2014-12-18 Vladimir Makarov <vmakarov@redhat.com>
[official-gcc.git] / gcc / lra-constraints.c
blob23fd44db54a3e2fc662b32a40f1f06240354b435
1 /* Code for RTL transformations to satisfy insn constraints.
2 Copyright (C) 2010-2014 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 /* This file contains code for 3 passes: constraint pass,
23 inheritance/split pass, and pass for undoing failed inheritance and
24 split.
26 The major goal of constraint pass is to transform RTL to satisfy
27 insn and address constraints by:
28 o choosing insn alternatives;
29 o generating *reload insns* (or reloads in brief) and *reload
30 pseudos* which will get necessary hard registers later;
31 o substituting pseudos with equivalent values and removing the
32 instructions that initialized those pseudos.
34 The constraint pass has biggest and most complicated code in LRA.
35 There are a lot of important details like:
36 o reuse of input reload pseudos to simplify reload pseudo
37 allocations;
38 o some heuristics to choose insn alternative to improve the
39 inheritance;
40 o early clobbers etc.
42 The pass is mimicking former reload pass in alternative choosing
43 because the reload pass is oriented to current machine description
44 model. It might be changed if the machine description model is
45 changed.
47 There is special code for preventing all LRA and this pass cycling
48 in case of bugs.
50 On the first iteration of the pass we process every instruction and
51 choose an alternative for each one. On subsequent iterations we try
52 to avoid reprocessing instructions if we can be sure that the old
53 choice is still valid.
55 The inheritance/spilt pass is to transform code to achieve
56 ineheritance and live range splitting. It is done on backward
57 traversal of EBBs.
59 The inheritance optimization goal is to reuse values in hard
60 registers. There is analogous optimization in old reload pass. The
61 inheritance is achieved by following transformation:
63 reload_p1 <- p reload_p1 <- p
64 ... new_p <- reload_p1
65 ... => ...
66 reload_p2 <- p reload_p2 <- new_p
68 where p is spilled and not changed between the insns. Reload_p1 is
69 also called *original pseudo* and new_p is called *inheritance
70 pseudo*.
72 The subsequent assignment pass will try to assign the same (or
73 another if it is not possible) hard register to new_p as to
74 reload_p1 or reload_p2.
76 If the assignment pass fails to assign a hard register to new_p,
77 this file will undo the inheritance and restore the original code.
78 This is because implementing the above sequence with a spilled
79 new_p would make the code much worse. The inheritance is done in
80 EBB scope. The above is just a simplified example to get an idea
81 of the inheritance as the inheritance is also done for non-reload
82 insns.
84 Splitting (transformation) is also done in EBB scope on the same
85 pass as the inheritance:
87 r <- ... or ... <- r r <- ... or ... <- r
88 ... s <- r (new insn -- save)
89 ... =>
90 ... r <- s (new insn -- restore)
91 ... <- r ... <- r
93 The *split pseudo* s is assigned to the hard register of the
94 original pseudo or hard register r.
96 Splitting is done:
97 o In EBBs with high register pressure for global pseudos (living
98 in at least 2 BBs) and assigned to hard registers when there
99 are more one reloads needing the hard registers;
100 o for pseudos needing save/restore code around calls.
102 If the split pseudo still has the same hard register as the
103 original pseudo after the subsequent assignment pass or the
104 original pseudo was split, the opposite transformation is done on
105 the same pass for undoing inheritance. */
107 #undef REG_OK_STRICT
109 #include "config.h"
110 #include "system.h"
111 #include "coretypes.h"
112 #include "tm.h"
113 #include "hard-reg-set.h"
114 #include "rtl.h"
115 #include "tm_p.h"
116 #include "regs.h"
117 #include "insn-config.h"
118 #include "insn-codes.h"
119 #include "recog.h"
120 #include "output.h"
121 #include "addresses.h"
122 #include "target.h"
123 #include "hashtab.h"
124 #include "hash-set.h"
125 #include "vec.h"
126 #include "machmode.h"
127 #include "input.h"
128 #include "function.h"
129 #include "expr.h"
130 #include "predict.h"
131 #include "dominance.h"
132 #include "cfg.h"
133 #include "cfgrtl.h"
134 #include "basic-block.h"
135 #include "except.h"
136 #include "optabs.h"
137 #include "df.h"
138 #include "ira.h"
139 #include "rtl-error.h"
140 #include "lra-int.h"
142 /* Value of LRA_CURR_RELOAD_NUM at the beginning of BB of the current
143 insn. Remember that LRA_CURR_RELOAD_NUM is the number of emitted
144 reload insns. */
145 static int bb_reload_num;
147 /* The current insn being processed and corresponding its single set
148 (NULL otherwise), its data (basic block, the insn data, the insn
149 static data, and the mode of each operand). */
150 static rtx_insn *curr_insn;
151 static rtx curr_insn_set;
152 static basic_block curr_bb;
153 static lra_insn_recog_data_t curr_id;
154 static struct lra_static_insn_data *curr_static_id;
155 static machine_mode curr_operand_mode[MAX_RECOG_OPERANDS];
159 /* Start numbers for new registers and insns at the current constraints
160 pass start. */
161 static int new_regno_start;
162 static int new_insn_uid_start;
164 /* If LOC is nonnull, strip any outer subreg from it. */
165 static inline rtx *
166 strip_subreg (rtx *loc)
168 return loc && GET_CODE (*loc) == SUBREG ? &SUBREG_REG (*loc) : loc;
171 /* Return hard regno of REGNO or if it is was not assigned to a hard
172 register, use a hard register from its allocno class. */
173 static int
174 get_try_hard_regno (int regno)
176 int hard_regno;
177 enum reg_class rclass;
179 if ((hard_regno = regno) >= FIRST_PSEUDO_REGISTER)
180 hard_regno = lra_get_regno_hard_regno (regno);
181 if (hard_regno >= 0)
182 return hard_regno;
183 rclass = lra_get_allocno_class (regno);
184 if (rclass == NO_REGS)
185 return -1;
186 return ira_class_hard_regs[rclass][0];
189 /* Return final hard regno (plus offset) which will be after
190 elimination. We do this for matching constraints because the final
191 hard regno could have a different class. */
192 static int
193 get_final_hard_regno (int hard_regno, int offset)
195 if (hard_regno < 0)
196 return hard_regno;
197 hard_regno = lra_get_elimination_hard_regno (hard_regno);
198 return hard_regno + offset;
201 /* Return hard regno of X after removing subreg and making
202 elimination. If X is not a register or subreg of register, return
203 -1. For pseudo use its assignment. */
204 static int
205 get_hard_regno (rtx x)
207 rtx reg;
208 int offset, hard_regno;
210 reg = x;
211 if (GET_CODE (x) == SUBREG)
212 reg = SUBREG_REG (x);
213 if (! REG_P (reg))
214 return -1;
215 if ((hard_regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER)
216 hard_regno = lra_get_regno_hard_regno (hard_regno);
217 if (hard_regno < 0)
218 return -1;
219 offset = 0;
220 if (GET_CODE (x) == SUBREG)
221 offset += subreg_regno_offset (hard_regno, GET_MODE (reg),
222 SUBREG_BYTE (x), GET_MODE (x));
223 return get_final_hard_regno (hard_regno, offset);
226 /* If REGNO is a hard register or has been allocated a hard register,
227 return the class of that register. If REGNO is a reload pseudo
228 created by the current constraints pass, return its allocno class.
229 Return NO_REGS otherwise. */
230 static enum reg_class
231 get_reg_class (int regno)
233 int hard_regno;
235 if ((hard_regno = regno) >= FIRST_PSEUDO_REGISTER)
236 hard_regno = lra_get_regno_hard_regno (regno);
237 if (hard_regno >= 0)
239 hard_regno = get_final_hard_regno (hard_regno, 0);
240 return REGNO_REG_CLASS (hard_regno);
242 if (regno >= new_regno_start)
243 return lra_get_allocno_class (regno);
244 return NO_REGS;
247 /* Return true if REG satisfies (or will satisfy) reg class constraint
248 CL. Use elimination first if REG is a hard register. If REG is a
249 reload pseudo created by this constraints pass, assume that it will
250 be allocated a hard register from its allocno class, but allow that
251 class to be narrowed to CL if it is currently a superset of CL.
253 If NEW_CLASS is nonnull, set *NEW_CLASS to the new allocno class of
254 REGNO (reg), or NO_REGS if no change in its class was needed. */
255 static bool
256 in_class_p (rtx reg, enum reg_class cl, enum reg_class *new_class)
258 enum reg_class rclass, common_class;
259 machine_mode reg_mode;
260 int class_size, hard_regno, nregs, i, j;
261 int regno = REGNO (reg);
263 if (new_class != NULL)
264 *new_class = NO_REGS;
265 if (regno < FIRST_PSEUDO_REGISTER)
267 rtx final_reg = reg;
268 rtx *final_loc = &final_reg;
270 lra_eliminate_reg_if_possible (final_loc);
271 return TEST_HARD_REG_BIT (reg_class_contents[cl], REGNO (*final_loc));
273 reg_mode = GET_MODE (reg);
274 rclass = get_reg_class (regno);
275 if (regno < new_regno_start
276 /* Do not allow the constraints for reload instructions to
277 influence the classes of new pseudos. These reloads are
278 typically moves that have many alternatives, and restricting
279 reload pseudos for one alternative may lead to situations
280 where other reload pseudos are no longer allocatable. */
281 || (INSN_UID (curr_insn) >= new_insn_uid_start
282 && curr_insn_set != NULL
283 && ((OBJECT_P (SET_SRC (curr_insn_set))
284 && ! CONSTANT_P (SET_SRC (curr_insn_set)))
285 || (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
286 && OBJECT_P (SUBREG_REG (SET_SRC (curr_insn_set)))
287 && ! CONSTANT_P (SUBREG_REG (SET_SRC (curr_insn_set)))))))
288 /* When we don't know what class will be used finally for reload
289 pseudos, we use ALL_REGS. */
290 return ((regno >= new_regno_start && rclass == ALL_REGS)
291 || (rclass != NO_REGS && ira_class_subset_p[rclass][cl]
292 && ! hard_reg_set_subset_p (reg_class_contents[cl],
293 lra_no_alloc_regs)));
294 else
296 common_class = ira_reg_class_subset[rclass][cl];
297 if (new_class != NULL)
298 *new_class = common_class;
299 if (hard_reg_set_subset_p (reg_class_contents[common_class],
300 lra_no_alloc_regs))
301 return false;
302 /* Check that there are enough allocatable regs. */
303 class_size = ira_class_hard_regs_num[common_class];
304 for (i = 0; i < class_size; i++)
306 hard_regno = ira_class_hard_regs[common_class][i];
307 nregs = hard_regno_nregs[hard_regno][reg_mode];
308 if (nregs == 1)
309 return true;
310 for (j = 0; j < nregs; j++)
311 if (TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno + j)
312 || ! TEST_HARD_REG_BIT (reg_class_contents[common_class],
313 hard_regno + j))
314 break;
315 if (j >= nregs)
316 return true;
318 return false;
322 /* Return true if REGNO satisfies a memory constraint. */
323 static bool
324 in_mem_p (int regno)
326 return get_reg_class (regno) == NO_REGS;
329 /* Return 1 if ADDR is a valid memory address for mode MODE in address
330 space AS, and check that each pseudo has the proper kind of hard
331 reg. */
332 static int
333 valid_address_p (machine_mode mode ATTRIBUTE_UNUSED,
334 rtx addr, addr_space_t as)
336 #ifdef GO_IF_LEGITIMATE_ADDRESS
337 lra_assert (ADDR_SPACE_GENERIC_P (as));
338 GO_IF_LEGITIMATE_ADDRESS (mode, addr, win);
339 return 0;
341 win:
342 return 1;
343 #else
344 return targetm.addr_space.legitimate_address_p (mode, addr, 0, as);
345 #endif
348 namespace {
349 /* Temporarily eliminates registers in an address (for the lifetime of
350 the object). */
351 class address_eliminator {
352 public:
353 address_eliminator (struct address_info *ad);
354 ~address_eliminator ();
356 private:
357 struct address_info *m_ad;
358 rtx *m_base_loc;
359 rtx m_base_reg;
360 rtx *m_index_loc;
361 rtx m_index_reg;
365 address_eliminator::address_eliminator (struct address_info *ad)
366 : m_ad (ad),
367 m_base_loc (strip_subreg (ad->base_term)),
368 m_base_reg (NULL_RTX),
369 m_index_loc (strip_subreg (ad->index_term)),
370 m_index_reg (NULL_RTX)
372 if (m_base_loc != NULL)
374 m_base_reg = *m_base_loc;
375 lra_eliminate_reg_if_possible (m_base_loc);
376 if (m_ad->base_term2 != NULL)
377 *m_ad->base_term2 = *m_ad->base_term;
379 if (m_index_loc != NULL)
381 m_index_reg = *m_index_loc;
382 lra_eliminate_reg_if_possible (m_index_loc);
386 address_eliminator::~address_eliminator ()
388 if (m_base_loc && *m_base_loc != m_base_reg)
390 *m_base_loc = m_base_reg;
391 if (m_ad->base_term2 != NULL)
392 *m_ad->base_term2 = *m_ad->base_term;
394 if (m_index_loc && *m_index_loc != m_index_reg)
395 *m_index_loc = m_index_reg;
398 /* Return true if the eliminated form of AD is a legitimate target address. */
399 static bool
400 valid_address_p (struct address_info *ad)
402 address_eliminator eliminator (ad);
403 return valid_address_p (ad->mode, *ad->outer, ad->as);
406 /* Return true if the eliminated form of memory reference OP satisfies
407 extra memory constraint CONSTRAINT. */
408 static bool
409 satisfies_memory_constraint_p (rtx op, enum constraint_num constraint)
411 struct address_info ad;
413 decompose_mem_address (&ad, op);
414 address_eliminator eliminator (&ad);
415 return constraint_satisfied_p (op, constraint);
418 /* Return true if the eliminated form of address AD satisfies extra
419 address constraint CONSTRAINT. */
420 static bool
421 satisfies_address_constraint_p (struct address_info *ad,
422 enum constraint_num constraint)
424 address_eliminator eliminator (ad);
425 return constraint_satisfied_p (*ad->outer, constraint);
428 /* Return true if the eliminated form of address OP satisfies extra
429 address constraint CONSTRAINT. */
430 static bool
431 satisfies_address_constraint_p (rtx op, enum constraint_num constraint)
433 struct address_info ad;
435 decompose_lea_address (&ad, &op);
436 return satisfies_address_constraint_p (&ad, constraint);
439 /* Initiate equivalences for LRA. As we keep original equivalences
440 before any elimination, we need to make copies otherwise any change
441 in insns might change the equivalences. */
442 void
443 lra_init_equiv (void)
445 ira_expand_reg_equiv ();
446 for (int i = FIRST_PSEUDO_REGISTER; i < max_reg_num (); i++)
448 rtx res;
450 if ((res = ira_reg_equiv[i].memory) != NULL_RTX)
451 ira_reg_equiv[i].memory = copy_rtx (res);
452 if ((res = ira_reg_equiv[i].invariant) != NULL_RTX)
453 ira_reg_equiv[i].invariant = copy_rtx (res);
457 static rtx loc_equivalence_callback (rtx, const_rtx, void *);
459 /* Update equivalence for REGNO. We need to this as the equivalence
460 might contain other pseudos which are changed by their
461 equivalences. */
462 static void
463 update_equiv (int regno)
465 rtx x;
467 if ((x = ira_reg_equiv[regno].memory) != NULL_RTX)
468 ira_reg_equiv[regno].memory
469 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
470 NULL_RTX);
471 if ((x = ira_reg_equiv[regno].invariant) != NULL_RTX)
472 ira_reg_equiv[regno].invariant
473 = simplify_replace_fn_rtx (x, NULL_RTX, loc_equivalence_callback,
474 NULL_RTX);
477 /* If we have decided to substitute X with another value, return that
478 value, otherwise return X. */
479 static rtx
480 get_equiv (rtx x)
482 int regno;
483 rtx res;
485 if (! REG_P (x) || (regno = REGNO (x)) < FIRST_PSEUDO_REGISTER
486 || ! ira_reg_equiv[regno].defined_p
487 || ! ira_reg_equiv[regno].profitable_p
488 || lra_get_regno_hard_regno (regno) >= 0)
489 return x;
490 if ((res = ira_reg_equiv[regno].memory) != NULL_RTX)
491 return res;
492 if ((res = ira_reg_equiv[regno].constant) != NULL_RTX)
493 return res;
494 if ((res = ira_reg_equiv[regno].invariant) != NULL_RTX)
495 return res;
496 gcc_unreachable ();
499 /* If we have decided to substitute X with the equivalent value,
500 return that value after elimination for INSN, otherwise return
501 X. */
502 static rtx
503 get_equiv_with_elimination (rtx x, rtx_insn *insn)
505 rtx res = get_equiv (x);
507 if (x == res || CONSTANT_P (res))
508 return res;
509 return lra_eliminate_regs_1 (insn, res, GET_MODE (res),
510 0, false, false, true);
513 /* Set up curr_operand_mode. */
514 static void
515 init_curr_operand_mode (void)
517 int nop = curr_static_id->n_operands;
518 for (int i = 0; i < nop; i++)
520 machine_mode mode = GET_MODE (*curr_id->operand_loc[i]);
521 if (mode == VOIDmode)
523 /* The .md mode for address operands is the mode of the
524 addressed value rather than the mode of the address itself. */
525 if (curr_id->icode >= 0 && curr_static_id->operand[i].is_address)
526 mode = Pmode;
527 else
528 mode = curr_static_id->operand[i].mode;
530 curr_operand_mode[i] = mode;
536 /* The page contains code to reuse input reloads. */
538 /* Structure describes input reload of the current insns. */
539 struct input_reload
541 /* Reloaded value. */
542 rtx input;
543 /* Reload pseudo used. */
544 rtx reg;
547 /* The number of elements in the following array. */
548 static int curr_insn_input_reloads_num;
549 /* Array containing info about input reloads. It is used to find the
550 same input reload and reuse the reload pseudo in this case. */
551 static struct input_reload curr_insn_input_reloads[LRA_MAX_INSN_RELOADS];
553 /* Initiate data concerning reuse of input reloads for the current
554 insn. */
555 static void
556 init_curr_insn_input_reloads (void)
558 curr_insn_input_reloads_num = 0;
561 /* Create a new pseudo using MODE, RCLASS, ORIGINAL or reuse already
562 created input reload pseudo (only if TYPE is not OP_OUT). Don't
563 reuse pseudo if IN_SUBREG_P is true and the reused pseudo should be
564 wrapped up in SUBREG. The result pseudo is returned through
565 RESULT_REG. Return TRUE if we created a new pseudo, FALSE if we
566 reused the already created input reload pseudo. Use TITLE to
567 describe new registers for debug purposes. */
568 static bool
569 get_reload_reg (enum op_type type, machine_mode mode, rtx original,
570 enum reg_class rclass, bool in_subreg_p,
571 const char *title, rtx *result_reg)
573 int i, regno;
574 enum reg_class new_class;
576 if (type == OP_OUT)
578 *result_reg
579 = lra_create_new_reg_with_unique_value (mode, original, rclass, title);
580 return true;
582 /* Prevent reuse value of expression with side effects,
583 e.g. volatile memory. */
584 if (! side_effects_p (original))
585 for (i = 0; i < curr_insn_input_reloads_num; i++)
586 if (rtx_equal_p (curr_insn_input_reloads[i].input, original)
587 && in_class_p (curr_insn_input_reloads[i].reg, rclass, &new_class))
589 rtx reg = curr_insn_input_reloads[i].reg;
590 regno = REGNO (reg);
591 /* If input is equal to original and both are VOIDmode,
592 GET_MODE (reg) might be still different from mode.
593 Ensure we don't return *result_reg with wrong mode. */
594 if (GET_MODE (reg) != mode)
596 if (in_subreg_p)
597 continue;
598 if (GET_MODE_SIZE (GET_MODE (reg)) < GET_MODE_SIZE (mode))
599 continue;
600 reg = lowpart_subreg (mode, reg, GET_MODE (reg));
601 if (reg == NULL_RTX || GET_CODE (reg) != SUBREG)
602 continue;
604 *result_reg = reg;
605 if (lra_dump_file != NULL)
607 fprintf (lra_dump_file, " Reuse r%d for reload ", regno);
608 dump_value_slim (lra_dump_file, original, 1);
610 if (new_class != lra_get_allocno_class (regno))
611 lra_change_class (regno, new_class, ", change to", false);
612 if (lra_dump_file != NULL)
613 fprintf (lra_dump_file, "\n");
614 return false;
616 *result_reg = lra_create_new_reg (mode, original, rclass, title);
617 lra_assert (curr_insn_input_reloads_num < LRA_MAX_INSN_RELOADS);
618 curr_insn_input_reloads[curr_insn_input_reloads_num].input = original;
619 curr_insn_input_reloads[curr_insn_input_reloads_num++].reg = *result_reg;
620 return true;
625 /* The page contains code to extract memory address parts. */
627 /* Wrapper around REGNO_OK_FOR_INDEX_P, to allow pseudos. */
628 static inline bool
629 ok_for_index_p_nonstrict (rtx reg)
631 unsigned regno = REGNO (reg);
633 return regno >= FIRST_PSEUDO_REGISTER || REGNO_OK_FOR_INDEX_P (regno);
636 /* A version of regno_ok_for_base_p for use here, when all pseudos
637 should count as OK. Arguments as for regno_ok_for_base_p. */
638 static inline bool
639 ok_for_base_p_nonstrict (rtx reg, machine_mode mode, addr_space_t as,
640 enum rtx_code outer_code, enum rtx_code index_code)
642 unsigned regno = REGNO (reg);
644 if (regno >= FIRST_PSEUDO_REGISTER)
645 return true;
646 return ok_for_base_p_1 (regno, mode, as, outer_code, index_code);
651 /* The page contains major code to choose the current insn alternative
652 and generate reloads for it. */
654 /* Return the offset from REGNO of the least significant register
655 in (reg:MODE REGNO).
657 This function is used to tell whether two registers satisfy
658 a matching constraint. (reg:MODE1 REGNO1) matches (reg:MODE2 REGNO2) if:
660 REGNO1 + lra_constraint_offset (REGNO1, MODE1)
661 == REGNO2 + lra_constraint_offset (REGNO2, MODE2) */
663 lra_constraint_offset (int regno, machine_mode mode)
665 lra_assert (regno < FIRST_PSEUDO_REGISTER);
666 if (WORDS_BIG_ENDIAN && GET_MODE_SIZE (mode) > UNITS_PER_WORD
667 && SCALAR_INT_MODE_P (mode))
668 return hard_regno_nregs[regno][mode] - 1;
669 return 0;
672 /* Like rtx_equal_p except that it allows a REG and a SUBREG to match
673 if they are the same hard reg, and has special hacks for
674 auto-increment and auto-decrement. This is specifically intended for
675 process_alt_operands to use in determining whether two operands
676 match. X is the operand whose number is the lower of the two.
678 It is supposed that X is the output operand and Y is the input
679 operand. Y_HARD_REGNO is the final hard regno of register Y or
680 register in subreg Y as we know it now. Otherwise, it is a
681 negative value. */
682 static bool
683 operands_match_p (rtx x, rtx y, int y_hard_regno)
685 int i;
686 RTX_CODE code = GET_CODE (x);
687 const char *fmt;
689 if (x == y)
690 return true;
691 if ((code == REG || (code == SUBREG && REG_P (SUBREG_REG (x))))
692 && (REG_P (y) || (GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y)))))
694 int j;
696 i = get_hard_regno (x);
697 if (i < 0)
698 goto slow;
700 if ((j = y_hard_regno) < 0)
701 goto slow;
703 i += lra_constraint_offset (i, GET_MODE (x));
704 j += lra_constraint_offset (j, GET_MODE (y));
706 return i == j;
709 /* If two operands must match, because they are really a single
710 operand of an assembler insn, then two post-increments are invalid
711 because the assembler insn would increment only once. On the
712 other hand, a post-increment matches ordinary indexing if the
713 post-increment is the output operand. */
714 if (code == POST_DEC || code == POST_INC || code == POST_MODIFY)
715 return operands_match_p (XEXP (x, 0), y, y_hard_regno);
717 /* Two pre-increments are invalid because the assembler insn would
718 increment only once. On the other hand, a pre-increment matches
719 ordinary indexing if the pre-increment is the input operand. */
720 if (GET_CODE (y) == PRE_DEC || GET_CODE (y) == PRE_INC
721 || GET_CODE (y) == PRE_MODIFY)
722 return operands_match_p (x, XEXP (y, 0), -1);
724 slow:
726 if (code == REG && GET_CODE (y) == SUBREG && REG_P (SUBREG_REG (y))
727 && x == SUBREG_REG (y))
728 return true;
729 if (GET_CODE (y) == REG && code == SUBREG && REG_P (SUBREG_REG (x))
730 && SUBREG_REG (x) == y)
731 return true;
733 /* Now we have disposed of all the cases in which different rtx
734 codes can match. */
735 if (code != GET_CODE (y))
736 return false;
738 /* (MULT:SI x y) and (MULT:HI x y) are NOT equivalent. */
739 if (GET_MODE (x) != GET_MODE (y))
740 return false;
742 switch (code)
744 CASE_CONST_UNIQUE:
745 return false;
747 case LABEL_REF:
748 return LABEL_REF_LABEL (x) == LABEL_REF_LABEL (y);
749 case SYMBOL_REF:
750 return XSTR (x, 0) == XSTR (y, 0);
752 default:
753 break;
756 /* Compare the elements. If any pair of corresponding elements fail
757 to match, return false for the whole things. */
759 fmt = GET_RTX_FORMAT (code);
760 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
762 int val, j;
763 switch (fmt[i])
765 case 'w':
766 if (XWINT (x, i) != XWINT (y, i))
767 return false;
768 break;
770 case 'i':
771 if (XINT (x, i) != XINT (y, i))
772 return false;
773 break;
775 case 'e':
776 val = operands_match_p (XEXP (x, i), XEXP (y, i), -1);
777 if (val == 0)
778 return false;
779 break;
781 case '0':
782 break;
784 case 'E':
785 if (XVECLEN (x, i) != XVECLEN (y, i))
786 return false;
787 for (j = XVECLEN (x, i) - 1; j >= 0; --j)
789 val = operands_match_p (XVECEXP (x, i, j), XVECEXP (y, i, j), -1);
790 if (val == 0)
791 return false;
793 break;
795 /* It is believed that rtx's at this level will never
796 contain anything but integers and other rtx's, except for
797 within LABEL_REFs and SYMBOL_REFs. */
798 default:
799 gcc_unreachable ();
802 return true;
805 /* True if X is a constant that can be forced into the constant pool.
806 MODE is the mode of the operand, or VOIDmode if not known. */
807 #define CONST_POOL_OK_P(MODE, X) \
808 ((MODE) != VOIDmode \
809 && CONSTANT_P (X) \
810 && GET_CODE (X) != HIGH \
811 && !targetm.cannot_force_const_mem (MODE, X))
813 /* True if C is a non-empty register class that has too few registers
814 to be safely used as a reload target class. */
815 #define SMALL_REGISTER_CLASS_P(C) \
816 (ira_class_hard_regs_num [(C)] == 1 \
817 || (ira_class_hard_regs_num [(C)] >= 1 \
818 && targetm.class_likely_spilled_p (C)))
820 /* If REG is a reload pseudo, try to make its class satisfying CL. */
821 static void
822 narrow_reload_pseudo_class (rtx reg, enum reg_class cl)
824 enum reg_class rclass;
826 /* Do not make more accurate class from reloads generated. They are
827 mostly moves with a lot of constraints. Making more accurate
828 class may results in very narrow class and impossibility of find
829 registers for several reloads of one insn. */
830 if (INSN_UID (curr_insn) >= new_insn_uid_start)
831 return;
832 if (GET_CODE (reg) == SUBREG)
833 reg = SUBREG_REG (reg);
834 if (! REG_P (reg) || (int) REGNO (reg) < new_regno_start)
835 return;
836 if (in_class_p (reg, cl, &rclass) && rclass != cl)
837 lra_change_class (REGNO (reg), rclass, " Change to", true);
840 /* Generate reloads for matching OUT and INS (array of input operand
841 numbers with end marker -1) with reg class GOAL_CLASS. Add input
842 and output reloads correspondingly to the lists *BEFORE and *AFTER.
