2018-11-11 Richard Biener <rguenther@suse.de>
[official-gcc.git] / gcc / lra.c
blob5d58d90f3a6b1fbcee9c78a60339db9da066beac
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. CLOBBER_HIGH marks if reference is a clobber
539 high. */
540 static struct lra_insn_reg *
541 new_insn_reg (rtx_insn *insn, int regno, enum op_type type,
542 machine_mode mode,
543 bool subreg_p, bool early_clobber,
544 alternative_mask early_clobber_alts,
545 struct lra_insn_reg *next, bool clobber_high)
547 lra_insn_reg *ir = lra_insn_reg_pool.allocate ();
548 ir->type = type;
549 ir->biggest_mode = mode;
550 if (NONDEBUG_INSN_P (insn)
551 && partial_subreg_p (lra_reg_info[regno].biggest_mode, mode))
552 lra_reg_info[regno].biggest_mode = mode;
553 ir->subreg_p = subreg_p;
554 ir->early_clobber = early_clobber;
555 ir->early_clobber_alts = early_clobber_alts;
556 ir->clobber_high = clobber_high;
557 ir->regno = regno;
558 ir->next = next;
559 return ir;
562 /* Free insn reg info list IR. */
563 static void
564 free_insn_regs (struct lra_insn_reg *ir)
566 struct lra_insn_reg *next_ir;
568 for (; ir != NULL; ir = next_ir)
570 next_ir = ir->next;
571 lra_insn_reg_pool.remove (ir);
575 /* Finish pool for insn reg info. */
576 static void
577 finish_insn_regs (void)
579 lra_insn_reg_pool.release ();
584 /* This page contains code dealing LRA insn info (or in other words
585 LRA internal insn representation). */
587 /* Map INSN_CODE -> the static insn data. This info is valid during
588 all translation unit. */
589 struct lra_static_insn_data *insn_code_data[NUM_INSN_CODES];
591 /* Debug insns are represented as a special insn with one input
592 operand which is RTL expression in var_location. */
594 /* The following data are used as static insn operand data for all
595 debug insns. If structure lra_operand_data is changed, the
596 initializer should be changed too. */
597 static struct lra_operand_data debug_operand_data =
599 NULL, /* alternative */
600 0, /* early_clobber_alts */
601 E_VOIDmode, /* We are not interesting in the operand mode. */
602 OP_IN,
603 0, 0, 0, 0
606 /* The following data are used as static insn data for all debug
607 bind insns. If structure lra_static_insn_data is changed, the
608 initializer should be changed too. */
609 static struct lra_static_insn_data debug_bind_static_data =
611 &debug_operand_data,
612 0, /* Duplication operands #. */
613 -1, /* Commutative operand #. */
614 1, /* Operands #. There is only one operand which is debug RTL
615 expression. */
616 0, /* Duplications #. */
617 0, /* Alternatives #. We are not interesting in alternatives
618 because we does not proceed debug_insns for reloads. */
619 NULL, /* Hard registers referenced in machine description. */
620 NULL /* Descriptions of operands in alternatives. */
623 /* The following data are used as static insn data for all debug
624 marker insns. If structure lra_static_insn_data is changed, the
625 initializer should be changed too. */
626 static struct lra_static_insn_data debug_marker_static_data =
628 &debug_operand_data,
629 0, /* Duplication operands #. */
630 -1, /* Commutative operand #. */
631 0, /* Operands #. There isn't any operand. */
632 0, /* Duplications #. */
633 0, /* Alternatives #. We are not interesting in alternatives
634 because we does not proceed debug_insns for reloads. */
635 NULL, /* Hard registers referenced in machine description. */
636 NULL /* Descriptions of operands in alternatives. */
639 /* Called once per compiler work to initialize some LRA data related
640 to insns. */
641 static void
642 init_insn_code_data_once (void)
644 memset (insn_code_data, 0, sizeof (insn_code_data));
647 /* Called once per compiler work to finalize some LRA data related to
648 insns. */
649 static void
650 finish_insn_code_data_once (void)
652 for (unsigned int i = 0; i < NUM_INSN_CODES; i++)
654 if (insn_code_data[i] != NULL)
655 free (insn_code_data[i]);
659 /* Return static insn data, allocate and setup if necessary. Although
660 dup_num is static data (it depends only on icode), to set it up we
661 need to extract insn first. So recog_data should be valid for
662 normal insn (ICODE >= 0) before the call. */
663 static struct lra_static_insn_data *
664 get_static_insn_data (int icode, int nop, int ndup, int nalt)
666 struct lra_static_insn_data *data;
667 size_t n_bytes;
669 lra_assert (icode < (int) NUM_INSN_CODES);
670 if (icode >= 0 && (data = insn_code_data[icode]) != NULL)
671 return data;
672 lra_assert (nop >= 0 && ndup >= 0 && nalt >= 0);
673 n_bytes = sizeof (struct lra_static_insn_data)
674 + sizeof (struct lra_operand_data) * nop
675 + sizeof (int) * ndup;
676 data = XNEWVAR (struct lra_static_insn_data, n_bytes);
677 data->operand_alternative = NULL;
678 data->n_operands = nop;
679 data->n_dups = ndup;
680 data->n_alternatives = nalt;
681 data->operand = ((struct lra_operand_data *)
682 ((char *) data + sizeof (struct lra_static_insn_data)));
683 data->dup_num = ((int *) ((char *) data->operand
684 + sizeof (struct lra_operand_data) * nop));
685 if (icode >= 0)
687 int i;
689 insn_code_data[icode] = data;
690 for (i = 0; i < nop; i++)
692 data->operand[i].constraint
693 = insn_data[icode].operand[i].constraint;
694 data->operand[i].mode = insn_data[icode].operand[i].mode;
695 data->operand[i].strict_low = insn_data[icode].operand[i].strict_low;
696 data->operand[i].is_operator
697 = insn_data[icode].operand[i].is_operator;
698 data->operand[i].type
699 = (data->operand[i].constraint[0] == '=' ? OP_OUT
700 : data->operand[i].constraint[0] == '+' ? OP_INOUT
701 : OP_IN);
702 data->operand[i].is_address = false;
704 for (i = 0; i < ndup; i++)
705 data->dup_num[i] = recog_data.dup_num[i];
707 return data;
710 /* The current length of the following array. */
711 int lra_insn_recog_data_len;
713 /* Map INSN_UID -> the insn recog data (NULL if unknown). */
714 lra_insn_recog_data_t *lra_insn_recog_data;
716 /* Initialize LRA data about insns. */
717 static void
718 init_insn_recog_data (void)
720 lra_insn_recog_data_len = 0;
721 lra_insn_recog_data = NULL;
724 /* Expand, if necessary, LRA data about insns. */
725 static void
726 check_and_expand_insn_recog_data (int index)
728 int i, old;
730 if (lra_insn_recog_data_len > index)
731 return;
732 old = lra_insn_recog_data_len;
733 lra_insn_recog_data_len = index * 3 / 2 + 1;
734 lra_insn_recog_data = XRESIZEVEC (lra_insn_recog_data_t,
735 lra_insn_recog_data,
736 lra_insn_recog_data_len);
737 for (i = old; i < lra_insn_recog_data_len; i++)
738 lra_insn_recog_data[i] = NULL;
741 /* Finish LRA DATA about insn. */
742 static void
743 free_insn_recog_data (lra_insn_recog_data_t data)
745 if (data->operand_loc != NULL)
746 free (data->operand_loc);
747 if (data->dup_loc != NULL)
748 free (data->dup_loc);
749 if (data->arg_hard_regs != NULL)
750 free (data->arg_hard_regs);
751 if (data->icode < 0 && NONDEBUG_INSN_P (data->insn))
753 if (data->insn_static_data->operand_alternative != NULL)
754 free (const_cast <operand_alternative *>
755 (data->insn_static_data->operand_alternative));
756 free_insn_regs (data->insn_static_data->hard_regs);
757 free (data->insn_static_data);
759 free_insn_regs (data->regs);
760 data->regs = NULL;
761 free (data);
764 /* Pools for copies. */
765 static object_allocator<lra_copy> lra_copy_pool ("lra copies");
767 /* Finish LRA data about all insns. */
768 static void
769 finish_insn_recog_data (void)
771 int i;
772 lra_insn_recog_data_t data;
774 for (i = 0; i < lra_insn_recog_data_len; i++)
775 if ((data = lra_insn_recog_data[i]) != NULL)
776 free_insn_recog_data (data);
777 finish_insn_regs ();
778 lra_copy_pool.release ();
779 lra_insn_reg_pool.release ();
780 free (lra_insn_recog_data);
783 /* Setup info about operands in alternatives of LRA DATA of insn. */
