head lra_subreg_reload_pseudos;
2279 /* File used for output of LRA debug information. */
2280 FILE *lra_dump_file;
2282 /* True if we should try spill into registers of different classes
2283 instead of memory. */
2284 bool lra_reg_spill_p;
2286 /* Set up value LRA_REG_SPILL_P. */
2287 static void
2288 setup_reg_spill_flag (void)
2290 int cl, mode;
2292 if (targetm.spill_class != NULL)
2293 for (cl = 0; cl < (int) LIM_REG_CLASSES; cl++)
2294 for (mode = 0; mode < MAX_MACHINE_MODE; mode++)
2295 if (targetm.spill_class ((enum reg_class) cl,
2296 (machine_mode) mode) != NO_REGS)
2298 lra_reg_spill_p = true;
2299 return;
2301 lra_reg_spill_p = false;
2304 /* True if the current function is too big to use regular algorithms
2305 in LRA. In other words, we should use simpler and faster algorithms
2306 in LRA. It also means we should not worry about generation code
2307 for caller saves. The value is set up in IRA. */
2308 bool lra_simple_p;
2310 /* Major LRA entry function. F is a file should be used to dump LRA
2311 debug info. */
2312 void
2313 lra (FILE *f)
2315 int i;
2316 bool live_p, inserted_p;
2318 lra_dump_file = f;
2320 timevar_push (TV_LRA);
2322 /* Make sure that the last insn is a note. Some subsequent passes
2323 need it. */
2324 emit_note (NOTE_INSN_DELETED);
2326 COPY_HARD_REG_SET (lra_no_alloc_regs, ira_no_alloc_regs);
2328 init_reg_info ();
2329 expand_reg_info ();
2331 init_insn_recog_data ();
2333 /* Some quick check on RTL generated by previous passes. */
2334 if (flag_checking)
2335 check_rtl (false);
2337 lra_in_progress = 1;
2339 lra_live_range_iter = lra_coalesce_iter = lra_constraint_iter = 0;
2340 lra_assignment_iter = lra_assignment_iter_after_spill = 0;
2341 lra_inheritance_iter = lra_undo_inheritance_iter = 0;
2342 lra_rematerialization_iter = 0;
2344 setup_reg_spill_flag ();
2346 /* Function remove_scratches can creates new pseudos for clobbers --
2347 so set up lra_constraint_new_regno_start before its call to
2348 permit changing reg classes for pseudos created by this
2349 simplification. */
2350 lra_constraint_new_regno_start = lra_new_regno_start = max_reg_num ();
2351 lra_bad_spill_regno_start = INT_MAX;
2352 remove_scratches ();
2354 /* A function that has a non-local label that can reach the exit
2355 block via non-exceptional paths must save all call-saved
2356 registers. */
2357 if (cfun->has_nonlocal_label && has_nonexceptional_receiver ())
2358 crtl->saves_all_registers = 1;
2360 if (crtl->saves_all_registers)
2361 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2362 if (! call_used_regs[i] && ! fixed_regs[i] && ! LOCAL_REGNO (i))
2363 df_set_regs_ever_live (i, true);
2365 /* We don't DF from now and avoid its using because it is to
2366 expensive when a lot of RTL changes are made. */
2367 df_set_flags (DF_NO_INSN_RESCAN);
2368 lra_constraint_insn_stack.create (get_max_uid ());
2369 lra_constraint_insn_stack_bitmap = sbitmap_alloc (get_max_uid ());
2370 bitmap_clear (lra_constraint_insn_stack_bitmap);
2371 lra_live_ranges_init ();
2372 lra_constraints_init ();
2373 lra_curr_reload_num = 0;
2374 push_insns (get_last_insn (), NULL);
2375 /* It is needed for the 1st coalescing. */
2376 bitmap_initialize (&lra_inheritance_pseudos, &reg_obstack);
2377 bitmap_initialize (&lra_split_regs, &reg_obstack);
2378 bitmap_initialize (&lra_optional_reload_pseudos, &reg_obstack);
2379 bitmap_initialize (&lra_subreg_reload_pseudos, &reg_obstack);
2380 live_p = false;
2381 if (get_frame_size () != 0 && crtl->stack_alignment_needed)
2382 /* If we have a stack frame, we must align it now. The stack size
2383 may be a part of the offset computation for register
2384 elimination. */
2385 assign_stack_local (BLKmode, 0, crtl->stack_alignment_needed);
2386 lra_init_equiv ();
2387 for (;;)
2389 for (;;)
2391 bool reloads_p = lra_constraints (lra_constraint_iter == 0);
2392 /* Constraint transformations may result in that eliminable
2393 hard regs become uneliminable and pseudos which use them
2394 should be spilled. It is better to do it before pseudo
2395 assignments.
2397 For example, rs6000 can make
2398 RS6000_PIC_OFFSET_TABLE_REGNUM uneliminable if we started
2399 to use a constant pool. */
2400 lra_eliminate (false, false);
2401 /* We should try to assign hard registers to scratches even
2402 if there were no RTL transformations in lra_constraints.