843 OUT might be negative. In this case we generate input reloads for
844 matched input operands INS. */
845 static void
846 match_reload (signed char out, signed char *ins, enum reg_class goal_class,
847 rtx_insn **before, rtx_insn **after)
849 int i, in;
850 rtx new_in_reg, new_out_reg, reg, clobber;
851 machine_mode inmode, outmode;
852 rtx in_rtx = *curr_id->operand_loc[ins[0]];
853 rtx out_rtx = out < 0 ? in_rtx : *curr_id->operand_loc[out];
855 inmode = curr_operand_mode[ins[0]];
856 outmode = out < 0 ? inmode : curr_operand_mode[out];
857 push_to_sequence (*before);
858 if (inmode != outmode)
860 if (GET_MODE_SIZE (inmode) > GET_MODE_SIZE (outmode))
862 reg = new_in_reg
863 = lra_create_new_reg_with_unique_value (inmode, in_rtx,
864 goal_class, "");
865 if (SCALAR_INT_MODE_P (inmode))
866 new_out_reg = gen_lowpart_SUBREG (outmode, reg);
867 else
868 new_out_reg = gen_rtx_SUBREG (outmode, reg, 0);
869 LRA_SUBREG_P (new_out_reg) = 1;
870 /* If the input reg is dying here, we can use the same hard
871 register for REG and IN_RTX. We do it only for original
872 pseudos as reload pseudos can die although original
873 pseudos still live where reload pseudos dies. */
874 if (REG_P (in_rtx) && (int) REGNO (in_rtx) < lra_new_regno_start
875 && find_regno_note (curr_insn, REG_DEAD, REGNO (in_rtx)))
876 lra_assign_reg_val (REGNO (in_rtx), REGNO (reg));
878 else
880 reg = new_out_reg
881 = lra_create_new_reg_with_unique_value (outmode, out_rtx,
882 goal_class, "");
883 if (SCALAR_INT_MODE_P (outmode))
884 new_in_reg = gen_lowpart_SUBREG (inmode, reg);
885 else
886 new_in_reg = gen_rtx_SUBREG (inmode, reg, 0);
887 /* NEW_IN_REG is non-paradoxical subreg. We don't want
888 NEW_OUT_REG living above. We add clobber clause for
889 this. This is just a temporary clobber. We can remove
890 it at the end of LRA work. */
891 clobber = emit_clobber (new_out_reg);
892 LRA_TEMP_CLOBBER_P (PATTERN (clobber)) = 1;
893 LRA_SUBREG_P (new_in_reg) = 1;
894 if (GET_CODE (in_rtx) == SUBREG)
896 rtx subreg_reg = SUBREG_REG (in_rtx);
898 /* If SUBREG_REG is dying here and sub-registers IN_RTX
899 and NEW_IN_REG are similar, we can use the same hard
900 register for REG and SUBREG_REG. */
901 if (REG_P (subreg_reg)
902 && (int) REGNO (subreg_reg) < lra_new_regno_start
903 && GET_MODE (subreg_reg) == outmode
904 && SUBREG_BYTE (in_rtx) == SUBREG_BYTE (new_in_reg)
905 && find_regno_note (curr_insn, REG_DEAD, REGNO (subreg_reg)))
906 lra_assign_reg_val (REGNO (subreg_reg), REGNO (reg));
910 else
912 /* Pseudos have values -- see comments for lra_reg_info.
913 Different pseudos with the same value do not conflict even if
914 they live in the same place. When we create a pseudo we
915 assign value of original pseudo (if any) from which we
916 created the new pseudo. If we create the pseudo from the
917 input pseudo, the new pseudo will no conflict with the input
918 pseudo which is wrong when the input pseudo lives after the
919 insn and as the new pseudo value is changed by the insn
920 output. Therefore we create the new pseudo from the output.
922 We cannot reuse the current output register because we might
923 have a situation like "a <- a op b", where the constraints
924 force the second input operand ("b") to match the output
925 operand ("a"). "b" must then be copied into a new register
926 so that it doesn't clobber the current value of "a". */
928 new_in_reg = new_out_reg
929 = lra_create_new_reg_with_unique_value (outmode, out_rtx,
930 goal_class, "");
932 /* In operand can be got from transformations before processing insn
933 constraints. One example of such transformations is subreg
934 reloading (see function simplify_operand_subreg). The new
935 pseudos created by the transformations might have inaccurate
936 class (ALL_REGS) and we should make their classes more
937 accurate. */
938 narrow_reload_pseudo_class (in_rtx, goal_class);
939 lra_emit_move (copy_rtx (new_in_reg), in_rtx);
940 *before = get_insns ();
941 end_sequence ();
942 for (i = 0; (in = ins[i]) >= 0; i++)
944 lra_assert
945 (GET_MODE (*curr_id->operand_loc[in]) == VOIDmode
946 || GET_MODE (new_in_reg) == GET_MODE (*curr_id->operand_loc[in]));
947 *curr_id->operand_loc[in] = new_in_reg;
949 lra_update_dups (curr_id, ins);
950 if (out < 0)
951 return;
952 /* See a comment for the input operand above. */
953 narrow_reload_pseudo_class (out_rtx, goal_class);
954 if (find_reg_note (curr_insn, REG_UNUSED, out_rtx) == NULL_RTX)
956 start_sequence ();
957 lra_emit_move (out_rtx, copy_rtx (new_out_reg));
958 emit_insn (*after);
959 *after = get_insns ();
960 end_sequence ();
962 *curr_id->operand_loc[out] = new_out_reg;
963 lra_update_dup (curr_id, out);
966 /* Return register class which is union of all reg classes in insn
967 constraint alternative string starting with P. */
968 static enum reg_class
969 reg_class_from_constraints (const char *p)
971 int c, len;
972 enum reg_class op_class = NO_REGS;
975 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
977 case '#':
978 case ',':
979 return op_class;
981 case 'g':
982 op_class = reg_class_subunion[op_class][GENERAL_REGS];
983 break;
985 default:
986 enum constraint_num cn = lookup_constraint (p);
987 enum reg_class cl = reg_class_for_constraint (cn);
988 if (cl == NO_REGS)
990 if (insn_extra_address_constraint (cn))
991 op_class
992 = (reg_class_subunion
993 [op_class][base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
994 ADDRESS, SCRATCH)]);
995 break;
998 op_class = reg_class_subunion[op_class][cl];
999 break;
1001 while ((p += len), c);
1002 return op_class;
1005 /* If OP is a register, return the class of the register as per
1006 get_reg_class, otherwise return NO_REGS. */
1007 static inline enum reg_class
1008 get_op_class (rtx op)
1010 return REG_P (op) ? get_reg_class (REGNO (op)) : NO_REGS;
1013 /* Return generated insn mem_pseudo:=val if TO_P or val:=mem_pseudo
1014 otherwise. If modes of MEM_PSEUDO and VAL are different, use
1015 SUBREG for VAL to make them equal. */
1016 static rtx_insn *
1017 emit_spill_move (bool to_p, rtx mem_pseudo, rtx val)
1019 if (GET_MODE (mem_pseudo) != GET_MODE (val))
1021 /* Usually size of mem_pseudo is greater than val size but in
1022 rare cases it can be less as it can be defined by target
1023 dependent macro HARD_REGNO_CALLER_SAVE_MODE. */
1024 if (! MEM_P (val))
1026 val = gen_rtx_SUBREG (GET_MODE (mem_pseudo),
1027 GET_CODE (val) == SUBREG ? SUBREG_REG (val) : val,
1029 LRA_SUBREG_P (val) = 1;
1031 else
1033 mem_pseudo = gen_lowpart_SUBREG (GET_MODE (val), mem_pseudo);
1034 LRA_SUBREG_P (mem_pseudo) = 1;
1037 return as_a <rtx_insn *> (to_p
1038 ? gen_move_insn (mem_pseudo, val)
1039 : gen_move_insn (val, mem_pseudo));
1042 /* Process a special case insn (register move), return true if we
1043 don't need to process it anymore. INSN should be a single set
1044 insn. Set up that RTL was changed through CHANGE_P and macro
1045 SECONDARY_MEMORY_NEEDED says to use secondary memory through
1046 SEC_MEM_P. */
1047 static bool
1048 check_and_process_move (bool *change_p, bool *sec_mem_p ATTRIBUTE_UNUSED)
1050 int sregno, dregno;
1051 rtx dest, src, dreg, sreg, old_sreg, new_reg, scratch_reg;
1052 rtx_insn *before;
1053 enum reg_class dclass, sclass, secondary_class;
1054 machine_mode sreg_mode;
1055 secondary_reload_info sri;
1057 lra_assert (curr_insn_set != NULL_RTX);
1058 dreg = dest = SET_DEST (curr_insn_set);
1059 sreg = src = SET_SRC (curr_insn_set);
1060 if (GET_CODE (dest) == SUBREG)
1061 dreg = SUBREG_REG (dest);
1062 if (GET_CODE (src) == SUBREG)
1063 sreg = SUBREG_REG (src);
1064 if (! (REG_P (dreg) || MEM_P (dreg)) || ! (REG_P (sreg) || MEM_P (sreg)))
1065 return false;
1066 sclass = dclass = NO_REGS;
1067 if (REG_P (dreg))
1068 dclass = get_reg_class (REGNO (dreg));
1069 if (dclass == ALL_REGS)
1070 /* ALL_REGS is used for new pseudos created by transformations
1071 like reload of SUBREG_REG (see function
1072 simplify_operand_subreg). We don't know their class yet. We
1073 should figure out the class from processing the insn
1074 constraints not in this fast path function. Even if ALL_REGS
1075 were a right class for the pseudo, secondary_... hooks usually
1076 are not define for ALL_REGS. */
1077 return false;
1078 sreg_mode = GET_MODE (sreg);
1079 old_sreg = sreg;
1080 if (REG_P (sreg))
1081 sclass = get_reg_class (REGNO (sreg));
1082 if (sclass == ALL_REGS)
1083 /* See comments above. */
1084 return false;
1085 if (sclass == NO_REGS && dclass == NO_REGS)
1086 return false;
1087 #ifdef SECONDARY_MEMORY_NEEDED
1088 if (SECONDARY_MEMORY_NEEDED (sclass, dclass, GET_MODE (src))
1089 #ifdef SECONDARY_MEMORY_NEEDED_MODE
1090 && ((sclass != NO_REGS && dclass != NO_REGS)
1091 || GET_MODE (src) != SECONDARY_MEMORY_NEEDED_MODE (GET_MODE (src)))
1092 #endif
1095 *sec_mem_p = true;
1096 return false;
1098 #endif
1099 if (! REG_P (dreg) || ! REG_P (sreg))
1100 return false;
1101 sri.prev_sri = NULL;
1102 sri.icode = CODE_FOR_nothing;
1103 sri.extra_cost = 0;
1104 secondary_class = NO_REGS;
1105 /* Set up hard register for a reload pseudo for hook
1106 secondary_reload because some targets just ignore unassigned
1107 pseudos in the hook. */
1108 if (dclass != NO_REGS && lra_get_regno_hard_regno (REGNO (dreg)) < 0)
1110 dregno = REGNO (dreg);
1111 reg_renumber[dregno] = ira_class_hard_regs[dclass][0];
1113 else
1114 dregno = -1;
1115 if (sclass != NO_REGS && lra_get_regno_hard_regno (REGNO (sreg)) < 0)
1117 sregno = REGNO (sreg);
1118 reg_renumber[sregno] = ira_class_hard_regs[sclass][0];
1120 else
1121 sregno = -1;
1122 if (sclass != NO_REGS)
1123 secondary_class
1124 = (enum reg_class) targetm.secondary_reload (false, dest,
1125 (reg_class_t) sclass,
1126 GET_MODE (src), &sri);
1127 if (sclass == NO_REGS
1128 || ((secondary_class != NO_REGS || sri.icode != CODE_FOR_nothing)
1129 && dclass != NO_REGS))
1131 enum reg_class old_sclass = secondary_class;
1132 secondary_reload_info old_sri = sri;
1134 sri.prev_sri = NULL;
1135 sri.icode = CODE_FOR_nothing;
1136 sri.extra_cost = 0;
1137 secondary_class
1138 = (enum reg_class) targetm.secondary_reload (true, sreg,
1139 (reg_class_t) dclass,
1140 sreg_mode, &sri);
1141 /* Check the target hook consistency. */
1142 lra_assert
1143 ((secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1144 || (old_sclass == NO_REGS && old_sri.icode == CODE_FOR_nothing)
1145 || (secondary_class == old_sclass && sri.icode == old_sri.icode));
1147 if (sregno >= 0)
1148 reg_renumber [sregno] = -1;
1149 if (dregno >= 0)
1150 reg_renumber [dregno] = -1;
1151 if (secondary_class == NO_REGS && sri.icode == CODE_FOR_nothing)
1152 return false;
1153 *change_p = true;
1154 new_reg = NULL_RTX;
1155 if (secondary_class != NO_REGS)
1156 new_reg = lra_create_new_reg_with_unique_value (sreg_mode, NULL_RTX,
1157 secondary_class,
1158 "secondary");
1159 start_sequence ();
1160 if (old_sreg != sreg)
1161 sreg = copy_rtx (sreg);
1162 if (sri.icode == CODE_FOR_nothing)
1163 lra_emit_move (new_reg, sreg);
1164 else
1166 enum reg_class scratch_class;
1168 scratch_class = (reg_class_from_constraints
1169 (insn_data[sri.icode].operand[2].constraint));
1170 scratch_reg = (lra_create_new_reg_with_unique_value
1171 (insn_data[sri.icode].operand[2].mode, NULL_RTX,
1172 scratch_class, "scratch"));
1173 emit_insn (GEN_FCN (sri.icode) (new_reg != NULL_RTX ? new_reg : dest,
1174 sreg, scratch_reg));
1176 before = get_insns ();
1177 end_sequence ();
1178 lra_process_new_insns (curr_insn, before, NULL, "Inserting the move");
1179 if (new_reg != NULL_RTX)
1181 if (GET_CODE (src) == SUBREG)
1182 SUBREG_REG (src) = new_reg;
1183 else
1184 SET_SRC (curr_insn_set) = new_reg;
1186 else
1188 if (lra_dump_file != NULL)
1190 fprintf (lra_dump_file, "Deleting move %u\n", INSN_UID (curr_insn));
1191 dump_insn_slim (lra_dump_file, curr_insn);
1193 lra_set_insn_deleted (curr_insn);
1194 return true;
1196 return false;
1199 /* The following data describe the result of process_alt_operands.
1200 The data are used in curr_insn_transform to generate reloads. */
1202 /* The chosen reg classes which should be used for the corresponding
1203 operands. */
1204 static enum reg_class goal_alt[MAX_RECOG_OPERANDS];
1205 /* True if the operand should be the same as another operand and that
1206 other operand does not need a reload. */
1207 static bool goal_alt_match_win[MAX_RECOG_OPERANDS];
1208 /* True if the operand does not need a reload. */
1209 static bool goal_alt_win[MAX_RECOG_OPERANDS];
1210 /* True if the operand can be offsetable memory. */
1211 static bool goal_alt_offmemok[MAX_RECOG_OPERANDS];
1212 /* The number of an operand to which given operand can be matched to. */
1213 static int goal_alt_matches[MAX_RECOG_OPERANDS];
1214 /* The number of elements in the following array. */
1215 static int goal_alt_dont_inherit_ops_num;
1216 /* Numbers of operands whose reload pseudos should not be inherited. */
1217 static int goal_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1218 /* True if the insn commutative operands should be swapped. */
1219 static bool goal_alt_swapped;
1220 /* The chosen insn alternative. */
1221 static int goal_alt_number;
1223 /* The following five variables are used to choose the best insn
1224 alternative. They reflect final characteristics of the best
1225 alternative. */
1227 /* Number of necessary reloads and overall cost reflecting the
1228 previous value and other unpleasantness of the best alternative. */
1229 static int best_losers, best_overall;
1230 /* Overall number hard registers used for reloads. For example, on
1231 some targets we need 2 general registers to reload DFmode and only
1232 one floating point register. */
1233 static int best_reload_nregs;
1234 /* Overall number reflecting distances of previous reloading the same
1235 value. The distances are counted from the current BB start. It is
1236 used to improve inheritance chances. */
1237 static int best_reload_sum;
1239 /* True if the current insn should have no correspondingly input or
1240 output reloads. */
1241 static bool no_input_reloads_p, no_output_reloads_p;
1243 /* True if we swapped the commutative operands in the current
1244 insn. */
1245 static int curr_swapped;
1247 /* if CHECK_ONLY_P is false, arrange for address element *LOC to be a
1248 register of class CL. Add any input reloads to list BEFORE. AFTER
1249 is nonnull if *LOC is an automodified value; handle that case by
1250 adding the required output reloads to list AFTER. Return true if
1251 the RTL was changed.
1253 if CHECK_ONLY_P is true, check that the *LOC is a correct address
1254 register. Return false if the address register is correct. */
1255 static bool
1256 process_addr_reg (rtx *loc, bool check_only_p, rtx_insn **before, rtx_insn **after,
1257 enum reg_class cl)
1259 int regno;
1260 enum reg_class rclass, new_class;
1261 rtx reg;
1262 rtx new_reg;
1263 machine_mode mode;
1264 bool subreg_p, before_p = false;
1266 subreg_p = GET_CODE (*loc) == SUBREG;
1267 if (subreg_p)
1268 loc = &SUBREG_REG (*loc);
1269 reg = *loc;
1270 mode = GET_MODE (reg);
1271 if (! REG_P (reg))
1273 if (check_only_p)
1274 return true;
1275 /* Always reload memory in an address even if the target supports
1276 such addresses. */
1277 new_reg = lra_create_new_reg_with_unique_value (mode, reg, cl, "address");
1278 before_p = true;
1280 else
1282 regno = REGNO (reg);
1283 rclass = get_reg_class (regno);
1284 if (! check_only_p
1285 && (*loc = get_equiv_with_elimination (reg, curr_insn)) != reg)
1287 if (lra_dump_file != NULL)
1289 fprintf (lra_dump_file,
1290 "Changing pseudo %d in address of insn %u on equiv ",
1291 REGNO (reg), INSN_UID (curr_insn));
1292 dump_value_slim (lra_dump_file, *loc, 1);
1293 fprintf (lra_dump_file, "\n");
1295 *loc = copy_rtx (*loc);
1297 if (*loc != reg || ! in_class_p (reg, cl, &new_class))
1299 if (check_only_p)
1300 return true;
1301 reg = *loc;
1302 if (get_reload_reg (after == NULL ? OP_IN : OP_INOUT,
1303 mode, reg, cl, subreg_p, "address", &new_reg))
1304 before_p = true;
1306 else if (new_class != NO_REGS && rclass != new_class)
1308 if (check_only_p)
1309 return true;
1310 lra_change_class (regno, new_class, " Change to", true);
1311 return false;
1313 else
1314 return false;
1316 if (before_p)
1318 push_to_sequence (*before);
1319 lra_emit_move (new_reg, reg);
1320 *before = get_insns ();
1321 end_sequence ();
1323 *loc = new_reg;
1324 if (after != NULL)
1326 start_sequence ();
1327 lra_emit_move (reg, new_reg);
1328 emit_insn (*after);
1329 *after = get_insns ();
1330 end_sequence ();
1332 return true;
1335 /* Insert move insn in simplify_operand_subreg. BEFORE returns
1336 the insn to be inserted before curr insn. AFTER returns the
1337 the insn to be inserted after curr insn. ORIGREG and NEWREG
1338 are the original reg and new reg for reload. */
1339 static void
1340 insert_move_for_subreg (rtx_insn **before, rtx_insn **after, rtx origreg,
1341 rtx newreg)
1343 if (before)
1345 push_to_sequence (*before);
1346 lra_emit_move (newreg, origreg);
1347 *before = get_insns ();
1348 end_sequence ();
1350 if (after)
1352 start_sequence ();
1353 lra_emit_move (origreg, newreg);
1354 emit_insn (*after);
1355 *after = get_insns ();
1356 end_sequence ();
1360 static int valid_address_p (machine_mode mode, rtx addr, addr_space_t as);
1362 /* Make reloads for subreg in operand NOP with internal subreg mode
1363 REG_MODE, add new reloads for further processing. Return true if
1364 any reload was generated. */
1365 static bool
1366 simplify_operand_subreg (int nop, machine_mode reg_mode)
1368 int hard_regno;
1369 rtx_insn *before, *after;
1370 machine_mode mode;
1371 rtx reg, new_reg;
1372 rtx operand = *curr_id->operand_loc[nop];
1373 enum reg_class regclass;
1374 enum op_type type;
1376 before = after = NULL;
1378 if (GET_CODE (operand) != SUBREG)
1379 return false;
1381 mode = GET_MODE (operand);
1382 reg = SUBREG_REG (operand);
1383 type = curr_static_id->operand[nop].type;
1384 /* If we change address for paradoxical subreg of memory, the
1385 address might violate the necessary alignment or the access might
1386 be slow. So take this into consideration. We should not worry
1387 about access beyond allocated memory for paradoxical memory
1388 subregs as we don't substitute such equiv memory (see processing
1389 equivalences in function lra_constraints) and because for spilled
1390 pseudos we allocate stack memory enough for the biggest
1391 corresponding paradoxical subreg. */
1392 if (MEM_P (reg)
1393 && (! SLOW_UNALIGNED_ACCESS (mode, MEM_ALIGN (reg))
1394 || MEM_ALIGN (reg) >= GET_MODE_ALIGNMENT (mode)))
1396 rtx subst, old = *curr_id->operand_loc[nop];
1398 alter_subreg (curr_id->operand_loc[nop], false);
1399 subst = *curr_id->operand_loc[nop];
1400 lra_assert (MEM_P (subst));
1401 if (! valid_address_p (GET_MODE (reg), XEXP (reg, 0),
1402 MEM_ADDR_SPACE (reg))
1403 || valid_address_p (GET_MODE (subst), XEXP (subst, 0),
1404 MEM_ADDR_SPACE (subst)))
1405 return true;
1406 /* If the address was valid and became invalid, prefer to reload
1407 the memory. Typical case is when the index scale should
1408 correspond the memory. */
1409 *curr_id->operand_loc[nop] = old;
1411 else if (REG_P (reg) && REGNO (reg) < FIRST_PSEUDO_REGISTER)
1413 alter_subreg (curr_id->operand_loc[nop], false);
1414 return true;
1416 /* Put constant into memory when we have mixed modes. It generates
1417 a better code in most cases as it does not need a secondary
1418 reload memory. It also prevents LRA looping when LRA is using
1419 secondary reload memory again and again. */
1420 if (CONSTANT_P (reg) && CONST_POOL_OK_P (reg_mode, reg)
1421 && SCALAR_INT_MODE_P (reg_mode) != SCALAR_INT_MODE_P (mode))
1423 SUBREG_REG (operand) = force_const_mem (reg_mode, reg);
1424 alter_subreg (curr_id->operand_loc[nop], false);
1425 return true;
1427 /* Force a reload of the SUBREG_REG if this is a constant or PLUS or
1428 if there may be a problem accessing OPERAND in the outer
1429 mode. */
1430 if ((REG_P (reg)
1431 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1432 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1433 /* Don't reload paradoxical subregs because we could be looping
1434 having repeatedly final regno out of hard regs range. */
1435 && (hard_regno_nregs[hard_regno][GET_MODE (reg)]
1436 >= hard_regno_nregs[hard_regno][mode])
1437 && simplify_subreg_regno (hard_regno, GET_MODE (reg),
1438 SUBREG_BYTE (operand), mode) < 0
1439 /* Don't reload subreg for matching reload. It is actually
1440 valid subreg in LRA. */
1441 && ! LRA_SUBREG_P (operand))
1442 || CONSTANT_P (reg) || GET_CODE (reg) == PLUS || MEM_P (reg))
1444 enum reg_class rclass;
1446 if (REG_P (reg))
1447 /* There is a big probability that we will get the same class
1448 for the new pseudo and we will get the same insn which
1449 means infinite looping. So spill the new pseudo. */
1450 rclass = NO_REGS;
1451 else
1452 /* The class will be defined later in curr_insn_transform. */
1453 rclass
1454 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1456 if (get_reload_reg (curr_static_id->operand[nop].type, reg_mode, reg,
1457 rclass, TRUE, "subreg reg", &new_reg))
1459 bool insert_before, insert_after;
1460 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1462 insert_before = (type != OP_OUT
1463 || GET_MODE_SIZE (GET_MODE (reg)) > GET_MODE_SIZE (mode));
1464 insert_after = (type != OP_IN);
1465 insert_move_for_subreg (insert_before ? &before : NULL,
1466 insert_after ? &after : NULL,
1467 reg, new_reg);
1469 SUBREG_REG (operand) = new_reg;
1470 lra_process_new_insns (curr_insn, before, after,
1471 "Inserting subreg reload");
1472 return true;
1474 /* Force a reload for a paradoxical subreg. For paradoxical subreg,
1475 IRA allocates hardreg to the inner pseudo reg according to its mode
1476 instead of the outermode, so the size of the hardreg may not be enough
1477 to contain the outermode operand, in that case we may need to insert
1478 reload for the reg. For the following two types of paradoxical subreg,
1479 we need to insert reload:
1480 1. If the op_type is OP_IN, and the hardreg could not be paired with
1481 other hardreg to contain the outermode operand
1482 (checked by in_hard_reg_set_p), we need to insert the reload.
1483 2. If the op_type is OP_OUT or OP_INOUT.
1485 Here is a paradoxical subreg example showing how the reload is generated:
1487 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1488 (subreg:TI (reg:DI 107 [ __comp ]) 0)) {*movti_internal_rex64}
1490 In IRA, reg107 is allocated to a DImode hardreg. We use x86-64 as example
1491 here, if reg107 is assigned to hardreg R15, because R15 is the last
1492 hardreg, compiler cannot find another hardreg to pair with R15 to
1493 contain TImode data. So we insert a TImode reload reg180 for it.