784 static void
785 setup_operand_alternative (lra_insn_recog_data_t data,
786 const operand_alternative *op_alt)
788 int i, j, nop, nalt;
789 int icode = data->icode;
790 struct lra_static_insn_data *static_data = data->insn_static_data;
792 static_data->commutative = -1;
793 nop = static_data->n_operands;
794 nalt = static_data->n_alternatives;
795 static_data->operand_alternative = op_alt;
796 for (i = 0; i < nop; i++)
798 static_data->operand[i].early_clobber_alts = 0;
799 static_data->operand[i].early_clobber = false;
800 static_data->operand[i].is_address = false;
801 if (static_data->operand[i].constraint[0] == '%')
803 /* We currently only support one commutative pair of operands. */
804 if (static_data->commutative < 0)
805 static_data->commutative = i;
806 else
807 lra_assert (icode < 0); /* Asm */
808 /* The last operand should not be marked commutative. */
809 lra_assert (i != nop - 1);
812 for (j = 0; j < nalt; j++)
813 for (i = 0; i < nop; i++, op_alt++)
815 static_data->operand[i].early_clobber |= op_alt->earlyclobber;
816 if (op_alt->earlyclobber)
817 static_data->operand[i].early_clobber_alts |= (alternative_mask) 1 << j;
818 static_data->operand[i].is_address |= op_alt->is_address;
822 /* Recursively process X and collect info about registers, which are
823 not the insn operands, in X with TYPE (in/out/inout) and flag that
824 it is early clobbered in the insn (EARLY_CLOBBER) and add the info
825 to LIST. X is a part of insn given by DATA. Return the result
826 list. CLOBBER_HIGH marks if X is a clobber high. */
827 static struct lra_insn_reg *
828 collect_non_operand_hard_regs (rtx_insn *insn, rtx *x,
829 lra_insn_recog_data_t data,
830 struct lra_insn_reg *list,
831 enum op_type type, bool early_clobber,
832 bool clobber_high)
834 int i, j, regno, last;
835 bool subreg_p;
836 machine_mode mode;
837 struct lra_insn_reg *curr;
838 rtx op = *x;
839 enum rtx_code code = GET_CODE (op);
840 const char *fmt = GET_RTX_FORMAT (code);
842 for (i = 0; i < data->insn_static_data->n_operands; i++)
843 if (! data->insn_static_data->operand[i].is_operator
844 && x == data->operand_loc[i])
845 /* It is an operand loc. Stop here. */
846 return list;
847 for (i = 0; i < data->insn_static_data->n_dups; i++)
848 if (x == data->dup_loc[i])
849 /* It is a dup loc. Stop here. */
850 return list;
851 mode = GET_MODE (op);
852 subreg_p = false;
853 if (code == SUBREG)
855 mode = wider_subreg_mode (op);
856 if (read_modify_subreg_p (op))
857 subreg_p = true;
858 op = SUBREG_REG (op);
859 code = GET_CODE (op);
861 if (REG_P (op))
863 if ((regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER)
864 return list;
865 /* Process all regs even unallocatable ones as we need info
866 about all regs for rematerialization pass. */
867 for (last = end_hard_regno (mode, regno); regno < last; regno++)
869 for (curr = list; curr != NULL; curr = curr->next)
870 if (curr->regno == regno && curr->subreg_p == subreg_p
871 && curr->biggest_mode == mode)
873 if (curr->type != type)
874 curr->type = OP_INOUT;
875 if (early_clobber)
877 curr->early_clobber = true;
878 curr->early_clobber_alts = ALL_ALTERNATIVES;
880 break;
882 if (curr == NULL)
884 /* This is a new hard regno or the info can not be
885 integrated into the found structure. */
886 #ifdef STACK_REGS
887 early_clobber
888 = (early_clobber
889 /* This clobber is to inform popping floating
890 point stack only. */
891 && ! (FIRST_STACK_REG <= regno
892 && regno <= LAST_STACK_REG));
893 #endif
894 list = new_insn_reg (data->insn, regno, type, mode, subreg_p,
895 early_clobber,
896 early_clobber ? ALL_ALTERNATIVES : 0, list,
897 clobber_high);
900 return list;
902 switch (code)
904 case SET:
905 list = collect_non_operand_hard_regs (insn, &SET_DEST (op), data,
906 list, OP_OUT, false, false);
907 list = collect_non_operand_hard_regs (insn, &SET_SRC (op), data,
908 list, OP_IN, false, false);
909 break;
910 case CLOBBER:
911 /* We treat clobber of non-operand hard registers as early clobber. */
912 list = collect_non_operand_hard_regs (insn, &XEXP (op, 0), data,
913 list, OP_OUT, true, false);
914 break;
915 case CLOBBER_HIGH:
916 /* Clobber high should always span exactly one register. */
917 gcc_assert (REG_NREGS (XEXP (op, 0)) == 1);
918 /* We treat clobber of non-operand hard registers as early clobber. */
919 list = collect_non_operand_hard_regs (insn, &XEXP (op, 0), data,
920 list, OP_OUT, true, true);
921 break;
922 case PRE_INC: case PRE_DEC: case POST_INC: case POST_DEC:
923 list = collect_non_operand_hard_regs (insn, &XEXP (op, 0), data,
924 list, OP_INOUT, false, false);
925 break;
926 case PRE_MODIFY: case POST_MODIFY:
927 list = collect_non_operand_hard_regs (insn, &XEXP (op, 0), data,
928 list, OP_INOUT, false, false);
929 list = collect_non_operand_hard_regs (insn, &XEXP (op, 1), data,
930 list, OP_IN, false, false);
931 break;
932 default:
933 fmt = GET_RTX_FORMAT (code);
934 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
936 if (fmt[i] == 'e')
937 list = collect_non_operand_hard_regs (insn, &XEXP (op, i), data,
938 list, OP_IN, false, false);
939 else if (fmt[i] == 'E')
940 for (j = XVECLEN (op, i) - 1; j >= 0; j--)
941 list = collect_non_operand_hard_regs (insn, &XVECEXP (op, i, j),
942 data, list, OP_IN, false,
943 false);
946 return list;
949 /* Set up and return info about INSN. Set up the info if it is not set up
950 yet. */
951 lra_insn_recog_data_t
952 lra_set_insn_recog_data (rtx_insn *insn)
954 lra_insn_recog_data_t data;
955 int i, n, icode;
956 rtx **locs;
957 unsigned int uid = INSN_UID (insn);
958 struct lra_static_insn_data *insn_static_data;
960 check_and_expand_insn_recog_data (uid);
961 if (DEBUG_INSN_P (insn))
962 icode = -1;
963 else
965 icode = INSN_CODE (insn);
966 if (icode < 0)
967 /* It might be a new simple insn which is not recognized yet. */
968 INSN_CODE (insn) = icode = recog_memoized (insn);
970 data = XNEW (struct lra_insn_recog_data);
971 lra_insn_recog_data[uid] = data;
972 data->insn = insn;
973 data->used_insn_alternative = LRA_UNKNOWN_ALT;
974 data->icode = icode;
975 data->regs = NULL;
976 if (DEBUG_INSN_P (insn))
978 data->dup_loc = NULL;
979 data->arg_hard_regs = NULL;
980 data->preferred_alternatives = ALL_ALTERNATIVES;
981 if (DEBUG_BIND_INSN_P (insn))
983 data->insn_static_data = &debug_bind_static_data;
984 data->operand_loc = XNEWVEC (rtx *, 1);
985 data->operand_loc[0] = &INSN_VAR_LOCATION_LOC (insn);
987 else if (DEBUG_MARKER_INSN_P (insn))
989 data->insn_static_data = &debug_marker_static_data;
990 data->operand_loc = NULL;
992 return data;
994 if (icode < 0)
996 int nop, nalt;
997 machine_mode operand_mode[MAX_RECOG_OPERANDS];
998 const char *constraints[MAX_RECOG_OPERANDS];
1000 nop = asm_noperands (PATTERN (insn));
1001 data->operand_loc = data->dup_loc = NULL;
1002 nalt = 1;
1003 if (nop < 0)
1005 /* It is a special insn like USE or CLOBBER. We should
1006 recognize any regular insn otherwise LRA can do nothing
1007 with this insn. */
1008 gcc_assert (GET_CODE (PATTERN (insn)) == USE
1009 || GET_CODE (PATTERN (insn)) == CLOBBER
1010 || GET_CODE (PATTERN (insn)) == ASM_INPUT);
1011 data->insn_static_data = insn_static_data
1012 = get_static_insn_data (-1, 0, 0, nalt);
1014 else
1016 /* expand_asm_operands makes sure there aren't too many
1017 operands. */
1018 lra_assert (nop <= MAX_RECOG_OPERANDS);
1019 if (nop != 0)
1020 data->operand_loc = XNEWVEC (rtx *, nop);
1021 /* Now get the operand values and constraints out of the
1022 insn. */
1023 decode_asm_operands (PATTERN (insn), NULL,
1024 data->operand_loc,
1025 constraints, operand_mode, NULL);
1026 if (nop > 0)
1028 const char *p = recog_data.constraints[0];
1030 for (p = constraints[0]; *p; p++)
1031 nalt += *p == ',';
1033 data->insn_static_data = insn_static_data