2403 Also we should check IRA assignments on the first
2404 iteration as they can be wrong because of early clobbers
2405 operands which are ignored in IRA. */
2406 if (! reloads_p && lra_constraint_iter > 1)
2408 /* Stack is not empty here only when there are changes
2409 during the elimination sub-pass. */
2410 if (bitmap_empty_p (lra_constraint_insn_stack_bitmap))
2411 break;
2412 else
2413 /* If there are no reloads but changing due
2414 elimination, restart the constraint sub-pass
2415 first. */
2416 continue;
2418 /* Do inheritance only for regular algorithms. */
2419 if (! lra_simple_p)
2421 if (flag_ipa_ra)
2423 if (live_p)
2424 lra_clear_live_ranges ();
2425 /* As a side-effect of lra_create_live_ranges, we calculate
2426 actual_call_used_reg_set, which is needed during
2427 lra_inheritance. */
2428 lra_create_live_ranges (true, true);
2429 live_p = true;
2431 lra_inheritance ();
2433 if (live_p)
2434 lra_clear_live_ranges ();
2435 /* We need live ranges for lra_assign -- so build them. But
2436 don't remove dead insns or change global live info as we
2437 can undo inheritance transformations after inheritance
2438 pseudo assigning. */
2439 lra_create_live_ranges (true, false);
2440 live_p = true;
2441 /* If we don't spill non-reload and non-inheritance pseudos,
2442 there is no sense to run memory-memory move coalescing.
2443 If inheritance pseudos were spilled, the memory-memory
2444 moves involving them will be removed by pass undoing
2445 inheritance. */
2446 if (lra_simple_p)
2447 lra_assign ();
2448 else
2450 bool spill_p = !lra_assign ();
2452 if (lra_undo_inheritance ())
2453 live_p = false;
2454 if (spill_p)
2456 if (! live_p)
2458 lra_create_live_ranges (true, true);
2459 live_p = true;
2461 if (lra_coalesce ())
2462 live_p = false;
2464 if (! live_p)
2465 lra_clear_live_ranges ();
2468 /* Don't clear optional reloads bitmap until all constraints are
2469 satisfied as we need to differ them from regular reloads. */
2470 bitmap_clear (&lra_optional_reload_pseudos);
2471 bitmap_clear (&lra_subreg_reload_pseudos);
2472 bitmap_clear (&lra_inheritance_pseudos);
2473 bitmap_clear (&lra_split_regs);
2474 if (! live_p)
2476 /* We need full live info for spilling pseudos into
2477 registers instead of memory. */
2478 lra_create_live_ranges (lra_reg_spill_p, true);
2479 live_p = true;
2481 /* We should check necessity for spilling here as the above live
2482 range pass can remove spilled pseudos. */
2483 if (! lra_need_for_spills_p ())
2484 break;
2485 /* Now we know what pseudos should be spilled. Try to
2486 rematerialize them first. */
2487 if (lra_remat ())
2489 /* We need full live info -- see the comment above. */
2490 lra_create_live_ranges (lra_reg_spill_p, true);
2491 live_p = true;
2492 if (! lra_need_for_spills_p ())
2493 break;
2495 lra_spill ();
2496 /* Assignment of stack slots changes elimination offsets for
2497 some eliminations. So update the offsets here. */
2498 lra_eliminate (false, false);
2499 lra_constraint_new_regno_start = max_reg_num ();
2500 if (lra_bad_spill_regno_start == INT_MAX
2501 && lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES
2502 && lra_rematerialization_iter > LRA_MAX_REMATERIALIZATION_PASSES)
2503 /* After switching off inheritance and rematerialization
2504 passes, avoid spilling reload pseudos will be created to
2505 prevent LRA cycling in some complicated cases. */
2506 lra_bad_spill_regno_start = lra_constraint_new_regno_start;
2507 lra_assignment_iter_after_spill = 0;
2509 restore_scratches ();
2510 lra_eliminate (true, false);
2511 lra_final_code_change ();
2512 lra_in_progress = 0;
2513 if (live_p)
2514 lra_clear_live_ranges ();
2515 lra_live_ranges_finish ();
2516 lra_constraints_finish ();
2517 finish_reg_info ();
2518 sbitmap_free (lra_constraint_insn_stack_bitmap);
2519 lra_constraint_insn_stack.release ();
2520 finish_insn_recog_data ();
2521 regstat_free_n_sets_and_refs ();
2522 regstat_free_ri ();
2523 reload_completed = 1;
2524 update_inc_notes ();
2526 inserted_p = fixup_abnormal_edges ();
2528 /* We've possibly turned single trapping insn into multiple ones. */
2529 if (cfun->can_throw_non_call_exceptions)
2531 auto_sbitmap blocks (last_basic_block_for_fn (cfun));
2532 bitmap_ones (blocks);
2533 find_many_sub_basic_blocks (blocks);
2536 if (inserted_p)
2537 commit_edge_insertions ();
2539 /* Replacing pseudos with their memory equivalents might have
2540 created shared rtx. Subsequent passes would get confused
2541 by this, so unshare everything here. */
2542 unshare_all_rtl_again (get_insns ());
2544 if (flag_checking)
2545 check_rtl (true);
2547 timevar_pop (TV_LRA);
2550 /* Called once per compiler to initialize LRA data once. */
2551 void
2552 lra_init_once (void)
2554 init_insn_code_data_once ();
2557 /* Called once per compiler to finish LRA data which are initialize
2558 once. */
2559 void
2560 lra_finish_once (void)
2562 finish_insn_code_data_once ();