1494 After reload is inserted:
1496 (insn 283 0 0 (set (subreg:DI (reg:TI 180 [orig:107 __comp ] [107]) 0)
1497 (reg:DI 107 [ __comp ])) -1
1498 (insn 5 4 7 2 (set (reg:TI 106 [ __comp ])
1499 (subreg:TI (reg:TI 180 [orig:107 __comp ] [107]) 0)) {*movti_internal_rex64}
1501 Two reload hard registers will be allocated to reg180 to save TImode data
1502 in LRA_assign. */
1503 else if (REG_P (reg)
1504 && REGNO (reg) >= FIRST_PSEUDO_REGISTER
1505 && (hard_regno = lra_get_regno_hard_regno (REGNO (reg))) >= 0
1506 && (hard_regno_nregs[hard_regno][GET_MODE (reg)]
1507 < hard_regno_nregs[hard_regno][mode])
1508 && (regclass = lra_get_allocno_class (REGNO (reg)))
1509 && (type != OP_IN
1510 || !in_hard_reg_set_p (reg_class_contents[regclass],
1511 mode, hard_regno)))
1513 /* The class will be defined later in curr_insn_transform. */
1514 enum reg_class rclass
1515 = (enum reg_class) targetm.preferred_reload_class (reg, ALL_REGS);
1517 if (get_reload_reg (curr_static_id->operand[nop].type, mode, reg,
1518 rclass, TRUE, "paradoxical subreg", &new_reg))
1520 rtx subreg;
1521 bool insert_before, insert_after;
1523 PUT_MODE (new_reg, mode);
1524 subreg = simplify_gen_subreg (GET_MODE (reg), new_reg, mode, 0);
1525 bitmap_set_bit (&lra_subreg_reload_pseudos, REGNO (new_reg));
1527 insert_before = (type != OP_OUT);
1528 insert_after = (type != OP_IN);
1529 insert_move_for_subreg (insert_before ? &before : NULL,
1530 insert_after ? &after : NULL,
1531 reg, subreg);
1533 SUBREG_REG (operand) = new_reg;
1534 lra_process_new_insns (curr_insn, before, after,
1535 "Inserting paradoxical subreg reload");
1536 return true;
1538 return false;
1541 /* Return TRUE if X refers for a hard register from SET. */
1542 static bool
1543 uses_hard_regs_p (rtx x, HARD_REG_SET set)
1545 int i, j, x_hard_regno;
1546 machine_mode mode;
1547 const char *fmt;
1548 enum rtx_code code;
1550 if (x == NULL_RTX)
1551 return false;
1552 code = GET_CODE (x);
1553 mode = GET_MODE (x);
1554 if (code == SUBREG)
1556 x = SUBREG_REG (x);
1557 code = GET_CODE (x);
1558 if (GET_MODE_SIZE (GET_MODE (x)) > GET_MODE_SIZE (mode))
1559 mode = GET_MODE (x);
1562 if (REG_P (x))
1564 x_hard_regno = get_hard_regno (x);
1565 return (x_hard_regno >= 0
1566 && overlaps_hard_reg_set_p (set, mode, x_hard_regno));
1568 if (MEM_P (x))
1570 struct address_info ad;
1572 decompose_mem_address (&ad, x);
1573 if (ad.base_term != NULL && uses_hard_regs_p (*ad.base_term, set))
1574 return true;
1575 if (ad.index_term != NULL && uses_hard_regs_p (*ad.index_term, set))
1576 return true;
1578 fmt = GET_RTX_FORMAT (code);
1579 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
1581 if (fmt[i] == 'e')
1583 if (uses_hard_regs_p (XEXP (x, i), set))
1584 return true;
1586 else if (fmt[i] == 'E')
1588 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
1589 if (uses_hard_regs_p (XVECEXP (x, i, j), set))
1590 return true;
1593 return false;
1596 /* Return true if OP is a spilled pseudo. */
1597 static inline bool
1598 spilled_pseudo_p (rtx op)
1600 return (REG_P (op)
1601 && REGNO (op) >= FIRST_PSEUDO_REGISTER && in_mem_p (REGNO (op)));
1604 /* Return true if X is a general constant. */
1605 static inline bool
1606 general_constant_p (rtx x)
1608 return CONSTANT_P (x) && (! flag_pic || LEGITIMATE_PIC_OPERAND_P (x));
1611 static bool
1612 reg_in_class_p (rtx reg, enum reg_class cl)
1614 if (cl == NO_REGS)
1615 return get_reg_class (REGNO (reg)) == NO_REGS;
1616 return in_class_p (reg, cl, NULL);
1619 /* Major function to choose the current insn alternative and what
1620 operands should be reloaded and how. If ONLY_ALTERNATIVE is not
1621 negative we should consider only this alternative. Return false if
1622 we can not choose the alternative or find how to reload the
1623 operands. */
1624 static bool
1625 process_alt_operands (int only_alternative)
1627 bool ok_p = false;
1628 int nop, overall, nalt;
1629 int n_alternatives = curr_static_id->n_alternatives;
1630 int n_operands = curr_static_id->n_operands;
1631 /* LOSERS counts the operands that don't fit this alternative and
1632 would require loading. */
1633 int losers;
1634 /* REJECT is a count of how undesirable this alternative says it is
1635 if any reloading is required. If the alternative matches exactly
1636 then REJECT is ignored, but otherwise it gets this much counted
1637 against it in addition to the reloading needed. */
1638 int reject;
1639 /* The number of elements in the following array. */
1640 int early_clobbered_regs_num;
1641 /* Numbers of operands which are early clobber registers. */
1642 int early_clobbered_nops[MAX_RECOG_OPERANDS];
1643 enum reg_class curr_alt[MAX_RECOG_OPERANDS];
1644 HARD_REG_SET curr_alt_set[MAX_RECOG_OPERANDS];
1645 bool curr_alt_match_win[MAX_RECOG_OPERANDS];
1646 bool curr_alt_win[MAX_RECOG_OPERANDS];
1647 bool curr_alt_offmemok[MAX_RECOG_OPERANDS];
1648 int curr_alt_matches[MAX_RECOG_OPERANDS];
1649 /* The number of elements in the following array. */
1650 int curr_alt_dont_inherit_ops_num;
1651 /* Numbers of operands whose reload pseudos should not be inherited. */
1652 int curr_alt_dont_inherit_ops[MAX_RECOG_OPERANDS];
1653 rtx op;
1654 /* The register when the operand is a subreg of register, otherwise the
1655 operand itself. */
1656 rtx no_subreg_reg_operand[MAX_RECOG_OPERANDS];
1657 /* The register if the operand is a register or subreg of register,
1658 otherwise NULL. */
1659 rtx operand_reg[MAX_RECOG_OPERANDS];
1660 int hard_regno[MAX_RECOG_OPERANDS];
1661 machine_mode biggest_mode[MAX_RECOG_OPERANDS];
1662 int reload_nregs, reload_sum;
1663 bool costly_p;
1664 enum reg_class cl;
1666 /* Calculate some data common for all alternatives to speed up the
1667 function. */
1668 for (nop = 0; nop < n_operands; nop++)
1670 rtx reg;
1672 op = no_subreg_reg_operand[nop] = *curr_id->operand_loc[nop];
1673 /* The real hard regno of the operand after the allocation. */
1674 hard_regno[nop] = get_hard_regno (op);
1676 operand_reg[nop] = reg = op;
1677 biggest_mode[nop] = GET_MODE (op);
1678 if (GET_CODE (op) == SUBREG)
1680 operand_reg[nop] = reg = SUBREG_REG (op);
1681 if (GET_MODE_SIZE (biggest_mode[nop])
1682 < GET_MODE_SIZE (GET_MODE (reg)))
1683 biggest_mode[nop] = GET_MODE (reg);
1685 if (! REG_P (reg))
1686 operand_reg[nop] = NULL_RTX;
1687 else if (REGNO (reg) >= FIRST_PSEUDO_REGISTER
1688 || ((int) REGNO (reg)
1689 == lra_get_elimination_hard_regno (REGNO (reg))))
1690 no_subreg_reg_operand[nop] = reg;
1691 else
1692 operand_reg[nop] = no_subreg_reg_operand[nop]
1693 /* Just use natural mode for elimination result. It should
1694 be enough for extra constraints hooks. */
1695 = regno_reg_rtx[hard_regno[nop]];
1698 /* The constraints are made of several alternatives. Each operand's
1699 constraint looks like foo,bar,... with commas separating the
1700 alternatives. The first alternatives for all operands go
1701 together, the second alternatives go together, etc.
1703 First loop over alternatives. */
1704 alternative_mask preferred = curr_id->preferred_alternatives;
1705 if (only_alternative >= 0)
1706 preferred &= ALTERNATIVE_BIT (only_alternative);
1708 for (nalt = 0; nalt < n_alternatives; nalt++)
1710 /* Loop over operands for one constraint alternative. */
1711 if (!TEST_BIT (preferred, nalt))
1712 continue;
1714 overall = losers = reject = reload_nregs = reload_sum = 0;
1715 for (nop = 0; nop < n_operands; nop++)
1717 int inc = (curr_static_id
1718 ->operand_alternative[nalt * n_operands + nop].reject);
1719 if (lra_dump_file != NULL && inc != 0)
1720 fprintf (lra_dump_file,
1721 " Staticly defined alt reject+=%d\n", inc);
1722 reject += inc;
1724 early_clobbered_regs_num = 0;
1726 for (nop = 0; nop < n_operands; nop++)
1728 const char *p;
1729 char *end;
1730 int len, c, m, i, opalt_num, this_alternative_matches;
1731 bool win, did_match, offmemok, early_clobber_p;
1732 /* false => this operand can be reloaded somehow for this
1733 alternative. */
1734 bool badop;
1735 /* true => this operand can be reloaded if the alternative
1736 allows regs. */
1737 bool winreg;
1738 /* True if a constant forced into memory would be OK for
1739 this operand. */
1740 bool constmemok;
1741 enum reg_class this_alternative, this_costly_alternative;
1742 HARD_REG_SET this_alternative_set, this_costly_alternative_set;
1743 bool this_alternative_match_win, this_alternative_win;
1744 bool this_alternative_offmemok;
1745 bool scratch_p;
1746 machine_mode mode;
1747 enum constraint_num cn;
1749 opalt_num = nalt * n_operands + nop;
1750 if (curr_static_id->operand_alternative[opalt_num].anything_ok)
1752 /* Fast track for no constraints at all. */
1753 curr_alt[nop] = NO_REGS;
1754 CLEAR_HARD_REG_SET (curr_alt_set[nop]);
1755 curr_alt_win[nop] = true;
1756 curr_alt_match_win[nop] = false;
1757 curr_alt_offmemok[nop] = false;
1758 curr_alt_matches[nop] = -1;
1759 continue;
1762 op = no_subreg_reg_operand[nop];
1763 mode = curr_operand_mode[nop];
1765 win = did_match = winreg = offmemok = constmemok = false;
1766 badop = true;
1768 early_clobber_p = false;
1769 p = curr_static_id->operand_alternative[opalt_num].constraint;
1771 this_costly_alternative = this_alternative = NO_REGS;
1772 /* We update set of possible hard regs besides its class
1773 because reg class might be inaccurate. For example,
1774 union of LO_REGS (l), HI_REGS(h), and STACK_REG(k) in ARM
1775 is translated in HI_REGS because classes are merged by
1776 pairs and there is no accurate intermediate class. */
1777 CLEAR_HARD_REG_SET (this_alternative_set);
1778 CLEAR_HARD_REG_SET (this_costly_alternative_set);
1779 this_alternative_win = false;
1780 this_alternative_match_win = false;
1781 this_alternative_offmemok = false;
1782 this_alternative_matches = -1;
1784 /* An empty constraint should be excluded by the fast
1785 track. */
1786 lra_assert (*p != 0 && *p != ',');
1788 /* Scan this alternative's specs for this operand; set WIN
1789 if the operand fits any letter in this alternative.
1790 Otherwise, clear BADOP if this operand could fit some
1791 letter after reloads, or set WINREG if this operand could
1792 fit after reloads provided the constraint allows some
1793 registers. */
1794 costly_p = false;
1797 switch ((c = *p, len = CONSTRAINT_LEN (c, p)), c)
1799 case '\0':
1800 len = 0;
1801 break;
1802 case ',':
1803 c = '\0';
1804 break;
1806 case '&':
1807 early_clobber_p = true;
1808 break;
1810 case '#':
1811 /* Ignore rest of this alternative. */
1812 c = '\0';
1813 break;
1815 case '0': case '1': case '2': case '3': case '4':
1816 case '5': case '6': case '7': case '8': case '9':
1818 int m_hregno;
1819 bool match_p;
1821 m = strtoul (p, &end, 10);
1822 p = end;
1823 len = 0;
1824 lra_assert (nop > m);
1826 this_alternative_matches = m;
1827 m_hregno = get_hard_regno (*curr_id->operand_loc[m]);
1828 /* We are supposed to match a previous operand.
1829 If we do, we win if that one did. If we do
1830 not, count both of the operands as losers.
1831 (This is too conservative, since most of the
1832 time only a single reload insn will be needed
1833 to make the two operands win. As a result,
1834 this alternative may be rejected when it is
1835 actually desirable.) */
1836 match_p = false;
1837 if (operands_match_p (*curr_id->operand_loc[nop],
1838 *curr_id->operand_loc[m], m_hregno))
1840 /* We should reject matching of an early
1841 clobber operand if the matching operand is
1842 not dying in the insn. */
1843 if (! curr_static_id->operand[m].early_clobber
1844 || operand_reg[nop] == NULL_RTX
1845 || (find_regno_note (curr_insn, REG_DEAD,
1846 REGNO (op))
1847 || REGNO (op) == REGNO (operand_reg[m])))
1848 match_p = true;
1850 if (match_p)
1852 /* If we are matching a non-offsettable
1853 address where an offsettable address was
1854 expected, then we must reject this
1855 combination, because we can't reload
1856 it. */
1857 if (curr_alt_offmemok[m]
1858 && MEM_P (*curr_id->operand_loc[m])
1859 && curr_alt[m] == NO_REGS && ! curr_alt_win[m])
1860 continue;
1862 else
1864 /* Operands don't match. Both operands must
1865 allow a reload register, otherwise we
1866 cannot make them match. */
1867 if (curr_alt[m] == NO_REGS)
1868 break;
1869 /* Retroactively mark the operand we had to
1870 match as a loser, if it wasn't already and
1871 it wasn't matched to a register constraint
1872 (e.g it might be matched by memory). */
1873 if (curr_alt_win[m]
1874 && (operand_reg[m] == NULL_RTX
1875 || hard_regno[m] < 0))
1877 losers++;
1878 reload_nregs
1879 += (ira_reg_class_max_nregs[curr_alt[m]]
1880 [GET_MODE (*curr_id->operand_loc[m])]);
1883 /* Prefer matching earlyclobber alternative as
1884 it results in less hard regs required for
1885 the insn than a non-matching earlyclobber
1886 alternative. */
1887 if (curr_static_id->operand[m].early_clobber)
1889 if (lra_dump_file != NULL)
1890 fprintf
1891 (lra_dump_file,
1892 " %d Matching earlyclobber alt:"
1893 " reject--\n",
1894 nop);
1895 reject--;
1897 /* Otherwise we prefer no matching
1898 alternatives because it gives more freedom
1899 in RA. */
1900 else if (operand_reg[nop] == NULL_RTX
1901 || (find_regno_note (curr_insn, REG_DEAD,
1902 REGNO (operand_reg[nop]))
1903 == NULL_RTX))
1905 if (lra_dump_file != NULL)
1906 fprintf
1907 (lra_dump_file,
1908 " %d Matching alt: reject+=2\n",
1909 nop);
1910 reject += 2;
1913 /* If we have to reload this operand and some
1914 previous operand also had to match the same
1915 thing as this operand, we don't know how to do
1916 that. */
1917 if (!match_p || !curr_alt_win[m])
1919 for (i = 0; i < nop; i++)
1920 if (curr_alt_matches[i] == m)
1921 break;
1922 if (i < nop)
1923 break;
1925 else
1926 did_match = true;
1928 /* This can be fixed with reloads if the operand
1929 we are supposed to match can be fixed with
1930 reloads. */
1931 badop = false;
1932 this_alternative = curr_alt[m];
1933 COPY_HARD_REG_SET (this_alternative_set, curr_alt_set[m]);
1934 winreg = this_alternative != NO_REGS;
1935 break;
1938 case 'g':
1939 if (MEM_P (op)
1940 || general_constant_p (op)
1941 || spilled_pseudo_p (op))
1942 win = true;
1943 cl = GENERAL_REGS;
1944 goto reg;
1946 default:
1947 cn = lookup_constraint (p);
1948 switch (get_constraint_type (cn))
1950 case CT_REGISTER:
1951 cl = reg_class_for_constraint (cn);
1952 if (cl != NO_REGS)
1953 goto reg;
1954 break;
1956 case CT_CONST_INT:
1957 if (CONST_INT_P (op)
1958 && insn_const_int_ok_for_constraint (INTVAL (op), cn))
1959 win = true;
1960 break;
1962 case CT_MEMORY:
1963 if (MEM_P (op)
1964 && satisfies_memory_constraint_p (op, cn))
1965 win = true;
1966 else if (spilled_pseudo_p (op))
1967 win = true;
1969 /* If we didn't already win, we can reload constants
1970 via force_const_mem or put the pseudo value into
1971 memory, or make other memory by reloading the
1972 address like for 'o'. */
1973 if (CONST_POOL_OK_P (mode, op)
1974 || MEM_P (op) || REG_P (op))
1975 badop = false;
1976 constmemok = true;
1977 offmemok = true;
1978 break;
1980 case CT_ADDRESS:
1981 /* If we didn't already win, we can reload the address
1982 into a base register. */
1983 if (satisfies_address_constraint_p (op, cn))
1984 win = true;
1985 cl = base_reg_class (VOIDmode, ADDR_SPACE_GENERIC,
1986 ADDRESS, SCRATCH);
1987 badop = false;
1988 goto reg;
1990 case CT_FIXED_FORM:
1991 if (constraint_satisfied_p (op, cn))
1992 win = true;
1993 break;
1995 break;
1997 reg:
1998 this_alternative = reg_class_subunion[this_alternative][cl];
1999 IOR_HARD_REG_SET (this_alternative_set,
2000 reg_class_contents[cl]);
2001 if (costly_p)
2003 this_costly_alternative
2004 = reg_class_subunion[this_costly_alternative][cl];
2005 IOR_HARD_REG_SET (this_costly_alternative_set,
2006 reg_class_contents[cl]);
2008 if (mode == BLKmode)
2009 break;
2010 winreg = true;
2011 if (REG_P (op))
2013 if (hard_regno[nop] >= 0
2014 && in_hard_reg_set_p (this_alternative_set,
2015 mode, hard_regno[nop]))
2016 win = true;
2017 else if (hard_regno[nop] < 0
2018 && in_class_p (op, this_alternative, NULL))
2019 win = true;
2021 break;
2023 if (c != ' ' && c != '\t')
2024 costly_p = c == '*';
2026 while ((p += len), c);
2028 scratch_p = (operand_reg[nop] != NULL_RTX
2029 && lra_former_scratch_p (REGNO (operand_reg[nop])));
2030 /* Record which operands fit this alternative. */
2031 if (win)
2033 this_alternative_win = true;
2034 if (operand_reg[nop] != NULL_RTX)
2036 if (hard_regno[nop] >= 0)
2038 if (in_hard_reg_set_p (this_costly_alternative_set,
2039 mode, hard_regno[nop]))
2041 if (lra_dump_file != NULL)
2042 fprintf (lra_dump_file,
2043 " %d Costly set: reject++\n",
2044 nop);
2045 reject++;
2048 else
2050 /* Prefer won reg to spilled pseudo under other
2051 equal conditions for possibe inheritance. */
2052 if (! scratch_p)
2054 if (lra_dump_file != NULL)
2055 fprintf
2056 (lra_dump_file,
2057 " %d Non pseudo reload: reject++\n",
2058 nop);
2059 reject++;
2061 if (in_class_p (operand_reg[nop],
2062 this_costly_alternative, NULL))
2064 if (lra_dump_file != NULL)
2065 fprintf
2066 (lra_dump_file,
2067 " %d Non pseudo costly reload:"
2068 " reject++\n",
2069 nop);
2070 reject++;
2073 /* We simulate the behaviour of old reload here.
2074 Although scratches need hard registers and it
2075 might result in spilling other pseudos, no reload
2076 insns are generated for the scratches. So it
2077 might cost something but probably less than old
2078 reload pass believes. */
2079 if (scratch_p)
2081 if (lra_dump_file != NULL)
2082 fprintf (lra_dump_file,
2083 " %d Scratch win: reject+=2\n",
2084 nop);
2085 reject += 2;
2089 else if (did_match)
2090 this_alternative_match_win = true;
2091 else
2093 int const_to_mem = 0;
2094 bool no_regs_p;
2096 /* Never do output reload of stack pointer. It makes
2097 impossible to do elimination when SP is changed in
2098 RTL. */
2099 if (op == stack_pointer_rtx && ! frame_pointer_needed
2100 && curr_static_id->operand[nop].type != OP_IN)
2101 goto fail;
2103 /* If this alternative asks for a specific reg class, see if there
2104 is at least one allocatable register in that class. */
2105 no_regs_p
2106 = (this_alternative == NO_REGS
2107 || (hard_reg_set_subset_p
2108 (reg_class_contents[this_alternative],
2109 lra_no_alloc_regs)));
2111 /* For asms, verify that the class for this alternative is possible
2112 for the mode that is specified. */
2113 if (!no_regs_p && INSN_CODE (curr_insn) < 0)
2115 int i;
2116 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
2117 if (HARD_REGNO_MODE_OK (i, mode)
2118 && in_hard_reg_set_p (reg_class_contents[this_alternative],
2119 mode, i))
2120 break;
2121 if (i == FIRST_PSEUDO_REGISTER)
2122 winreg = false;
2125 /* If this operand accepts a register, and if the
2126 register class has at least one allocatable register,
2127 then this operand can be reloaded. */
2128 if (winreg && !no_regs_p)
2129 badop = false;
2131 if (badop)
2133 if (lra_dump_file != NULL)
2134 fprintf (lra_dump_file,
2135 " alt=%d: Bad operand -- refuse\n",
2136 nalt);
2137 goto fail;
2140 /* If not assigned pseudo has a class which a subset of
2141 required reg class, it is a less costly alternative
2142 as the pseudo still can get a hard reg of necessary
2143 class. */
2144 if (! no_regs_p && REG_P (op) && hard_regno[nop] < 0
2145 && (cl = get_reg_class (REGNO (op))) != NO_REGS
2146 && ira_class_subset_p[this_alternative][cl])
2148 if (lra_dump_file != NULL)
2149 fprintf
2150 (lra_dump_file,
2151 " %d Super set class reg: reject-=3\n", nop);
2152 reject -= 3;
2155 this_alternative_offmemok = offmemok;
2156 if (this_costly_alternative != NO_REGS)
2158 if (lra_dump_file != NULL)
2159 fprintf (lra_dump_file,
2160 " %d Costly loser: reject++\n", nop);
2161 reject++;
2163 /* If the operand is dying, has a matching constraint,
2164 and satisfies constraints of the matched operand
2165 which failed to satisfy the own constraints, most probably
2166 the reload for this operand will be gone. */
2167 if (this_alternative_matches >= 0
2168 && !curr_alt_win[this_alternative_matches]
2169 && REG_P (op)
2170 && find_regno_note (curr_insn, REG_DEAD, REGNO (op))
2171 && (hard_regno[nop] >= 0
2172 ? in_hard_reg_set_p (this_alternative_set,
2173 mode, hard_regno[nop])
2174 : in_class_p (op, this_alternative, NULL)))
2176 if (lra_dump_file != NULL)
2177 fprintf
2178 (lra_dump_file,
2179 " %d Dying matched operand reload: reject++\n",
2180 nop);
2181 reject++;
2183 else
2185 /* Strict_low_part requires to reload the register
2186 not the sub-register. In this case we should
2187 check that a final reload hard reg can hold the
2188 value mode. */
2189 if (curr_static_id->operand[nop].strict_low
2190 && REG_P (op)
2191 && hard_regno[nop] < 0
2192 && GET_CODE (*curr_id->operand_loc[nop]) == SUBREG
2193 && ira_class_hard_regs_num[this_alternative] > 0
2194 && ! HARD_REGNO_MODE_OK (ira_class_hard_regs
2195 [this_alternative][0],
2196 GET_MODE
2197 (*curr_id->operand_loc[nop])))
2199 if (lra_dump_file != NULL)
2200 fprintf
2201 (lra_dump_file,
2202 " alt=%d: Strict low subreg reload -- refuse\n",
2203 nalt);
2204 goto fail;
2206 losers++;
2208 if (operand_reg[nop] != NULL_RTX
2209 /* Output operands and matched input operands are
2210 not inherited. The following conditions do not
2211 exactly describe the previous statement but they
2212 are pretty close. */
2213 && curr_static_id->operand[nop].type != OP_OUT
2214 && (this_alternative_matches < 0
2215 || curr_static_id->operand[nop].type != OP_IN))
2217 int last_reload = (lra_reg_info[ORIGINAL_REGNO
2218 (operand_reg[nop])]
2219 .last_reload);
2221 /* The value of reload_sum has sense only if we
2222 process insns in their order. It happens only on
2223 the first constraints sub-pass when we do most of
2224 reload work. */
2225 if (lra_constraint_iter == 1 && last_reload > bb_reload_num)
2226 reload_sum += last_reload - bb_reload_num;
2228 /* If this is a constant that is reloaded into the
2229 desired class by copying it to memory first, count
2230 that as another reload. This is consistent with
2231 other code and is required to avoid choosing another
2232 alternative when the constant is moved into memory.
2233 Note that the test here is precisely the same as in
2234 the code below that calls force_const_mem. */
2235 if (CONST_POOL_OK_P (mode, op)
2236 && ((targetm.preferred_reload_class
2237 (op, this_alternative) == NO_REGS)
2238 || no_input_reloads_p))
2240 const_to_mem = 1;
2241 if (! no_regs_p)
2242 losers++;
2245 /* Alternative loses if it requires a type of reload not
2246 permitted for this insn. We can always reload
2247 objects with a REG_UNUSED note. */
2248 if ((curr_static_id->operand[nop].type != OP_IN
2249 && no_output_reloads_p
2250 && ! find_reg_note (curr_insn, REG_UNUSED, op))
2251 || (curr_static_id->operand[nop].type != OP_OUT
2252 && no_input_reloads_p && ! const_to_mem)
2253 || (this_alternative_matches >= 0
2254 && (no_input_reloads_p
2255 || (no_output_reloads_p
2256 && (curr_static_id->operand
2257 [this_alternative_matches].type != OP_IN)
2258 && ! find_reg_note (curr_insn, REG_UNUSED,
2259 no_subreg_reg_operand
2260 [this_alternative_matches])))))
2262 if (lra_dump_file != NULL)
2263 fprintf
2264 (lra_dump_file,
2265 " alt=%d: No input/otput reload -- refuse\n",
2266 nalt);
2267 goto fail;
2270 /* Alternative loses if it required class pseudo can not
2271 hold value of required mode. Such insns can be
2272 described by insn definitions with mode iterators. */
2273 if (GET_MODE (*curr_id->operand_loc[nop]) != VOIDmode
2274 && ! hard_reg_set_empty_p (this_alternative_set)
2275 /* It is common practice for constraints to use a
2276 class which does not have actually enough regs to
2277 hold the value (e.g. x86 AREG for mode requiring
2278 more one general reg). Therefore we have 2
2279 conditions to check that the reload pseudo can
2280 not hold the mode value. */
2281 && ! HARD_REGNO_MODE_OK (ira_class_hard_regs
2282 [this_alternative][0],
2283 GET_MODE (*curr_id->operand_loc[nop])))
2285 HARD_REG_SET temp;
2287 COPY_HARD_REG_SET (temp, this_alternative_set);
2288 AND_COMPL_HARD_REG_SET (temp, lra_no_alloc_regs);
2289 /* The above condition is not enough as the first
2290 reg in ira_class_hard_regs can be not aligned for
2291 multi-words mode values. */
2292 if (hard_reg_set_subset_p (temp,
2293 ira_prohibited_class_mode_regs
2294 [this_alternative]
2295 [GET_MODE (*curr_id->operand_loc[nop])]))
2297 if (lra_dump_file != NULL)
2298 fprintf
2299 (lra_dump_file,
2300 " alt=%d: reload pseudo for op %d "
2301 " can not hold the mode value -- refuse\n",
2302 nalt, nop);
2303 goto fail;
2307 /* Check strong discouragement of reload of non-constant
2308 into class THIS_ALTERNATIVE. */
2309 if (! CONSTANT_P (op) && ! no_regs_p
2310 && (targetm.preferred_reload_class
2311 (op, this_alternative) == NO_REGS
2312 || (curr_static_id->operand[nop].type == OP_OUT
2313 && (targetm.preferred_output_reload_class
2314 (op, this_alternative) == NO_REGS))))
2316 if (lra_dump_file != NULL)
2317 fprintf (lra_dump_file,
2318 " %d Non-prefered reload: reject+=%d\n",
2319 nop, LRA_MAX_REJECT);
2320 reject += LRA_MAX_REJECT;
2323 if (! (MEM_P (op) && offmemok)
2324 && ! (const_to_mem && constmemok))
2326 /* We prefer to reload pseudos over reloading other
2327 things, since such reloads may be able to be
2328 eliminated later. So bump REJECT in other cases.
2329 Don't do this in the case where we are forcing a
2330 constant into memory and it will then win since
2331 we don't want to have a different alternative
2332 match then. */
2333 if (! (REG_P (op) && REGNO (op) >= FIRST_PSEUDO_REGISTER))
2335 if (lra_dump_file != NULL)
2336 fprintf
2337 (lra_dump_file,
2338 " %d Non-pseudo reload: reject+=2\n",
2339 nop);
2340 reject += 2;
2343 if (! no_regs_p)
2344 reload_nregs
2345 += ira_reg_class_max_nregs[this_alternative][mode];
2347 if (SMALL_REGISTER_CLASS_P (this_alternative))
2349 if (lra_dump_file != NULL)
2350 fprintf
2351 (lra_dump_file,
2352 " %d Small class reload: reject+=%d\n",
2353 nop, LRA_LOSER_COST_FACTOR / 2);
2354 reject += LRA_LOSER_COST_FACTOR / 2;
2358 /* We are trying to spill pseudo into memory. It is
2359 usually more costly than moving to a hard register
2360 although it might takes the same number of
2361 reloads. */
2362 if (no_regs_p && REG_P (op) && hard_regno[nop] >= 0)
2364 if (lra_dump_file != NULL)
2365 fprintf
2366 (lra_dump_file,
2367 " %d Spill pseudo into memory: reject+=3\n",
2368 nop);
2369 reject += 3;
2370 if (VECTOR_MODE_P (mode))
2372 /* Spilling vectors into memory is usually more
2373 costly as they contain big values. */
2374 if (lra_dump_file != NULL)
2375 fprintf
2376 (lra_dump_file,
2377 " %d Spill vector pseudo: reject+=2\n",
2378 nop);
2379 reject += 2;
2383 #ifdef SECONDARY_MEMORY_NEEDED
2384 /* If reload requires moving value through secondary
2385 memory, it will need one more insn at least. */
2386 if (this_alternative != NO_REGS
2387 && REG_P (op) && (cl = get_reg_class (REGNO (op))) != NO_REGS
2388 && ((curr_static_id->operand[nop].type != OP_OUT
2389 && SECONDARY_MEMORY_NEEDED (cl, this_alternative,
2390 GET_MODE (op)))
2391 || (curr_static_id->operand[nop].type != OP_IN
2392 && SECONDARY_MEMORY_NEEDED (this_alternative, cl,
2393 GET_MODE (op)))))
2394 losers++;
2395 #endif
2396 /* Input reloads can be inherited more often than output
2397 reloads can be removed, so penalize output
2398 reloads. */
2399 if (!REG_P (op) || curr_static_id->operand[nop].type != OP_IN)
2401 if (lra_dump_file != NULL)
2402 fprintf
2403 (lra_dump_file,
2404 " %d Non input pseudo reload: reject++\n",
2405 nop);
2406 reject++;
2410 if (early_clobber_p && ! scratch_p)
2412 if (lra_dump_file != NULL)
2413 fprintf (lra_dump_file,
2414 " %d Early clobber: reject++\n", nop);
2415 reject++;
2417 /* ??? We check early clobbers after processing all operands
2418 (see loop below) and there we update the costs more.