1034 = get_static_insn_data (-1, nop, 0, nalt);
1035 for (i = 0; i < nop; i++)
1037 insn_static_data->operand[i].mode = operand_mode[i];
1038 insn_static_data->operand[i].constraint = constraints[i];
1039 insn_static_data->operand[i].strict_low = false;
1040 insn_static_data->operand[i].is_operator = false;
1041 insn_static_data->operand[i].is_address = false;
1044 for (i = 0; i < insn_static_data->n_operands; i++)
1045 insn_static_data->operand[i].type
1046 = (insn_static_data->operand[i].constraint[0] == '=' ? OP_OUT
1047 : insn_static_data->operand[i].constraint[0] == '+' ? OP_INOUT
1048 : OP_IN);
1049 data->preferred_alternatives = ALL_ALTERNATIVES;
1050 if (nop > 0)
1052 operand_alternative *op_alt = XCNEWVEC (operand_alternative,
1053 nalt * nop);
1054 preprocess_constraints (nop, nalt, constraints, op_alt,
1055 data->operand_loc);
1056 setup_operand_alternative (data, op_alt);
1059 else
1061 insn_extract (insn);
1062 data->insn_static_data = insn_static_data
1063 = get_static_insn_data (icode, insn_data[icode].n_operands,
1064 insn_data[icode].n_dups,
1065 insn_data[icode].n_alternatives);
1066 n = insn_static_data->n_operands;
1067 if (n == 0)
1068 locs = NULL;
1069 else
1071 locs = XNEWVEC (rtx *, n);
1072 memcpy (locs, recog_data.operand_loc, n * sizeof (rtx *));
1074 data->operand_loc = locs;
1075 n = insn_static_data->n_dups;
1076 if (n == 0)
1077 locs = NULL;
1078 else
1080 locs = XNEWVEC (rtx *, n);
1081 memcpy (locs, recog_data.dup_loc, n * sizeof (rtx *));
1083 data->dup_loc = locs;
1084 data->preferred_alternatives = get_preferred_alternatives (insn);
1085 const operand_alternative *op_alt = preprocess_insn_constraints (icode);
1086 if (!insn_static_data->operand_alternative)
1087 setup_operand_alternative (data, op_alt);
1088 else if (op_alt != insn_static_data->operand_alternative)
1089 insn_static_data->operand_alternative = op_alt;
1091 if (GET_CODE (PATTERN (insn)) == CLOBBER || GET_CODE (PATTERN (insn)) == USE)
1092 insn_static_data->hard_regs = NULL;
1093 else
1094 insn_static_data->hard_regs
1095 = collect_non_operand_hard_regs (insn, &PATTERN (insn), data,
1096 NULL, OP_IN, false, false);
1097 data->arg_hard_regs = NULL;
1098 if (CALL_P (insn))
1100 bool use_p;
1101 rtx link;
1102 int n_hard_regs, regno, arg_hard_regs[FIRST_PSEUDO_REGISTER];
1104 n_hard_regs = 0;
1105 /* Finding implicit hard register usage. We believe it will be
1106 not changed whatever transformations are used. Call insns
1107 are such example. */
1108 for (link = CALL_INSN_FUNCTION_USAGE (insn);
1109 link != NULL_RTX;
1110 link = XEXP (link, 1))
1111 if (((use_p = GET_CODE (XEXP (link, 0)) == USE)
1112 || GET_CODE (XEXP (link, 0)) == CLOBBER)
1113 && REG_P (XEXP (XEXP (link, 0), 0)))
1115 regno = REGNO (XEXP (XEXP (link, 0), 0));
1116 lra_assert (regno < FIRST_PSEUDO_REGISTER);
1117 /* It is an argument register. */
1118 for (i = REG_NREGS (XEXP (XEXP (link, 0), 0)) - 1; i >= 0; i--)
1119 arg_hard_regs[n_hard_regs++]
1120 = regno + i + (use_p ? 0 : FIRST_PSEUDO_REGISTER);
1122 else if (GET_CODE (XEXP (link, 0)) == CLOBBER_HIGH)
1123 /* We could support CLOBBER_HIGH and treat it in the same way as
1124 HARD_REGNO_CALL_PART_CLOBBERED, but no port needs that yet. */
1125 gcc_unreachable ();
1127 if (n_hard_regs != 0)
1129 arg_hard_regs[n_hard_regs++] = -1;
1130 data->arg_hard_regs = XNEWVEC (int, n_hard_regs);
1131 memcpy (data->arg_hard_regs, arg_hard_regs,
1132 sizeof (int) * n_hard_regs);
1135 /* Some output operand can be recognized only from the context not
1136 from the constraints which are empty in this case. Call insn may
1137 contain a hard register in set destination with empty constraint
1138 and extract_insn treats them as an input. */
1139 for (i = 0; i < insn_static_data->n_operands; i++)
1141 int j;
1142 rtx pat, set;
1143 struct lra_operand_data *operand = &insn_static_data->operand[i];
1145 /* ??? Should we treat 'X' the same way. It looks to me that
1146 'X' means anything and empty constraint means we do not
1147 care. */
1148 if (operand->type != OP_IN || *operand->constraint != '\0'
1149 || operand->is_operator)
1150 continue;
1151 pat = PATTERN (insn);
1152 if (GET_CODE (pat) == SET)
1154 if (data->operand_loc[i] != &SET_DEST (pat))
1155 continue;
1157 else if (GET_CODE (pat) == PARALLEL)
1159 for (j = XVECLEN (pat, 0) - 1; j >= 0; j--)
1161 set = XVECEXP (PATTERN (insn), 0, j);
1162 if (GET_CODE (set) == SET
1163 && &SET_DEST (set) == data->operand_loc[i])
1164 break;
1166 if (j < 0)
1167 continue;
1169 else
1170 continue;
1171 operand->type = OP_OUT;
1173 return data;
1176 /* Return info about insn give by UID. The info should be already set
1177 up. */
1178 static lra_insn_recog_data_t
1179 get_insn_recog_data_by_uid (int uid)
1181 lra_insn_recog_data_t data;
1183 data = lra_insn_recog_data[uid];
1184 lra_assert (data != NULL);
1185 return data;
1188 /* Invalidate all info about insn given by its UID. */
1189 static void
1190 invalidate_insn_recog_data (int uid)
1192 lra_insn_recog_data_t data;
1194 data = lra_insn_recog_data[uid];
1195 lra_assert (data != NULL);
1196 free_insn_recog_data (data);
1197 lra_insn_recog_data[uid] = NULL;
1200 /* Update all the insn info about INSN. It is usually called when
1201 something in the insn was changed. Return the updated info. */
1202 lra_insn_recog_data_t
1203 lra_update_insn_recog_data (rtx_insn *insn)
1205 lra_insn_recog_data_t data;
1206 int n;
1207 unsigned int uid = INSN_UID (insn);
1208 struct lra_static_insn_data *insn_static_data;
1209 poly_int64 sp_offset = 0;
1211 check_and_expand_insn_recog_data (uid);
1212 if ((data = lra_insn_recog_data[uid]) != NULL
1213 && data->icode != INSN_CODE (insn))
1215 sp_offset = data->sp_offset;
1216 invalidate_insn_data_regno_info (data, insn, get_insn_freq (insn));
1217 invalidate_insn_recog_data (uid);
1218 data = NULL;
1220 if (data == NULL)
1222 data = lra_get_insn_recog_data (insn);
1223 /* Initiate or restore SP offset. */
1224 data->sp_offset = sp_offset;
1225 return data;
1227 insn_static_data = data->insn_static_data;
1228 data->used_insn_alternative = LRA_UNKNOWN_ALT;
1229 if (DEBUG_INSN_P (insn))
1230 return data;
1231 if (data->icode < 0)
1233 int nop;
1234 machine_mode operand_mode[MAX_RECOG_OPERANDS];
1235 const char *constraints[MAX_RECOG_OPERANDS];
1237 nop = asm_noperands (PATTERN (insn));
1238 if (nop >= 0)
1240 lra_assert (nop == data->insn_static_data->n_operands);
1241 /* Now get the operand values and constraints out of the
1242 insn. */
1243 decode_asm_operands (PATTERN (insn), NULL,
1244 data->operand_loc,
1245 constraints, operand_mode, NULL);
1247 if (flag_checking)
1248 for (int i = 0; i < nop; i++)
1249 lra_assert
1250 (insn_static_data->operand[i].mode == operand_mode[i]
1251 && insn_static_data->operand[i].constraint == constraints[i]
1252 && ! insn_static_data->operand[i].is_operator);
1255 if (flag_checking)
1256 for (int i = 0; i < insn_static_data->n_operands; i++)
1257 lra_assert
1258 (insn_static_data->operand[i].type
1259 == (insn_static_data->operand[i].constraint[0] == '=' ? OP_OUT
1260 : insn_static_data->operand[i].constraint[0] == '+' ? OP_INOUT
1261 : OP_IN));
1263 else
1265 insn_extract (insn);
1266 n = insn_static_data->n_operands;
1267 if (n != 0)
1268 memcpy (data->operand_loc, recog_data.operand_loc, n * sizeof (rtx *));
1269 n = insn_static_data->n_dups;
1270 if (n != 0)
1271 memcpy (data->dup_loc, recog_data.dup_loc, n * sizeof (rtx *));
1272 lra_assert (check_bool_attrs (insn));
1274 return data;
1277 /* Set up that INSN is using alternative ALT now. */
1278 void
1279 lra_set_used_insn_alternative (rtx_insn *insn, int alt)
1281 lra_insn_recog_data_t data;
1283 data = lra_get_insn_recog_data (insn);
1284 data->used_insn_alternative = alt;
1287 /* Set up that insn with UID is using alternative ALT now. The insn