2419 Should we update the cost (may be approximately) here
2420 because of early clobber register reloads or it is a rare
2421 or non-important thing to be worth to do it. */
2422 overall = losers * LRA_LOSER_COST_FACTOR + reject;
2423 if ((best_losers == 0 || losers != 0) && best_overall < overall)
2425 if (lra_dump_file != NULL)
2426 fprintf (lra_dump_file,
2427 " alt=%d,overall=%d,losers=%d -- refuse\n",
2428 nalt, overall, losers);
2429 goto fail;
2432 curr_alt[nop] = this_alternative;
2433 COPY_HARD_REG_SET (curr_alt_set[nop], this_alternative_set);
2434 curr_alt_win[nop] = this_alternative_win;
2435 curr_alt_match_win[nop] = this_alternative_match_win;
2436 curr_alt_offmemok[nop] = this_alternative_offmemok;
2437 curr_alt_matches[nop] = this_alternative_matches;
2439 if (this_alternative_matches >= 0
2440 && !did_match && !this_alternative_win)
2441 curr_alt_win[this_alternative_matches] = false;
2443 if (early_clobber_p && operand_reg[nop] != NULL_RTX)
2444 early_clobbered_nops[early_clobbered_regs_num++] = nop;
2446 if (curr_insn_set != NULL_RTX && n_operands == 2
2447 /* Prevent processing non-move insns. */
2448 && (GET_CODE (SET_SRC (curr_insn_set)) == SUBREG
2449 || SET_SRC (curr_insn_set) == no_subreg_reg_operand[1])
2450 && ((! curr_alt_win[0] && ! curr_alt_win[1]
2451 && REG_P (no_subreg_reg_operand[0])
2452 && REG_P (no_subreg_reg_operand[1])
2453 && (reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2454 || reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0])))
2455 || (! curr_alt_win[0] && curr_alt_win[1]
2456 && REG_P (no_subreg_reg_operand[1])
2457 && reg_in_class_p (no_subreg_reg_operand[1], curr_alt[0]))
2458 || (curr_alt_win[0] && ! curr_alt_win[1]
2459 && REG_P (no_subreg_reg_operand[0])
2460 && reg_in_class_p (no_subreg_reg_operand[0], curr_alt[1])
2461 && (! CONST_POOL_OK_P (curr_operand_mode[1],
2462 no_subreg_reg_operand[1])
2463 || (targetm.preferred_reload_class
2464 (no_subreg_reg_operand[1],
2465 (enum reg_class) curr_alt[1]) != NO_REGS))
2466 /* If it is a result of recent elimination in move
2467 insn we can transform it into an add still by
2468 using this alternative. */
2469 && GET_CODE (no_subreg_reg_operand[1]) != PLUS)))
2471 /* We have a move insn and a new reload insn will be similar
2472 to the current insn. We should avoid such situation as it
2473 results in LRA cycling. */
2474 overall += LRA_MAX_REJECT;
2476 ok_p = true;
2477 curr_alt_dont_inherit_ops_num = 0;
2478 for (nop = 0; nop < early_clobbered_regs_num; nop++)
2480 int i, j, clobbered_hard_regno, first_conflict_j, last_conflict_j;
2481 HARD_REG_SET temp_set;
2483 i = early_clobbered_nops[nop];
2484 if ((! curr_alt_win[i] && ! curr_alt_match_win[i])
2485 || hard_regno[i] < 0)
2486 continue;
2487 lra_assert (operand_reg[i] != NULL_RTX);
2488 clobbered_hard_regno = hard_regno[i];
2489 CLEAR_HARD_REG_SET (temp_set);
2490 add_to_hard_reg_set (&temp_set, biggest_mode[i], clobbered_hard_regno);
2491 first_conflict_j = last_conflict_j = -1;
2492 for (j = 0; j < n_operands; j++)
2493 if (j == i
2494 /* We don't want process insides of match_operator and
2495 match_parallel because otherwise we would process
2496 their operands once again generating a wrong
2497 code. */
2498 || curr_static_id->operand[j].is_operator)
2499 continue;
2500 else if ((curr_alt_matches[j] == i && curr_alt_match_win[j])
2501 || (curr_alt_matches[i] == j && curr_alt_match_win[i]))
2502 continue;
2503 /* If we don't reload j-th operand, check conflicts. */
2504 else if ((curr_alt_win[j] || curr_alt_match_win[j])
2505 && uses_hard_regs_p (*curr_id->operand_loc[j], temp_set))
2507 if (first_conflict_j < 0)
2508 first_conflict_j = j;
2509 last_conflict_j = j;
2511 if (last_conflict_j < 0)
2512 continue;
2513 /* If earlyclobber operand conflicts with another
2514 non-matching operand which is actually the same register
2515 as the earlyclobber operand, it is better to reload the
2516 another operand as an operand matching the earlyclobber
2517 operand can be also the same. */
2518 if (first_conflict_j == last_conflict_j
2519 && operand_reg[last_conflict_j]
2520 != NULL_RTX && ! curr_alt_match_win[last_conflict_j]
2521 && REGNO (operand_reg[i]) == REGNO (operand_reg[last_conflict_j]))
2523 curr_alt_win[last_conflict_j] = false;
2524 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++]
2525 = last_conflict_j;
2526 losers++;
2527 /* Early clobber was already reflected in REJECT. */
2528 lra_assert (reject > 0);
2529 if (lra_dump_file != NULL)
2530 fprintf
2531 (lra_dump_file,
2532 " %d Conflict early clobber reload: reject--\n",
2534 reject--;
2535 overall += LRA_LOSER_COST_FACTOR - 1;
2537 else
2539 /* We need to reload early clobbered register and the
2540 matched registers. */
2541 for (j = 0; j < n_operands; j++)
2542 if (curr_alt_matches[j] == i)
2544 curr_alt_match_win[j] = false;
2545 losers++;
2546 overall += LRA_LOSER_COST_FACTOR;
2548 if (! curr_alt_match_win[i])
2549 curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++] = i;
2550 else
2552 /* Remember pseudos used for match reloads are never
2553 inherited. */
2554 lra_assert (curr_alt_matches[i] >= 0);
2555 curr_alt_win[curr_alt_matches[i]] = false;
2557 curr_alt_win[i] = curr_alt_match_win[i] = false;
2558 losers++;
2559 /* Early clobber was already reflected in REJECT. */
2560 lra_assert (reject > 0);
2561 if (lra_dump_file != NULL)
2562 fprintf
2563 (lra_dump_file,
2564 " %d Matched conflict early clobber reloads:"
2565 "reject--\n",
2567 reject--;
2568 overall += LRA_LOSER_COST_FACTOR - 1;
2571 if (lra_dump_file != NULL)
2572 fprintf (lra_dump_file, " alt=%d,overall=%d,losers=%d,rld_nregs=%d\n",
2573 nalt, overall, losers, reload_nregs);
2575 /* If this alternative can be made to work by reloading, and it
2576 needs less reloading than the others checked so far, record
2577 it as the chosen goal for reloading. */
2578 if ((best_losers != 0 && losers == 0)
2579 || (((best_losers == 0 && losers == 0)
2580 || (best_losers != 0 && losers != 0))
2581 && (best_overall > overall
2582 || (best_overall == overall
2583 /* If the cost of the reloads is the same,
2584 prefer alternative which requires minimal
2585 number of reload regs. */
2586 && (reload_nregs < best_reload_nregs
2587 || (reload_nregs == best_reload_nregs
2588 && (best_reload_sum < reload_sum
2589 || (best_reload_sum == reload_sum
2590 && nalt < goal_alt_number))))))))
2592 for (nop = 0; nop < n_operands; nop++)
2594 goal_alt_win[nop] = curr_alt_win[nop];
2595 goal_alt_match_win[nop] = curr_alt_match_win[nop];
2596 goal_alt_matches[nop] = curr_alt_matches[nop];
2597 goal_alt[nop] = curr_alt[nop];
2598 goal_alt_offmemok[nop] = curr_alt_offmemok[nop];
2600 goal_alt_dont_inherit_ops_num = curr_alt_dont_inherit_ops_num;
2601 for (nop = 0; nop < curr_alt_dont_inherit_ops_num; nop++)
2602 goal_alt_dont_inherit_ops[nop] = curr_alt_dont_inherit_ops[nop];
2603 goal_alt_swapped = curr_swapped;
2604 best_overall = overall;
2605 best_losers = losers;
2606 best_reload_nregs = reload_nregs;
2607 best_reload_sum = reload_sum;
2608 goal_alt_number = nalt;
2610 if (losers == 0)
2611 /* Everything is satisfied. Do not process alternatives
2612 anymore. */
2613 break;
2614 fail:
2617 return ok_p;
2620 /* Make reload base reg from address AD. */
2621 static rtx
2622 base_to_reg (struct address_info *ad)
2624 enum reg_class cl;
2625 int code = -1;
2626 rtx new_inner = NULL_RTX;
2627 rtx new_reg = NULL_RTX;
2628 rtx_insn *insn;
2629 rtx_insn *last_insn = get_last_insn();
2631 lra_assert (ad->base == ad->base_term && ad->disp == ad->disp_term);
2632 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
2633 get_index_code (ad));
2634 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
2635 cl, "base");
2636 new_inner = simplify_gen_binary (PLUS, GET_MODE (new_reg), new_reg,
2637 ad->disp_term == NULL
2638 ? gen_int_mode (0, ad->mode)
2639 : *ad->disp_term);
2640 if (!valid_address_p (ad->mode, new_inner, ad->as))
2641 return NULL_RTX;
2642 insn = emit_insn (gen_rtx_SET (ad->mode, new_reg, *ad->base_term));
2643 code = recog_memoized (insn);
2644 if (code < 0)
2646 delete_insns_since (last_insn);
2647 return NULL_RTX;
2650 return new_inner;
2653 /* Make reload base reg + disp from address AD. Return the new pseudo. */
2654 static rtx
2655 base_plus_disp_to_reg (struct address_info *ad)
2657 enum reg_class cl;
2658 rtx new_reg;
2660 lra_assert (ad->base == ad->base_term && ad->disp == ad->disp_term);
2661 cl = base_reg_class (ad->mode, ad->as, ad->base_outer_code,
2662 get_index_code (ad));
2663 new_reg = lra_create_new_reg (GET_MODE (*ad->base_term), NULL_RTX,
2664 cl, "base + disp");
2665 lra_emit_add (new_reg, *ad->base_term, *ad->disp_term);
2666 return new_reg;
2669 /* Make reload of index part of address AD. Return the new
2670 pseudo. */
2671 static rtx
2672 index_part_to_reg (struct address_info *ad)
2674 rtx new_reg;
2676 new_reg = lra_create_new_reg (GET_MODE (*ad->index), NULL_RTX,
2677 INDEX_REG_CLASS, "index term");
2678 expand_mult (GET_MODE (*ad->index), *ad->index_term,
2679 GEN_INT (get_index_scale (ad)), new_reg, 1);
2680 return new_reg;
2683 /* Return true if we can add a displacement to address AD, even if that
2684 makes the address invalid. The fix-up code requires any new address
2685 to be the sum of the BASE_TERM, INDEX and DISP_TERM fields. */
2686 static bool
2687 can_add_disp_p (struct address_info *ad)
2689 return (!ad->autoinc_p
2690 && ad->segment == NULL
2691 && ad->base == ad->base_term
2692 && ad->disp == ad->disp_term);
2695 /* Make equiv substitution in address AD. Return true if a substitution
2696 was made. */
2697 static bool
2698 equiv_address_substitution (struct address_info *ad)
2700 rtx base_reg, new_base_reg, index_reg, new_index_reg, *base_term, *index_term;
2701 HOST_WIDE_INT disp, scale;
2702 bool change_p;
2704 base_term = strip_subreg (ad->base_term);
2705 if (base_term == NULL)
2706 base_reg = new_base_reg = NULL_RTX;
2707 else
2709 base_reg = *base_term;
2710 new_base_reg = get_equiv_with_elimination (base_reg, curr_insn);
2712 index_term = strip_subreg (ad->index_term);
2713 if (index_term == NULL)
2714 index_reg = new_index_reg = NULL_RTX;
2715 else
2717 index_reg = *index_term;
2718 new_index_reg = get_equiv_with_elimination (index_reg, curr_insn);
2720 if (base_reg == new_base_reg && index_reg == new_index_reg)
2721 return false;
2722 disp = 0;
2723 change_p = false;
2724 if (lra_dump_file != NULL)
2726 fprintf (lra_dump_file, "Changing address in insn %d ",
2727 INSN_UID (curr_insn));
2728 dump_value_slim (lra_dump_file, *ad->outer, 1);
2730 if (base_reg != new_base_reg)
2732 if (REG_P (new_base_reg))
2734 *base_term = new_base_reg;
2735 change_p = true;
2737 else if (GET_CODE (new_base_reg) == PLUS
2738 && REG_P (XEXP (new_base_reg, 0))
2739 && CONST_INT_P (XEXP (new_base_reg, 1))
2740 && can_add_disp_p (ad))
2742 disp += INTVAL (XEXP (new_base_reg, 1));
2743 *base_term = XEXP (new_base_reg, 0);
2744 change_p = true;
2746 if (ad->base_term2 != NULL)
2747 *ad->base_term2 = *ad->base_term;
2749 if (index_reg != new_index_reg)
2751 if (REG_P (new_index_reg))
2753 *index_term = new_index_reg;
2754 change_p = true;
2756 else if (GET_CODE (new_index_reg) == PLUS
2757 && REG_P (XEXP (new_index_reg, 0))
2758 && CONST_INT_P (XEXP (new_index_reg, 1))
2759 && can_add_disp_p (ad)
2760 && (scale = get_index_scale (ad)))
2762 disp += INTVAL (XEXP (new_index_reg, 1)) * scale;
2763 *index_term = XEXP (new_index_reg, 0);
2764 change_p = true;
2767 if (disp != 0)
2769 if (ad->disp != NULL)
2770 *ad->disp = plus_constant (GET_MODE (*ad->inner), *ad->disp, disp);
2771 else
2773 *ad->inner = plus_constant (GET_MODE (*ad->inner), *ad->inner, disp);
2774 update_address (ad);
2776 change_p = true;
2778 if (lra_dump_file != NULL)
2780 if (! change_p)
2781 fprintf (lra_dump_file, " -- no change\n");
2782 else
2784 fprintf (lra_dump_file, " on equiv ");
2785 dump_value_slim (lra_dump_file, *ad->outer, 1);
2786 fprintf (lra_dump_file, "\n");
2789 return change_p;
2792 /* Major function to make reloads for an address in operand NOP or
2793 check its correctness (If CHECK_ONLY_P is true). The supported
2794 cases are:
2796 1) an address that existed before LRA started, at which point it
2797 must have been valid. These addresses are subject to elimination
2798 and may have become invalid due to the elimination offset being out
2799 of range.
2801 2) an address created by forcing a constant to memory
2802 (force_const_to_mem). The initial form of these addresses might
2803 not be valid, and it is this function's job to make them valid.
2805 3) a frame address formed from a register and a (possibly zero)
2806 constant offset. As above, these addresses might not be valid and
2807 this function must make them so.
2809 Add reloads to the lists *BEFORE and *AFTER. We might need to add
2810 reloads to *AFTER because of inc/dec, {pre, post} modify in the
2811 address. Return true for any RTL change.
2813 The function is a helper function which does not produce all
2814 transformations (when CHECK_ONLY_P is false) which can be
2815 necessary. It does just basic steps. To do all necessary
2816 transformations use function process_address. */
2817 static bool
2818 process_address_1 (int nop, bool check_only_p,
2819 rtx_insn **before, rtx_insn **after)
2821 struct address_info ad;
2822 rtx new_reg;
2823 rtx op = *curr_id->operand_loc[nop];
2824 const char *constraint = curr_static_id->operand[nop].constraint;
2825 enum constraint_num cn = lookup_constraint (constraint);
2826 bool change_p = false;
2828 if (insn_extra_address_constraint (cn))
2829 decompose_lea_address (&ad, curr_id->operand_loc[nop]);
2830 else if (MEM_P (op))
2831 decompose_mem_address (&ad, op);
2832 else if (GET_CODE (op) == SUBREG
2833 && MEM_P (SUBREG_REG (op)))
2834 decompose_mem_address (&ad, SUBREG_REG (op));
2835 else
2836 return false;
2837 if (! check_only_p)
2838 change_p = equiv_address_substitution (&ad);
2839 if (ad.base_term != NULL
2840 && (process_addr_reg
2841 (ad.base_term, check_only_p, before,
2842 (ad.autoinc_p
2843 && !(REG_P (*ad.base_term)
2844 && find_regno_note (curr_insn, REG_DEAD,
2845 REGNO (*ad.base_term)) != NULL_RTX)
2846 ? after : NULL),
2847 base_reg_class (ad.mode, ad.as, ad.base_outer_code,
2848 get_index_code (&ad)))))
2850 change_p = true;
2851 if (ad.base_term2 != NULL)
2852 *ad.base_term2 = *ad.base_term;
2854 if (ad.index_term != NULL
2855 && process_addr_reg (ad.index_term, check_only_p,
2856 before, NULL, INDEX_REG_CLASS))
2857 change_p = true;
2859 /* Target hooks sometimes don't treat extra-constraint addresses as
2860 legitimate address_operands, so handle them specially. */
2861 if (insn_extra_address_constraint (cn)
2862 && satisfies_address_constraint_p (&ad, cn))
2863 return change_p;
2865 if (check_only_p)
2866 return change_p;
2868 /* There are three cases where the shape of *AD.INNER may now be invalid:
2870 1) the original address was valid, but either elimination or
2871 equiv_address_substitution was applied and that made
2872 the address invalid.
2874 2) the address is an invalid symbolic address created by
2875 force_const_to_mem.
2877 3) the address is a frame address with an invalid offset.
2879 4) the address is a frame address with an invalid base.
2881 All these cases involve a non-autoinc address, so there is no
2882 point revalidating other types. */
2883 if (ad.autoinc_p || valid_address_p (&ad))
2884 return change_p;
2886 /* Any index existed before LRA started, so we can assume that the
2887 presence and shape of the index is valid. */
2888 push_to_sequence (*before);
2889 lra_assert (ad.disp == ad.disp_term);
2890 if (ad.base == NULL)
2892 if (ad.index == NULL)
2894 int code = -1;
2895 enum reg_class cl = base_reg_class (ad.mode, ad.as,
2896 SCRATCH, SCRATCH);
2897 rtx addr = *ad.inner;
2899 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "addr");
2900 #ifdef HAVE_lo_sum
2902 rtx_insn *insn;
2903 rtx_insn *last = get_last_insn ();
2905 /* addr => lo_sum (new_base, addr), case (2) above. */
2906 insn = emit_insn (gen_rtx_SET
2907 (VOIDmode, new_reg,
2908 gen_rtx_HIGH (Pmode, copy_rtx (addr))));
2909 code = recog_memoized (insn);
2910 if (code >= 0)
2912 *ad.inner = gen_rtx_LO_SUM (Pmode, new_reg, addr);
2913 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
2915 /* Try to put lo_sum into register. */
2916 insn = emit_insn (gen_rtx_SET
2917 (VOIDmode, new_reg,
2918 gen_rtx_LO_SUM (Pmode, new_reg, addr)));
2919 code = recog_memoized (insn);
2920 if (code >= 0)
2922 *ad.inner = new_reg;
2923 if (! valid_address_p (ad.mode, *ad.outer, ad.as))
2925 *ad.inner = addr;
2926 code = -1;
2932 if (code < 0)
2933 delete_insns_since (last);
2935 #endif
2936 if (code < 0)
2938 /* addr => new_base, case (2) above. */
2939 lra_emit_move (new_reg, addr);
2940 *ad.inner = new_reg;
2943 else
2945 /* index * scale + disp => new base + index * scale,
2946 case (1) above. */
2947 enum reg_class cl = base_reg_class (ad.mode, ad.as, PLUS,
2948 GET_CODE (*ad.index));
2950 lra_assert (INDEX_REG_CLASS != NO_REGS);
2951 new_reg = lra_create_new_reg (Pmode, NULL_RTX, cl, "disp");
2952 lra_emit_move (new_reg, *ad.disp);
2953 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
2954 new_reg, *ad.index);
2957 else if (ad.index == NULL)
2959 int regno;
2960 enum reg_class cl;
2961 rtx set;
2962 rtx_insn *insns, *last_insn;
2963 /* Try to reload base into register only if the base is invalid
2964 for the address but with valid offset, case (4) above. */
2965 start_sequence ();
2966 new_reg = base_to_reg (&ad);
2968 /* base + disp => new base, cases (1) and (3) above. */
2969 /* Another option would be to reload the displacement into an
2970 index register. However, postreload has code to optimize
2971 address reloads that have the same base and different
2972 displacements, so reloading into an index register would
2973 not necessarily be a win. */
2974 if (new_reg == NULL_RTX)
2975 new_reg = base_plus_disp_to_reg (&ad);
2976 insns = get_insns ();
2977 last_insn = get_last_insn ();
2978 /* If we generated at least two insns, try last insn source as
2979 an address. If we succeed, we generate one less insn. */
2980 if (last_insn != insns && (set = single_set (last_insn)) != NULL_RTX
2981 && GET_CODE (SET_SRC (set)) == PLUS
2982 && REG_P (XEXP (SET_SRC (set), 0))
2983 && CONSTANT_P (XEXP (SET_SRC (set), 1)))
2985 *ad.inner = SET_SRC (set);
2986 if (valid_address_p (ad.mode, *ad.outer, ad.as))
2988 *ad.base_term = XEXP (SET_SRC (set), 0);
2989 *ad.disp_term = XEXP (SET_SRC (set), 1);
2990 cl = base_reg_class (ad.mode, ad.as, ad.base_outer_code,
2991 get_index_code (&ad));
2992 regno = REGNO (*ad.base_term);
2993 if (regno >= FIRST_PSEUDO_REGISTER
2994 && cl != lra_get_allocno_class (regno))
2995 lra_change_class (regno, cl, " Change to", true);
2996 new_reg = SET_SRC (set);
2997 delete_insns_since (PREV_INSN (last_insn));
3000 end_sequence ();
3001 emit_insn (insns);
3002 *ad.inner = new_reg;
3004 else if (ad.disp_term != NULL)
3006 /* base + scale * index + disp => new base + scale * index,
3007 case (1) above. */
3008 new_reg = base_plus_disp_to_reg (&ad);
3009 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3010 new_reg, *ad.index);
3012 else if (get_index_scale (&ad) == 1)
3014 /* The last transformation to one reg will be made in
3015 curr_insn_transform function. */
3016 end_sequence ();
3017 return false;
3019 else
3021 /* base + scale * index => base + new_reg,
3022 case (1) above.
3023 Index part of address may become invalid. For example, we
3024 changed pseudo on the equivalent memory and a subreg of the
3025 pseudo onto the memory of different mode for which the scale is
3026 prohibitted. */
3027 new_reg = index_part_to_reg (&ad);
3028 *ad.inner = simplify_gen_binary (PLUS, GET_MODE (new_reg),
3029 *ad.base_term, new_reg);
3031 *before = get_insns ();
3032 end_sequence ();
3033 return true;
3036 /* If CHECK_ONLY_P is false, do address reloads until it is necessary.
3037 Use process_address_1 as a helper function. Return true for any
3038 RTL changes.
3040 If CHECK_ONLY_P is true, just check address correctness. Return
3041 false if the address correct. */
3042 static bool
3043 process_address (int nop, bool check_only_p,
3044 rtx_insn **before, rtx_insn **after)
3046 bool res = false;
3048 while (process_address_1 (nop, check_only_p, before, after))
3050 if (check_only_p)
3051 return true;
3052 res = true;
3054 return res;
3057 /* Emit insns to reload VALUE into a new register. VALUE is an
3058 auto-increment or auto-decrement RTX whose operand is a register or
3059 memory location; so reloading involves incrementing that location.
3060 IN is either identical to VALUE, or some cheaper place to reload
3061 value being incremented/decremented from.
3063 INC_AMOUNT is the number to increment or decrement by (always
3064 positive and ignored for POST_MODIFY/PRE_MODIFY).
3066 Return pseudo containing the result. */
3067 static rtx
3068 emit_inc (enum reg_class new_rclass, rtx in, rtx value, int inc_amount)
3070 /* REG or MEM to be copied and incremented. */
3071 rtx incloc = XEXP (value, 0);
3072 /* Nonzero if increment after copying. */
3073 int post = (GET_CODE (value) == POST_DEC || GET_CODE (value) == POST_INC
3074 || GET_CODE (value) == POST_MODIFY);
3075 rtx_insn *last;
3076 rtx inc;
3077 rtx_insn *add_insn;
3078 int code;
3079 rtx real_in = in == value ? incloc : in;
3080 rtx result;
3081 bool plus_p = true;
3083 if (GET_CODE (value) == PRE_MODIFY || GET_CODE (value) == POST_MODIFY)
3085 lra_assert (GET_CODE (XEXP (value, 1)) == PLUS
3086 || GET_CODE (XEXP (value, 1)) == MINUS);
3087 lra_assert (rtx_equal_p (XEXP (XEXP (value, 1), 0), XEXP (value, 0)));
3088 plus_p = GET_CODE (XEXP (value, 1)) == PLUS;
3089 inc = XEXP (XEXP (value, 1), 1);
3091 else
3093 if (GET_CODE (value) == PRE_DEC || GET_CODE (value) == POST_DEC)
3094 inc_amount = -inc_amount;
3096 inc = GEN_INT (inc_amount);
3099 if (! post && REG_P (incloc))
3100 result = incloc;
3101 else
3102 result = lra_create_new_reg (GET_MODE (value), value, new_rclass,
3103 "INC/DEC result");
3105 if (real_in != result)
3107 /* First copy the location to the result register. */
3108 lra_assert (REG_P (result));
3109 emit_insn (gen_move_insn (result, real_in));
3112 /* We suppose that there are insns to add/sub with the constant
3113 increment permitted in {PRE/POST)_{DEC/INC/MODIFY}. At least the
3114 old reload worked with this assumption. If the assumption
3115 becomes wrong, we should use approach in function
3116 base_plus_disp_to_reg. */
3117 if (in == value)
3119 /* See if we can directly increment INCLOC. */
3120 last = get_last_insn ();
3121 add_insn = emit_insn (plus_p
3122 ? gen_add2_insn (incloc, inc)
3123 : gen_sub2_insn (incloc, inc));
3125 code = recog_memoized (add_insn);
3126 if (code >= 0)
3128 if (! post && result != incloc)
3129 emit_insn (gen_move_insn (result, incloc));
3130 return result;
3132 delete_insns_since (last);
3135 /* If couldn't do the increment directly, must increment in RESULT.
3136 The way we do this depends on whether this is pre- or
3137 post-increment. For pre-increment, copy INCLOC to the reload
3138 register, increment it there, then save back. */
3139 if (! post)
3141 if (real_in != result)
3142 emit_insn (gen_move_insn (result, real_in));
3143 if (plus_p)
3144 emit_insn (gen_add2_insn (result, inc));
3145 else
3146 emit_insn (gen_sub2_insn (result, inc));
3147 if (result != incloc)
3148 emit_insn (gen_move_insn (incloc, result));
3150 else
3152 /* Post-increment.