1288 info should be already set up. */
1289 void
1290 lra_set_used_insn_alternative_by_uid (int uid, int alt)
1292 lra_insn_recog_data_t data;
1294 check_and_expand_insn_recog_data (uid);
1295 data = lra_insn_recog_data[uid];
1296 lra_assert (data != NULL);
1297 data->used_insn_alternative = alt;
1302 /* This page contains code dealing with common register info and
1303 pseudo copies. */
1305 /* The size of the following array. */
1306 static int reg_info_size;
1307 /* Common info about each register. */
1308 struct lra_reg *lra_reg_info;
1310 HARD_REG_SET hard_regs_spilled_into;
1312 /* Last register value. */
1313 static int last_reg_value;
1315 /* Return new register value. */
1316 static int
1317 get_new_reg_value (void)
1319 return ++last_reg_value;
1322 /* Vec referring to pseudo copies. */
1323 static vec<lra_copy_t> copy_vec;
1325 /* Initialize I-th element of lra_reg_info. */
1326 static inline void
1327 initialize_lra_reg_info_element (int i)
1329 bitmap_initialize (&lra_reg_info[i].insn_bitmap, &reg_obstack);
1330 #ifdef STACK_REGS
1331 lra_reg_info[i].no_stack_p = false;
1332 #endif
1333 CLEAR_HARD_REG_SET (lra_reg_info[i].conflict_hard_regs);
1334 CLEAR_HARD_REG_SET (lra_reg_info[i].actual_call_used_reg_set);
1335 lra_reg_info[i].preferred_hard_regno1 = -1;
1336 lra_reg_info[i].preferred_hard_regno2 = -1;
1337 lra_reg_info[i].preferred_hard_regno_profit1 = 0;
1338 lra_reg_info[i].preferred_hard_regno_profit2 = 0;
1339 lra_reg_info[i].biggest_mode = VOIDmode;
1340 lra_reg_info[i].live_ranges = NULL;
1341 lra_reg_info[i].nrefs = lra_reg_info[i].freq = 0;
1342 lra_reg_info[i].last_reload = 0;
1343 lra_reg_info[i].restore_rtx = NULL_RTX;
1344 lra_reg_info[i].val = get_new_reg_value ();
1345 lra_reg_info[i].offset = 0;
1346 lra_reg_info[i].copies = NULL;
1349 /* Initialize common reg info and copies. */
1350 static void
1351 init_reg_info (void)
1353 int i;
1355 last_reg_value = 0;
1356 reg_info_size = max_reg_num () * 3 / 2 + 1;
1357 lra_reg_info = XNEWVEC (struct lra_reg, reg_info_size);
1358 for (i = 0; i < reg_info_size; i++)
1359 initialize_lra_reg_info_element (i);
1360 copy_vec.truncate (0);
1361 CLEAR_HARD_REG_SET (hard_regs_spilled_into);
1365 /* Finish common reg info and copies. */
1366 static void
1367 finish_reg_info (void)
1369 int i;
1371 for (i = 0; i < reg_info_size; i++)
1372 bitmap_clear (&lra_reg_info[i].insn_bitmap);
1373 free (lra_reg_info);
1374 reg_info_size = 0;
1377 /* Expand common reg info if it is necessary. */
1378 static void
1379 expand_reg_info (void)
1381 int i, old = reg_info_size;
1383 if (reg_info_size > max_reg_num ())
1384 return;
1385 reg_info_size = max_reg_num () * 3 / 2 + 1;
1386 lra_reg_info = XRESIZEVEC (struct lra_reg, lra_reg_info, reg_info_size);
1387 for (i = old; i < reg_info_size; i++)
1388 initialize_lra_reg_info_element (i);
1391 /* Free all copies. */
1392 void
1393 lra_free_copies (void)
1395 lra_copy_t cp;
1397 while (copy_vec.length () != 0)
1399 cp = copy_vec.pop ();
1400 lra_reg_info[cp->regno1].copies = lra_reg_info[cp->regno2].copies = NULL;
1401 lra_copy_pool.remove (cp);
1405 /* Create copy of two pseudos REGNO1 and REGNO2. The copy execution
1406 frequency is FREQ. */
1407 void
1408 lra_create_copy (int regno1, int regno2, int freq)
1410 bool regno1_dest_p;
1411 lra_copy_t cp;
1413 lra_assert (regno1 != regno2);
1414 regno1_dest_p = true;
1415 if (regno1 > regno2)
1417 std::swap (regno1, regno2);
1418 regno1_dest_p = false;
1420 cp = lra_copy_pool.allocate ();
1421 copy_vec.safe_push (cp);
1422 cp->regno1_dest_p = regno1_dest_p;
1423 cp->freq = freq;
1424 cp->regno1 = regno1;
1425 cp->regno2 = regno2;
1426 cp->regno1_next = lra_reg_info[regno1].copies;
1427 lra_reg_info[regno1].copies = cp;
1428 cp->regno2_next = lra_reg_info[regno2].copies;
1429 lra_reg_info[regno2].copies = cp;
1430 if (lra_dump_file != NULL)
1431 fprintf (lra_dump_file, " Creating copy r%d%sr%d@%d\n",
1432 regno1, regno1_dest_p ? "<-" : "->", regno2, freq);
1435 /* Return N-th (0, 1, ...) copy. If there is no copy, return
1436 NULL. */
1437 lra_copy_t
1438 lra_get_copy (int n)
1440 if (n >= (int) copy_vec.length ())
1441 return NULL;
1442 return copy_vec[n];
1447 /* This page contains code dealing with info about registers in
1448 insns. */
1450 /* Process X of INSN recursively and add info (operand type is
1451 given by TYPE, flag of that it is early clobber is EARLY_CLOBBER)
1452 about registers in X to the insn DATA. If X can be early clobbered,
1453 alternatives in which it can be early clobbered are given by
1454 EARLY_CLOBBER_ALTS. */
1455 static void
1456 add_regs_to_insn_regno_info (lra_insn_recog_data_t data, rtx x,
1457 rtx_insn *insn,
1458 enum op_type type, bool early_clobber,
1459 alternative_mask early_clobber_alts)
1461 int i, j, regno;
1462 bool subreg_p;
1463 machine_mode mode;
1464 const char *fmt;
1465 enum rtx_code code;
1466 struct lra_insn_reg *curr;
1468 code = GET_CODE (x);
1469 mode = GET_MODE (x);
1470 subreg_p = false;
1471 if (GET_CODE (x) == SUBREG)
1473 mode = wider_subreg_mode (x);
1474 if (read_modify_subreg_p (x))
1475 subreg_p = true;
1476 x = SUBREG_REG (x);
1477 code = GET_CODE (x);
1479 if (REG_P (x))
1481 regno = REGNO (x);
1482 /* Process all regs even unallocatable ones as we need info about
1483 all regs for rematerialization pass. */
1484 expand_reg_info ();
1485 if (bitmap_set_bit (&lra_reg_info[regno].insn_bitmap, INSN_UID (insn)))
1487 data->regs = new_insn_reg (data->insn, regno, type, mode, subreg_p,
1488 early_clobber, early_clobber_alts,
1489 data->regs, false);
1490 return;
1492 else
1494 for (curr = data->regs; curr != NULL; curr = curr->next)
1495 if (curr->regno == regno)
1497 if (curr->subreg_p != subreg_p || curr->biggest_mode != mode)
1498 /* The info can not be integrated into the found
1499 structure. */
1500 data->regs = new_insn_reg (data->insn, regno, type, mode,
1501 subreg_p, early_clobber,
1502 early_clobber_alts, data->regs,
1503 false);
1504 else
1506 if (curr->type != type)
1507 curr->type = OP_INOUT;
1508 if (curr->early_clobber != early_clobber)
1509 curr->early_clobber = true;
1510 curr->early_clobber_alts |= early_clobber_alts;
1512 return;
1514 gcc_unreachable ();
1518 switch (code)
1520 case SET:
1521 add_regs_to_insn_regno_info (data, SET_DEST (x), insn, OP_OUT, false, 0);
1522 add_regs_to_insn_regno_info (data, SET_SRC (x), insn, OP_IN, false, 0);
1523 break;
1524 case CLOBBER:
1525 /* We treat clobber of non-operand hard registers as early
1526 clobber. */
1527 add_regs_to_insn_regno_info (data, XEXP (x, 0), insn, OP_OUT,
1528 true, ALL_ALTERNATIVES);
1529 break;
1530 case CLOBBER_HIGH:
1531 gcc_unreachable ();
1532 case PRE_INC: case PRE_DEC: case POST_INC: case POST_DEC:
1533 add_regs_to_insn_regno_info (data, XEXP (x, 0), insn, OP_INOUT, false, 0);
1534 break;
1535 case PRE_MODIFY: case POST_MODIFY:
1536 add_regs_to_insn_regno_info (data, XEXP (x, 0), insn, OP_INOUT, false, 0);
1537 add_regs_to_insn_regno_info (data, XEXP (x, 1), insn, OP_IN, false, 0);
1538 break;
1539 default:
1540 if ((code != PARALLEL && code != EXPR_LIST) || type != OP_OUT)
1541 /* Some targets place small structures in registers for return
1542 values of functions, and those registers are wrapped in
1543 PARALLEL that we may see as the destination of a SET. Here
1544 is an example:
1546 (call_insn 13 12 14 2 (set (parallel:BLK [
1547 (expr_list:REG_DEP_TRUE (reg:DI 0 ax)
1548 (const_int 0 [0]))
1549 (expr_list:REG_DEP_TRUE (reg:DI 1 dx)
1550 (const_int 8 [0x8]))
1552 (call (mem:QI (symbol_ref:DI (... */
1553 type = OP_IN;
1554 fmt = GET_RTX_FORMAT (code);
1555 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1557 if (fmt[i] == 'e')
1558 add_regs_to_insn_regno_info (data, XEXP (x, i), insn, type, false, 0);
1559 else if (fmt[i] == 'E')
1561 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1562 add_regs_to_insn_regno_info (data, XVECEXP (x, i, j), insn,
1563 type, false, 0);
1569 /* Return execution frequency of INSN. */
1570 static int
1571 get_insn_freq (rtx_insn *insn)
1573 basic_block bb = BLOCK_FOR_INSN (insn);
1575 gcc_checking_assert (bb != NULL);
1576 return REG_FREQ_FROM_BB (bb);
1579 /* Invalidate all reg info of INSN with DATA and execution frequency
1580 FREQ. Update common info about the invalidated registers. */
1581 static void
1582 invalidate_insn_data_regno_info (lra_insn_recog_data_t data, rtx_insn *insn,
1583 int freq)
1585 int uid;
1586 bool debug_p;
1587 unsigned int i;
1588 struct lra_insn_reg *ir, *next_ir;
1590 uid = INSN_UID (insn);
1591 debug_p = DEBUG_INSN_P (insn);
1592 for (ir = data->regs; ir != NULL; ir = next_ir)
1594 i = ir->regno;
1595 next_ir = ir->next;
1596 lra_insn_reg_pool.remove (ir);
1597 bitmap_clear_bit (&lra_reg_info[i].insn_bitmap, uid);
1598 if (i >= FIRST_PSEUDO_REGISTER && ! debug_p)
1600 lra_reg_info[i].nrefs--;
1601 lra_reg_info[i].freq -= freq;
1602 lra_assert (lra_reg_info[i].nrefs >= 0 && lra_reg_info[i].freq >= 0);
1605 data->regs = NULL;
1608 /* Invalidate all reg info of INSN. Update common info about the
1609 invalidated registers. */
1610 void
1611 lra_invalidate_insn_regno_info (rtx_insn *insn)
1613 invalidate_insn_data_regno_info (lra_get_insn_recog_data (insn), insn,
1614 get_insn_freq (insn));
1617 /* Update common reg info from reg info of insn given by its DATA and
1618 execution frequency FREQ. */
1619 static void
1620 setup_insn_reg_info (lra_insn_recog_data_t data, int freq)
1622 unsigned int i;
1623 struct lra_insn_reg *ir;
1625 for (ir = data->regs; ir != NULL; ir = ir->next)
1626 if ((i = ir->regno) >= FIRST_PSEUDO_REGISTER)
1628 lra_reg_info[i].nrefs++;
1629 lra_reg_info[i].freq += freq;
1633 /* Set up insn reg info of INSN. Update common reg info from reg info
1634 of INSN. */
1635 void
1636 lra_update_insn_regno_info (rtx_insn *insn)
1638 int i, freq;
1639 lra_insn_recog_data_t data;
1640 struct lra_static_insn_data *static_data;
1641 enum rtx_code code;
1642 rtx link;
1644 if (! INSN_P (insn))
1645 return;
1646 data = lra_get_insn_recog_data (insn);
1647 static_data = data->insn_static_data;
1648 freq = NONDEBUG_INSN_P (insn) ? get_insn_freq (insn) : 0;
1649 invalidate_insn_data_regno_info (data, insn, freq);
1650 for (i = static_data->n_operands - 1; i >= 0; i--)
1651 add_regs_to_insn_regno_info (data, *data->operand_loc[i], insn,
1652 static_data->operand[i].type,
1653 static_data->operand[i].early_clobber,
1654 static_data->operand[i].early_clobber_alts);
1655 if ((code = GET_CODE (PATTERN (insn))) == CLOBBER || code == USE)
1656 add_regs_to_insn_regno_info (data, XEXP (PATTERN (insn), 0), insn,
1657 code == USE ? OP_IN : OP_OUT, false, 0);
1658 if (CALL_P (insn))
1659 /* On some targets call insns can refer to pseudos in memory in
1660 CALL_INSN_FUNCTION_USAGE list. Process them in order to
1661 consider their occurrences in calls for different
1662 transformations (e.g. inheritance) with given pseudos. */
1663 for (link = CALL_INSN_FUNCTION_USAGE (insn);
1664 link != NULL_RTX;
1665 link = XEXP (link, 1))
1667 code = GET_CODE (XEXP (link, 0));
1668 /* We could support CLOBBER_HIGH and treat it in the same way as
1669 HARD_REGNO_CALL_PART_CLOBBERED, but no port needs that yet. */
1670 gcc_assert (code != CLOBBER_HIGH);
1671 if ((code == USE || code == CLOBBER)
1672 && MEM_P (XEXP (XEXP (link, 0), 0)))
1673 add_regs_to_insn_regno_info (data, XEXP (XEXP (link, 0), 0), insn,
1674 code == USE ? OP_IN : OP_OUT, false, 0);
1676 if (NONDEBUG_INSN_P (insn))
1677 setup_insn_reg_info (data, freq);
1680 /* Return reg info of insn given by it UID. */
1681 struct lra_insn_reg *
1682 lra_get_insn_regs (int uid)
1684 lra_insn_recog_data_t data;
1686 data = get_insn_recog_data_by_uid (uid);
1687 return data->regs;
1692 /* Recursive hash function for RTL X. */
1693 hashval_t
1694 lra_rtx_hash (rtx x)
1696 int i, j;
1697 enum rtx_code code;
1698 const char *fmt;
1699 hashval_t val = 0;
1701 if (x == 0)
1702 return val;
1704 code = GET_CODE (x);
1705 val += (int) code + 4095;
1707 /* Some RTL can be compared nonrecursively. */
1708 switch (code)
1710 case REG:
1711 return val + REGNO (x);
1713 case LABEL_REF:
1714 return iterative_hash_object (XEXP (x, 0), val);
1716 case SYMBOL_REF:
1717 return iterative_hash_object (XSTR (x, 0), val);
1719 case SCRATCH:
1720 case CONST_DOUBLE:
1721 case CONST_INT:
1722 case CONST_VECTOR:
1723 return val;
1725 default:
1726 break;
1729 /* Hash the elements. */
1730 fmt = GET_RTX_FORMAT (code);
1731 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1733 switch (fmt[i])
1735 case 'w':
1736 val += XWINT (x, i);
1737 break;
1739 case 'n':
1740 case 'i':
1741 val += XINT (x, i);
1742 break;
1744 case 'V':
1745 case 'E':
1746 val += XVECLEN (x, i);
1748 for (j = 0; j < XVECLEN (x, i); j++)
1749 val += lra_rtx_hash (XVECEXP (x, i, j));
1750 break;
1752 case 'e':
1753 val += lra_rtx_hash (XEXP (x, i));
1754 break;
1756 case 'S':
1757 case 's':
1758 val += htab_hash_string (XSTR (x, i));
1759 break;
1761 case 'u':
1762 case '0':
1763 case 't':
1764 break;
1766 /* It is believed that rtx's at this level will never
1767 contain anything but integers and other rtx's, except for
1768 within LABEL_REFs and SYMBOL_REFs. */
1769 default:
1770 abort ();
1773 return val;
1778 /* This page contains code dealing with stack of the insns which
1779 should be processed by the next constraint pass. */
1781 /* Bitmap used to put an insn on the stack only in one exemplar. */
1782 static sbitmap lra_constraint_insn_stack_bitmap;
1784 /* The stack itself. */
1785 vec<rtx_insn *> lra_constraint_insn_stack;
1787 /* Put INSN on the stack. If ALWAYS_UPDATE is true, always update the reg
1788 info for INSN, otherwise only update it if INSN is not already on the
1789 stack. */
1790 static inline void
1791 lra_push_insn_1 (rtx_insn *insn, bool always_update)
1793 unsigned int uid = INSN_UID (insn);
1794 if (always_update)
1795 lra_update_insn_regno_info (insn);
1796 if (uid >= SBITMAP_SIZE (lra_constraint_insn_stack_bitmap))
1797 lra_constraint_insn_stack_bitmap =
1798 sbitmap_resize (lra_constraint_insn_stack_bitmap, 3 * uid / 2, 0);
1799 if (bitmap_bit_p (lra_constraint_insn_stack_bitmap, uid))
1800 return;
1801 bitmap_set_bit (lra_constraint_insn_stack_bitmap, uid);
1802 if (! always_update)
1803 lra_update_insn_regno_info (insn);
1804 lra_constraint_insn_stack.safe_push (insn);
1807 /* Put INSN on the stack. */
1808 void
1809 lra_push_insn (rtx_insn *insn)
1811 lra_push_insn_1 (insn, false);
1814 /* Put INSN on the stack and update its reg info. */
1815 void
1816 lra_push_insn_and_update_insn_regno_info (rtx_insn *insn)
1818 lra_push_insn_1 (insn, true);
1821 /* Put insn with UID on the stack. */
1822 void
1823 lra_push_insn_by_uid (unsigned int uid)
1825 lra_push_insn (lra_insn_recog_data[uid]->insn);
1828 /* Take the last-inserted insns off the stack and return it. */
1829 rtx_insn *
1830 lra_pop_insn (void)
1832 rtx_insn *insn = lra_constraint_insn_stack.pop ();
1833 bitmap_clear_bit (lra_constraint_insn_stack_bitmap, INSN_UID (insn));
1834 return insn;
1837 /* Return the current size of the insn stack. */
1838 unsigned int
1839 lra_insn_stack_length (void)
1841 return lra_constraint_insn_stack.length ();
1844 /* Push insns FROM to TO (excluding it) going in reverse order. */
1845 static void
1846 push_insns (rtx_insn *from, rtx_insn *to)
1848 rtx_insn *insn;
1850 if (from == NULL_RTX)
1851 return;
1852 for (insn = from; insn != to; insn = PREV_INSN (insn))
1853 if (INSN_P (insn))
1854 lra_push_insn (insn);
1857 /* Set up sp offset for insn in range [FROM, LAST]. The offset is