3154 Because this might be a jump insn or a compare, and because
3155 RESULT may not be available after the insn in an input
3156 reload, we must do the incrementing before the insn being
3157 reloaded for.
3159 We have already copied IN to RESULT. Increment the copy in
3160 RESULT, save that back, then decrement RESULT so it has
3161 the original value. */
3162 if (plus_p)
3163 emit_insn (gen_add2_insn (result, inc));
3164 else
3165 emit_insn (gen_sub2_insn (result, inc));
3166 emit_insn (gen_move_insn (incloc, result));
3167 /* Restore non-modified value for the result. We prefer this
3168 way because it does not require an additional hard
3169 register. */
3170 if (plus_p)
3172 if (CONST_INT_P (inc))
3173 emit_insn (gen_add2_insn (result,
3174 gen_int_mode (-INTVAL (inc),
3175 GET_MODE (result))));
3176 else
3177 emit_insn (gen_sub2_insn (result, inc));
3179 else
3180 emit_insn (gen_add2_insn (result, inc));
3182 return result;
3185 /* Return true if the current move insn does not need processing as we
3186 already know that it satisfies its constraints. */
3187 static bool
3188 simple_move_p (void)
3190 rtx dest, src;
3191 enum reg_class dclass, sclass;
3193 lra_assert (curr_insn_set != NULL_RTX);
3194 dest = SET_DEST (curr_insn_set);
3195 src = SET_SRC (curr_insn_set);
3196 return ((dclass = get_op_class (dest)) != NO_REGS
3197 && (sclass = get_op_class (src)) != NO_REGS
3198 /* The backend guarantees that register moves of cost 2
3199 never need reloads. */
3200 && targetm.register_move_cost (GET_MODE (src), sclass, dclass) == 2);
3203 /* Swap operands NOP and NOP + 1. */
3204 static inline void
3205 swap_operands (int nop)
3207 machine_mode mode = curr_operand_mode[nop];
3208 curr_operand_mode[nop] = curr_operand_mode[nop + 1];
3209 curr_operand_mode[nop + 1] = mode;
3210 rtx x = *curr_id->operand_loc[nop];
3211 *curr_id->operand_loc[nop] = *curr_id->operand_loc[nop + 1];
3212 *curr_id->operand_loc[nop + 1] = x;
3213 /* Swap the duplicates too. */
3214 lra_update_dup (curr_id, nop);
3215 lra_update_dup (curr_id, nop + 1);
3218 /* Main entry point of the constraint code: search the body of the
3219 current insn to choose the best alternative. It is mimicking insn
3220 alternative cost calculation model of former reload pass. That is
3221 because machine descriptions were written to use this model. This
3222 model can be changed in future. Make commutative operand exchange
3223 if it is chosen.
3225 if CHECK_ONLY_P is false, do RTL changes to satisfy the
3226 constraints. Return true if any change happened during function
3227 call.
3229 If CHECK_ONLY_P is true then don't do any transformation. Just
3230 check that the insn satisfies all constraints. If the insn does
3231 not satisfy any constraint, return true. */
3232 static bool
3233 curr_insn_transform (bool check_only_p)
3235 int i, j, k;
3236 int n_operands;
3237 int n_alternatives;
3238 int commutative;
3239 signed char goal_alt_matched[MAX_RECOG_OPERANDS][MAX_RECOG_OPERANDS];
3240 signed char match_inputs[MAX_RECOG_OPERANDS + 1];
3241 rtx_insn *before, *after;
3242 bool alt_p = false;
3243 /* Flag that the insn has been changed through a transformation. */
3244 bool change_p;
3245 bool sec_mem_p;
3246 #ifdef SECONDARY_MEMORY_NEEDED
3247 bool use_sec_mem_p;
3248 #endif
3249 int max_regno_before;
3250 int reused_alternative_num;
3252 curr_insn_set = single_set (curr_insn);
3253 if (curr_insn_set != NULL_RTX && simple_move_p ())
3254 return false;
3256 no_input_reloads_p = no_output_reloads_p = false;
3257 goal_alt_number = -1;
3258 change_p = sec_mem_p = false;
3259 /* JUMP_INSNs and CALL_INSNs are not allowed to have any output
3260 reloads; neither are insns that SET cc0. Insns that use CC0 are
3261 not allowed to have any input reloads. */
3262 if (JUMP_P (curr_insn) || CALL_P (curr_insn))
3263 no_output_reloads_p = true;
3265 #ifdef HAVE_cc0
3266 if (reg_referenced_p (cc0_rtx, PATTERN (curr_insn)))
3267 no_input_reloads_p = true;
3268 if (reg_set_p (cc0_rtx, PATTERN (curr_insn)))
3269 no_output_reloads_p = true;
3270 #endif
3272 n_operands = curr_static_id->n_operands;
3273 n_alternatives = curr_static_id->n_alternatives;
3275 /* Just return "no reloads" if insn has no operands with
3276 constraints. */
3277 if (n_operands == 0 || n_alternatives == 0)
3278 return false;
3280 max_regno_before = max_reg_num ();
3282 for (i = 0; i < n_operands; i++)
3284 goal_alt_matched[i][0] = -1;
3285 goal_alt_matches[i] = -1;
3288 commutative = curr_static_id->commutative;
3290 /* Now see what we need for pseudos that didn't get hard regs or got
3291 the wrong kind of hard reg. For this, we must consider all the
3292 operands together against the register constraints. */
3294 best_losers = best_overall = INT_MAX;
3295 best_reload_sum = 0;
3297 curr_swapped = false;
3298 goal_alt_swapped = false;
3300 if (! check_only_p)
3301 /* Make equivalence substitution and memory subreg elimination
3302 before address processing because an address legitimacy can
3303 depend on memory mode. */
3304 for (i = 0; i < n_operands; i++)
3306 rtx op = *curr_id->operand_loc[i];
3307 rtx subst, old = op;
3308 bool op_change_p = false;
3310 if (GET_CODE (old) == SUBREG)
3311 old = SUBREG_REG (old);
3312 subst = get_equiv_with_elimination (old, curr_insn);
3313 if (subst != old)
3315 subst = copy_rtx (subst);
3316 lra_assert (REG_P (old));
3317 if (GET_CODE (op) == SUBREG)
3318 SUBREG_REG (op) = subst;
3319 else
3320 *curr_id->operand_loc[i] = subst;
3321 if (lra_dump_file != NULL)
3323 fprintf (lra_dump_file,
3324 "Changing pseudo %d in operand %i of insn %u on equiv ",
3325 REGNO (old), i, INSN_UID (curr_insn));
3326 dump_value_slim (lra_dump_file, subst, 1);
3327 fprintf (lra_dump_file, "\n");
3329 op_change_p = change_p = true;
3331 if (simplify_operand_subreg (i, GET_MODE (old)) || op_change_p)
3333 change_p = true;
3334 lra_update_dup (curr_id, i);
3338 /* Reload address registers and displacements. We do it before
3339 finding an alternative because of memory constraints. */
3340 before = after = NULL;
3341 for (i = 0; i < n_operands; i++)
3342 if (! curr_static_id->operand[i].is_operator
3343 && process_address (i, check_only_p, &before, &after))
3345 if (check_only_p)
3346 return true;
3347 change_p = true;
3348 lra_update_dup (curr_id, i);
3351 if (change_p)
3352 /* If we've changed the instruction then any alternative that
3353 we chose previously may no longer be valid. */
3354 lra_set_used_insn_alternative (curr_insn, -1);
3356 if (! check_only_p && curr_insn_set != NULL_RTX
3357 && check_and_process_move (&change_p, &sec_mem_p))
3358 return change_p;
3360 try_swapped:
3362 reused_alternative_num = check_only_p ? -1 : curr_id->used_insn_alternative;
3363 if (lra_dump_file != NULL && reused_alternative_num >= 0)
3364 fprintf (lra_dump_file, "Reusing alternative %d for insn #%u\n",
3365 reused_alternative_num, INSN_UID (curr_insn));
3367 if (process_alt_operands (reused_alternative_num))
3368 alt_p = true;
3370 if (check_only_p)
3371 return ! alt_p || best_losers != 0;
3373 /* If insn is commutative (it's safe to exchange a certain pair of
3374 operands) then we need to try each alternative twice, the second
3375 time matching those two operands as if we had exchanged them. To
3376 do this, really exchange them in operands.
3378 If we have just tried the alternatives the second time, return
3379 operands to normal and drop through. */
3381 if (reused_alternative_num < 0 && commutative >= 0)
3383 curr_swapped = !curr_swapped;
3384 if (curr_swapped)
3386 swap_operands (commutative);
3387 goto try_swapped;
3389 else
3390 swap_operands (commutative);
3393 if (! alt_p && ! sec_mem_p)
3395 /* No alternative works with reloads?? */
3396 if (INSN_CODE (curr_insn) >= 0)
3397 fatal_insn ("unable to generate reloads for:", curr_insn);
3398 error_for_asm (curr_insn,
3399 "inconsistent operand constraints in an %<asm%>");
3400 /* Avoid further trouble with this insn. */
3401 PATTERN (curr_insn) = gen_rtx_USE (VOIDmode, const0_rtx);
3402 lra_invalidate_insn_data (curr_insn);
3403 return true;
3406 /* If the best alternative is with operands 1 and 2 swapped, swap
3407 them. Update the operand numbers of any reloads already
3408 pushed. */
3410 if (goal_alt_swapped)
3412 if (lra_dump_file != NULL)
3413 fprintf (lra_dump_file, " Commutative operand exchange in insn %u\n",
3414 INSN_UID (curr_insn));
3416 /* Swap the duplicates too. */
3417 swap_operands (commutative);
3418 change_p = true;
3421 #ifdef SECONDARY_MEMORY_NEEDED
3422 /* Some target macros SECONDARY_MEMORY_NEEDED (e.g. x86) are defined
3423 too conservatively. So we use the secondary memory only if there
3424 is no any alternative without reloads. */
3425 use_sec_mem_p = false;
3426 if (! alt_p)
3427 use_sec_mem_p = true;
3428 else if (sec_mem_p)
3430 for (i = 0; i < n_operands; i++)
3431 if (! goal_alt_win[i] && ! goal_alt_match_win[i])
3432 break;
3433 use_sec_mem_p = i < n_operands;
3436 if (use_sec_mem_p)
3438 rtx new_reg, src, dest, rld;
3439 machine_mode sec_mode, rld_mode;
3441 lra_assert (sec_mem_p);
3442 lra_assert (curr_static_id->operand[0].type == OP_OUT
3443 && curr_static_id->operand[1].type == OP_IN);
3444 dest = *curr_id->operand_loc[0];
3445 src = *curr_id->operand_loc[1];
3446 rld = (GET_MODE_SIZE (GET_MODE (dest)) <= GET_MODE_SIZE (GET_MODE (src))
3447 ? dest : src);
3448 rld_mode = GET_MODE (rld);
3449 #ifdef SECONDARY_MEMORY_NEEDED_MODE
3450 sec_mode = SECONDARY_MEMORY_NEEDED_MODE (rld_mode);
3451 #else
3452 sec_mode = rld_mode;
3453 #endif
3454 new_reg = lra_create_new_reg (sec_mode, NULL_RTX,
3455 NO_REGS, "secondary");
3456 /* If the mode is changed, it should be wider. */
3457 lra_assert (GET_MODE_SIZE (sec_mode) >= GET_MODE_SIZE (rld_mode));
3458 if (sec_mode != rld_mode)
3460 /* If the target says specifically to use another mode for
3461 secondary memory moves we can not reuse the original
3462 insn. */
3463 after = emit_spill_move (false, new_reg, dest);
3464 lra_process_new_insns (curr_insn, NULL, after,
3465 "Inserting the sec. move");
3466 /* We may have non null BEFORE here (e.g. after address
3467 processing. */
3468 push_to_sequence (before);
3469 before = emit_spill_move (true, new_reg, src);
3470 emit_insn (before);
3471 before = get_insns ();
3472 end_sequence ();
3473 lra_process_new_insns (curr_insn, before, NULL, "Changing on");
3474 lra_set_insn_deleted (curr_insn);
3476 else if (dest == rld)
3478 *curr_id->operand_loc[0] = new_reg;
3479 after = emit_spill_move (false, new_reg, dest);
3480 lra_process_new_insns (curr_insn, NULL, after,
3481 "Inserting the sec. move");
3483 else
3485 *curr_id->operand_loc[1] = new_reg;
3486 /* See comments above. */
3487 push_to_sequence (before);
3488 before = emit_spill_move (true, new_reg, src);
3489 emit_insn (before);
3490 before = get_insns ();
3491 end_sequence ();
3492 lra_process_new_insns (curr_insn, before, NULL,
3493 "Inserting the sec. move");
3495 lra_update_insn_regno_info (curr_insn);
3496 return true;
3498 #endif
3500 lra_assert (goal_alt_number >= 0);
3501 lra_set_used_insn_alternative (curr_insn, goal_alt_number);
3503 if (lra_dump_file != NULL)
3505 const char *p;
3507 fprintf (lra_dump_file, " Choosing alt %d in insn %u:",
3508 goal_alt_number, INSN_UID (curr_insn));
3509 for (i = 0; i < n_operands; i++)
3511 p = (curr_static_id->operand_alternative
3512 [goal_alt_number * n_operands + i].constraint);
3513 if (*p == '\0')
3514 continue;
3515 fprintf (lra_dump_file, " (%d) ", i);
3516 for (; *p != '\0' && *p != ',' && *p != '#'; p++)
3517 fputc (*p, lra_dump_file);
3519 if (INSN_CODE (curr_insn) >= 0
3520 && (p = get_insn_name (INSN_CODE (curr_insn))) != NULL)
3521 fprintf (lra_dump_file, " {%s}", p);
3522 if (curr_id->sp_offset != 0)
3523 fprintf (lra_dump_file, " (sp_off=%" HOST_WIDE_INT_PRINT "d)",
3524 curr_id->sp_offset);
3525 fprintf (lra_dump_file, "\n");
3528 /* Right now, for any pair of operands I and J that are required to
3529 match, with J < I, goal_alt_matches[I] is J. Add I to
3530 goal_alt_matched[J]. */
3532 for (i = 0; i < n_operands; i++)
3533 if ((j = goal_alt_matches[i]) >= 0)
3535 for (k = 0; goal_alt_matched[j][k] >= 0; k++)
3537 /* We allow matching one output operand and several input
3538 operands. */
3539 lra_assert (k == 0
3540 || (curr_static_id->operand[j].type == OP_OUT
3541 && curr_static_id->operand[i].type == OP_IN
3542 && (curr_static_id->operand
3543 [goal_alt_matched[j][0]].type == OP_IN)));
3544 goal_alt_matched[j][k] = i;
3545 goal_alt_matched[j][k + 1] = -1;
3548 for (i = 0; i < n_operands; i++)
3549 goal_alt_win[i] |= goal_alt_match_win[i];
3551 /* Any constants that aren't allowed and can't be reloaded into
3552 registers are here changed into memory references. */
3553 for (i = 0; i < n_operands; i++)
3554 if (goal_alt_win[i])
3556 int regno;
3557 enum reg_class new_class;
3558 rtx reg = *curr_id->operand_loc[i];
3560 if (GET_CODE (reg) == SUBREG)
3561 reg = SUBREG_REG (reg);
3563 if (REG_P (reg) && (regno = REGNO (reg)) >= FIRST_PSEUDO_REGISTER)
3565 bool ok_p = in_class_p (reg, goal_alt[i], &new_class);
3567 if (new_class != NO_REGS && get_reg_class (regno) != new_class)
3569 lra_assert (ok_p);
3570 lra_change_class (regno, new_class, " Change to", true);
3574 else
3576 const char *constraint;
3577 char c;
3578 rtx op = *curr_id->operand_loc[i];
3579 rtx subreg = NULL_RTX;
3580 machine_mode mode = curr_operand_mode[i];
3582 if (GET_CODE (op) == SUBREG)
3584 subreg = op;
3585 op = SUBREG_REG (op);
3586 mode = GET_MODE (op);
3589 if (CONST_POOL_OK_P (mode, op)
3590 && ((targetm.preferred_reload_class
3591 (op, (enum reg_class) goal_alt[i]) == NO_REGS)
3592 || no_input_reloads_p))
3594 rtx tem = force_const_mem (mode, op);
3596 change_p = true;
3597 if (subreg != NULL_RTX)
3598 tem = gen_rtx_SUBREG (mode, tem, SUBREG_BYTE (subreg));
3600 *curr_id->operand_loc[i] = tem;
3601 lra_update_dup (curr_id, i);
3602 process_address (i, false, &before, &after);
3604 /* If the alternative accepts constant pool refs directly
3605 there will be no reload needed at all. */
3606 if (subreg != NULL_RTX)
3607 continue;
3608 /* Skip alternatives before the one requested. */
3609 constraint = (curr_static_id->operand_alternative
3610 [goal_alt_number * n_operands + i].constraint);
3611 for (;
3612 (c = *constraint) && c != ',' && c != '#';
3613 constraint += CONSTRAINT_LEN (c, constraint))
3615 enum constraint_num cn = lookup_constraint (constraint);
3616 if (insn_extra_memory_constraint (cn)
3617 && satisfies_memory_constraint_p (tem, cn))
3618 break;
3620 if (c == '\0' || c == ',' || c == '#')
3621 continue;
3623 goal_alt_win[i] = true;
3627 for (i = 0; i < n_operands; i++)
3629 int regno;
3630 bool optional_p = false;
3631 rtx old, new_reg;
3632 rtx op = *curr_id->operand_loc[i];
3634 if (goal_alt_win[i])
3636 if (goal_alt[i] == NO_REGS
3637 && REG_P (op)
3638 /* When we assign NO_REGS it means that we will not
3639 assign a hard register to the scratch pseudo by
3640 assigment pass and the scratch pseudo will be
3641 spilled. Spilled scratch pseudos are transformed
3642 back to scratches at the LRA end. */
3643 && lra_former_scratch_operand_p (curr_insn, i))
3645 int regno = REGNO (op);
3646 lra_change_class (regno, NO_REGS, " Change to", true);
3647 if (lra_get_regno_hard_regno (regno) >= 0)
3648 /* We don't have to mark all insn affected by the
3649 spilled pseudo as there is only one such insn, the
3650 current one. */
3651 reg_renumber[regno] = -1;
3653 /* We can do an optional reload. If the pseudo got a hard
3654 reg, we might improve the code through inheritance. If
3655 it does not get a hard register we coalesce memory/memory
3656 moves later. Ignore move insns to avoid cycling. */
3657 if (! lra_simple_p
3658 && lra_undo_inheritance_iter < LRA_MAX_INHERITANCE_PASSES
3659 && goal_alt[i] != NO_REGS && REG_P (op)
3660 && (regno = REGNO (op)) >= FIRST_PSEUDO_REGISTER
3661 && regno < new_regno_start
3662 && ! lra_former_scratch_p (regno)
3663 && reg_renumber[regno] < 0
3664 && (curr_insn_set == NULL_RTX
3665 || !((REG_P (SET_SRC (curr_insn_set))
3666 || MEM_P (SET_SRC (curr_insn_set))
3667 || GET_CODE (SET_SRC (curr_insn_set)) == SUBREG)
3668 && (REG_P (SET_DEST (curr_insn_set))
3669 || MEM_P (SET_DEST (curr_insn_set))
3670 || GET_CODE (SET_DEST (curr_insn_set)) == SUBREG))))
3671 optional_p = true;
3672 else
3673 continue;
3676 /* Operands that match previous ones have already been handled. */
3677 if (goal_alt_matches[i] >= 0)
3678 continue;
3680 /* We should not have an operand with a non-offsettable address
3681 appearing where an offsettable address will do. It also may
3682 be a case when the address should be special in other words
3683 not a general one (e.g. it needs no index reg). */
3684 if (goal_alt_matched[i][0] == -1 && goal_alt_offmemok[i] && MEM_P (op))
3686 enum reg_class rclass;
3687 rtx *loc = &XEXP (op, 0);
3688 enum rtx_code code = GET_CODE (*loc);
3690 push_to_sequence (before);
3691 rclass = base_reg_class (GET_MODE (op), MEM_ADDR_SPACE (op),
3692 MEM, SCRATCH);
3693 if (GET_RTX_CLASS (code) == RTX_AUTOINC)
3694 new_reg = emit_inc (rclass, *loc, *loc,
3695 /* This value does not matter for MODIFY. */
3696 GET_MODE_SIZE (GET_MODE (op)));
3697 else if (get_reload_reg (OP_IN, Pmode, *loc, rclass, FALSE,
3698 "offsetable address", &new_reg))
3699 lra_emit_move (new_reg, *loc);
3700 before = get_insns ();
3701 end_sequence ();
3702 *loc = new_reg;
3703 lra_update_dup (curr_id, i);
3705 else if (goal_alt_matched[i][0] == -1)
3707 machine_mode mode;
3708 rtx reg, *loc;
3709 int hard_regno, byte;
3710 enum op_type type = curr_static_id->operand[i].type;
3712 loc = curr_id->operand_loc[i];
3713 mode = curr_operand_mode[i];
3714 if (GET_CODE (*loc) == SUBREG)
3716 reg = SUBREG_REG (*loc);
3717 byte = SUBREG_BYTE (*loc);
3718 if (REG_P (reg)
3719 /* Strict_low_part requires reload the register not
3720 the sub-register. */
3721 && (curr_static_id->operand[i].strict_low
3722 || (GET_MODE_SIZE (mode)
3723 <= GET_MODE_SIZE (GET_MODE (reg))
3724 && (hard_regno
3725 = get_try_hard_regno (REGNO (reg))) >= 0
3726 && (simplify_subreg_regno
3727 (hard_regno,
3728 GET_MODE (reg), byte, mode) < 0)
3729 && (goal_alt[i] == NO_REGS
3730 || (simplify_subreg_regno
3731 (ira_class_hard_regs[goal_alt[i]][0],
3732 GET_MODE (reg), byte, mode) >= 0)))))
3734 loc = &SUBREG_REG (*loc);
3735 mode = GET_MODE (*loc);
3738 old = *loc;
3739 if (get_reload_reg (type, mode, old, goal_alt[i],
3740 loc != curr_id->operand_loc[i], "", &new_reg)
3741 && type != OP_OUT)
3743 push_to_sequence (before);
3744 lra_emit_move (new_reg, old);
3745 before = get_insns ();
3746 end_sequence ();
3748 *loc = new_reg;
3749 if (type != OP_IN
3750 && find_reg_note (curr_insn, REG_UNUSED, old) == NULL_RTX)
3752 start_sequence ();
3753 lra_emit_move (type == OP_INOUT ? copy_rtx (old) : old, new_reg);
3754 emit_insn (after);
3755 after = get_insns ();
3756 end_sequence ();
3757 *loc = new_reg;
3759 for (j = 0; j < goal_alt_dont_inherit_ops_num; j++)
3760 if (goal_alt_dont_inherit_ops[j] == i)
3762 lra_set_regno_unique_value (REGNO (new_reg));
3763 break;
3765 lra_update_dup (curr_id, i);
3767 else if (curr_static_id->operand[i].type == OP_IN
3768 && (curr_static_id->operand[goal_alt_matched[i][0]].type
3769 == OP_OUT))
3771 /* generate reloads for input and matched outputs. */
3772 match_inputs[0] = i;
3773 match_inputs[1] = -1;
3774 match_reload (goal_alt_matched[i][0], match_inputs,
3775 goal_alt[i], &before, &after);
3777 else if (curr_static_id->operand[i].type == OP_OUT
3778 && (curr_static_id->operand[goal_alt_matched[i][0]].type
3779 == OP_IN))
3780 /* Generate reloads for output and matched inputs. */
3781 match_reload (i, goal_alt_matched[i], goal_alt[i], &before, &after);
3782 else if (curr_static_id->operand[i].type == OP_IN
3783 && (curr_static_id->operand[goal_alt_matched[i][0]].type
3784 == OP_IN))
3786 /* Generate reloads for matched inputs. */
3787 match_inputs[0] = i;
3788 for (j = 0; (k = goal_alt_matched[i][j]) >= 0; j++)
3789 match_inputs[j + 1] = k;
3790 match_inputs[j + 1] = -1;
3791 match_reload (-1, match_inputs, goal_alt[i], &before, &after);
3793 else
3794 /* We must generate code in any case when function
3795 process_alt_operands decides that it is possible. */
3796 gcc_unreachable ();
3797 if (optional_p)
3799 lra_assert (REG_P (op));
3800 regno = REGNO (op);
3801 op = *curr_id->operand_loc[i]; /* Substitution. */
3802 if (GET_CODE (op) == SUBREG)
3803 op = SUBREG_REG (op);
3804 gcc_assert (REG_P (op) && (int) REGNO (op) >= new_regno_start);
3805 bitmap_set_bit (&lra_optional_reload_pseudos, REGNO (op));
3806 lra_reg_info[REGNO (op)].restore_regno = regno;
3807 if (lra_dump_file != NULL)
3808 fprintf (lra_dump_file,
3809 " Making reload reg %d for reg %d optional\n",
3810 REGNO (op), regno);
3813 if (before != NULL_RTX || after != NULL_RTX
3814 || max_regno_before != max_reg_num ())
3815 change_p = true;
3816 if (change_p)
3818 lra_update_operator_dups (curr_id);
3819 /* Something changes -- process the insn. */
3820 lra_update_insn_regno_info (curr_insn);
3822 lra_process_new_insns (curr_insn, before, after, "Inserting insn reload");
3823 return change_p;
3826 /* Return true if INSN satisfies all constraints. In other words, no
3827 reload insns are needed. */
3828 bool
3829 lra_constrain_insn (rtx_insn *insn)
3831 int saved_new_regno_start = new_regno_start;
3832 int saved_new_insn_uid_start = new_insn_uid_start;
3833 bool change_p;
3835 curr_insn = insn;
3836 curr_id = lra_get_insn_recog_data (curr_insn);
3837 curr_static_id = curr_id->insn_static_data;
3838 new_insn_uid_start = get_max_uid ();
3839 new_regno_start = max_reg_num ();
3840 change_p = curr_insn_transform (true);
3841 new_regno_start = saved_new_regno_start;
3842 new_insn_uid_start = saved_new_insn_uid_start;
3843 return ! change_p;
3846 /* Return true if X is in LIST. */
3847 static bool
3848 in_list_p (rtx x, rtx list)
3850 for (; list != NULL_RTX; list = XEXP (list, 1))
3851 if (XEXP (list, 0) == x)
3852 return true;
3853 return false;
3856 /* Return true if X contains an allocatable hard register (if
3857 HARD_REG_P) or a (spilled if SPILLED_P) pseudo. */
3858 static bool
3859 contains_reg_p (rtx x, bool hard_reg_p, bool spilled_p)
3861 int i, j;
3862 const char *fmt;
3863 enum rtx_code code;
3865 code = GET_CODE (x);
3866 if (REG_P (x))
3868 int regno = REGNO (x);
3869 HARD_REG_SET alloc_regs;
3871 if (hard_reg_p)
3873 if (regno >= FIRST_PSEUDO_REGISTER)
3874 regno = lra_get_regno_hard_regno (regno);
3875 if (regno < 0)
3876 return false;
3877 COMPL_HARD_REG_SET (alloc_regs, lra_no_alloc_regs);
3878 return overlaps_hard_reg_set_p (alloc_regs, GET_MODE (x), regno);
3880 else
3882 if (regno < FIRST_PSEUDO_REGISTER)
3883 return false;
3884 if (! spilled_p)
3885 return true;
3886 return lra_get_regno_hard_regno (regno) < 0;
3889 fmt = GET_RTX_FORMAT (code);
3890 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3892 if (fmt[i] == 'e')
3894 if (contains_reg_p (XEXP (x, i), hard_reg_p, spilled_p))
3895 return true;
3897 else if (fmt[i] == 'E')
3899 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3900 if (contains_reg_p (XVECEXP (x, i, j), hard_reg_p, spilled_p))
3901 return true;
3904 return false;
3907 /* Return true if X contains a symbol reg. */
3908 static bool
3909 contains_symbol_ref_p (rtx x)
3911 int i, j;
3912 const char *fmt;
3913 enum rtx_code code;
3915 code = GET_CODE (x);
3916 if (code == SYMBOL_REF)
3917 return true;
3918 fmt = GET_RTX_FORMAT (code);
3919 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3921 if (fmt[i] == 'e')
3923 if (contains_symbol_ref_p (XEXP (x, i)))
3924 return true;
3926 else if (fmt[i] == 'E')
3928 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3929 if (contains_symbol_ref_p (XVECEXP (x, i, j)))
3930 return true;
3933 return false;
3936 /* Process all regs in location *LOC and change them on equivalent
3937 substitution. Return true if any change was done. */
3938 static bool
3939 loc_equivalence_change_p (rtx *loc)
3941 rtx subst, reg, x = *loc;
3942 bool result = false;
3943 enum rtx_code code = GET_CODE (x);
3944 const char *fmt;
3945 int i, j;
3947 if (code == SUBREG)
3949 reg = SUBREG_REG (x);
3950 if ((subst = get_equiv_with_elimination (reg, curr_insn)) != reg
3951 && GET_MODE (subst) == VOIDmode)
3953 /* We cannot reload debug location. Simplify subreg here
3954 while we know the inner mode. */
3955 *loc = simplify_gen_subreg (GET_MODE (x), subst,
3956 GET_MODE (reg), SUBREG_BYTE (x));
3957 return true;
3960 if (code == REG && (subst = get_equiv_with_elimination (x, curr_insn)) != x)
3962 *loc = subst;
3963 return true;
3966 /* Scan all the operand sub-expressions. */
3967 fmt = GET_RTX_FORMAT (code);
3968 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
3970 if (fmt[i] == 'e')
3971 result = loc_equivalence_change_p (&XEXP (x, i)) || result;
3972 else if (fmt[i] == 'E')
3973 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
3974 result
3975 = loc_equivalence_change_p (&XVECEXP (x, i, j)) || result;
3977 return result;
3980 /* Similar to loc_equivalence_change_p, but for use as
3981 simplify_replace_fn_rtx callback. DATA is insn for which the
3982 elimination is done. If it null we don't do the elimination. */
3983 static rtx
3984 loc_equivalence_callback (rtx loc, const_rtx, void *data)
3986 if (!REG_P (loc))
3987 return NULL_RTX;
3989 rtx subst = (data == NULL
3990 ? get_equiv (loc) : get_equiv_with_elimination (loc, (rtx_insn *) data));
3991 if (subst != loc)
3992 return subst;
3994 return NULL_RTX;
3997 /* Maximum number of generated reload insns per an insn. It is for