1858 taken from the next BB insn after LAST or zero if there in such
1859 insn. */
1860 static void
1861 setup_sp_offset (rtx_insn *from, rtx_insn *last)
1863 rtx_insn *before = next_nonnote_nondebug_insn_bb (last);
1864 poly_int64 offset = (before == NULL_RTX || ! INSN_P (before)
1865 ? 0 : lra_get_insn_recog_data (before)->sp_offset);
1867 for (rtx_insn *insn = from; insn != NEXT_INSN (last); insn = NEXT_INSN (insn))
1868 lra_get_insn_recog_data (insn)->sp_offset = offset;
1871 /* Emit insns BEFORE before INSN and insns AFTER after INSN. Put the
1872 insns onto the stack. Print about emitting the insns with
1873 TITLE. */
1874 void
1875 lra_process_new_insns (rtx_insn *insn, rtx_insn *before, rtx_insn *after,
1876 const char *title)
1878 rtx_insn *last;
1880 if (before == NULL_RTX && after == NULL_RTX)
1881 return;
1882 if (lra_dump_file != NULL)
1884 dump_insn_slim (lra_dump_file, insn);
1885 if (before != NULL_RTX)
1887 fprintf (lra_dump_file," %s before:\n", title);
1888 dump_rtl_slim (lra_dump_file, before, NULL, -1, 0);
1890 if (after != NULL_RTX)
1892 fprintf (lra_dump_file, " %s after:\n", title);
1893 dump_rtl_slim (lra_dump_file, after, NULL, -1, 0);
1895 fprintf (lra_dump_file, "\n");
1897 if (before != NULL_RTX)
1899 if (cfun->can_throw_non_call_exceptions)
1900 copy_reg_eh_region_note_forward (insn, before, NULL);
1901 emit_insn_before (before, insn);
1902 push_insns (PREV_INSN (insn), PREV_INSN (before));
1903 setup_sp_offset (before, PREV_INSN (insn));
1905 if (after != NULL_RTX)
1907 if (cfun->can_throw_non_call_exceptions)
1908 copy_reg_eh_region_note_forward (insn, after, NULL);
1909 for (last = after; NEXT_INSN (last) != NULL_RTX; last = NEXT_INSN (last))
1911 emit_insn_after (after, insn);
1912 push_insns (last, insn);
1913 setup_sp_offset (after, last);
1915 if (cfun->can_throw_non_call_exceptions)
1917 rtx note = find_reg_note (insn, REG_EH_REGION, NULL_RTX);
1918 if (note && !insn_could_throw_p (insn))
1919 remove_note (insn, note);
1924 /* Replace all references to register OLD_REGNO in *LOC with pseudo
1925 register NEW_REG. Try to simplify subreg of constant if SUBREG_P.
1926 DEBUG_P is if LOC is within a DEBUG_INSN. Return true if any
1927 change was made. */
1928 bool
1929 lra_substitute_pseudo (rtx *loc, int old_regno, rtx new_reg, bool subreg_p,
1930 bool debug_p)
1932 rtx x = *loc;
1933 bool result = false;
1934 enum rtx_code code;
1935 const char *fmt;
1936 int i, j;
1938 if (x == NULL_RTX)
1939 return false;
1941 code = GET_CODE (x);
1942 if (code == SUBREG && subreg_p)
1944 rtx subst, inner = SUBREG_REG (x);
1945 /* Transform subreg of constant while we still have inner mode
1946 of the subreg. The subreg internal should not be an insn
1947 operand. */
1948 if (REG_P (inner) && (int) REGNO (inner) == old_regno
1949 && CONSTANT_P (new_reg)
1950 && (subst = simplify_subreg (GET_MODE (x), new_reg, GET_MODE (inner),
1951 SUBREG_BYTE (x))) != NULL_RTX)
1953 *loc = subst;
1954 return true;
1958 else if (code == REG && (int) REGNO (x) == old_regno)
1960 machine_mode mode = GET_MODE (x);
1961 machine_mode inner_mode = GET_MODE (new_reg);
1963 if (mode != inner_mode
1964 && ! (CONST_SCALAR_INT_P (new_reg) && SCALAR_INT_MODE_P (mode)))
1966 poly_uint64 offset = 0;
1967 if (partial_subreg_p (mode, inner_mode)
1968 && SCALAR_INT_MODE_P (inner_mode))
1969 offset = subreg_lowpart_offset (mode, inner_mode);
1970 if (debug_p)
1971 new_reg = gen_rtx_raw_SUBREG (mode, new_reg, offset);
1972 else
1973 new_reg = gen_rtx_SUBREG (mode, new_reg, offset);
1975 *loc = new_reg;
1976 return true;
1979 /* Scan all the operand sub-expressions. */
1980 fmt = GET_RTX_FORMAT (code);
1981 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1983 if (fmt[i] == 'e')
1985 if (lra_substitute_pseudo (&XEXP (x, i), old_regno,
1986 new_reg, subreg_p, debug_p))
1987 result = true;
1989 else if (fmt[i] == 'E')
1991 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1992 if (lra_substitute_pseudo (&XVECEXP (x, i, j), old_regno,
1993 new_reg, subreg_p, debug_p))
1994 result = true;
1997 return result;
2000 /* Call lra_substitute_pseudo within an insn. Try to simplify subreg
2001 of constant if SUBREG_P. This won't update the insn ptr, just the
2002 contents of the insn. */
2003 bool
2004 lra_substitute_pseudo_within_insn (rtx_insn *insn, int old_regno,
2005 rtx new_reg, bool subreg_p)
2007 rtx loc = insn;
2008 return lra_substitute_pseudo (&loc, old_regno, new_reg, subreg_p,
2009 DEBUG_INSN_P (insn));
2014 /* This page contains code dealing with scratches (changing them onto
2015 pseudos and restoring them from the pseudos).
2017 We change scratches into pseudos at the beginning of LRA to
2018 simplify dealing with them (conflicts, hard register assignments).
2020 If the pseudo denoting scratch was spilled it means that we do need
2021 a hard register for it. Such pseudos are transformed back to
2022 scratches at the end of LRA. */
2024 /* Description of location of a former scratch operand. */
2025 struct sloc
2027 rtx_insn *insn; /* Insn where the scratch was. */
2028 int nop; /* Number of the operand which was a scratch. */
2031 typedef struct sloc *sloc_t;
2033 /* Locations of the former scratches. */
2034 static vec<sloc_t> scratches;
2036 /* Bitmap of scratch regnos. */
2037 static bitmap_head scratch_bitmap;
2039 /* Bitmap of scratch operands. */
2040 static bitmap_head scratch_operand_bitmap;
2042 /* Return true if pseudo REGNO is made of SCRATCH. */
2043 bool
2044 lra_former_scratch_p (int regno)
2046 return bitmap_bit_p (&scratch_bitmap, regno);
2049 /* Return true if the operand NOP of INSN is a former scratch. */
2050 bool
2051 lra_former_scratch_operand_p (rtx_insn *insn, int nop)
2053 return bitmap_bit_p (&scratch_operand_bitmap,
2054 INSN_UID (insn) * MAX_RECOG_OPERANDS + nop) != 0;
2057 /* Register operand NOP in INSN as a former scratch. It will be
2058 changed to scratch back, if it is necessary, at the LRA end. */
2059 void
2060 lra_register_new_scratch_op (rtx_insn *insn, int nop)
2062 lra_insn_recog_data_t id = lra_get_insn_recog_data (insn);
2063 rtx op = *id->operand_loc[nop];
2064 sloc_t loc = XNEW (struct sloc);
2065 lra_assert (REG_P (op));
2066 loc->insn = insn;
2067 loc->nop = nop;
2068 scratches.safe_push (loc);
2069 bitmap_set_bit (&scratch_bitmap, REGNO (op));
2070 bitmap_set_bit (&scratch_operand_bitmap,
2071 INSN_UID (insn) * MAX_RECOG_OPERANDS + nop);
2072 add_reg_note (insn, REG_UNUSED, op);
2075 /* Change scratches onto pseudos and save their location. */
2076 static void
2077 remove_scratches (void)
2079 int i;
2080 bool insn_changed_p;
2081 basic_block bb;
2082 rtx_insn *insn;
2083 rtx reg;
2084 lra_insn_recog_data_t id;
2085 struct lra_static_insn_data *static_id;
2087 scratches.create (get_max_uid ());
2088 bitmap_initialize (&scratch_bitmap, &reg_obstack);
2089 bitmap_initialize (&scratch_operand_bitmap, &reg_obstack);
2090 FOR_EACH_BB_FN (bb, cfun)
2091 FOR_BB_INSNS (bb, insn)
2092 if (INSN_P (insn))
2094 id = lra_get_insn_recog_data (insn);
2095 static_id = id->insn_static_data;
2096 insn_changed_p = false;
2097 for (i = 0; i < static_id->n_operands; i++)
2098 if (GET_CODE (*id->operand_loc[i]) == SCRATCH
2099 && GET_MODE (*id->operand_loc[i]) != VOIDmode)
2101 insn_changed_p = true;
2102 *id->operand_loc[i] = reg
2103 = lra_create_new_reg (static_id->operand[i].mode,
2104 *id->operand_loc[i], ALL_REGS, NULL);
2105 lra_register_new_scratch_op (insn, i);
2106 if (lra_dump_file != NULL)
2107 fprintf (lra_dump_file,
2108 "Removing SCRATCH in insn #%u (nop %d)\n",
2109 INSN_UID (insn), i);
2111 if (insn_changed_p)
2112 /* Because we might use DF right after caller-saves sub-pass
2113 we need to keep DF info up to date. */
2114 df_insn_rescan (insn);
2118 /* Changes pseudos created by function remove_scratches onto scratches. */