3998 preventing this pass cycling in a bug case. */
3999 #define MAX_RELOAD_INSNS_NUMBER LRA_MAX_INSN_RELOADS
4001 /* The current iteration number of this LRA pass. */
4002 int lra_constraint_iter;
4004 /* True if we substituted equiv which needs checking register
4005 allocation correctness because the equivalent value contains
4006 allocatable hard registers or when we restore multi-register
4007 pseudo. */
4008 bool lra_risky_transformations_p;
4010 /* Return true if REGNO is referenced in more than one block. */
4011 static bool
4012 multi_block_pseudo_p (int regno)
4014 basic_block bb = NULL;
4015 unsigned int uid;
4016 bitmap_iterator bi;
4018 if (regno < FIRST_PSEUDO_REGISTER)
4019 return false;
4021 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi)
4022 if (bb == NULL)
4023 bb = BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn);
4024 else if (BLOCK_FOR_INSN (lra_insn_recog_data[uid]->insn) != bb)
4025 return true;
4026 return false;
4029 /* Return true if LIST contains a deleted insn. */
4030 static bool
4031 contains_deleted_insn_p (rtx_insn_list *list)
4033 for (; list != NULL_RTX; list = list->next ())
4034 if (NOTE_P (list->insn ())
4035 && NOTE_KIND (list->insn ()) == NOTE_INSN_DELETED)
4036 return true;
4037 return false;
4040 /* Return true if X contains a pseudo dying in INSN. */
4041 static bool
4042 dead_pseudo_p (rtx x, rtx insn)
4044 int i, j;
4045 const char *fmt;
4046 enum rtx_code code;
4048 if (REG_P (x))
4049 return (insn != NULL_RTX
4050 && find_regno_note (insn, REG_DEAD, REGNO (x)) != NULL_RTX);
4051 code = GET_CODE (x);
4052 fmt = GET_RTX_FORMAT (code);
4053 for (i = GET_RTX_LENGTH (code) - 1; i >= 0; i--)
4055 if (fmt[i] == 'e')
4057 if (dead_pseudo_p (XEXP (x, i), insn))
4058 return true;
4060 else if (fmt[i] == 'E')
4062 for (j = XVECLEN (x, i) - 1; j >= 0; j--)
4063 if (dead_pseudo_p (XVECEXP (x, i, j), insn))
4064 return true;
4067 return false;
4070 /* Return true if INSN contains a dying pseudo in INSN right hand
4071 side. */
4072 static bool
4073 insn_rhs_dead_pseudo_p (rtx_insn *insn)
4075 rtx set = single_set (insn);
4077 gcc_assert (set != NULL);
4078 return dead_pseudo_p (SET_SRC (set), insn);
4081 /* Return true if any init insn of REGNO contains a dying pseudo in
4082 insn right hand side. */
4083 static bool
4084 init_insn_rhs_dead_pseudo_p (int regno)
4086 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4088 if (insns == NULL)
4089 return false;
4090 for (; insns != NULL_RTX; insns = insns->next ())
4091 if (insn_rhs_dead_pseudo_p (insns->insn ()))
4092 return true;
4093 return false;
4096 /* Return TRUE if REGNO has a reverse equivalence. The equivalence is
4097 reverse only if we have one init insn with given REGNO as a
4098 source. */
4099 static bool
4100 reverse_equiv_p (int regno)
4102 rtx_insn_list *insns = ira_reg_equiv[regno].init_insns;
4103 rtx set;
4105 if (insns == NULL)
4106 return false;
4107 if (! INSN_P (insns->insn ())
4108 || insns->next () != NULL)
4109 return false;
4110 if ((set = single_set (insns->insn ())) == NULL_RTX)
4111 return false;
4112 return REG_P (SET_SRC (set)) && (int) REGNO (SET_SRC (set)) == regno;
4115 /* Return TRUE if REGNO was reloaded in an equivalence init insn. We
4116 call this function only for non-reverse equivalence. */
4117 static bool
4118 contains_reloaded_insn_p (int regno)
4120 rtx set;
4121 rtx_insn_list *list = ira_reg_equiv[regno].init_insns;
4123 for (; list != NULL; list = list->next ())
4124 if ((set = single_set (list->insn ())) == NULL_RTX
4125 || ! REG_P (SET_DEST (set))
4126 || (int) REGNO (SET_DEST (set)) != regno)
4127 return true;
4128 return false;
4131 /* Entry function of LRA constraint pass. Return true if the
4132 constraint pass did change the code. */
4133 bool
4134 lra_constraints (bool first_p)
4136 bool changed_p;
4137 int i, hard_regno, new_insns_num;
4138 unsigned int min_len, new_min_len, uid;
4139 rtx set, x, reg, dest_reg;
4140 basic_block last_bb;
4141 bitmap_head equiv_insn_bitmap;
4142 bitmap_iterator bi;
4144 lra_constraint_iter++;
4145 if (lra_dump_file != NULL)
4146 fprintf (lra_dump_file, "\n********** Local #%d: **********\n\n",
4147 lra_constraint_iter);
4148 changed_p = false;
4149 if (pic_offset_table_rtx
4150 && REGNO (pic_offset_table_rtx) >= FIRST_PSEUDO_REGISTER)
4151 lra_risky_transformations_p = true;
4152 else
4153 lra_risky_transformations_p = false;
4154 new_insn_uid_start = get_max_uid ();
4155 new_regno_start = first_p ? lra_constraint_new_regno_start : max_reg_num ();
4156 /* Mark used hard regs for target stack size calulations. */
4157 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4158 if (lra_reg_info[i].nrefs != 0
4159 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4161 int j, nregs;
4163 nregs = hard_regno_nregs[hard_regno][lra_reg_info[i].biggest_mode];
4164 for (j = 0; j < nregs; j++)
4165 df_set_regs_ever_live (hard_regno + j, true);
4167 /* Do elimination before the equivalence processing as we can spill
4168 some pseudos during elimination. */
4169 lra_eliminate (false, first_p);
4170 bitmap_initialize (&equiv_insn_bitmap, &reg_obstack);
4171 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4172 if (lra_reg_info[i].nrefs != 0)
4174 ira_reg_equiv[i].profitable_p = true;
4175 reg = regno_reg_rtx[i];
4176 if (lra_get_regno_hard_regno (i) < 0 && (x = get_equiv (reg)) != reg)
4178 bool pseudo_p = contains_reg_p (x, false, false);
4180 /* After RTL transformation, we can not guarantee that
4181 pseudo in the substitution was not reloaded which might
4182 make equivalence invalid. For example, in reverse
4183 equiv of p0
4185 p0 <- ...
4187 equiv_mem <- p0
4189 the memory address register was reloaded before the 2nd
4190 insn. */
4191 if ((! first_p && pseudo_p)
4192 /* We don't use DF for compilation speed sake. So it
4193 is problematic to update live info when we use an
4194 equivalence containing pseudos in more than one
4195 BB. */
4196 || (pseudo_p && multi_block_pseudo_p (i))
4197 /* If an init insn was deleted for some reason, cancel
4198 the equiv. We could update the equiv insns after
4199 transformations including an equiv insn deletion
4200 but it is not worthy as such cases are extremely
4201 rare. */
4202 || contains_deleted_insn_p (ira_reg_equiv[i].init_insns)
4203 /* If it is not a reverse equivalence, we check that a
4204 pseudo in rhs of the init insn is not dying in the
4205 insn. Otherwise, the live info at the beginning of
4206 the corresponding BB might be wrong after we
4207 removed the insn. When the equiv can be a
4208 constant, the right hand side of the init insn can
4209 be a pseudo. */
4210 || (! reverse_equiv_p (i)
4211 && (init_insn_rhs_dead_pseudo_p (i)
4212 /* If we reloaded the pseudo in an equivalence
4213 init insn, we can not remove the equiv init
4214 insns and the init insns might write into
4215 const memory in this case. */
4216 || contains_reloaded_insn_p (i)))
4217 /* Prevent access beyond equivalent memory for
4218 paradoxical subregs. */
4219 || (MEM_P (x)
4220 && (GET_MODE_SIZE (lra_reg_info[i].biggest_mode)
4221 > GET_MODE_SIZE (GET_MODE (x))))
4222 || (pic_offset_table_rtx
4223 && ((CONST_POOL_OK_P (PSEUDO_REGNO_MODE (i), x)
4224 && (targetm.preferred_reload_class
4225 (x, lra_get_allocno_class (i)) == NO_REGS))
4226 || contains_symbol_ref_p (x))))
4227 ira_reg_equiv[i].defined_p = false;
4228 if (contains_reg_p (x, false, true))
4229 ira_reg_equiv[i].profitable_p = false;
4230 if (get_equiv (reg) != reg)
4231 bitmap_ior_into (&equiv_insn_bitmap, &lra_reg_info[i].insn_bitmap);
4234 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4235 update_equiv (i);
4236 /* We should add all insns containing pseudos which should be
4237 substituted by their equivalences. */
4238 EXECUTE_IF_SET_IN_BITMAP (&equiv_insn_bitmap, 0, uid, bi)
4239 lra_push_insn_by_uid (uid);
4240 min_len = lra_insn_stack_length ();
4241 new_insns_num = 0;
4242 last_bb = NULL;
4243 changed_p = false;
4244 while ((new_min_len = lra_insn_stack_length ()) != 0)
4246 curr_insn = lra_pop_insn ();
4247 --new_min_len;
4248 curr_bb = BLOCK_FOR_INSN (curr_insn);
4249 if (curr_bb != last_bb)
4251 last_bb = curr_bb;
4252 bb_reload_num = lra_curr_reload_num;
4254 if (min_len > new_min_len)
4256 min_len = new_min_len;
4257 new_insns_num = 0;
4259 if (new_insns_num > MAX_RELOAD_INSNS_NUMBER)
4260 internal_error
4261 ("Max. number of generated reload insns per insn is achieved (%d)\n",
4262 MAX_RELOAD_INSNS_NUMBER);
4263 new_insns_num++;
4264 if (DEBUG_INSN_P (curr_insn))
4266 /* We need to check equivalence in debug insn and change
4267 pseudo to the equivalent value if necessary. */
4268 curr_id = lra_get_insn_recog_data (curr_insn);
4269 if (bitmap_bit_p (&equiv_insn_bitmap, INSN_UID (curr_insn)))
4271 rtx old = *curr_id->operand_loc[0];
4272 *curr_id->operand_loc[0]
4273 = simplify_replace_fn_rtx (old, NULL_RTX,
4274 loc_equivalence_callback, curr_insn);
4275 if (old != *curr_id->operand_loc[0])
4277 lra_update_insn_regno_info (curr_insn);
4278 changed_p = true;
4282 else if (INSN_P (curr_insn))
4284 if ((set = single_set (curr_insn)) != NULL_RTX)
4286 dest_reg = SET_DEST (set);
4287 /* The equivalence pseudo could be set up as SUBREG in a
4288 case when it is a call restore insn in a mode
4289 different from the pseudo mode. */
4290 if (GET_CODE (dest_reg) == SUBREG)
4291 dest_reg = SUBREG_REG (dest_reg);
4292 if ((REG_P (dest_reg)
4293 && (x = get_equiv (dest_reg)) != dest_reg
4294 /* Remove insns which set up a pseudo whose value
4295 can not be changed. Such insns might be not in
4296 init_insns because we don't update equiv data
4297 during insn transformations.
4299 As an example, let suppose that a pseudo got
4300 hard register and on the 1st pass was not
4301 changed to equivalent constant. We generate an
4302 additional insn setting up the pseudo because of
4303 secondary memory movement. Then the pseudo is
4304 spilled and we use the equiv constant. In this
4305 case we should remove the additional insn and
4306 this insn is not init_insns list. */
4307 && (! MEM_P (x) || MEM_READONLY_P (x)
4308 /* Check that this is actually an insn setting
4309 up the equivalence. */
4310 || in_list_p (curr_insn,
4311 ira_reg_equiv
4312 [REGNO (dest_reg)].init_insns)))
4313 || (((x = get_equiv (SET_SRC (set))) != SET_SRC (set))
4314 && in_list_p (curr_insn,
4315 ira_reg_equiv
4316 [REGNO (SET_SRC (set))].init_insns)))
4318 /* This is equiv init insn of pseudo which did not get a
4319 hard register -- remove the insn. */
4320 if (lra_dump_file != NULL)
4322 fprintf (lra_dump_file,
4323 " Removing equiv init insn %i (freq=%d)\n",
4324 INSN_UID (curr_insn),
4325 REG_FREQ_FROM_BB (BLOCK_FOR_INSN (curr_insn)));
4326 dump_insn_slim (lra_dump_file, curr_insn);
4328 if (contains_reg_p (x, true, false))
4329 lra_risky_transformations_p = true;
4330 lra_set_insn_deleted (curr_insn);
4331 continue;
4334 curr_id = lra_get_insn_recog_data (curr_insn);
4335 curr_static_id = curr_id->insn_static_data;
4336 init_curr_insn_input_reloads ();
4337 init_curr_operand_mode ();
4338 if (curr_insn_transform (false))
4339 changed_p = true;
4340 /* Check non-transformed insns too for equiv change as USE
4341 or CLOBBER don't need reloads but can contain pseudos
4342 being changed on their equivalences. */
4343 else if (bitmap_bit_p (&equiv_insn_bitmap, INSN_UID (curr_insn))
4344 && loc_equivalence_change_p (&PATTERN (curr_insn)))
4346 lra_update_insn_regno_info (curr_insn);
4347 changed_p = true;
4351 bitmap_clear (&equiv_insn_bitmap);
4352 /* If we used a new hard regno, changed_p should be true because the
4353 hard reg is assigned to a new pseudo. */
4354 #ifdef ENABLE_CHECKING
4355 if (! changed_p)
4357 for (i = FIRST_PSEUDO_REGISTER; i < new_regno_start; i++)
4358 if (lra_reg_info[i].nrefs != 0
4359 && (hard_regno = lra_get_regno_hard_regno (i)) >= 0)
4361 int j, nregs = hard_regno_nregs[hard_regno][PSEUDO_REGNO_MODE (i)];
4363 for (j = 0; j < nregs; j++)
4364 lra_assert (df_regs_ever_live_p (hard_regno + j));
4367 #endif
4368 return changed_p;
4371 /* Initiate the LRA constraint pass. It is done once per
4372 function. */
4373 void
4374 lra_constraints_init (void)
4378 /* Finalize the LRA constraint pass. It is done once per
4379 function. */
4380 void
4381 lra_constraints_finish (void)
4387 /* This page contains code to do inheritance/split
4388 transformations. */
4390 /* Number of reloads passed so far in current EBB. */
4391 static int reloads_num;
4393 /* Number of calls passed so far in current EBB. */
4394 static int calls_num;
4396 /* Current reload pseudo check for validity of elements in
4397 USAGE_INSNS. */
4398 static int curr_usage_insns_check;
4400 /* Info about last usage of registers in EBB to do inheritance/split
4401 transformation. Inheritance transformation is done from a spilled
4402 pseudo and split transformations from a hard register or a pseudo
4403 assigned to a hard register. */
4404 struct usage_insns
4406 /* If the value is equal to CURR_USAGE_INSNS_CHECK, then the member
4407 value INSNS is valid. The insns is chain of optional debug insns
4408 and a finishing non-debug insn using the corresponding reg. The
4409 value is also used to mark the registers which are set up in the
4410 current insn. The negated insn uid is used for this. */
4411 int check;
4412 /* Value of global reloads_num at the last insn in INSNS. */
4413 int reloads_num;
4414 /* Value of global reloads_nums at the last insn in INSNS. */
4415 int calls_num;
4416 /* It can be true only for splitting. And it means that the restore
4417 insn should be put after insn given by the following member. */
4418 bool after_p;
4419 /* Next insns in the current EBB which use the original reg and the
4420 original reg value is not changed between the current insn and
4421 the next insns. In order words, e.g. for inheritance, if we need
4422 to use the original reg value again in the next insns we can try
4423 to use the value in a hard register from a reload insn of the
4424 current insn. */
4425 rtx insns;
4428 /* Map: regno -> corresponding pseudo usage insns. */
4429 static struct usage_insns *usage_insns;
4431 static void
4432 setup_next_usage_insn (int regno, rtx insn, int reloads_num, bool after_p)
4434 usage_insns[regno].check = curr_usage_insns_check;
4435 usage_insns[regno].insns = insn;
4436 usage_insns[regno].reloads_num = reloads_num;
4437 usage_insns[regno].calls_num = calls_num;
4438 usage_insns[regno].after_p = after_p;
4441 /* The function is used to form list REGNO usages which consists of
4442 optional debug insns finished by a non-debug insn using REGNO.
4443 RELOADS_NUM is current number of reload insns processed so far. */
4444 static void
4445 add_next_usage_insn (int regno, rtx insn, int reloads_num)
4447 rtx next_usage_insns;
4449 if (usage_insns[regno].check == curr_usage_insns_check
4450 && (next_usage_insns = usage_insns[regno].insns) != NULL_RTX
4451 && DEBUG_INSN_P (insn))
4453 /* Check that we did not add the debug insn yet. */
4454 if (next_usage_insns != insn
4455 && (GET_CODE (next_usage_insns) != INSN_LIST
4456 || XEXP (next_usage_insns, 0) != insn))
4457 usage_insns[regno].insns = gen_rtx_INSN_LIST (VOIDmode, insn,
4458 next_usage_insns);
4460 else if (NONDEBUG_INSN_P (insn))
4461 setup_next_usage_insn (regno, insn, reloads_num, false);
4462 else
4463 usage_insns[regno].check = 0;
4466 /* Return first non-debug insn in list USAGE_INSNS. */
4467 static rtx_insn *
4468 skip_usage_debug_insns (rtx usage_insns)
4470 rtx insn;
4472 /* Skip debug insns. */
4473 for (insn = usage_insns;
4474 insn != NULL_RTX && GET_CODE (insn) == INSN_LIST;
4475 insn = XEXP (insn, 1))
4477 return safe_as_a <rtx_insn *> (insn);
4480 /* Return true if we need secondary memory moves for insn in
4481 USAGE_INSNS after inserting inherited pseudo of class INHER_CL
4482 into the insn. */
4483 static bool
4484 check_secondary_memory_needed_p (enum reg_class inher_cl ATTRIBUTE_UNUSED,
4485 rtx usage_insns ATTRIBUTE_UNUSED)
4487 #ifndef SECONDARY_MEMORY_NEEDED
4488 return false;
4489 #else
4490 rtx_insn *insn;
4491 rtx set, dest;
4492 enum reg_class cl;
4494 if (inher_cl == ALL_REGS
4495 || (insn = skip_usage_debug_insns (usage_insns)) == NULL_RTX)
4496 return false;
4497 lra_assert (INSN_P (insn));
4498 if ((set = single_set (insn)) == NULL_RTX || ! REG_P (SET_DEST (set)))
4499 return false;
4500 dest = SET_DEST (set);
4501 if (! REG_P (dest))
4502 return false;
4503 lra_assert (inher_cl != NO_REGS);
4504 cl = get_reg_class (REGNO (dest));
4505 return (cl != NO_REGS && cl != ALL_REGS
4506 && SECONDARY_MEMORY_NEEDED (inher_cl, cl, GET_MODE (dest)));
4507 #endif
4510 /* Registers involved in inheritance/split in the current EBB
4511 (inheritance/split pseudos and original registers). */
4512 static bitmap_head check_only_regs;
4514 /* Do inheritance transformations for insn INSN, which defines (if
4515 DEF_P) or uses ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which
4516 instruction in the EBB next uses ORIGINAL_REGNO; it has the same
4517 form as the "insns" field of usage_insns. Return true if we
4518 succeed in such transformation.
4520 The transformations look like:
4522 p <- ... i <- ...
4523 ... p <- i (new insn)
4524 ... =>
4525 <- ... p ... <- ... i ...
4527 ... i <- p (new insn)
4528 <- ... p ... <- ... i ...
4529 ... =>
4530 <- ... p ... <- ... i ...
4531 where p is a spilled original pseudo and i is a new inheritance pseudo.
4534 The inheritance pseudo has the smallest class of two classes CL and
4535 class of ORIGINAL REGNO. */
4536 static bool
4537 inherit_reload_reg (bool def_p, int original_regno,
4538 enum reg_class cl, rtx_insn *insn, rtx next_usage_insns)
4540 if (optimize_function_for_size_p (cfun))
4541 return false;
4543 enum reg_class rclass = lra_get_allocno_class (original_regno);
4544 rtx original_reg = regno_reg_rtx[original_regno];
4545 rtx new_reg, usage_insn;
4546 rtx_insn *new_insns;
4548 lra_assert (! usage_insns[original_regno].after_p);
4549 if (lra_dump_file != NULL)
4550 fprintf (lra_dump_file,
4551 " <<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<\n");
4552 if (! ira_reg_classes_intersect_p[cl][rclass])
4554 if (lra_dump_file != NULL)
4556 fprintf (lra_dump_file,
4557 " Rejecting inheritance for %d "
4558 "because of disjoint classes %s and %s\n",
4559 original_regno, reg_class_names[cl],
4560 reg_class_names[rclass]);
4561 fprintf (lra_dump_file,
4562 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
4564 return false;
4566 if ((ira_class_subset_p[cl][rclass] && cl != rclass)
4567 /* We don't use a subset of two classes because it can be
4568 NO_REGS. This transformation is still profitable in most
4569 cases even if the classes are not intersected as register
4570 move is probably cheaper than a memory load. */
4571 || ira_class_hard_regs_num[cl] < ira_class_hard_regs_num[rclass])
4573 if (lra_dump_file != NULL)
4574 fprintf (lra_dump_file, " Use smallest class of %s and %s\n",
4575 reg_class_names[cl], reg_class_names[rclass]);
4577 rclass = cl;
4579 if (check_secondary_memory_needed_p (rclass, next_usage_insns))
4581 /* Reject inheritance resulting in secondary memory moves.
4582 Otherwise, there is a danger in LRA cycling. Also such
4583 transformation will be unprofitable. */
4584 if (lra_dump_file != NULL)
4586 rtx_insn *insn = skip_usage_debug_insns (next_usage_insns);
4587 rtx set = single_set (insn);
4589 lra_assert (set != NULL_RTX);
4591 rtx dest = SET_DEST (set);
4593 lra_assert (REG_P (dest));
4594 fprintf (lra_dump_file,
4595 " Rejecting inheritance for insn %d(%s)<-%d(%s) "
4596 "as secondary mem is needed\n",
4597 REGNO (dest), reg_class_names[get_reg_class (REGNO (dest))],
4598 original_regno, reg_class_names[rclass]);
4599 fprintf (lra_dump_file,
4600 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
4602 return false;
4604 new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg,
4605 rclass, "inheritance");
4606 start_sequence ();
4607 if (def_p)
4608 lra_emit_move (original_reg, new_reg);
4609 else
4610 lra_emit_move (new_reg, original_reg);
4611 new_insns = get_insns ();
4612 end_sequence ();
4613 if (NEXT_INSN (new_insns) != NULL_RTX)
4615 if (lra_dump_file != NULL)
4617 fprintf (lra_dump_file,
4618 " Rejecting inheritance %d->%d "
4619 "as it results in 2 or more insns:\n",
4620 original_regno, REGNO (new_reg));
4621 dump_rtl_slim (lra_dump_file, new_insns, NULL, -1, 0);
4622 fprintf (lra_dump_file,
4623 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
4625 return false;
4627 lra_substitute_pseudo_within_insn (insn, original_regno, new_reg);
4628 lra_update_insn_regno_info (insn);
4629 if (! def_p)
4630 /* We now have a new usage insn for original regno. */
4631 setup_next_usage_insn (original_regno, new_insns, reloads_num, false);
4632 if (lra_dump_file != NULL)
4633 fprintf (lra_dump_file, " Original reg change %d->%d (bb%d):\n",
4634 original_regno, REGNO (new_reg), BLOCK_FOR_INSN (insn)->index);
4635 lra_reg_info[REGNO (new_reg)].restore_regno = original_regno;
4636 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
4637 bitmap_set_bit (&check_only_regs, original_regno);
4638 bitmap_set_bit (&lra_inheritance_pseudos, REGNO (new_reg));
4639 if (def_p)
4640 lra_process_new_insns (insn, NULL, new_insns,
4641 "Add original<-inheritance");
4642 else
4643 lra_process_new_insns (insn, new_insns, NULL,
4644 "Add inheritance<-original");
4645 while (next_usage_insns != NULL_RTX)
4647 if (GET_CODE (next_usage_insns) != INSN_LIST)
4649 usage_insn = next_usage_insns;
4650 lra_assert (NONDEBUG_INSN_P (usage_insn));
4651 next_usage_insns = NULL;
4653 else
4655 usage_insn = XEXP (next_usage_insns, 0);
4656 lra_assert (DEBUG_INSN_P (usage_insn));
4657 next_usage_insns = XEXP (next_usage_insns, 1);
4659 lra_substitute_pseudo (&usage_insn, original_regno, new_reg);
4660 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
4661 if (lra_dump_file != NULL)
4663 fprintf (lra_dump_file,
4664 " Inheritance reuse change %d->%d (bb%d):\n",
4665 original_regno, REGNO (new_reg),
4666 BLOCK_FOR_INSN (usage_insn)->index);
4667 dump_insn_slim (lra_dump_file, usage_insn);
4670 if (lra_dump_file != NULL)
4671 fprintf (lra_dump_file,
4672 " >>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>\n");
4673 return true;
4676 /* Return true if we need a caller save/restore for pseudo REGNO which
4677 was assigned to a hard register. */
4678 static inline bool
4679 need_for_call_save_p (int regno)
4681 lra_assert (regno >= FIRST_PSEUDO_REGISTER && reg_renumber[regno] >= 0);
4682 return (usage_insns[regno].calls_num < calls_num
4683 && (overlaps_hard_reg_set_p
4684 ((flag_ipa_ra &&
4685 ! hard_reg_set_empty_p (lra_reg_info[regno].actual_call_used_reg_set))
4686 ? lra_reg_info[regno].actual_call_used_reg_set
4687 : call_used_reg_set,
4688 PSEUDO_REGNO_MODE (regno), reg_renumber[regno])
4689 || HARD_REGNO_CALL_PART_CLOBBERED (reg_renumber[regno],
4690 PSEUDO_REGNO_MODE (regno))));
4693 /* Global registers occurring in the current EBB. */
4694 static bitmap_head ebb_global_regs;
4696 /* Return true if we need a split for hard register REGNO or pseudo
4697 REGNO which was assigned to a hard register.
4698 POTENTIAL_RELOAD_HARD_REGS contains hard registers which might be
4699 used for reloads since the EBB end. It is an approximation of the
4700 used hard registers in the split range. The exact value would
4701 require expensive calculations. If we were aggressive with
4702 splitting because of the approximation, the split pseudo will save
4703 the same hard register assignment and will be removed in the undo
4704 pass. We still need the approximation because too aggressive
4705 splitting would result in too inaccurate cost calculation in the
4706 assignment pass because of too many generated moves which will be
4707 probably removed in the undo pass. */
4708 static inline bool
4709 need_for_split_p (HARD_REG_SET potential_reload_hard_regs, int regno)
4711 int hard_regno = regno < FIRST_PSEUDO_REGISTER ? regno : reg_renumber[regno];
4713 lra_assert (hard_regno >= 0);
4714 return ((TEST_HARD_REG_BIT (potential_reload_hard_regs, hard_regno)
4715 /* Don't split eliminable hard registers, otherwise we can
4716 split hard registers like hard frame pointer, which
4717 lives on BB start/end according to DF-infrastructure,
4718 when there is a pseudo assigned to the register and
4719 living in the same BB. */
4720 && (regno >= FIRST_PSEUDO_REGISTER
4721 || ! TEST_HARD_REG_BIT (eliminable_regset, hard_regno))
4722 && ! TEST_HARD_REG_BIT (lra_no_alloc_regs, hard_regno)
4723 /* Don't split call clobbered hard regs living through
4724 calls, otherwise we might have a check problem in the
4725 assign sub-pass as in the most cases (exception is a
4726 situation when lra_risky_transformations_p value is
4727 true) the assign pass assumes that all pseudos living
4728 through calls are assigned to call saved hard regs. */
4729 && (regno >= FIRST_PSEUDO_REGISTER
4730 || ! TEST_HARD_REG_BIT (call_used_reg_set, regno)
4731 || usage_insns[regno].calls_num == calls_num)
4732 /* We need at least 2 reloads to make pseudo splitting
4733 profitable. We should provide hard regno splitting in
4734 any case to solve 1st insn scheduling problem when
4735 moving hard register definition up might result in
4736 impossibility to find hard register for reload pseudo of
4737 small register class. */
4738 && (usage_insns[regno].reloads_num
4739 + (regno < FIRST_PSEUDO_REGISTER ? 0 : 3) < reloads_num)
4740 && (regno < FIRST_PSEUDO_REGISTER
4741 /* For short living pseudos, spilling + inheritance can
4742 be considered a substitution for splitting.
4743 Therefore we do not splitting for local pseudos. It
4744 decreases also aggressiveness of splitting. The
4745 minimal number of references is chosen taking into
4746 account that for 2 references splitting has no sense
4747 as we can just spill the pseudo. */
4748 || (regno >= FIRST_PSEUDO_REGISTER
4749 && lra_reg_info[regno].nrefs > 3
4750 && bitmap_bit_p (&ebb_global_regs, regno))))
4751 || (regno >= FIRST_PSEUDO_REGISTER && need_for_call_save_p (regno)));
4754 /* Return class for the split pseudo created from original pseudo with
4755 ALLOCNO_CLASS and MODE which got a hard register HARD_REGNO. We
4756 choose subclass of ALLOCNO_CLASS which contains HARD_REGNO and
4757 results in no secondary memory movements. */
4758 static enum reg_class
4759 choose_split_class (enum reg_class allocno_class,
4760 int hard_regno ATTRIBUTE_UNUSED,
4761 machine_mode mode ATTRIBUTE_UNUSED)
4763 #ifndef SECONDARY_MEMORY_NEEDED
4764 return allocno_class;
4765 #else
4766 int i;
4767 enum reg_class cl, best_cl = NO_REGS;
4768 enum reg_class hard_reg_class ATTRIBUTE_UNUSED
4769 = REGNO_REG_CLASS (hard_regno);
4771 if (! SECONDARY_MEMORY_NEEDED (allocno_class, allocno_class, mode)
4772 && TEST_HARD_REG_BIT (reg_class_contents[allocno_class], hard_regno))
4773 return allocno_class;
4774 for (i = 0;
4775 (cl = reg_class_subclasses[allocno_class][i]) != LIM_REG_CLASSES;
4776 i++)
4777 if (! SECONDARY_MEMORY_NEEDED (cl, hard_reg_class, mode)
4778 && ! SECONDARY_MEMORY_NEEDED (hard_reg_class, cl, mode)
4779 && TEST_HARD_REG_BIT (reg_class_contents[cl], hard_regno)
4780 && (best_cl == NO_REGS
4781 || ira_class_hard_regs_num[best_cl] < ira_class_hard_regs_num[cl]))
4782 best_cl = cl;
4783 return best_cl;
4784 #endif
4787 /* Do split transformations for insn INSN, which defines or uses
4788 ORIGINAL_REGNO. NEXT_USAGE_INSNS specifies which instruction in
4789 the EBB next uses ORIGINAL_REGNO; it has the same form as the
4790 "insns" field of usage_insns.
4792 The transformations look like:
4794 p <- ... p <- ...
4795 ... s <- p (new insn -- save)
4796 ... =>
4797 ... p <- s (new insn -- restore)
4798 <- ... p ... <- ... p ...
4800 <- ... p ... <- ... p ...
4801 ... s <- p (new insn -- save)
4802 ... =>
4803 ... p <- s (new insn -- restore)
4804 <- ... p ... <- ... p ...
4806 where p is an original pseudo got a hard register or a hard
4807 register and s is a new split pseudo. The save is put before INSN
4808 if BEFORE_P is true. Return true if we succeed in such
4809 transformation. */
4810 static bool
4811 split_reg (bool before_p, int original_regno, rtx_insn *insn,
4812 rtx next_usage_insns)
4814 enum reg_class rclass;
4815 rtx original_reg;
4816 int hard_regno, nregs;
4817 rtx new_reg, usage_insn;
4818 rtx_insn *restore, *save;
4819 bool after_p;
4820 bool call_save_p;
4822 if (original_regno < FIRST_PSEUDO_REGISTER)
4824 rclass = ira_allocno_class_translate[REGNO_REG_CLASS (original_regno)];
4825 hard_regno = original_regno;
4826 call_save_p = false;
4827 nregs = 1;
4829 else
4831 hard_regno = reg_renumber[original_regno];
4832 nregs = hard_regno_nregs[hard_regno][PSEUDO_REGNO_MODE (original_regno)];
4833 rclass = lra_get_allocno_class (original_regno);
4834 original_reg = regno_reg_rtx[original_regno];
4835 call_save_p = need_for_call_save_p (original_regno);
4837 original_reg = regno_reg_rtx[original_regno];
4838 lra_assert (hard_regno >= 0);
4839 if (lra_dump_file != NULL)
4840 fprintf (lra_dump_file,
4841 " ((((((((((((((((((((((((((((((((((((((((((((((((\n");
4842 if (call_save_p)
4844 machine_mode mode = GET_MODE (original_reg);
4846 mode = HARD_REGNO_CALLER_SAVE_MODE (hard_regno,
4847 hard_regno_nregs[hard_regno][mode],
4848 mode);
4849 new_reg = lra_create_new_reg (mode, NULL_RTX, NO_REGS, "save");
4851 else
4853 rclass = choose_split_class (rclass, hard_regno,
4854 GET_MODE (original_reg));
4855 if (rclass == NO_REGS)
4857 if (lra_dump_file != NULL)
4859 fprintf (lra_dump_file,
4860 " Rejecting split of %d(%s): "
4861 "no good reg class for %d(%s)\n",
4862 original_regno,
4863 reg_class_names[lra_get_allocno_class (original_regno)],
4864 hard_regno,
4865 reg_class_names[REGNO_REG_CLASS (hard_regno)]);
4866 fprintf
4867 (lra_dump_file,
4868 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
4870 return false;
4872 new_reg = lra_create_new_reg (GET_MODE (original_reg), original_reg,
4873 rclass, "split");
4874 reg_renumber[REGNO (new_reg)] = hard_regno;
4876 save = emit_spill_move (true, new_reg, original_reg);
4877 if (NEXT_INSN (save) != NULL_RTX)
4879 lra_assert (! call_save_p);
4880 if (lra_dump_file != NULL)
4882 fprintf
4883 (lra_dump_file,
4884 " Rejecting split %d->%d resulting in > 2 %s save insns:\n",
4885 original_regno, REGNO (new_reg), call_save_p ? "call" : "");
4886 dump_rtl_slim (lra_dump_file, save, NULL, -1, 0);
4887 fprintf (lra_dump_file,
4888 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
4890 return false;
4892 restore = emit_spill_move (false, new_reg, original_reg);
4893 if (NEXT_INSN (restore) != NULL_RTX)
4895 lra_assert (! call_save_p);
4896 if (lra_dump_file != NULL)
4898 fprintf (lra_dump_file,
4899 " Rejecting split %d->%d "
4900 "resulting in > 2 %s restore insns:\n",
4901 original_regno, REGNO (new_reg), call_save_p ? "call" : "");
4902 dump_rtl_slim (lra_dump_file, restore, NULL, -1, 0);
4903 fprintf (lra_dump_file,
4904 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
4906 return false;
4908 after_p = usage_insns[original_regno].after_p;
4909 lra_reg_info[REGNO (new_reg)].restore_regno = original_regno;
4910 bitmap_set_bit (&check_only_regs, REGNO (new_reg));
4911 bitmap_set_bit (&check_only_regs, original_regno);
4912 bitmap_set_bit (&lra_split_regs, REGNO (new_reg));
4913 for (;;)
4915 if (GET_CODE (next_usage_insns) != INSN_LIST)
4917 usage_insn = next_usage_insns;
4918 break;
4920 usage_insn = XEXP (next_usage_insns, 0);
4921 lra_assert (DEBUG_INSN_P (usage_insn));
4922 next_usage_insns = XEXP (next_usage_insns, 1);
4923 lra_substitute_pseudo (&usage_insn, original_regno, new_reg);
4924 lra_update_insn_regno_info (as_a <rtx_insn *> (usage_insn));
4925 if (lra_dump_file != NULL)
4927 fprintf (lra_dump_file, " Split reuse change %d->%d:\n",
4928 original_regno, REGNO (new_reg));
4929 dump_insn_slim (lra_dump_file, usage_insn);
4932 lra_assert (NOTE_P (usage_insn) || NONDEBUG_INSN_P (usage_insn));
4933 lra_assert (usage_insn != insn || (after_p && before_p));
4934 lra_process_new_insns (as_a <rtx_insn *> (usage_insn),
4935 after_p ? NULL : restore,
4936 after_p ? restore : NULL,
4937 call_save_p
4938 ? "Add reg<-save" : "Add reg<-split");
4939 lra_process_new_insns (insn, before_p ? save : NULL,
4940 before_p ? NULL : save,
4941 call_save_p
4942 ? "Add save<-reg" : "Add split<-reg");
4943 if (nregs > 1)
4944 /* If we are trying to split multi-register. We should check
4945 conflicts on the next assignment sub-pass. IRA can allocate on
4946 sub-register levels, LRA do this on pseudos level right now and
4947 this discrepancy may create allocation conflicts after
4948 splitting. */
4949 lra_risky_transformations_p = true;
4950 if (lra_dump_file != NULL)
4951 fprintf (lra_dump_file,
4952 " ))))))))))))))))))))))))))))))))))))))))))))))))\n");
4953 return true;
4956 /* Recognize that we need a split transformation for insn INSN, which
4957 defines or uses REGNO in its insn biggest MODE (we use it only if
4958 REGNO is a hard register). POTENTIAL_RELOAD_HARD_REGS contains
4959 hard registers which might be used for reloads since the EBB end.
4960 Put the save before INSN if BEFORE_P is true. MAX_UID is maximla
4961 uid before starting INSN processing. Return true if we succeed in
4962 such transformation. */
4963 static bool
4964 split_if_necessary (int regno, machine_mode mode,
4965 HARD_REG_SET potential_reload_hard_regs,
4966 bool before_p, rtx_insn *insn, int max_uid)
4968 bool res = false;
4969 int i, nregs = 1;
4970 rtx next_usage_insns;
4972 if (regno < FIRST_PSEUDO_REGISTER)
4973 nregs = hard_regno_nregs[regno][mode];
4974 for (i = 0; i < nregs; i++)
4975 if (usage_insns[regno + i].check == curr_usage_insns_check
4976 && (next_usage_insns = usage_insns[regno + i].insns) != NULL_RTX
4977 /* To avoid processing the register twice or more. */
4978 && ((GET_CODE (next_usage_insns) != INSN_LIST
4979 && INSN_UID (next_usage_insns) < max_uid)
4980 || (GET_CODE (next_usage_insns) == INSN_LIST
4981 && (INSN_UID (XEXP (next_usage_insns, 0)) < max_uid)))
4982 && need_for_split_p (potential_reload_hard_regs, regno + i)
4983 && split_reg (before_p, regno + i, insn, next_usage_insns))
4984 res = true;
4985 return res;
4988 /* Check only registers living at the current program point in the
4989 current EBB. */
4990 static bitmap_head live_regs;
4992 /* Update live info in EBB given by its HEAD and TAIL insns after
4993 inheritance/split transformation. The function removes dead moves
4994 too. */
4995 static void
4996 update_ebb_live_info (rtx_insn *head, rtx_insn *tail)
4998 unsigned int j;
4999 int i, regno;
5000 bool live_p;
5001 rtx_insn *prev_insn;
5002 rtx set;
5003 bool remove_p;
5004 basic_block last_bb, prev_bb, curr_bb;
5005 bitmap_iterator bi;
5006 struct lra_insn_reg *reg;
5007 edge e;
5008 edge_iterator ei;
5010 last_bb = BLOCK_FOR_INSN (tail);
5011 prev_bb = NULL;
5012 for (curr_insn = tail;
5013 curr_insn != PREV_INSN (head);
5014 curr_insn = prev_insn)
5016 prev_insn = PREV_INSN (curr_insn);
5017 /* We need to process empty blocks too. They contain
5018 NOTE_INSN_BASIC_BLOCK referring for the basic block. */
5019 if (NOTE_P (curr_insn) && NOTE_KIND (curr_insn) != NOTE_INSN_BASIC_BLOCK)
5020 continue;
5021 curr_bb = BLOCK_FOR_INSN (curr_insn);
5022 if (curr_bb != prev_bb)
5024 if (prev_bb != NULL)
5026 /* Update df_get_live_in (prev_bb): */
5027 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5028 if (bitmap_bit_p (&live_regs, j))
5029 bitmap_set_bit (df_get_live_in (prev_bb), j);
5030 else
5031 bitmap_clear_bit (df_get_live_in (prev_bb), j);
5033 if (curr_bb != last_bb)
5035 /* Update df_get_live_out (curr_bb): */
5036 EXECUTE_IF_SET_IN_BITMAP (&check_only_regs, 0, j, bi)
5038 live_p = bitmap_bit_p (&live_regs, j);
5039 if (! live_p)
5040 FOR_EACH_EDGE (e, ei, curr_bb->succs)
5041 if (bitmap_bit_p (df_get_live_in (e->dest), j))
5043 live_p = true;
5044 break;
5046 if (live_p)
5047 bitmap_set_bit (df_get_live_out (curr_bb), j);
5048 else
5049 bitmap_clear_bit (df_get_live_out (curr_bb), j);
5052 prev_bb = curr_bb;
5053 bitmap_and (&live_regs, &check_only_regs, df_get_live_out (curr_bb));
5055 if (! NONDEBUG_INSN_P (curr_insn))
5056 continue;
5057 curr_id = lra_get_insn_recog_data (curr_insn);
5058 curr_static_id = curr_id->insn_static_data;
5059 remove_p = false;
5060 if ((set = single_set (curr_insn)) != NULL_RTX && REG_P (SET_DEST (set))
5061 && (regno = REGNO (SET_DEST (set))) >= FIRST_PSEUDO_REGISTER
5062 && bitmap_bit_p (&check_only_regs, regno)
5063 && ! bitmap_bit_p (&live_regs, regno))
5064 remove_p = true;
5065 /* See which defined values die here. */
5066 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5067 if (reg->type == OP_OUT && ! reg->subreg_p)
5068 bitmap_clear_bit (&live_regs, reg->regno);
5069 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5070 if (reg->type == OP_OUT && ! reg->subreg_p)
5071 bitmap_clear_bit (&live_regs, reg->regno);
5072 /* Mark each used value as live. */
5073 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5074 if (reg->type != OP_OUT
5075 && bitmap_bit_p (&check_only_regs, reg->regno))
5076 bitmap_set_bit (&live_regs, reg->regno);
5077 for (reg = curr_static_id->hard_regs; reg != NULL; reg = reg->next)
5078 if (reg->type != OP_OUT
5079 && bitmap_bit_p (&check_only_regs, reg->regno))
5080 bitmap_set_bit (&live_regs, reg->regno);
5081 if (curr_id->arg_hard_regs != NULL)
5082 /* Make argument hard registers live. */
5083 for (i = 0; (regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5084 if (bitmap_bit_p (&check_only_regs, regno))
5085 bitmap_set_bit (&live_regs, regno);
5086 /* It is quite important to remove dead move insns because it
5087 means removing dead store. We don't need to process them for
5088 constraints. */
5089 if (remove_p)
5091 if (lra_dump_file != NULL)
5093 fprintf (lra_dump_file, " Removing dead insn:\n ");
5094 dump_insn_slim (lra_dump_file, curr_insn);
5096 lra_set_insn_deleted (curr_insn);
5101 /* The structure describes info to do an inheritance for the current
5102 insn. We need to collect such info first before doing the
5103 transformations because the transformations change the insn
5104 internal representation. */
5105 struct to_inherit
5107 /* Original regno. */
5108 int regno;
5109 /* Subsequent insns which can inherit original reg value. */
5110 rtx insns;
5113 /* Array containing all info for doing inheritance from the current
5114 insn. */
5115 static struct to_inherit to_inherit[LRA_MAX_INSN_RELOADS];
5117 /* Number elements in the previous array. */
5118 static int to_inherit_num;
5120 /* Add inheritance info REGNO and INSNS. Their meaning is described in
5121 structure to_inherit. */
5122 static void
5123 add_to_inherit (int regno, rtx insns)
5125 int i;
5127 for (i = 0; i < to_inherit_num; i++)
5128 if (to_inherit[i].regno == regno)
5129 return;
5130 lra_assert (to_inherit_num < LRA_MAX_INSN_RELOADS);
5131 to_inherit[to_inherit_num].regno = regno;
5132 to_inherit[to_inherit_num++].insns = insns;
5135 /* Return the last non-debug insn in basic block BB, or the block begin
5136 note if none. */
5137 static rtx_insn *
5138 get_last_insertion_point (basic_block bb)
5140 rtx_insn *insn;
5142 FOR_BB_INSNS_REVERSE (bb, insn)
5143 if (NONDEBUG_INSN_P (insn) || NOTE_INSN_BASIC_BLOCK_P (insn))
5144 return insn;
5145 gcc_unreachable ();
5148 /* Set up RES by registers living on edges FROM except the edge (FROM,
5149 TO) or by registers set up in a jump insn in BB FROM. */
5150 static void
5151 get_live_on_other_edges (basic_block from, basic_block to, bitmap res)
5153 rtx_insn *last;
5154 struct lra_insn_reg *reg;
5155 edge e;
5156 edge_iterator ei;
5158 lra_assert (to != NULL);
5159 bitmap_clear (res);
5160 FOR_EACH_EDGE (e, ei, from->succs)
5161 if (e->dest != to)
5162 bitmap_ior_into (res, df_get_live_in (e->dest));
5163 last = get_last_insertion_point (from);
5164 if (! JUMP_P (last))
5165 return;
5166 curr_id = lra_get_insn_recog_data (last);
5167 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5168 if (reg->type != OP_IN)
5169 bitmap_set_bit (res, reg->regno);
5172 /* Used as a temporary results of some bitmap calculations. */
5173 static bitmap_head temp_bitmap;
5175 /* We split for reloads of small class of hard regs. The following
5176 defines how many hard regs the class should have to be qualified as
5177 small. The code is mostly oriented to x86/x86-64 architecture
5178 where some insns need to use only specific register or pair of
5179 registers and these register can live in RTL explicitly, e.g. for
5180 parameter passing. */
5181 static const int max_small_class_regs_num = 2;
5183 /* Do inheritance/split transformations in EBB starting with HEAD and
5184 finishing on TAIL. We process EBB insns in the reverse order.
5185 Return true if we did any inheritance/split transformation in the
5186 EBB.
5188 We should avoid excessive splitting which results in worse code
5189 because of inaccurate cost calculations for spilling new split
5190 pseudos in such case. To achieve this we do splitting only if
5191 register pressure is high in given basic block and there are reload
5192 pseudos requiring hard registers. We could do more register
5193 pressure calculations at any given program point to avoid necessary
5194 splitting even more but it is to expensive and the current approach
5195 works well enough. */
5196 static bool
5197 inherit_in_ebb (rtx_insn *head, rtx_insn *tail)
5199 int i, src_regno, dst_regno, nregs;
5200 bool change_p, succ_p, update_reloads_num_p;
5201 rtx_insn *prev_insn, *last_insn;
5202 rtx next_usage_insns, set;
5203 enum reg_class cl;
5204 struct lra_insn_reg *reg;
5205 basic_block last_processed_bb, curr_bb = NULL;
5206 HARD_REG_SET potential_reload_hard_regs, live_hard_regs;
5207 bitmap to_process;
5208 unsigned int j;
5209 bitmap_iterator bi;
5210 bool head_p, after_p;
5212 change_p = false;
5213 curr_usage_insns_check++;
5214 reloads_num = calls_num = 0;
5215 bitmap_clear (&check_only_regs);
5216 last_processed_bb = NULL;
5217 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
5218 COPY_HARD_REG_SET (live_hard_regs, eliminable_regset);
5219 IOR_HARD_REG_SET (live_hard_regs, lra_no_alloc_regs);
5220 /* We don't process new insns generated in the loop. */
5221 for (curr_insn = tail; curr_insn != PREV_INSN (head); curr_insn = prev_insn)
5223 prev_insn = PREV_INSN (curr_insn);
5224 if (BLOCK_FOR_INSN (curr_insn) != NULL)
5225 curr_bb = BLOCK_FOR_INSN (curr_insn);
5226 if (last_processed_bb != curr_bb)
5228 /* We are at the end of BB. Add qualified living
5229 pseudos for potential splitting. */
5230 to_process = df_get_live_out (curr_bb);
5231 if (last_processed_bb != NULL)
5233 /* We are somewhere in the middle of EBB. */
5234 get_live_on_other_edges (curr_bb, last_processed_bb,
5235 &temp_bitmap);
5236 to_process = &temp_bitmap;
5238 last_processed_bb = curr_bb;
5239 last_insn = get_last_insertion_point (curr_bb);
5240 after_p = (! JUMP_P (last_insn)
5241 && (! CALL_P (last_insn)
5242 || (find_reg_note (last_insn,
5243 REG_NORETURN, NULL_RTX) == NULL_RTX
5244 && ! SIBLING_CALL_P (last_insn))));
5245 CLEAR_HARD_REG_SET (potential_reload_hard_regs);
5246 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
5248 if ((int) j >= lra_constraint_new_regno_start)
5249 break;
5250 if (j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
5252 if (j < FIRST_PSEUDO_REGISTER)
5253 SET_HARD_REG_BIT (live_hard_regs, j);
5254 else
5255 add_to_hard_reg_set (&live_hard_regs,
5256 PSEUDO_REGNO_MODE (j),
5257 reg_renumber[j]);
5258 setup_next_usage_insn (j, last_insn, reloads_num, after_p);
5262 src_regno = dst_regno = -1;
5263 if (NONDEBUG_INSN_P (curr_insn)
5264 && (set = single_set (curr_insn)) != NULL_RTX
5265 && REG_P (SET_DEST (set)) && REG_P (SET_SRC (set)))
5267 src_regno = REGNO (SET_SRC (set));
5268 dst_regno = REGNO (SET_DEST (set));
5270 update_reloads_num_p = true;
5271 if (src_regno < lra_constraint_new_regno_start
5272 && src_regno >= FIRST_PSEUDO_REGISTER
5273 && reg_renumber[src_regno] < 0
5274 && dst_regno >= lra_constraint_new_regno_start
5275 && (cl = lra_get_allocno_class (dst_regno)) != NO_REGS)
5277 /* 'reload_pseudo <- original_pseudo'. */
5278 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
5279 reloads_num++;
5280 update_reloads_num_p = false;
5281 succ_p = false;
5282 if (usage_insns[src_regno].check == curr_usage_insns_check
5283 && (next_usage_insns = usage_insns[src_regno].insns) != NULL_RTX)
5284 succ_p = inherit_reload_reg (false, src_regno, cl,
5285 curr_insn, next_usage_insns);
5286 if (succ_p)
5287 change_p = true;
5288 else
5289 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
5290 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
5291 IOR_HARD_REG_SET (potential_reload_hard_regs,
5292 reg_class_contents[cl]);
5294 else if (src_regno >= lra_constraint_new_regno_start
5295 && dst_regno < lra_constraint_new_regno_start
5296 && dst_regno >= FIRST_PSEUDO_REGISTER
5297 && reg_renumber[dst_regno] < 0
5298 && (cl = lra_get_allocno_class (src_regno)) != NO_REGS
5299 && usage_insns[dst_regno].check == curr_usage_insns_check
5300 && (next_usage_insns
5301 = usage_insns[dst_regno].insns) != NULL_RTX)
5303 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
5304 reloads_num++;
5305 update_reloads_num_p = false;
5306 /* 'original_pseudo <- reload_pseudo'. */
5307 if (! JUMP_P (curr_insn)
5308 && inherit_reload_reg (true, dst_regno, cl,
5309 curr_insn, next_usage_insns))
5310 change_p = true;
5311 /* Invalidate. */
5312 usage_insns[dst_regno].check = 0;
5313 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
5314 IOR_HARD_REG_SET (potential_reload_hard_regs,
5315 reg_class_contents[cl]);
5317 else if (INSN_P (curr_insn))
5319 int iter;
5320 int max_uid = get_max_uid ();
5322 curr_id = lra_get_insn_recog_data (curr_insn);
5323 curr_static_id = curr_id->insn_static_data;
5324 to_inherit_num = 0;
5325 /* Process insn definitions. */
5326 for (iter = 0; iter < 2; iter++)
5327 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
5328 reg != NULL;
5329 reg = reg->next)
5330 if (reg->type != OP_IN
5331 && (dst_regno = reg->regno) < lra_constraint_new_regno_start)
5333 if (dst_regno >= FIRST_PSEUDO_REGISTER && reg->type == OP_OUT
5334 && reg_renumber[dst_regno] < 0 && ! reg->subreg_p
5335 && usage_insns[dst_regno].check == curr_usage_insns_check
5336 && (next_usage_insns
5337 = usage_insns[dst_regno].insns) != NULL_RTX)
5339 struct lra_insn_reg *r;
5341 for (r = curr_id->regs; r != NULL; r = r->next)
5342 if (r->type != OP_OUT && r->regno == dst_regno)
5343 break;
5344 /* Don't do inheritance if the pseudo is also
5345 used in the insn. */
5346 if (r == NULL)
5347 /* We can not do inheritance right now
5348 because the current insn reg info (chain
5349 regs) can change after that. */
5350 add_to_inherit (dst_regno, next_usage_insns);
5352 /* We can not process one reg twice here because of
5353 usage_insns invalidation. */
5354 if ((dst_regno < FIRST_PSEUDO_REGISTER
5355 || reg_renumber[dst_regno] >= 0)
5356 && ! reg->subreg_p && reg->type != OP_IN)
5358 HARD_REG_SET s;
5360 if (split_if_necessary (dst_regno, reg->biggest_mode,
5361 potential_reload_hard_regs,
5362 false, curr_insn, max_uid))
5363 change_p = true;
5364 CLEAR_HARD_REG_SET (s);
5365 if (dst_regno < FIRST_PSEUDO_REGISTER)
5366 add_to_hard_reg_set (&s, reg->biggest_mode, dst_regno);
5367 else
5368 add_to_hard_reg_set (&s, PSEUDO_REGNO_MODE (dst_regno),
5369 reg_renumber[dst_regno]);
5370 AND_COMPL_HARD_REG_SET (live_hard_regs, s);
5372 /* We should invalidate potential inheritance or
5373 splitting for the current insn usages to the next
5374 usage insns (see code below) as the output pseudo
5375 prevents this. */
5376 if ((dst_regno >= FIRST_PSEUDO_REGISTER
5377 && reg_renumber[dst_regno] < 0)
5378 || (reg->type == OP_OUT && ! reg->subreg_p
5379 && (dst_regno < FIRST_PSEUDO_REGISTER
5380 || reg_renumber[dst_regno] >= 0)))
5382 /* Invalidate and mark definitions. */
5383 if (dst_regno >= FIRST_PSEUDO_REGISTER)
5384 usage_insns[dst_regno].check = -(int) INSN_UID (curr_insn);
5385 else
5387 nregs = hard_regno_nregs[dst_regno][reg->biggest_mode];
5388 for (i = 0; i < nregs; i++)
5389 usage_insns[dst_regno + i].check
5390 = -(int) INSN_UID (curr_insn);
5394 if (! JUMP_P (curr_insn))
5395 for (i = 0; i < to_inherit_num; i++)
5396 if (inherit_reload_reg (true, to_inherit[i].regno,
5397 ALL_REGS, curr_insn,
5398 to_inherit[i].insns))
5399 change_p = true;
5400 if (CALL_P (curr_insn))
5402 rtx cheap, pat, dest;
5403 rtx_insn *restore;
5404 int regno, hard_regno;
5406 calls_num++;
5407 if ((cheap = find_reg_note (curr_insn,
5408 REG_RETURNED, NULL_RTX)) != NULL_RTX
5409 && ((cheap = XEXP (cheap, 0)), true)
5410 && (regno = REGNO (cheap)) >= FIRST_PSEUDO_REGISTER
5411 && (hard_regno = reg_renumber[regno]) >= 0
5412 /* If there are pending saves/restores, the
5413 optimization is not worth. */
5414 && usage_insns[regno].calls_num == calls_num - 1
5415 && TEST_HARD_REG_BIT (call_used_reg_set, hard_regno))
5417 /* Restore the pseudo from the call result as
5418 REG_RETURNED note says that the pseudo value is
5419 in the call result and the pseudo is an argument
5420 of the call. */
5421 pat = PATTERN (curr_insn);
5422 if (GET_CODE (pat) == PARALLEL)
5423 pat = XVECEXP (pat, 0, 0);
5424 dest = SET_DEST (pat);
5425 /* For multiple return values dest is PARALLEL.