2119 static void
2120 restore_scratches (void)
2122 int regno;
2123 unsigned i;
2124 sloc_t loc;
2125 rtx_insn *last = NULL;
2126 lra_insn_recog_data_t id = NULL;
2128 for (i = 0; scratches.iterate (i, &loc); i++)
2130 /* Ignore already deleted insns. */
2131 if (NOTE_P (loc->insn)
2132 && NOTE_KIND (loc->insn) == NOTE_INSN_DELETED)
2133 continue;
2134 if (last != loc->insn)
2136 last = loc->insn;
2137 id = lra_get_insn_recog_data (last);
2139 if (REG_P (*id->operand_loc[loc->nop])
2140 && ((regno = REGNO (*id->operand_loc[loc->nop]))
2141 >= FIRST_PSEUDO_REGISTER)
2142 && lra_get_regno_hard_regno (regno) < 0)
2144 /* It should be only case when scratch register with chosen
2145 constraint 'X' did not get memory or hard register. */
2146 lra_assert (lra_former_scratch_p (regno));
2147 *id->operand_loc[loc->nop]
2148 = gen_rtx_SCRATCH (GET_MODE (*id->operand_loc[loc->nop]));
2149 lra_update_dup (id, loc->nop);
2150 if (lra_dump_file != NULL)
2151 fprintf (lra_dump_file, "Restoring SCRATCH in insn #%u(nop %d)\n",
2152 INSN_UID (loc->insn), loc->nop);
2155 for (i = 0; scratches.iterate (i, &loc); i++)
2156 free (loc);
2157 scratches.release ();
2158 bitmap_clear (&scratch_bitmap);
2159 bitmap_clear (&scratch_operand_bitmap);
2164 /* Function checks RTL for correctness. If FINAL_P is true, it is
2165 done at the end of LRA and the check is more rigorous. */
2166 static void
2167 check_rtl (bool final_p)
2169 basic_block bb;
2170 rtx_insn *insn;
2172 lra_assert (! final_p || reload_completed);
2173 FOR_EACH_BB_FN (bb, cfun)
2174 FOR_BB_INSNS (bb, insn)
2175 if (NONDEBUG_INSN_P (insn)
2176 && GET_CODE (PATTERN (insn)) != USE
2177 && GET_CODE (PATTERN (insn)) != CLOBBER
2178 && GET_CODE (PATTERN (insn)) != ASM_INPUT)
2180 if (final_p)
2182 extract_constrain_insn (insn);
2183 continue;
2185 /* LRA code is based on assumption that all addresses can be
2186 correctly decomposed. LRA can generate reloads for
2187 decomposable addresses. The decomposition code checks the
2188 correctness of the addresses. So we don't need to check
2189 the addresses here. Don't call insn_invalid_p here, it can
2190 change the code at this stage. */
2191 if (recog_memoized (insn) < 0 && asm_noperands (PATTERN (insn)) < 0)
2192 fatal_insn_not_found (insn);
2196 /* Determine if the current function has an exception receiver block
2197 that reaches the exit block via non-exceptional edges */
2198 static bool
2199 has_nonexceptional_receiver (void)
2201 edge e;
2202 edge_iterator ei;
2203 basic_block *tos, *worklist, bb;
2205 /* If we're not optimizing, then just err on the safe side. */
2206 if (!optimize)
2207 return true;
2209 /* First determine which blocks can reach exit via normal paths. */
2210 tos = worklist = XNEWVEC (basic_block, n_basic_blocks_for_fn (cfun) + 1);
2212 FOR_EACH_BB_FN (bb, cfun)
2213 bb->flags &= ~BB_REACHABLE;
2215 /* Place the exit block on our worklist. */
2216 EXIT_BLOCK_PTR_FOR_FN (cfun)->flags |= BB_REACHABLE;
2217 *tos++ = EXIT_BLOCK_PTR_FOR_FN (cfun);
2219 /* Iterate: find everything reachable from what we've already seen. */
2220 while (tos != worklist)
2222 bb = *--tos;
2224 FOR_EACH_EDGE (e, ei, bb->preds)
2225 if (e->flags & EDGE_ABNORMAL)
2227 free (worklist);
2228 return true;
2230 else
2232 basic_block src = e->src;
2234 if (!(src->flags & BB_REACHABLE))
2236 src->flags |= BB_REACHABLE;
2237 *tos++ = src;
2241 free (worklist);
2242 /* No exceptional block reached exit unexceptionally. */
2243 return false;
2247 /* Process recursively X of INSN and add REG_INC notes if necessary. */
2248 static void
2249 add_auto_inc_notes (rtx_insn *insn, rtx x)
2251 enum rtx_code code = GET_CODE (x);
2252 const char *fmt;
2253 int i, j;
2255 if (code == MEM && auto_inc_p (XEXP (x, 0)))
2257 add_reg_note (insn, REG_INC, XEXP (XEXP (x, 0), 0));
2258 return;
2261 /* Scan all X sub-expressions. */
2262 fmt = GET_RTX_FORMAT (code);
2263 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
2265 if (fmt[i] == 'e')
2266 add_auto_inc_notes (insn, XEXP (x, i));
2267 else if (fmt[i] == 'E')
2268 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
2269 add_auto_inc_notes (insn, XVECEXP (x, i, j));
2274 /* Remove all REG_DEAD and REG_UNUSED notes and regenerate REG_INC.
2275 We change pseudos by hard registers without notification of DF and
2276 that can make the notes obsolete. DF-infrastructure does not deal
2277 with REG_INC notes -- so we should regenerate them here. */
2278 static void
2279 update_inc_notes (void)
2281 rtx *pnote;
2282 basic_block bb;
2283 rtx_insn *insn;
2285 FOR_EACH_BB_FN (bb, cfun)
2286 FOR_BB_INSNS (bb, insn)
2287 if (NONDEBUG_INSN_P (insn))
2289 pnote = &REG_NOTES (insn);
2290 while (*pnote != 0)
2292 if (REG_NOTE_KIND (*pnote) == REG_DEAD
2293 || REG_NOTE_KIND (*pnote) == REG_UNUSED
2294 || REG_NOTE_KIND (*pnote) == REG_INC)
2295 *pnote = XEXP (*pnote, 1);
2296 else
2297 pnote = &XEXP (*pnote, 1);
2300 if (AUTO_INC_DEC)
2301 add_auto_inc_notes (insn, PATTERN (insn));
2305 /* Set to 1 while in lra. */
2306 int lra_in_progress;
2308 /* Start of pseudo regnos before the LRA. */
2309 int lra_new_regno_start;
2311 /* Start of reload pseudo regnos before the new spill pass. */
2312 int lra_constraint_new_regno_start;
2314 /* Avoid spilling pseudos with regno more than the following value if
2315 it is possible. */
2316 int lra_bad_spill_regno_start;
2318 /* Inheritance pseudo regnos before the new spill pass. */
2319 bitmap_head lra_inheritance_pseudos;
2321 /* Split regnos before the new spill pass. */
2322 bitmap_head lra_split_regs;
2324 /* Reload pseudo regnos before the new assignment pass which still can
2325 be spilled after the assignment pass as memory is also accepted in
2326 insns for the reload pseudos. */
2327 bitmap_head lra_optional_reload_pseudos;
2329 /* Pseudo regnos used for subreg reloads before the new assignment
2330 pass. Such pseudos still can be spilled after the assignment
2331 pass. */
2332 bitmap_head lra_subreg_reload_pseudos;
2334 /* File used for output of LRA debug information. */
2335 FILE *lra_dump_file;
2337 /* True if we should try spill into registers of different classes
2338 instead of memory. */
2339 bool lra_reg_spill_p;
2341 /* Set up value LRA_REG_SPILL_P. */
2342 static void
2343 setup_reg_spill_flag (void)
2345 int cl, mode;
2347 if (targetm.spill_class != NULL)
2348 for (cl = 0; cl < (int) LIM_REG_CLASSES; cl++)
2349 for (mode = 0; mode < MAX_MACHINE_MODE; mode++)
2350 if (targetm.spill_class ((enum reg_class) cl,
2351 (machine_mode) mode) != NO_REGS)
2353 lra_reg_spill_p = true;
2354 return;
2356 lra_reg_spill_p = false;
2359 /* True if the current function is too big to use regular algorithms
2360 in LRA. In other words, we should use simpler and faster algorithms
2361 in LRA. It also means we should not worry about generation code
2362 for caller saves. The value is set up in IRA. */
2363 bool lra_simple_p;
2365 /* Major LRA entry function. F is a file should be used to dump LRA
2366 debug info. */
2367 void
2368 lra (FILE *f)
2370 int i;
2371 bool live_p, inserted_p;
2373 lra_dump_file = f;
2375 timevar_push (TV_LRA);
2377 /* Make sure that the last insn is a note. Some subsequent passes
2378 need it. */
2379 emit_note (NOTE_INSN_DELETED);
2381 COPY_HARD_REG_SET (lra_no_alloc_regs, ira_no_alloc_regs);
2383 init_reg_info ();
2384 expand_reg_info ();
2386 init_insn_recog_data ();
2388 /* Some quick check on RTL generated by previous passes. */
2389 if (flag_checking)
2390 check_rtl (false);
2392 lra_in_progress = 1;
2394 lra_live_range_iter = lra_coalesce_iter = lra_constraint_iter = 0;
2395 lra_assignment_iter = lra_assignment_iter_after_spill = 0;
2396 lra_inheritance_iter = lra_undo_inheritance_iter = 0;
2397 lra_rematerialization_iter = 0;
2399 setup_reg_spill_flag ();