5426 Currently we handle only single return value case. */
5427 if (REG_P (dest))
5429 start_sequence ();
5430 emit_move_insn (cheap, copy_rtx (dest));
5431 restore = get_insns ();
5432 end_sequence ();
5433 lra_process_new_insns (curr_insn, NULL, restore,
5434 "Inserting call parameter restore");
5435 /* We don't need to save/restore of the pseudo from
5436 this call. */
5437 usage_insns[regno].calls_num = calls_num;
5438 bitmap_set_bit (&check_only_regs, regno);
5442 to_inherit_num = 0;
5443 /* Process insn usages. */
5444 for (iter = 0; iter < 2; iter++)
5445 for (reg = iter == 0 ? curr_id->regs : curr_static_id->hard_regs;
5446 reg != NULL;
5447 reg = reg->next)
5448 if ((reg->type != OP_OUT
5449 || (reg->type == OP_OUT && reg->subreg_p))
5450 && (src_regno = reg->regno) < lra_constraint_new_regno_start)
5452 if (src_regno >= FIRST_PSEUDO_REGISTER
5453 && reg_renumber[src_regno] < 0 && reg->type == OP_IN)
5455 if (usage_insns[src_regno].check == curr_usage_insns_check
5456 && (next_usage_insns
5457 = usage_insns[src_regno].insns) != NULL_RTX
5458 && NONDEBUG_INSN_P (curr_insn))
5459 add_to_inherit (src_regno, next_usage_insns);
5460 else if (usage_insns[src_regno].check
5461 != -(int) INSN_UID (curr_insn))
5462 /* Add usages but only if the reg is not set up
5463 in the same insn. */
5464 add_next_usage_insn (src_regno, curr_insn, reloads_num);
5466 else if (src_regno < FIRST_PSEUDO_REGISTER
5467 || reg_renumber[src_regno] >= 0)
5469 bool before_p;
5470 rtx use_insn = curr_insn;
5472 before_p = (JUMP_P (curr_insn)
5473 || (CALL_P (curr_insn) && reg->type == OP_IN));
5474 if (NONDEBUG_INSN_P (curr_insn)
5475 && (! JUMP_P (curr_insn) || reg->type == OP_IN)
5476 && split_if_necessary (src_regno, reg->biggest_mode,
5477 potential_reload_hard_regs,
5478 before_p, curr_insn, max_uid))
5480 if (reg->subreg_p)
5481 lra_risky_transformations_p = true;
5482 change_p = true;
5483 /* Invalidate. */
5484 usage_insns[src_regno].check = 0;
5485 if (before_p)
5486 use_insn = PREV_INSN (curr_insn);
5488 if (NONDEBUG_INSN_P (curr_insn))
5490 if (src_regno < FIRST_PSEUDO_REGISTER)
5491 add_to_hard_reg_set (&live_hard_regs,
5492 reg->biggest_mode, src_regno);
5493 else
5494 add_to_hard_reg_set (&live_hard_regs,
5495 PSEUDO_REGNO_MODE (src_regno),
5496 reg_renumber[src_regno]);
5498 add_next_usage_insn (src_regno, use_insn, reloads_num);
5501 /* Process call args. */
5502 if (curr_id->arg_hard_regs != NULL)
5503 for (i = 0; (src_regno = curr_id->arg_hard_regs[i]) >= 0; i++)
5504 if (src_regno < FIRST_PSEUDO_REGISTER)
5506 SET_HARD_REG_BIT (live_hard_regs, src_regno);
5507 add_next_usage_insn (src_regno, curr_insn, reloads_num);
5509 for (i = 0; i < to_inherit_num; i++)
5511 src_regno = to_inherit[i].regno;
5512 if (inherit_reload_reg (false, src_regno, ALL_REGS,
5513 curr_insn, to_inherit[i].insns))
5514 change_p = true;
5515 else
5516 setup_next_usage_insn (src_regno, curr_insn, reloads_num, false);
5519 if (update_reloads_num_p
5520 && NONDEBUG_INSN_P (curr_insn)
5521 && (set = single_set (curr_insn)) != NULL_RTX)
5523 int regno = -1;
5524 if ((REG_P (SET_DEST (set))
5525 && (regno = REGNO (SET_DEST (set))) >= lra_constraint_new_regno_start
5526 && reg_renumber[regno] < 0
5527 && (cl = lra_get_allocno_class (regno)) != NO_REGS)
5528 || (REG_P (SET_SRC (set))
5529 && (regno = REGNO (SET_SRC (set))) >= lra_constraint_new_regno_start
5530 && reg_renumber[regno] < 0
5531 && (cl = lra_get_allocno_class (regno)) != NO_REGS))
5533 if (ira_class_hard_regs_num[cl] <= max_small_class_regs_num)
5534 reloads_num++;
5535 if (hard_reg_set_subset_p (reg_class_contents[cl], live_hard_regs))
5536 IOR_HARD_REG_SET (potential_reload_hard_regs,
5537 reg_class_contents[cl]);
5540 /* We reached the start of the current basic block. */
5541 if (prev_insn == NULL_RTX || prev_insn == PREV_INSN (head)
5542 || BLOCK_FOR_INSN (prev_insn) != curr_bb)
5544 /* We reached the beginning of the current block -- do
5545 rest of spliting in the current BB. */
5546 to_process = df_get_live_in (curr_bb);
5547 if (BLOCK_FOR_INSN (head) != curr_bb)
5549 /* We are somewhere in the middle of EBB. */
5550 get_live_on_other_edges (EDGE_PRED (curr_bb, 0)->src,
5551 curr_bb, &temp_bitmap);
5552 to_process = &temp_bitmap;
5554 head_p = true;
5555 EXECUTE_IF_SET_IN_BITMAP (to_process, 0, j, bi)
5557 if ((int) j >= lra_constraint_new_regno_start)
5558 break;
5559 if (((int) j < FIRST_PSEUDO_REGISTER || reg_renumber[j] >= 0)
5560 && usage_insns[j].check == curr_usage_insns_check
5561 && (next_usage_insns = usage_insns[j].insns) != NULL_RTX)
5563 if (need_for_split_p (potential_reload_hard_regs, j))
5565 if (lra_dump_file != NULL && head_p)
5567 fprintf (lra_dump_file,
5568 " ----------------------------------\n");
5569 head_p = false;
5571 if (split_reg (false, j, bb_note (curr_bb),
5572 next_usage_insns))
5573 change_p = true;
5575 usage_insns[j].check = 0;
5580 return change_p;
5583 /* This value affects EBB forming. If probability of edge from EBB to
5584 a BB is not greater than the following value, we don't add the BB
5585 to EBB. */
5586 #define EBB_PROBABILITY_CUTOFF ((REG_BR_PROB_BASE * 50) / 100)
5588 /* Current number of inheritance/split iteration. */
5589 int lra_inheritance_iter;
5591 /* Entry function for inheritance/split pass. */
5592 void
5593 lra_inheritance (void)
5595 int i;
5596 basic_block bb, start_bb;
5597 edge e;
5599 lra_inheritance_iter++;
5600 if (lra_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
5601 return;
5602 timevar_push (TV_LRA_INHERITANCE);
5603 if (lra_dump_file != NULL)
5604 fprintf (lra_dump_file, "\n********** Inheritance #%d: **********\n\n",
5605 lra_inheritance_iter);
5606 curr_usage_insns_check = 0;
5607 usage_insns = XNEWVEC (struct usage_insns, lra_constraint_new_regno_start);
5608 for (i = 0; i < lra_constraint_new_regno_start; i++)
5609 usage_insns[i].check = 0;
5610 bitmap_initialize (&check_only_regs, &reg_obstack);
5611 bitmap_initialize (&live_regs, &reg_obstack);
5612 bitmap_initialize (&temp_bitmap, &reg_obstack);
5613 bitmap_initialize (&ebb_global_regs, &reg_obstack);
5614 FOR_EACH_BB_FN (bb, cfun)
5616 start_bb = bb;
5617 if (lra_dump_file != NULL)
5618 fprintf (lra_dump_file, "EBB");
5619 /* Form a EBB starting with BB. */
5620 bitmap_clear (&ebb_global_regs);
5621 bitmap_ior_into (&ebb_global_regs, df_get_live_in (bb));
5622 for (;;)
5624 if (lra_dump_file != NULL)
5625 fprintf (lra_dump_file, " %d", bb->index);
5626 if (bb->next_bb == EXIT_BLOCK_PTR_FOR_FN (cfun)
5627 || LABEL_P (BB_HEAD (bb->next_bb)))
5628 break;
5629 e = find_fallthru_edge (bb->succs);
5630 if (! e)
5631 break;
5632 if (e->probability <= EBB_PROBABILITY_CUTOFF)
5633 break;
5634 bb = bb->next_bb;
5636 bitmap_ior_into (&ebb_global_regs, df_get_live_out (bb));
5637 if (lra_dump_file != NULL)
5638 fprintf (lra_dump_file, "\n");
5639 if (inherit_in_ebb (BB_HEAD (start_bb), BB_END (bb)))
5640 /* Remember that the EBB head and tail can change in
5641 inherit_in_ebb. */
5642 update_ebb_live_info (BB_HEAD (start_bb), BB_END (bb));
5644 bitmap_clear (&ebb_global_regs);
5645 bitmap_clear (&temp_bitmap);
5646 bitmap_clear (&live_regs);
5647 bitmap_clear (&check_only_regs);
5648 free (usage_insns);
5650 timevar_pop (TV_LRA_INHERITANCE);
5655 /* This page contains code to undo failed inheritance/split
5656 transformations. */
5658 /* Current number of iteration undoing inheritance/split. */
5659 int lra_undo_inheritance_iter;
5661 /* Fix BB live info LIVE after removing pseudos created on pass doing
5662 inheritance/split which are REMOVED_PSEUDOS. */
5663 static void
5664 fix_bb_live_info (bitmap live, bitmap removed_pseudos)
5666 unsigned int regno;
5667 bitmap_iterator bi;
5669 EXECUTE_IF_SET_IN_BITMAP (removed_pseudos, 0, regno, bi)
5670 if (bitmap_clear_bit (live, regno))
5671 bitmap_set_bit (live, lra_reg_info[regno].restore_regno);
5674 /* Return regno of the (subreg of) REG. Otherwise, return a negative
5675 number. */
5676 static int
5677 get_regno (rtx reg)
5679 if (GET_CODE (reg) == SUBREG)
5680 reg = SUBREG_REG (reg);
5681 if (REG_P (reg))
5682 return REGNO (reg);
5683 return -1;
5686 /* Remove inheritance/split pseudos which are in REMOVE_PSEUDOS and
5687 return true if we did any change. The undo transformations for
5688 inheritance looks like
5689 i <- i2
5690 p <- i => p <- i2
5691 or removing
5692 p <- i, i <- p, and i <- i3
5693 where p is original pseudo from which inheritance pseudo i was
5694 created, i and i3 are removed inheritance pseudos, i2 is another
5695 not removed inheritance pseudo. All split pseudos or other
5696 occurrences of removed inheritance pseudos are changed on the
5697 corresponding original pseudos.
5699 The function also schedules insns changed and created during
5700 inheritance/split pass for processing by the subsequent constraint
5701 pass. */
5702 static bool
5703 remove_inheritance_pseudos (bitmap remove_pseudos)
5705 basic_block bb;
5706 int regno, sregno, prev_sregno, dregno, restore_regno;
5707 rtx set, prev_set;
5708 rtx_insn *prev_insn;
5709 bool change_p, done_p;
5711 change_p = ! bitmap_empty_p (remove_pseudos);
5712 /* We can not finish the function right away if CHANGE_P is true
5713 because we need to marks insns affected by previous
5714 inheritance/split pass for processing by the subsequent
5715 constraint pass. */
5716 FOR_EACH_BB_FN (bb, cfun)
5718 fix_bb_live_info (df_get_live_in (bb), remove_pseudos);
5719 fix_bb_live_info (df_get_live_out (bb), remove_pseudos);
5720 FOR_BB_INSNS_REVERSE (bb, curr_insn)
5722 if (! INSN_P (curr_insn))
5723 continue;
5724 done_p = false;
5725 sregno = dregno = -1;
5726 if (change_p && NONDEBUG_INSN_P (curr_insn)
5727 && (set = single_set (curr_insn)) != NULL_RTX)
5729 dregno = get_regno (SET_DEST (set));
5730 sregno = get_regno (SET_SRC (set));
5733 if (sregno >= 0 && dregno >= 0)
5735 if ((bitmap_bit_p (remove_pseudos, sregno)
5736 && (lra_reg_info[sregno].restore_regno == dregno
5737 || (bitmap_bit_p (remove_pseudos, dregno)
5738 && (lra_reg_info[sregno].restore_regno
5739 == lra_reg_info[dregno].restore_regno))))
5740 || (bitmap_bit_p (remove_pseudos, dregno)
5741 && lra_reg_info[dregno].restore_regno == sregno))
5742 /* One of the following cases:
5743 original <- removed inheritance pseudo
5744 removed inherit pseudo <- another removed inherit pseudo
5745 removed inherit pseudo <- original pseudo
5747 removed_split_pseudo <- original_reg
5748 original_reg <- removed_split_pseudo */
5750 if (lra_dump_file != NULL)
5752 fprintf (lra_dump_file, " Removing %s:\n",
5753 bitmap_bit_p (&lra_split_regs, sregno)
5754 || bitmap_bit_p (&lra_split_regs, dregno)
5755 ? "split" : "inheritance");
5756 dump_insn_slim (lra_dump_file, curr_insn);
5758 lra_set_insn_deleted (curr_insn);
5759 done_p = true;
5761 else if (bitmap_bit_p (remove_pseudos, sregno)
5762 && bitmap_bit_p (&lra_inheritance_pseudos, sregno))
5764 /* Search the following pattern:
5765 inherit_or_split_pseudo1 <- inherit_or_split_pseudo2
5766 original_pseudo <- inherit_or_split_pseudo1
5767 where the 2nd insn is the current insn and
5768 inherit_or_split_pseudo2 is not removed. If it is found,
5769 change the current insn onto:
5770 original_pseudo <- inherit_or_split_pseudo2. */
5771 for (prev_insn = PREV_INSN (curr_insn);
5772 prev_insn != NULL_RTX && ! NONDEBUG_INSN_P (prev_insn);
5773 prev_insn = PREV_INSN (prev_insn))
5775 if (prev_insn != NULL_RTX && BLOCK_FOR_INSN (prev_insn) == bb
5776 && (prev_set = single_set (prev_insn)) != NULL_RTX
5777 /* There should be no subregs in insn we are
5778 searching because only the original reg might
5779 be in subreg when we changed the mode of
5780 load/store for splitting. */
5781 && REG_P (SET_DEST (prev_set))
5782 && REG_P (SET_SRC (prev_set))
5783 && (int) REGNO (SET_DEST (prev_set)) == sregno
5784 && ((prev_sregno = REGNO (SET_SRC (prev_set)))
5785 >= FIRST_PSEUDO_REGISTER)
5786 /* As we consider chain of inheritance or
5787 splitting described in above comment we should
5788 check that sregno and prev_sregno were
5789 inheritance/split pseudos created from the
5790 same original regno. */
5791 && (lra_reg_info[sregno].restore_regno
5792 == lra_reg_info[prev_sregno].restore_regno)
5793 && ! bitmap_bit_p (remove_pseudos, prev_sregno))
5795 lra_assert (GET_MODE (SET_SRC (prev_set))
5796 == GET_MODE (regno_reg_rtx[sregno]));
5797 if (GET_CODE (SET_SRC (set)) == SUBREG)
5798 SUBREG_REG (SET_SRC (set)) = SET_SRC (prev_set);
5799 else
5800 SET_SRC (set) = SET_SRC (prev_set);
5801 /* As we are finishing with processing the insn
5802 here, check the destination too as it might
5803 inheritance pseudo for another pseudo. */
5804 if (bitmap_bit_p (remove_pseudos, dregno)
5805 && bitmap_bit_p (&lra_inheritance_pseudos, dregno)
5806 && (restore_regno
5807 = lra_reg_info[dregno].restore_regno) >= 0)
5809 if (GET_CODE (SET_DEST (set)) == SUBREG)
5810 SUBREG_REG (SET_DEST (set))
5811 = regno_reg_rtx[restore_regno];
5812 else
5813 SET_DEST (set) = regno_reg_rtx[restore_regno];
5815 lra_push_insn_and_update_insn_regno_info (curr_insn);
5816 lra_set_used_insn_alternative_by_uid
5817 (INSN_UID (curr_insn), -1);
5818 done_p = true;
5819 if (lra_dump_file != NULL)
5821 fprintf (lra_dump_file, " Change reload insn:\n");
5822 dump_insn_slim (lra_dump_file, curr_insn);
5827 if (! done_p)
5829 struct lra_insn_reg *reg;
5830 bool restored_regs_p = false;
5831 bool kept_regs_p = false;
5833 curr_id = lra_get_insn_recog_data (curr_insn);
5834 for (reg = curr_id->regs; reg != NULL; reg = reg->next)
5836 regno = reg->regno;
5837 restore_regno = lra_reg_info[regno].restore_regno;
5838 if (restore_regno >= 0)
5840 if (change_p && bitmap_bit_p (remove_pseudos, regno))
5842 lra_substitute_pseudo_within_insn (
5843 curr_insn, regno, regno_reg_rtx[restore_regno]);
5844 restored_regs_p = true;
5846 else
5847 kept_regs_p = true;
5850 if (NONDEBUG_INSN_P (curr_insn) && kept_regs_p)
5852 /* The instruction has changed since the previous
5853 constraints pass. */
5854 lra_push_insn_and_update_insn_regno_info (curr_insn);
5855 lra_set_used_insn_alternative_by_uid
5856 (INSN_UID (curr_insn), -1);
5858 else if (restored_regs_p)
5859 /* The instruction has been restored to the form that
5860 it had during the previous constraints pass. */
5861 lra_update_insn_regno_info (curr_insn);
5862 if (restored_regs_p && lra_dump_file != NULL)
5864 fprintf (lra_dump_file, " Insn after restoring regs:\n");
5865 dump_insn_slim (lra_dump_file, curr_insn);
5870 return change_p;
5873 /* If optional reload pseudos failed to get a hard register or was not
5874 inherited, it is better to remove optional reloads. We do this
5875 transformation after undoing inheritance to figure out necessity to
5876 remove optional reloads easier. Return true if we do any
5877 change. */
5878 static bool
5879 undo_optional_reloads (void)
5881 bool change_p, keep_p;
5882 unsigned int regno, uid;
5883 bitmap_iterator bi, bi2;
5884 rtx_insn *insn;
5885 rtx set, src, dest;
5886 bitmap_head removed_optional_reload_pseudos, insn_bitmap;
5888 bitmap_initialize (&removed_optional_reload_pseudos, &reg_obstack);
5889 bitmap_copy (&removed_optional_reload_pseudos, &lra_optional_reload_pseudos);
5890 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
5892 keep_p = false;
5893 /* Keep optional reloads from previous subpasses. */
5894 if (lra_reg_info[regno].restore_regno < 0
5895 /* If the original pseudo changed its allocation, just
5896 removing the optional pseudo is dangerous as the original
5897 pseudo will have longer live range. */
5898 || reg_renumber[lra_reg_info[regno].restore_regno] >= 0)
5899 keep_p = true;
5900 else if (reg_renumber[regno] >= 0)
5901 EXECUTE_IF_SET_IN_BITMAP (&lra_reg_info[regno].insn_bitmap, 0, uid, bi2)
5903 insn = lra_insn_recog_data[uid]->insn;
5904 if ((set = single_set (insn)) == NULL_RTX)
5905 continue;
5906 src = SET_SRC (set);
5907 dest = SET_DEST (set);
5908 if (! REG_P (src) || ! REG_P (dest))
5909 continue;
5910 if (REGNO (dest) == regno
5911 /* Ignore insn for optional reloads itself. */
5912 && lra_reg_info[regno].restore_regno != (int) REGNO (src)
5913 /* Check only inheritance on last inheritance pass. */
5914 && (int) REGNO (src) >= new_regno_start
5915 /* Check that the optional reload was inherited. */
5916 && bitmap_bit_p (&lra_inheritance_pseudos, REGNO (src)))
5918 keep_p = true;
5919 break;
5922 if (keep_p)
5924 bitmap_clear_bit (&removed_optional_reload_pseudos, regno);
5925 if (lra_dump_file != NULL)
5926 fprintf (lra_dump_file, "Keep optional reload reg %d\n", regno);
5929 change_p = ! bitmap_empty_p (&removed_optional_reload_pseudos);
5930 bitmap_initialize (&insn_bitmap, &reg_obstack);
5931 EXECUTE_IF_SET_IN_BITMAP (&removed_optional_reload_pseudos, 0, regno, bi)
5933 if (lra_dump_file != NULL)
5934 fprintf (lra_dump_file, "Remove optional reload reg %d\n", regno);
5935 bitmap_copy (&insn_bitmap, &lra_reg_info[regno].insn_bitmap);
5936 EXECUTE_IF_SET_IN_BITMAP (&insn_bitmap, 0, uid, bi2)
5938 insn = lra_insn_recog_data[uid]->insn;
5939 if ((set = single_set (insn)) != NULL_RTX)
5941 src = SET_SRC (set);
5942 dest = SET_DEST (set);
5943 if (REG_P (src) && REG_P (dest)
5944 && ((REGNO (src) == regno
5945 && (lra_reg_info[regno].restore_regno
5946 == (int) REGNO (dest)))
5947 || (REGNO (dest) == regno
5948 && (lra_reg_info[regno].restore_regno
5949 == (int) REGNO (src)))))
5951 if (lra_dump_file != NULL)
5953 fprintf (lra_dump_file, " Deleting move %u\n",
5954 INSN_UID (insn));
5955 dump_insn_slim (lra_dump_file, insn);
5957 lra_set_insn_deleted (insn);
5958 continue;
5960 /* We should not worry about generation memory-memory
5961 moves here as if the corresponding inheritance did
5962 not work (inheritance pseudo did not get a hard reg),
5963 we remove the inheritance pseudo and the optional
5964 reload. */
5966 lra_substitute_pseudo_within_insn (
5967 insn, regno,
5968 regno_reg_rtx[lra_reg_info[regno].restore_regno]);
5969 lra_update_insn_regno_info (insn);
5970 if (lra_dump_file != NULL)
5972 fprintf (lra_dump_file,
5973 " Restoring original insn:\n");
5974 dump_insn_slim (lra_dump_file, insn);
5978 /* Clear restore_regnos. */
5979 EXECUTE_IF_SET_IN_BITMAP (&lra_optional_reload_pseudos, 0, regno, bi)
5980 lra_reg_info[regno].restore_regno = -1;
5981 bitmap_clear (&insn_bitmap);
5982 bitmap_clear (&removed_optional_reload_pseudos);
5983 return change_p;
5986 /* Entry function for undoing inheritance/split transformation. Return true
5987 if we did any RTL change in this pass. */
5988 bool
5989 lra_undo_inheritance (void)
5991 unsigned int regno;
5992 int restore_regno, hard_regno;
5993 int n_all_inherit, n_inherit, n_all_split, n_split;
5994 bitmap_head remove_pseudos;
5995 bitmap_iterator bi;
5996 bool change_p;
5998 lra_undo_inheritance_iter++;
5999 if (lra_undo_inheritance_iter > LRA_MAX_INHERITANCE_PASSES)
6000 return false;
6001 if (lra_dump_file != NULL)
6002 fprintf (lra_dump_file,
6003 "\n********** Undoing inheritance #%d: **********\n\n",
6004 lra_undo_inheritance_iter);
6005 bitmap_initialize (&remove_pseudos, &reg_obstack);
6006 n_inherit = n_all_inherit = 0;
6007 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6008 if (lra_reg_info[regno].restore_regno >= 0)
6010 n_all_inherit++;
6011 if (reg_renumber[regno] < 0
6012 /* If the original pseudo changed its allocation, just
6013 removing inheritance is dangerous as for changing
6014 allocation we used shorter live-ranges. */
6015 && reg_renumber[lra_reg_info[regno].restore_regno] < 0)
6016 bitmap_set_bit (&remove_pseudos, regno);
6017 else
6018 n_inherit++;
6020 if (lra_dump_file != NULL && n_all_inherit != 0)
6021 fprintf (lra_dump_file, "Inherit %d out of %d (%.2f%%)\n",
6022 n_inherit, n_all_inherit,
6023 (double) n_inherit / n_all_inherit * 100);
6024 n_split = n_all_split = 0;
6025 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6026 if ((restore_regno = lra_reg_info[regno].restore_regno) >= 0)
6028 n_all_split++;
6029 hard_regno = (restore_regno >= FIRST_PSEUDO_REGISTER
6030 ? reg_renumber[restore_regno] : restore_regno);
6031 if (hard_regno < 0 || reg_renumber[regno] == hard_regno)
6032 bitmap_set_bit (&remove_pseudos, regno);
6033 else
6035 n_split++;
6036 if (lra_dump_file != NULL)
6037 fprintf (lra_dump_file, " Keep split r%d (orig=r%d)\n",
6038 regno, restore_regno);
6041 if (lra_dump_file != NULL && n_all_split != 0)
6042 fprintf (lra_dump_file, "Split %d out of %d (%.2f%%)\n",
6043 n_split, n_all_split,
6044 (double) n_split / n_all_split * 100);
6045 change_p = remove_inheritance_pseudos (&remove_pseudos);
6046 bitmap_clear (&remove_pseudos);
6047 /* Clear restore_regnos. */
6048 EXECUTE_IF_SET_IN_BITMAP (&lra_inheritance_pseudos, 0, regno, bi)
6049 lra_reg_info[regno].restore_regno = -1;
6050 EXECUTE_IF_SET_IN_BITMAP (&lra_split_regs, 0, regno, bi)
6051 lra_reg_info[regno].restore_regno = -1;
6052 change_p = undo_optional_reloads () || change_p;
6053 return change_p;