2401 /* Function remove_scratches can creates new pseudos for clobbers --
2402 so set up lra_constraint_new_regno_start before its call to
2403 permit changing reg classes for pseudos created by this
2404 simplification. */
2405 lra_constraint_new_regno_start = lra_new_regno_start = max_reg_num ();
2406 lra_bad_spill_regno_start = INT_MAX;
2407 remove_scratches ();
2409 /* A function that has a non-local label that can reach the exit
2410 block via non-exceptional paths must save all call-saved
2411 registers. */
2412 if (cfun->has_nonlocal_label && has_nonexceptional_receiver ())
2413 crtl->saves_all_registers = 1;
2415 if (crtl->saves_all_registers)
2416 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2417 if (! call_used_regs[i] && ! fixed_regs[i] && ! LOCAL_REGNO (i))
2418 df_set_regs_ever_live (i, true);
2420 /* We don't DF from now and avoid its using because it is to
2421 expensive when a lot of RTL changes are made. */
2422 df_set_flags (DF_NO_INSN_RESCAN);
2423 lra_constraint_insn_stack.create (get_max_uid ());
2424 lra_constraint_insn_stack_bitmap = sbitmap_alloc (get_max_uid ());
2425 bitmap_clear (lra_constraint_insn_stack_bitmap);
2426 lra_live_ranges_init ();
2427 lra_constraints_init ();
2428 lra_curr_reload_num = 0;
2429 push_insns (get_last_insn (), NULL);
2430 /* It is needed for the 1st coalescing. */
2431 bitmap_initialize (&lra_inheritance_pseudos, &reg_obstack);
2432 bitmap_initialize (&lra_split_regs, &reg_obstack);
2433 bitmap_initialize (&lra_optional_reload_pseudos, &reg_obstack);
2434 bitmap_initialize (&lra_subreg_reload_pseudos, &reg_obstack);
2435 live_p = false;
2436 if (maybe_ne (get_frame_size (), 0) && crtl->stack_alignment_needed)
2437 /* If we have a stack frame, we must align it now. The stack size
2438 may be a part of the offset computation for register
2439 elimination. */
2440 assign_stack_local (BLKmode, 0, crtl->stack_alignment_needed);
2441 lra_init_equiv ();
2442 for (;;)
2444 for (;;)
2446 bool reloads_p = lra_constraints (lra_constraint_iter == 0);
2447 /* Constraint transformations may result in that eliminable
2448 hard regs become uneliminable and pseudos which use them
2449 should be spilled. It is better to do it before pseudo
2450 assignments.
2452 For example, rs6000 can make
2453 RS6000_PIC_OFFSET_TABLE_REGNUM uneliminable if we started
2454 to use a constant pool. */
2455 lra_eliminate (false, false);
2456 /* We should try to assign hard registers to scratches even
2457 if there were no RTL transformations in lra_constraints.
2458 Also we should check IRA assignments on the first
2459 iteration as they can be wrong because of early clobbers
2460 operands which are ignored in IRA. */
2461 if (! reloads_p && lra_constraint_iter > 1)
2463 /* Stack is not empty here only when there are changes
2464 during the elimination sub-pass. */
2465 if (bitmap_empty_p (lra_constraint_insn_stack_bitmap))
2466 break;
2467 else
2468 /* If there are no reloads but changing due
2469 elimination, restart the constraint sub-pass
2470 first. */
2471 continue;
2473 /* Do inheritance only for regular algorithms. */
2474 if (! lra_simple_p)
2476 if (flag_ipa_ra)
2478 if (live_p)
2479 lra_clear_live_ranges ();
2480 /* As a side-effect of lra_create_live_ranges, we calculate
2481 actual_call_used_reg_set, which is needed during
2482 lra_inheritance. */
2483 lra_create_live_ranges (true, true);
2484 live_p = true;
2486 lra_inheritance ();
2488 if (live_p)
2489 lra_clear_live_ranges ();
2490 bool fails_p;
2493 /* We need live ranges for lra_assign -- so build them.
2494 But don't remove dead insns or change global live
2495 info as we can undo inheritance transformations after
2496 inheritance pseudo assigning. */
2497 lra_create_live_ranges (true, false);
2498 live_p = true;
2499 /* If we don't spill non-reload and non-inheritance
2500 pseudos, there is no sense to run memory-memory move
2501 coalescing. If inheritance pseudos were spilled, the
2502 memory-memory moves involving them will be removed by
2503 pass undoing inheritance. */
2504 if (lra_simple_p)
2505 lra_assign (fails_p);
2506 else
2508 bool spill_p = !lra_assign (fails_p);
2510 if (lra_undo_inheritance ())
2511 live_p = false;
2512 if (spill_p && ! fails_p)
2514 if (! live_p)
2516 lra_create_live_ranges (true, true);
2517 live_p = true;
2519 if (lra_coalesce ())
2520 live_p = false;
2522 if (! live_p)
2523 lra_clear_live_ranges ();
2525 if (fails_p)
2527 /* It is a very rare case. It is the last hope to
2528 split a hard regno live range for a reload
2529 pseudo. */
2530 if (live_p)
2531 lra_clear_live_ranges ();
2532 live_p = false;
2533 if (! lra_split_hard_reg_for ())
2534 break;
2537 while (fails_p);
2539 /* Don't clear optional reloads bitmap until all constraints are
2540 satisfied as we need to differ them from regular reloads. */
2541 bitmap_clear (&lra_optional_reload_pseudos);
2542 bitmap_clear (&lra_subreg_reload_pseudos);
2543 bitmap_clear (&lra_inheritance_pseudos);
2544 bitmap_clear (&lra_split_regs);
2545 if (! live_p)
2547 /* We need full live info for spilling pseudos into
2548 registers instead of memory. */
2549 lra_create_live_ranges (lra_reg_spill_p, true);
2550 live_p = true;
2552 /* We should check necessity for spilling here as the above live
2553 range pass can remove spilled pseudos. */
2554 if (! lra_need_for_spills_p ())
2555 break;
2556 /* Now we know what pseudos should be spilled. Try to
2557 rematerialize them first. */
2558 if (lra_remat ())
2560 /* We need full live info -- see the comment above. */
2561 lra_create_live_ranges (lra_reg_spill_p, true);
2562 live_p = true;
2563 if (! lra_need_for_spills_p ())
2564 break;
2566 lra_spill ();
2567 /* Assignment of stack slots changes elimination offsets for
2568 some eliminations. So update the offsets here. */
2569 lra_eliminate (false, false);
2570 lra_constraint_new_regno_start = max_reg_num ();
2571 if (lra_bad_spill_regno_start == INT_MAX
2572 && lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES
2573 && lra_rematerialization_iter > LRA_MAX_REMATERIALIZATION_PASSES)
2574 /* After switching off inheritance and rematerialization
2575 passes, avoid spilling reload pseudos will be created to
2576 prevent LRA cycling in some complicated cases. */
2577 lra_bad_spill_regno_start = lra_constraint_new_regno_start;
2578 lra_assignment_iter_after_spill = 0;
2580 restore_scratches ();
2581 lra_eliminate (true, false);
2582 lra_final_code_change ();
2583 lra_in_progress = 0;
2584 if (live_p)
2585 lra_clear_live_ranges ();
2586 lra_live_ranges_finish ();
2587 lra_constraints_finish ();
2588 finish_reg_info ();
2589 sbitmap_free (lra_constraint_insn_stack_bitmap);
2590 lra_constraint_insn_stack.release ();
2591 finish_insn_recog_data ();
2592 regstat_free_n_sets_and_refs ();
2593 regstat_free_ri ();
2594 reload_completed = 1;
2595 update_inc_notes ();
2597 inserted_p = fixup_abnormal_edges ();
2599 /* We've possibly turned single trapping insn into multiple ones. */
2600 if (cfun->can_throw_non_call_exceptions)
2602 auto_sbitmap blocks (last_basic_block_for_fn (cfun));
2603 bitmap_ones (blocks);
2604 find_many_sub_basic_blocks (blocks);
2607 if (inserted_p)
2608 commit_edge_insertions ();
2610 /* Replacing pseudos with their memory equivalents might have
2611 created shared rtx. Subsequent passes would get confused
2612 by this, so unshare everything here. */
2613 unshare_all_rtl_again (get_insns ());
2615 if (flag_checking)
2616 check_rtl (true);
2618 timevar_pop (TV_LRA);
2621 /* Called once per compiler to initialize LRA data once. */
2622 void
2623 lra_init_once (void)
2625 init_insn_code_data_once ();
2628 /* Called once per compiler to finish LRA data which are initialize
2629 once. */
2630 void
2631 lra_finish_once (void)
2633 finish_insn_code_data_once ();